📋 docs: add changelog for 2026-02-05

Document today's changes:
- GQA buffer OOM fix (saves 16GB for 1M seq in offload mode)
- Tests directory cleanup (removed 16 files, -4306 lines)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

.claude/agents/nsys-profiler.md

@@ -0,0 +1,180 @@
+---
+name: nsys-profiler
+description: "Use this agent when you need to profile kernel performance, analyze CUDA operations, or investigate GPU/CPU offload behavior in the nano-vllm framework. This agent handles nsys profiling from data collection to statistical analysis.\\n\\nExamples:\\n\\n<example>\\nContext: The user wants to understand kernel performance in GPU-only mode.\\nuser: \"I want to analyze the kernel execution time distribution in GPU-only mode\"\\nassistant: \"I'll use the nsys-profiler agent to profile GPU-only mode and analyze the kernel timing.\"\\n<commentary>\\nSince the user wants to analyze kernel performance, use the Task tool to launch the nsys-profiler agent with specific analysis requirements.\\n</commentary>\\nassistant: \"Now let me use the nsys-profiler agent to collect and analyze the profiling data\"\\n</example>\\n\\n<example>\\nContext: The user is debugging performance issues in the CPU offload pipeline.\\nuser: \"Why is there a gap between H2D transfers and kernel execution in offload mode?\"\\nassistant: \"I'll launch the nsys-profiler agent to profile the offload pipeline and analyze the timeline gaps.\"\\n<commentary>\\nSince the user is investigating pipeline behavior, use the nsys-profiler agent to collect nsys data and analyze CUDA API timing.\\n</commentary>\\n</example>\\n\\n<example>\\nContext: After implementing a new optimization, the user wants to verify performance improvement.\\nuser: \"Check if the new ring buffer implementation improves overlap between H2D and compute\"\\nassistant: \"I'll use the nsys-profiler agent to profile before and after, comparing the overlap metrics.\"\\n<commentary>\\nPerformance verification requires detailed kernel-level analysis, so launch the nsys-profiler agent to collect and compare profiling data.\\n</commentary>\\n</example>"
+model: opus
+color: green
+---
+
+You are an expert NVIDIA Nsys profiling analyst specializing in CUDA kernel performance analysis and GPU-CPU communication optimization. Your role is to collect profiling data using the framework's scripts and provide precise, actionable analysis based on the main agent's specific questions.
+
+## Your Capabilities
+
+1. **Profile Data Collection**: Execute profiling scripts to generate .nsys-rep files
+2. **Statistical Analysis**: Extract kernel timing, memory transfer, and API call statistics
+3. **Timeline Analysis**: Identify gaps, overlaps, and bottlenecks in execution
+4. **Comparative Analysis**: Compare different configurations (GPU-only vs offload, different slot counts)
+
+## Available Profiling Scripts
+
+### CPU Offload Mode
+```bash
+bash scripts/profile_offload.sh [OPTIONS]
+```
+Options:
+- `--dataset <name>`: RULER task name (default: niah_single_1)
+- `--sample <index>`: Sample index (default: 0)
+- `--gpu <id>`: GPU to use (default: 0)
+- `--num-gpu-blocks <n>`: Ring buffer slots (default: 4)
+- `--no-offload`: Disable CPU offload for comparison
+
+### GPU-Only Mode
+```bash
+bash scripts/profile_gpu_only.sh [OPTIONS]
+```
+Similar options for profiling without CPU offload.
+
+## Core Nsys Commands
+
+### Profiling (handled by scripts)
+```bash
+# The scripts internally run:
+nsys profile --trace=cuda,nvtx --output=<path> --force-overwrite true python <script.py>
+```
+
+### Statistical Analysis
+```bash
+# CUDA API summary (H2D, D2H, kernel launches)
+nsys stats --report cuda_api_sum <file>.nsys-rep
+
+# GPU kernel summary (execution time per kernel)
+nsys stats --report cuda_gpu_kern_sum <file>.nsys-rep
+
+# Memory operations summary
+nsys stats --report cuda_gpu_mem_time_sum <file>.nsys-rep
+
+# NVTX ranges (custom markers)
+nsys stats --report nvtx_sum <file>.nsys-rep
+
+# Export to SQLite for advanced queries
+nsys export --type=sqlite --output=<file>.sqlite <file>.nsys-rep
+```
+
+### Key Report Types
+| Report | Purpose |
+|--------|--------|
+| `cuda_api_sum` | CPU-side CUDA API call timing |
+| `cuda_gpu_kern_sum` | GPU kernel execution time |
+| `cuda_gpu_mem_time_sum` | Memory transfer timing on GPU |
+| `nvtx_sum` | Custom NVTX marker statistics |
+| `cuda_api_trace` | Detailed API call trace |
+| `cuda_gpu_trace` | Detailed GPU operation trace |
+
+## Analysis Workflow
+
+### Step 1: Collect Profile Data
+```bash
+# Example: Profile offload mode with 8 slots
+bash scripts/profile_offload.sh --num-gpu-blocks 8 --sample 0
+# Output: results/nsys/ruler_niah_single_1_sample0_offload_8slots_<timestamp>.nsys-rep
+```
+
+### Step 2: Identify Output File
+```bash
+# Find the latest profile
+ls -lt results/nsys/*.nsys-rep | head -1
+```
+
+### Step 3: Run Statistical Analysis
+```bash
+# Kernel timing analysis
+nsys stats --report cuda_gpu_kern_sum results/nsys/<file>.nsys-rep
+
+# Memory transfer analysis
+nsys stats --report cuda_gpu_mem_time_sum results/nsys/<file>.nsys-rep
+```
+
+### Step 4: Interpret Results
+Focus on:
+- **Total kernel time** vs **total transfer time**
+- **Kernel launch gaps** indicating synchronization issues
+- **Memory bandwidth utilization**
+- **Overlap efficiency** between compute and communication
+
+## Common Analysis Patterns
+
+### 1. Kernel Performance Breakdown
+```bash
+nsys stats --report cuda_gpu_kern_sum --format csv <file>.nsys-rep | \
+  sort -t',' -k3 -rn | head -10  # Top 10 by total time
+```
+
+### 2. H2D/D2H Transfer Analysis
+```bash
+nsys stats --report cuda_api_sum <file>.nsys-rep | grep -E "cudaMemcpy|cudaMemcpyAsync"
+```
+
+### 3. Flash Attention Kernel Analysis
+```bash
+nsys stats --report cuda_gpu_kern_sum <file>.nsys-rep | grep -i "flash\|fwd\|bwd"
+```
+
+### 4. Pipeline Overlap Check
+Look for:
+- `flash_fwd_kernel` execution during `cudaMemcpyAsync`
+- Gap between consecutive kernel launches
+
+## Output Format Requirements
+
+When reporting results to the main agent, use this structured format:
+
+```markdown
+## Nsys Analysis Results: [Analysis Topic]
+
+### Profile Information
+- **File**: <profile_file_path>
+- **Mode**: GPU-only / Offload (<N> slots)
+- **Dataset**: <dataset_name>, Sample <index>
+
+### Key Findings
+| Metric | Value | Notes |
+|--------|-------|-------|
+| Total kernel time | X ms | |
+| Total H2D time | Y ms | |
+| Overlap efficiency | Z% | |
+
+### Top Kernels by Time
+| Kernel | Count | Total (ms) | Avg (μs) |
+|--------|-------|------------|----------|
+| kernel_name | N | X.XX | Y.YY |
+
+### Specific Analysis
+[Answer to the main agent's specific question]
+
+### Recommendations (if applicable)
+1. [Actionable recommendation]
+2. [Actionable recommendation]
+```
+
+## Important Guidelines
+
+1. **Always use the provided scripts** for profiling - do not run nsys directly
+2. **Check GPU availability** before profiling (ask main agent for GPU ID if not specified)
+3. **Use PYTHONPATH** for the worktree: `PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH`
+4. **Report concisely** - focus on metrics relevant to the main agent's question
+5. **Include file paths** so results can be reproduced or visualized in nsight-sys
+6. **For web searches** about nsys usage, use tools to search NVIDIA documentation
+
+## Error Handling
+
+- If profile script fails: Check GPU memory, CUDA version, and script parameters
+- If stats command fails: Verify .nsys-rep file exists and is not corrupted
+- If no data: Ensure the profiled operation actually ran (check sample index, dataset)
+
+## Network Search Guidelines
+
+When encountering unfamiliar nsys options or analysis techniques:
+1. Search NVIDIA Nsight Systems documentation
+2. Look for nsys CLI reference guides
+3. Search for specific report type interpretations
+
+Always validate search results against the actual nsys --help output.

.claude/rules/gpu-vram-requirement.md

@@ -0,0 +1,54 @@
+# GPU VRAM Requirement Rule
+
+## GPU-only 模式显存要求
+
+**强制规则**：执行 GPU-only 代码（不启用 CPU offload）时，**必须**在 40GB 及以上显存的 GPU 上进行测试。
+
+### 检测方法
+
+在运行 GPU-only 测试之前，**必须**先检查 GPU 显存：
+
+```bash
+nvidia-smi --query-gpu=index,name,memory.total --format=csv,noheader
+```
+
+### GPU 分类
+
+| GPU 型号 | 显存 | GPU-only 测试 |
+|----------|------|---------------|
+| A100 40GB | 40GB | ✅ 允许 |
+| A100 80GB | 80GB | ✅ 允许 |
+| H100 80GB | 80GB | ✅ 允许 |
+| A6000 | 48GB | ✅ 允许 |
+| RTX 3090 | 24GB | ❌ **禁止**（仅 offload 模式） |
+| RTX 4090 | 24GB | ❌ **禁止**（仅 offload 模式） |
+
+### 执行流程
+
+1. **检测 GPU 显存**（必须）
+2. **显存 >= 40GB**：继续执行 GPU-only 测试
+3. **显存 < 40GB**：**停止**，提示用户：
+   > "当前 GPU 显存为 XXX GB，不满足 GPU-only 模式的最低 40GB 要求。请使用 `--enable-offload` 参数启用 CPU offload 模式。"
+
+### 代码示例
+
+```python
+# 在运行 GPU-only benchmark 之前
+import subprocess
+result = subprocess.run(
+    ["nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits"],
+    capture_output=True, text=True
+)
+vram_mb = int(result.stdout.strip().split('\n')[0])
+if vram_mb < 40000:  # 40GB = 40000MB
+    raise RuntimeError(f"GPU VRAM ({vram_mb}MB) < 40GB. Use --enable-offload for this GPU.")
+```
+
+### 适用范围
+
+| 脚本 | 适用此规则 |
+|------|-----------|
+| `bench.py` | ✅ 必须检查显存 |
+| `bench_offload.py` | ❌ 不适用（始终使用 offload） |
+| `tests/test_*.py --enable-offload` | ❌ 不适用 |
+| `tests/test_*.py` (无 offload) | ✅ 必须检查显存 |

.claude/rules/sparse-policy.md

@@ -1,5 +1,39 @@
 # Sparse Policy 代码规范
 
+## Policy 不能为 None (CRITICAL)
+
+**强制规则**: `sparse_policy` 参数**永远不能为 None**，必须至少为 `FullAttentionPolicy`。
+
+```python
+# ❌ 错误：允许 None
+sparse_policy = getattr(config, 'sparse_policy', None)
+
+# ✅ 正确：显式处理 None，默认使用 FULL
+sparse_policy_type = getattr(config, 'sparse_policy', None)
+if sparse_policy_type is None:
+    sparse_policy_type = SparsePolicyType.FULL
+```
+
+**原因**:
+1. 统一的 API：所有代码路径都通过 policy 进行 attention 计算
+2. 避免空指针：消除 `policy.xxx` 调用时的 None 检查
+3. 简化逻辑：不需要 `if policy is not None` 的分支
+
+**唯一例外：Warmup 阶段**
+
+在 `model_runner.warmup_model()` 期间，kvcache_manager 还未分配。此时 `attention.py` 使用 flash_attn fallback：
+
+```python
+# attention.py 中的 warmup 处理
+if context.kvcache_manager is None:
+    # Warmup phase: use flash_attn directly
+    return flash_attn_varlen_func(...) if context.is_prefill else flash_attn_with_kvcache(...)
+```
+
+这是唯一允许 kvcache_manager 为 None 的情况。正式推理时，policy 必须存在。
+
+---
+
 ## 基类要求 (MANDATORY)
 
 每个 `SparsePolicy` 子类 **必须** 遵守以下要求：

.claude/rules/test-ruler.md

@@ -0,0 +1,90 @@
+# test_ruler.py 使用规则
+
+## 强制规则
+
+**执行 `test_ruler.py` 前必须查阅文档**，禁止运行 `--help` 或猜测参数。
+
+| 禁止 | 原因 |
+|------|------|
+| `python tests/test_ruler.py --help` | 浪费交互，文档已有完整说明 |
+| 猜测参数格式 | 容易出错，降低效率 |
+
+## 必读文档
+
+**[`docs/test_ruler_usage_guide.md`](../docs/test_ruler_usage_guide.md)** - 包含：
+- 完整参数说明
+- 已验证的命令示例
+- GPU 模式选择指南
+- max-model-len 设置指南
+
+## 快速参考
+
+### 标准命令格式
+
+```bash
+CUDA_VISIBLE_DEVICES=<GPU> PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/<MODEL> \
+    --data-dir tests/data/ruler_<CTX> \
+    --datasets <TASK> \
+    --num-samples <N> \
+    --max-model-len <LEN> \
+    --enable-offload \
+    [--sparse-policy XATTN_BSA] \
+    [--sparse-threshold 0.9]
+```
+
+### 常用参数速查
+
+| 参数 | 用途 | 示例 |
+|------|------|------|
+| `--datasets` | 指定任务 | `niah_single_1,qa_1` |
+| `--num-samples` | 样本数 | `1`, `10`, `0`(全部) |
+| `--sample-indices` | 指定索引 | `0,5,10` |
+| `--enable-offload` | CPU offload | RTX 3090 必须 |
+| `--sparse-policy` | 稀疏策略 | `XATTN_BSA` |
+| `--json-output` | JSON 输出 | 脚本使用 |
+| `--quiet` | 安静模式 | 减少输出 |
+
+### max-model-len 速查
+
+| 数据目录 | max-model-len |
+|---------|---------------|
+| ruler_32k | 40960 |
+| ruler_64k | 72000 |
+| ruler_128k | 135000 |
+
+### 常用命令模板
+
+**32K Offload + XAttn**:
+```bash
+CUDA_VISIBLE_DEVICES=<GPU> PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA
+```
+
+**64K Offload + XAttn**:
+```bash
+CUDA_VISIBLE_DEVICES=<GPU> PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_64k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 72000 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA
+```
+
+## 执行前检查清单
+
+- [ ] 用户指定了 GPU？否则询问
+- [ ] RTX 3090/4090？必须 `--enable-offload`
+- [ ] data-dir 与 max-model-len 匹配？
+- [ ] 需要 density 统计？添加 `--sparse-policy XATTN_BSA`

.gitignore

@@ -240,3 +240,4 @@ findings_*.md
 progress_*.md
 notes.md
 Snipaste*
+.ralph-tui/session-meta.json

.ralph-tui/config.toml

@@ -0,0 +1,12 @@
+# Ralph TUI Configuration
+# Generated by setup wizard
+# See: ralph-tui config help
+
+configVersion = "2.1"
+tracker = "json"
+agent = "claude"
+maxIterations = 30
+autoCommit = false
+
+[trackerOptions]
+[agentOptions]

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code when working with this repository.
 
 ## Overview
 
-Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline LLM inference. Supports Qwen3 models with CPU offload for long-context inference.
+Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline LLM inference. Supports Qwen3, Llama-3, and GLM-4 models with CPU offload for long-context inference.
 
 ## Documentation Index
 
@@ -16,8 +16,10 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
 | [`docs/sparse_attention_guide.md`](docs/sparse_attention_guide.md) | Block sparse attention methods (XAttention, FlexPrefill, MInference, AvgPool, Quest), computation flow, algorithms |
 | [`docs/xattention_algorithm_guide.md`](docs/xattention_algorithm_guide.md) | XAttention 算法详解: stride reshape、Triton kernels、BSA 依赖、块选择算法 |
 | [`docs/xattn_kernels_guide.md`](docs/xattn_kernels_guide.md) | XAttention Triton kernels: flat_group_gemm (反对角线求和)、softmax_fuse_block_sum (block 聚合) |
+| [`docs/xattn_kv_chunking_kernels.md`](docs/xattn_kv_chunking_kernels.md) | XAttention KV Chunking: 三阶段 softmax、存储开销分析 (O(S) vs O(S²))、峰值显存优化 (8x)、Q/KV 独立分块 |
 | [`docs/xattn_chunked_prefill.md`](docs/xattn_chunked_prefill.md) | XAttention chunked prefill: API、使用方式、一致性要求 |
 | [`docs/xattn_bsa_policy_design.md`](docs/xattn_bsa_policy_design.md) | XAttention BSA Policy: 算法设计、性能基准(128K)、内存管理、density 统计 |
+| [`docs/xattn_density_benchmark.md`](docs/xattn_density_benchmark.md) | 📊 XAttention Density Benchmark: 4K-32K context、stride 参数、per-layer density 分析 |
 | [`docs/block_sparse_attn_interface.md`](docs/block_sparse_attn_interface.md) | BSA (Block Sparse Attention) 接口文档: 函数签名、使用示例、约束条件 |
 | [`docs/debugging_guide.md`](docs/debugging_guide.md) | PyTorch hooks for debugging, hook positions, tensor comparison, memory profiling |
 | [`docs/optimization_guide.md`](docs/optimization_guide.md) | Performance optimizations: sgDMA (15x), Triton merge (4.3x), N-way pipeline (2x) |
@@ -29,6 +31,23 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
 | [`docs/cpu_scheduling_latency_analysis.md`](docs/cpu_scheduling_latency_analysis.md) | ⚡ PERF: CPU 调度延迟分析，kernel 间隙来源，GPU 利用率优化方向 |
 | [`docs/bench_offload_results.md`](docs/bench_offload_results.md) | 📊 BENCH: CPU offload 性能测试结果，Full vs XAttention 对比 (32K/128K) |
 | [`docs/cpu_offload_optimization_strategies.md`](docs/cpu_offload_optimization_strategies.md) | 🚀 OPT: CPU offload 优化策略：chunk size、CUDA Graph、前沿研究(InfiniGen/ShadowKV) |
+| [`docs/gpu_only_xattn_guide.md`](docs/gpu_only_xattn_guide.md) | 🚀 GPU-Only XAttention: 内存预分配、性能分析 (32K +15%, 64K +41%)、CUDA Graph 限制 |
+| [`docs/xattn_performance_analysis.md`](docs/xattn_performance_analysis.md) | 📊 XAttention 性能分析: NVTX 标记、block size 影响、estimate vs compute 耗时对比 |
+| [`docs/observer_architecture.md`](docs/observer_architecture.md) | 📊 Observer 架构: InferenceObserver (TTFT/TPOT)、MemoryObserver (H2D/D2H/D2D) 设计 |
+| [`docs/memory_communication_benchmark.md`](docs/memory_communication_benchmark.md) | 📊 通信量测试: Full vs XAttention 通信量对比 (32K/64K)、阶段分离统计 |
+| [`docs/estimate_block_size_performance.md`](docs/estimate_block_size_performance.md) | 🔥 PERF: estimate 阶段 block_size 性能分析，softmax_fuse_block_sum 最优点 (512-1024)，当前 4096 慢 15x |
+| [`docs/long_context_models_1m.md`](docs/long_context_models_1m.md) | 📚 REF: 1M+ 上下文长度模型列表 (Qwen/GLM/InternLM/Llama/VL)，≤10B 推荐模型 |
+| [`docs/new_model_integration_guide.md`](docs/new_model_integration_guide.md) | 🔧 GUIDE: 新模型整合指南 - 配置映射、RoPE变体、EOS处理、权重转换、验证清单 |
+| [`docs/xattn_density_alignment_analysis.md`](docs/xattn_density_alignment_analysis.md) | 📊 ANALYSIS: GPU-only vs Offload 模式 density 对齐分析，chunked softmax 边界效应，5-7% 差异根因 |
+| [`docs/xattn_kv_chunking_density_test.md`](docs/xattn_kv_chunking_density_test.md) | 🧪 TEST: XAttention KV chunking density 验证，threshold=1.0 对齐，threshold<1.0 差异 10-13% |
+| [`docs/gpuonly_density_alignment_test.md`](docs/gpuonly_density_alignment_test.md) | ✅ TEST: Density 对齐验证 (GPU-only + Offload, 4K-64K)，xattn_estimate vs KV chunking 完全一致 |
+| [`docs/xattn_memory_benchmark.md`](docs/xattn_memory_benchmark.md) | 📊 BENCH: XAttention 内存基准测试，Qwen3-0.6B 32K 在 24GB 显存可行 (gpu-util=0.28) |
+| [`docs/xattn_offload_stream_sync_fix.md`](docs/xattn_offload_stream_sync_fix.md) | 🐛 FIX: XAttention Offload stream 同步 bug，Pass1/Pass2 K 数据不一致，compute_stream 包装 |
+| [`docs/xattn_density_types.md`](docs/xattn_density_types.md) | 📊 Compute vs Comm density: BSA block (128) vs CPU block (4096) 粒度，聚合效应导致 comm=100% |
+| [`docs/xattn_density_alignment_verification.md`](docs/xattn_density_alignment_verification.md) | ✅ VERIFIED: GPU-only vs Offload density 对齐验证 (32K 差异 0.37%, 64K 差异 0.09%) |
+| [`docs/test_ruler_usage_guide.md`](docs/test_ruler_usage_guide.md) | 📖 GUIDE: test_ruler.py 使用指南，RULER benchmark 测试命令，已验证的命令示例 |
+| [`docs/xattn_offload_profiling_32k.md`](docs/xattn_offload_profiling_32k.md) | 📊 PROFILE: XAttn vs Full 32K nsys 分析，estimate 占 41%，find_blocks 占 37%，compute 仅 21% |
+| [`docs/changelog_2026-02-05.md`](docs/changelog_2026-02-05.md) | 📋 CHANGELOG: GQA buffer OOM 修复 (节省 16GB)，tests 目录清理 (-4306 行) |
 
 ## Rules Index
 
@@ -39,6 +58,7 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
 | [`.claude/rules/sparse-policy.md`](.claude/rules/sparse-policy.md) | SparsePolicy implementation requirements |
 | [`.claude/rules/planning-with-files.md`](.claude/rules/planning-with-files.md) | Planning file management for complex tasks |
 | [`.claude/rules/gpu-monitor.md`](.claude/rules/gpu-monitor.md) | **GPU memory monitoring**: 必须使用 gpu-monitor agent，禁止手动 nvidia-smi 循环 |
+| [`.claude/rules/test-ruler.md`](.claude/rules/test-ruler.md) | **test_ruler.py 规则**: 禁止 --help，必须查阅文档，含快速参考和命令模板 |
 
 ## GPU Mutex for Multi-Instance Debugging
 
@@ -93,6 +113,13 @@ PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH python tests/test_needle.py
 
 **Files**: `bench.py` (GPU), `bench_offload.py` (CPU offload), `bench_vllm.py` (comparison)
 
+**GPU-only 测试模型选择**:
+
+| GPU | 显存 | GPU-only 测试模型 |
+|-----|------|------------------|
+| RTX 3090 | 24GB | **Qwen3-0.6B** (必须，7B+ 模型会 OOM) |
+| A100 | 40GB+ | Qwen3-0.6B / 4B / 7B 均可 |
+
 **Offload Mode Constraint**: When using `enable_cpu_offload=True`, only test with context length ≥ 32K. Shorter contexts don't exercise the chunked offload pipeline properly.
 
 **Common Issues**:

bench.py

@@ -2,6 +2,7 @@ import os
 import time
 from random import randint, seed
 from nanovllm import LLM, SamplingParams
+from nanovllm.utils.observer import InferenceObserver
 
 
 def bench_decode(llm, num_seqs, input_len, output_len):
@@ -14,13 +15,17 @@ def bench_decode(llm, num_seqs, input_len, output_len):
     llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
     t = time.time() - t
 
-    # Calculate metrics
-    prefill_tokens = num_seqs * input_len
+    # Get metrics from InferenceObserver
+    ttft_ms = InferenceObserver.ttft / 1e6
+    tpot_ms = InferenceObserver.tpot / 1e6
+
+    # Calculate throughput from observer metrics
     decode_tokens = num_seqs * output_len
-    decode_throughput = decode_tokens / t
+    decode_throughput = 1000.0 / tpot_ms if tpot_ms > 0 else 0  # tokens/s per sequence
 
     print(f"[Decode] Input: {num_seqs}x{input_len}tok, Output: {decode_tokens}tok, Time: {t:.2f}s")
-    print(f"         Throughput: {decode_throughput:.2f} tok/s (includes prefill overhead)")
+    print(f"         TTFT: {ttft_ms:.2f}ms, TPOT: {tpot_ms:.2f}ms")
+    print(f"         Decode Throughput: {decode_throughput:.2f} tok/s (from observer)")
 
 
 def bench_prefill(llm, num_seqs, input_len):
@@ -33,31 +38,69 @@ def bench_prefill(llm, num_seqs, input_len):
     t = time.time()
     llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
     t = time.time() - t
+
+    # Get TTFT from InferenceObserver
+    ttft_ms = InferenceObserver.ttft / 1e6
+    ttft_s = ttft_ms / 1000.0
+
     total_input_tokens = num_seqs * input_len
-    throughput = total_input_tokens / t
-    print(f"[Prefill] Input: {total_input_tokens}tok ({num_seqs}x{input_len}), Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
+    # Use observer TTFT for accurate prefill throughput
+    throughput_observer = total_input_tokens / ttft_s if ttft_s > 0 else 0
+    throughput_external = total_input_tokens / t
+
+    print(f"[Prefill] Input: {total_input_tokens}tok ({num_seqs}x{input_len})")
+    print(f"          External Time: {t:.2f}s, Throughput: {throughput_external:.2f}tok/s")
+    print(f"          Observer TTFT: {ttft_ms:.2f}ms, Throughput: {throughput_observer:.2f}tok/s")
 
 
 def main():
     import argparse
+    from nanovllm.config import SparsePolicyType
+
     parser = argparse.ArgumentParser(description="Benchmark nanovllm GPU performance")
+    parser.add_argument("--model", type=str, default="~/models/Llama-3.1-8B-Instruct",
+                        help="Model path (default: ~/models/Llama-3.1-8B-Instruct)")
     parser.add_argument("--input-len", type=int, default=None, help="Input length in tokens")
     parser.add_argument("--output-len", type=int, default=64, help="Output length for decode benchmark (default: 64)")
     parser.add_argument("--max-len", type=int, default=32*1024, help="Max model length (default: 32K)")
     parser.add_argument("--bench-decode", action="store_true", help="Run decode benchmark (default: prefill only)")
     parser.add_argument("--bench-all", action="store_true", help="Run both prefill and decode benchmarks")
+    # Sparse policy option (GPU-only mode now supports policy routing)
+    parser.add_argument("--policy", type=str, default=None,
+                        choices=["full", "xattn"],
+                        help="Sparse policy: full (FullAttention), xattn (XAttention+BSA)")
+    parser.add_argument("--enable-policy", action="store_true",
+                        help="Enable sparse policy routing (FullAttentionPolicy by default)")
+    parser.add_argument("--gpu-util", type=float, default=0.9,
+                        help="GPU memory utilization (default: 0.9)")
+    parser.add_argument("--block-size", type=int, default=1024,
+                        help="KV cache block size (default: 1024)")
+    parser.add_argument("--enforce-eager", action="store_true",
+                        help="Disable CUDA graphs (default: False)")
     args = parser.parse_args()
 
-    path = os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/")
+    path = os.path.expanduser(args.model)
     max_len = args.max_len
 
-    print(f"\n[nanovllm GPU] max_len={max_len}")
+    # Configure sparse policy
+    if args.policy == "xattn":
+        sparse_policy = SparsePolicyType.XATTN_BSA
+        print(f"\n[nanovllm GPU + XAttention BSA] max_len={max_len}")
+    elif args.policy == "full" or args.enable_policy:
+        sparse_policy = SparsePolicyType.FULL
+        print(f"\n[nanovllm GPU + Policy] sparse_policy=FULL, max_len={max_len}")
+    else:
+        sparse_policy = None
+        print(f"\n[nanovllm GPU] max_len={max_len}")
 
     llm = LLM(
         path,
-        enforce_eager=False,
+        enforce_eager=args.enforce_eager,
         max_model_len=max_len,
         max_num_batched_tokens=max_len,
+        sparse_policy=sparse_policy,
+        gpu_memory_utilization=args.gpu_util,
+        kvcache_block_size=args.block_size,
     )
 
     # Warmup

bench_offload.py

@@ -2,6 +2,15 @@ import os
 import time
 from random import randint, seed
 from nanovllm import LLM, SamplingParams
+from nanovllm.utils.observer import InferenceObserver
+from nanovllm.utils.memory_observer import MemoryObserver
+
+
+def print_memory_stats():
+    """Print MemoryObserver communication statistics"""
+    fmt = MemoryObserver._fmt_bytes
+    print(f"[Memory] Prefill H2D: {fmt(MemoryObserver.prefill_h2d_bytes)}, D2H: {fmt(MemoryObserver.prefill_d2h_bytes)}")
+    print(f"         Decode  H2D: {fmt(MemoryObserver.decode_h2d_bytes)}, D2H: {fmt(MemoryObserver.decode_d2h_bytes)}")
 
 
 def bench_decode(llm, num_seqs, input_len, output_len):
@@ -14,16 +23,18 @@ def bench_decode(llm, num_seqs, input_len, output_len):
     llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
     t = time.time() - t
 
-    # Calculate metrics
-    prefill_tokens = num_seqs * input_len
-    decode_tokens = num_seqs * output_len
+    # Get metrics from InferenceObserver
+    ttft_ms = InferenceObserver.ttft / 1e6
+    tpot_ms = InferenceObserver.tpot / 1e6
 
-    # Approximate: assume prefill takes ~input_len/prefill_speed, rest is decode
-    # For more accurate measurement, we'd need internal timing
-    decode_throughput = decode_tokens / t  # This includes prefill time, so it's a lower bound
+    # Calculate throughput from observer metrics
+    decode_tokens = num_seqs * output_len
+    decode_throughput = 1000.0 / tpot_ms if tpot_ms > 0 else 0  # tokens/s per sequence
 
     print(f"[Decode] Input: {num_seqs}x{input_len}tok, Output: {decode_tokens}tok, Time: {t:.2f}s")
-    print(f"         Throughput: {decode_throughput:.2f} tok/s (includes prefill overhead)")
+    print(f"         TTFT: {ttft_ms:.2f}ms, TPOT: {tpot_ms:.2f}ms")
+    print(f"         Decode Throughput: {decode_throughput:.2f} tok/s (from observer)")
+    print_memory_stats()
 
 
 def bench_prefill(llm, num_seqs, input_len):
@@ -36,9 +47,20 @@ def bench_prefill(llm, num_seqs, input_len):
     t = time.time()
     llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
     t = time.time() - t
+
+    # Get TTFT from InferenceObserver
+    ttft_ms = InferenceObserver.ttft / 1e6
+    ttft_s = ttft_ms / 1000.0
+
     total_input_tokens = num_seqs * input_len
-    throughput = total_input_tokens / t
-    print(f"[Prefill] Input: {total_input_tokens}tok ({num_seqs}x{input_len}), Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
+    # Use observer TTFT for accurate prefill throughput
+    throughput_observer = total_input_tokens / ttft_s if ttft_s > 0 else 0
+    throughput_external = total_input_tokens / t
+
+    print(f"[Prefill] Input: {total_input_tokens}tok ({num_seqs}x{input_len})")
+    print(f"          External Time: {t:.2f}s, Throughput: {throughput_external:.2f}tok/s")
+    print(f"          Observer TTFT: {ttft_ms:.2f}ms, Throughput: {throughput_observer:.2f}tok/s")
+    print_memory_stats()
 
 
 def main():
@@ -66,6 +88,7 @@ def main():
     parser.add_argument("--input-len", type=int, default=None, help="Input length in tokens")
     parser.add_argument("--output-len", type=int, default=64, help="Output length for decode benchmark (default: 64)")
     parser.add_argument("--num-gpu-blocks", type=int, default=4, help="Number of GPU blocks (default: 4)")
+    parser.add_argument("--block-size", type=int, default=1024, help="KV cache block size (default: 1024)")
     parser.add_argument("--max-len", type=int, default=32*1024, help="Max model length (default: 32K)")
     parser.add_argument("--bench-decode", action="store_true", help="Run decode benchmark (default: prefill only)")
     parser.add_argument("--bench-all", action="store_true", help="Run both prefill and decode benchmarks")
@@ -75,6 +98,9 @@ def main():
     path = os.path.expanduser(args.model)
     max_len = args.max_len
 
+    # Enable MemoryObserver for communication stats
+    MemoryObserver._enabled = True
+
     # Setup policy configuration
     if args.enable_quest:
         sparse_policy = SparsePolicyType.QUEST
@@ -86,7 +112,7 @@ def main():
         sparse_policy = SparsePolicyType.FULL
         print("\n[Full Attention] baseline (no sparse)")
 
-    print(f"[Config] max_len={max_len}, num_gpu_blocks={args.num_gpu_blocks}")
+    print(f"[Config] max_len={max_len}, num_gpu_blocks={args.num_gpu_blocks}, block_size={args.block_size}")
 
     llm = LLM(
         path,
@@ -95,6 +121,7 @@ def main():
         max_num_batched_tokens=max_len,
         enable_cpu_offload=True,
         num_gpu_blocks=args.num_gpu_blocks,
+        kvcache_block_size=args.block_size,
         sparse_policy=sparse_policy,
         # Quest parameters
         sparse_topk_blocks=args.topk,

bench_vllm.py

@@ -1,5 +1,14 @@
 import os
-os.environ["VLLM_USE_V1"] = "1"
+import sys
+
+# Parse --use-v1 flag before importing vllm
+use_v1 = "--use-v1" in sys.argv
+if use_v1:
+    os.environ["VLLM_USE_V1"] = "1"
+    sys.argv.remove("--use-v1")
+else:
+    os.environ["VLLM_USE_V1"] = "0"
+
 import time
 from random import randint, seed
 from vllm import LLM, SamplingParams
@@ -44,24 +53,28 @@ def bench_prefill(llm, num_seqs, input_len):
 def main():
     import argparse
     parser = argparse.ArgumentParser(description="Benchmark vLLM performance (for comparison)")
+    parser.add_argument("--model", type=str, default="~/models/Llama-3.1-8B-Instruct",
+                        help="Model path (default: ~/models/Llama-3.1-8B-Instruct)")
     parser.add_argument("--input-len", type=int, default=None, help="Input length in tokens")
     parser.add_argument("--output-len", type=int, default=64, help="Output length for decode benchmark (default: 64)")
     parser.add_argument("--max-len", type=int, default=32*1024, help="Max model length (default: 32K)")
+    parser.add_argument("--gpu-util", type=float, default=0.9, help="GPU memory utilization (default: 0.9)")
+    parser.add_argument("--enforce-eager", action="store_true", help="Disable CUDA Graphs (use eager mode)")
     parser.add_argument("--bench-decode", action="store_true", help="Run decode benchmark (default: prefill only)")
     parser.add_argument("--bench-all", action="store_true", help="Run both prefill and decode benchmarks")
     args = parser.parse_args()
 
-    path = os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/")
+    path = os.path.expanduser(args.model)
     max_len = args.max_len
 
-    print(f"\n[vLLM] max_len={max_len}")
+    print(f"\n[vLLM] max_len={max_len}, gpu_util={args.gpu_util}, enforce_eager={args.enforce_eager}")
 
     llm = LLM(
         path,
-        enforce_eager=False,
+        enforce_eager=args.enforce_eager,
         max_model_len=max_len,
         max_num_seqs=128,
-        gpu_memory_utilization=0.9,
+        gpu_memory_utilization=args.gpu_util,
     )
 
     # Warmup

docs/bench_offload_results.md

@@ -9,8 +9,6 @@
 | GPU | NVIDIA A100-SXM4-80GB |
 | 模型 | Llama-3.1-8B-Instruct |
 | GPU slots | 4 |
-| Block size | 1024 tokens |
-| Chunk size | 2048 tokens |
 
 ## Sparse Policy 配置
 
@@ -21,69 +19,181 @@
 
 ## 测试结果
 
-### 32K 上下文
+### Block Size 4096 (推荐)
 
-| 策略 | 输入长度 | 耗时 | 吞吐量 | 相对性能 |
-|------|----------|------|--------|----------|
-| Full Attention | 32767 tok | 20.64s | **1587.74 tok/s** | baseline |
-| XAttention BSA | 32767 tok | 27.95s | **1172.33 tok/s** | 0.74x |
+#### GPU-only 模式
 
-### 128K 上下文
+| 上下文 | Full Attention | XAttention | 相对性能 |
+|--------|----------------|------------|----------|
+| 32K | 4863 tok/s | 5587 tok/s | **+14.9%** ✅ |
+| 64K | 3373 tok/s | 4766 tok/s | **+41.3%** ✅ |
 
-| 策略 | 输入长度 | 耗时 | 吞吐量 | 相对性能 |
-|------|----------|------|--------|----------|
-| Full Attention | 131071 tok | 237.18s | **552.63 tok/s** | baseline |
-| XAttention BSA | 131071 tok | 281.17s | **466.17 tok/s** | 0.84x |
+#### CPU Offload 模式 (优化后, 2026-01-28)
 
-### KV Cache 配置
+| 上下文 | Full Attention | XAttention | 相对性能 |
+|--------|----------------|------------|----------|
+| 32K | 4678 tok/s | 4398 tok/s | **-6.0%** |
+| 64K | 3331 tok/s | 3203 tok/s | **-3.8%** |
+| 128K | 2144 tok/s | 2196 tok/s | **+2.4%** ✅ |
 
-| 上下文 | GPU Memory | CPU Memory | Total |
-|--------|------------|------------|-------|
-| 32K | 512 MB (4 blocks) | 4096 MB (32 blocks) | 4608 MB |
-| 128K | 512 MB (4 blocks) | 16384 MB (128 blocks) | 16896 MB |
+#### CPU Offload 模式 (优化前, 2026-01-27)
 
-## 分析
+| 上下文 | Full Attention | XAttention | 相对性能 |
+|--------|----------------|------------|----------|
+| 32K | 4648 tok/s | 4002 tok/s | **-13.9%** ❌ |
+| 64K | 3329 tok/s | 2642 tok/s | **-20.6%** ❌ |
+| 128K | 2122 tok/s | 867 tok/s | **-59.1%** ❌ |
 
-### XAttention 性能特点
+### Block Size 256 (小 block 测试)
 
-1. **32K 上下文**: XAttention 比 Full 慢 26%
-2. **128K 上下文**: XAttention 比 Full 慢 16%
+#### CPU Offload 模式 (64K)
 
-随着上下文增长，XAttention 的相对性能有所提升（74% → 84%），但仍未超过 Full Attention。
+| 策略 | 耗时 | 吞吐量 | 相对性能 |
+|------|------|--------|----------|
+| Full Attention | 401.04s | 163.41 tok/s | baseline |
+| XAttention BSA | 390.35s | 167.89 tok/s | **+2.7%** ✅ |
 
-### 原因分析
+### Block Size 1024 (历史测试)
 
-1. **tau=0.95 阈值较高**: 需要覆盖 95% 累积注意力，实际跳过的 block 较少
-2. **估计开销**: `xattn_estimate_chunked` 需要对每个 chunk 计算稀疏 mask
-3. **BSA kernel overhead**: Block sparse kernel 有额外的 mask 处理和索引开销
-4. **Offload 瓶颈**: CPU→GPU 传输是主要瓶颈，稀疏注意力节省的是计算而非传输
+#### CPU Offload 模式
 
-### 适用场景
+| 上下文 | Full Attention | XAttention | 相对性能 |
+|--------|----------------|------------|----------|
+| 32K | 1587.74 tok/s | 1172.33 tok/s | -26% |
+| 128K | 552.63 tok/s | 466.17 tok/s | -16% |
 
-XAttention BSA 更适合以下场景：
-- 更长的上下文（256K+），稀疏收益更明显
-- 计算密集型任务（非 offload 模式），传输不是瓶颈
-- 较低的 tau 阈值（如 0.8），增加稀疏性
+## 关键发现
+
+### 1. GPU-only vs CPU Offload 模式差异
+
+| 模式 | XAttention 效果 | 原因 |
+|------|-----------------|------|
+| **GPU-only** | ✅ 显著加速 (+15% ~ +41%) | 计算是瓶颈，稀疏注意力减少 FLOPs |
+| **CPU Offload (优化后)** | ✅ 长上下文略有收益 | estimate_block_size 优化减少估计开销 |
+| **CPU Offload (优化前)** | ❌ 性能下降 (-14% ~ -59%) | 传输是瓶颈，稀疏估计增加额外开销 |
+
+### 2. Block Size 对性能的影响
+
+| Block Size | 64K Full (Offload) | 特点 |
+|------------|-------------------|------|
+| 4096 | 3329 tok/s | ⭐ 最佳性能 |
+| 1024 | ~1500 tok/s | 中等 |
+| 256 | 163 tok/s | 极慢（20x 下降） |
+
+**原因**: 更小的 block = 更多的 blocks = 更多 H2D 传输开销
+
+### 3. XAttention 在小 Block Size 下反转
+
+当 block size = 256 时，XAttention 反而略有优势 (+2.7%)：
+- 256 个 blocks (vs 16 个 @ 4096)
+- 稀疏跳过的 blocks 比例更明显
+- 但绝对性能极差，不推荐使用
+
+### 4. estimate_block_size 优化效果 (2026-01-28)
+
+```
+Offload 模式 XAttention 相对性能变化:
+         优化前    优化后    改进
+32K:    -13.9%   -6.0%    +7.9pp
+64K:    -20.6%   -3.8%    +16.8pp
+128K:   -59.1%   +2.4%    +61.5pp ✅
+```
+
+优化内容：
+- `estimate_block_size` 从 4096 改为 1024
+- `softmax_fuse_block_sum` kernel 时间从 48% 降到 1% (44x 加速)
+- 选择策略从 mask + voting 改为 score + threshold
+
+优化后结论：
+- **128K 长上下文 XAttention 反超 Full Attention**
+- 短上下文仍有少量开销，但已显著减少
+
+## 结论
+
+### 推荐配置 (优化后, 2026-01-28)
+
+| 场景 | 推荐策略 | Block Size |
+|------|----------|------------|
+| GPU-only (VRAM 充足) | XAttention | 4096 |
+| CPU Offload (128K+) | XAttention | 4096 |
+| CPU Offload (32K-64K) | Full Attention 或 XAttention | 4096 |
+
+### XAttention 适用条件 (优化后)
+
+✅ **适合**:
+- GPU-only 模式（计算密集）
+- CPU Offload + 长上下文（128K+）有正向收益
+- 长上下文（64K+）收益更大
+
+⚠️ **中性**:
+- CPU Offload + 中等上下文（32K-64K）：略慢 3-6%，可接受
+
+❌ **不推荐**:
+- 短上下文（<32K）收益不明显
 
 ## 运行命令
 
 ```bash
-# Full Attention (32K)
-CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 32768
+# GPU-only 模式
+CUDA_VISIBLE_DEVICES=0 python bench.py --max-len 65536 --block-size 4096 --gpu-util 0.7
+CUDA_VISIBLE_DEVICES=0 python bench.py --max-len 65536 --block-size 4096 --gpu-util 0.7 --policy xattn
 
-# XAttention BSA (32K)
-CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 32768 --enable-xattn
+# CPU Offload 模式 (推荐 block-size 4096)
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 65536 --block-size 4096
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 65536 --block-size 4096 --enable-xattn
 
-# Full Attention (128K)
-CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 131072
-
-# XAttention BSA (128K)
-CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 131072 --enable-xattn
+# CPU Offload 模式 (小 block size 测试)
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 65536 --block-size 256
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 65536 --block-size 256 --enable-xattn
 
 # 调整 XAttention 参数
 CUDA_VISIBLE_DEVICES=0 python bench_offload.py --enable-xattn --xattn-threshold 0.8 --xattn-stride 16
 ```
 
+## FlashInfer Merge 优化 (2026-01-28)
+
+将 Triton 实现的 `merge_attention_outputs` 替换为 FlashInfer 的 `cascade.merge_state`。
+
+### 性能对比 (Full Attention, block-size 4096)
+
+| 上下文 | Triton merge | FlashInfer merge | 提升 |
+|--------|--------------|------------------|------|
+| 32K | 4678 tok/s | 4717 tok/s | **+0.8%** |
+| 64K | 3331 tok/s | 3411 tok/s | **+2.4%** |
+| 128K | 2144 tok/s | 2178 tok/s | **+1.6%** |
+
+### 关键发现
+
+1. **端到端提升有限**（0.8% ~ 2.4%）：merge 操作不是主要瓶颈
+   - H2D 传输占主导（64K 传输 64GB）
+   - Attention 计算是另一主要耗时
+   - Merge 在总耗时中占比很小
+
+2. **Merge kernel 单独对比**（长序列时 FlashInfer 优势明显）：
+
+| seq_len | heads | Triton (ms) | FlashInfer (ms) | Speedup |
+|---------|-------|-------------|-----------------|---------|
+| 4096 | 32 | 0.129 | 0.087 | **1.49x** |
+| 8192 | 32 | 0.251 | 0.147 | **1.70x** |
+| 16384 | 32 | 0.499 | 0.274 | **1.82x** |
+
+3. **短序列 FlashInfer 反而慢**：格式转换开销（squeeze, transpose, contiguous）
+
+### 技术细节
+
+- **LSE 格式差异**：FlashInfer 使用 log2，flash_attn 使用 ln
+- **转换系数**：`LOG2_E = 1.4427`（ln → log2），`LN_2 = 0.6931`（log2 → ln）
+- **FlashInfer attention JIT 问题**：CUDA 版本兼容性问题，仅使用 merge_state
+
+### 代码位置
+
+- `nanovllm/ops/chunked_attention.py`: `merge_attention_outputs_flashinfer()`
+- `nanovllm/kvcache/sparse/full_policy.py`: 3 处 import 更新
+- `nanovllm/kvcache/sparse/xattn_bsa.py`: 1 处 import 更新
+
 ## 更新记录
 
+- 2026-01-28: **FlashInfer merge 替换 Triton merge**，端到端提升 0.8% ~ 2.4%
+- 2026-01-28: **estimate_block_size 优化后重新测试**，128K XAttention 反超 Full (+2.4%)
+- 2026-01-27: 添加 GPU-only vs Offload 对比，block size 影响分析
 - 2026-01-27: 初始测试，Llama-3.1-8B-Instruct, A100 80GB

docs/changelog_2026-02-05.md

@@ -0,0 +1,94 @@
+# Changelog 2026-02-05
+
+## Bug Fixes
+
+### XAttention Offload GQA Buffer OOM Fix
+
+**Issue**: `docs/issue_xattn_offload_gqa_buffer_oom.md`
+
+**Problem**: 在 XAttention BSA + CPU Offload 模式下，`alloc_policy_metadata()` 分配了只有 GPU-only 模式才需要的 GQA expansion buffers (`_k_expanded`, `_v_expanded`)，导致 24GB GPU (RTX 3090) 上 OOM。
+
+**Root Cause**:
+- GQA buffer 大小: `2 × num_heads × max_seq_len × head_dim × dtype_size`
+- 对于 1M max_seq_len: 2 × 32 × 1048576 × 128 × 2 = **16 GB**
+- Offload 模式的 `compute_chunked_prefill()` 不需要这些 buffer
+
+**Fix** (commit `11a867f`):
+1. `nanovllm/kvcache/sparse/policy.py`: 基类添加 `enable_cpu_offload` 参数
+2. `nanovllm/kvcache/sparse/xattn_bsa.py`: offload 模式跳过 GQA buffer 分配
+3. `nanovllm/engine/model_runner.py`: 传入 `enable_cpu_offload` 参数
+
+**Memory Savings**:
+| max_model_len | 修复前 | 修复后 |
+|---------------|--------|--------|
+| 72K | +1.1 GB | 0 GB |
+| 1M | +16 GB | 0 GB |
+
+**Verification**:
+```bash
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_64k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 72000 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA
+```
+- 日志显示: `[XAttn] Offload mode: skipping GQA expansion buffers`
+- 测试结果: 100% 准确率
+
+---
+
+## Code Cleanup
+
+### Tests Directory Cleanup
+
+**Commits**: `a709551`, `2b61c5a`, `d35dd76`
+
+删除了 16 个冗余/过时的测试文件，保留核心测试：
+
+**保留的文件** (4 个):
+| 文件 | 用途 |
+|------|------|
+| `test_ruler.py` | 核心 RULER benchmark (13 tasks, 100 samples) |
+| `test_xattn_estimate_alignment.py` | XAttn kernel 一致性验证 |
+| `utils.py` | 共享工具函数 |
+| `__init__.py` | 包标记 |
+
+**删除的文件** (16 个, -4306 行):
+
+| 类别 | 文件 | 删除原因 |
+|------|------|----------|
+| XAttn 测试 | `test_xattn_bsa.py` | 功能被 test_ruler 覆盖 |
+| | `test_xattn_chunked.py` | 与 estimate_chunked 重复 |
+| | `test_xattn_estimate_chunked.py` | chunked prefill 验证 |
+| | `test_xattn_kernels.py` | Triton kernel 单元测试 |
+| | `test_xattn_kv_chunking_batch.py` | batch 验证 |
+| Needle 测试 | `test_needle.py` | 被 test_ruler NIAH 任务覆盖 |
+| | `test_needle_ref.py` | HF 参考实现 |
+| CUDA Graph | `test_chunk_attention_graph.py` | 被 graph_reuse 取代 |
+| | `test_chunk_attention_graph_reuse.py` | 实验性功能 |
+| | `test_cudagraph_memory.py` | 内存分析工具 |
+| 其他 | `test_gpuonly_density_alignment.py` | GPU-only 密度测试 |
+| | `test_hierarchical_estimate.py` | 分层估计测试 |
+| | `test_quest_policy.py` | Quest 策略测试 |
+| | `test_sequential.py` | 状态隔离测试 |
+| | `bench_estimate_block_size.py` | 性能 benchmark |
+| | `modeling_qwen3.py` | Qwen3 参考模型 |
+
+**Note**: 所有删除的文件可从 git 历史恢复：
+```bash
+git checkout <commit-hash>^ -- tests/<filename>
+```
+
+---
+
+## Summary
+
+| 类型 | 数量 | 影响 |
+|------|------|------|
+| Bug Fix | 1 | 节省 16GB 显存 (1M seq) |
+| 文件删除 | 16 | -4306 行代码 |
+| 新增文档 | 1 | 本文件 |

docs/estimate_block_size_performance.md

@@ -0,0 +1,258 @@
+# Estimate Block Size 性能分析
+
+本文档记录 XAttention estimate 阶段中 `block_size` 参数对 `softmax_fuse_block_sum` kernel 性能的影响。
+
+## 问题背景
+
+当前 `select_blocks` 中的 estimate 过程使用全局的 `kvcache_block_size`（通常为 4096）：
+
+```python
+# xattn_bsa.py: select_blocks()
+block_size = ctx.block_size  # 来自 kvcache_manager.block_size (4096)
+reshaped_block_size = block_size // self.stride  # 4096/8 = 512
+
+block_sums = softmax_fuse_block_sum(
+    attn_scores,
+    reshaped_block_size,  # 512 - 性能最差点！
+    ...
+)
+```
+
+这导致 `softmax_fuse_block_sum` kernel 使用 `reshaped_block_size=512`，而这正是性能曲线的最差点。
+
+## Benchmark 结果
+
+### 测试配置
+
+- GPU: NVIDIA A100-SXM4-80GB
+- NUM_HEADS: 32
+- HEAD_DIM: 128
+- STRIDE: 8
+- 测试脚本: `tests/bench_estimate_block_size.py`
+
+### softmax_fuse_block_sum 性能数据
+
+| block_size | reshaped | 16K context | 32K context | 64K context |
+|------------|----------|-------------|-------------|-------------|
+| 64 | 8 | 4.86ms | 18.36ms | 70.83ms |
+| 128 | 16 | 0.83ms | 3.12ms | 16.83ms |
+| 256 | 32 | 0.63ms | 2.41ms | 11.24ms |
+| 512 | 64 | **0.38ms** | **1.52ms** | 9.54ms |
+| 1024 | 128 | 0.42ms | 1.54ms | **6.01ms** |
+| 2048 | 256 | 1.08ms | 3.24ms | 12.81ms |
+| **4096** | **512** | 9.66ms | 25.36ms | **95.32ms** |
+
+### 性能曲线
+
+```
+softmax_fuse_block_sum 耗时 (64K context):
+
+block_size=64   ████████████████████████████████████ 70.83ms
+block_size=128  ████████ 16.83ms
+block_size=256  █████ 11.24ms
+block_size=512  ████ 9.54ms
+block_size=1024 ███ 6.01ms  ◀── 最优点
+block_size=2048 ██████ 12.81ms
+block_size=4096 ████████████████████████████████████████████████ 95.32ms ◀── 当前使用
+```
+
+### 关键发现
+
+1. **性能呈 U 型曲线**：太小和太大的 block_size 都会导致性能下降
+2. **最优点在 512-1024**：对应 `reshaped_block_size` 64-128
+3. **当前配置 (4096) 是最差点**：95.32ms vs 最优 6.01ms，**慢 15.85x**
+
+## 性能曲线解释
+
+```
+Performance (耗时)
+    │
+    │  ▲ 太小:
+    │ /   - output blocks 数量多 (q_len / block_size)
+    │/    - grid 调度开销大
+    │     - 每个 thread block 工作量小
+    │   ┌─────────┐
+    │  /   最优    \
+    │ /    区域     \ ▲ 太大:
+    │/               \   - block_size 作为 tl.constexpr
+    │                 \  - 寄存器压力增大 (可能 spill)
+    │                  \ - shared memory 不足
+    │                   \- L1 cache 效率下降
+    └──────────────────────────────────→ block_size
+      64  128  256  512 1024 2048 4096
+                    ↑
+              最优点 (512-1024)
+```
+
+### Triton Kernel 内部分析
+
+`softmax_fuse_block_sum_kernel` 中的关键约束：
+
+```python
+# 每个 thread block 处理的数据
+offs_q = tl.arange(0, block_size)  # block_size 个元素
+m_i = tl.zeros([block_size], dtype=tl.float32)  # 寄存器分配
+
+# reshape 操作
+X = tl.reshape(X, (block_size, segment_size // block_size, block_size))
+# 当 block_size=512, segment_size=512 时 → (512, 1, 512) 的 3D tensor
+```
+
+当 `block_size` 过大时：
+- 每个 thread block 需要更多寄存器
+- `tl.arange(0, block_size)` 生成更大的向量
+- reshape 操作的内存访问模式变差
+
+## 优化建议
+
+### 方案 1: 固定 estimate block_size
+
+在 `select_blocks` 中使用固定的小 block_size 进行估计：
+
+```python
+# 建议修改
+ESTIMATE_BLOCK_SIZE = 1024  # 或 512，而非 ctx.block_size
+
+reshaped_block_size = ESTIMATE_BLOCK_SIZE // self.stride  # 128
+```
+
+**优点**：简单直接，预期提升 15x
+**缺点**：estimate 的 block 粒度与 CPU block 不一致，需要映射
+
+### 方案 2: 两级 block 结构
+
+- 外层使用 `kvcache_block_size` (4096) 管理 CPU blocks
+- 内层使用 `estimate_block_size` (1024) 进行估计
+- 估计结果聚合回 CPU block 粒度
+
+### 方案 3: 自适应 block_size
+
+根据 context length 动态选择 estimate block_size：
+
+| Context Length | Recommended block_size |
+|----------------|------------------------|
+| < 16K | 512 |
+| 16K - 64K | 1024 |
+| > 64K | 1024 |
+
+## 与实际 Profiling 的对比
+
+### Nsys Profiling 数据 (64K context, block_size=4096)
+
+| 阶段 | 时间占比 | 说明 |
+|------|----------|------|
+| softmax_fuse_block_sum | **48.1%** | 最后一个 chunk |
+| flash_fwd_kernel | 30.7% | 实际 attention 计算 |
+| flat_group_gemm | 3.5% | estimate GEMM |
+
+### 预期优化效果
+
+如果将 estimate block_size 从 4096 改为 1024：
+
+| 指标 | 当前 (4096) | 优化后 (1024) | 提升 |
+|------|-------------|---------------|------|
+| softmax kernel | 95.32ms | 6.01ms | **15.85x** |
+| estimate 阶段占比 | 48.1% | ~5% | 显著降低 |
+| 总体 prefill 时间 | ~2s (最后chunk) | ~1.1s | ~1.8x |
+
+## 测试命令
+
+```bash
+# 运行 benchmark
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/bench_estimate_block_size.py --gpu 0
+
+# 指定单个 context length
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/bench_estimate_block_size.py --gpu 0 --ctx-len 65536
+```
+
+## 相关文件
+
+| 文件 | 说明 |
+|------|------|
+| `nanovllm/kvcache/sparse/xattn_bsa.py` | XAttention BSA Policy 实现 |
+| `nanovllm/ops/xattn.py` | Triton kernels |
+| `tests/bench_estimate_block_size.py` | 性能测试脚本 |
+| `docs/xattn_performance_analysis.md` | XAttention 整体性能分析 |
+
+## 分级求和方案 (Hierarchical Block Sum)
+
+使用小的 `estimate_block_size=1024` 计算细粒度 block_sums，然后聚合到 CPU block 级别 (4096)。
+
+### 数学等价性
+
+```
+方案1 (block_size=4096): softmax_fuse_block_sum → [1, heads, 1, 1]
+方案2 (block_size=1024): softmax_fuse_block_sum → [1, heads, 4, 4] → sum → [1, heads]
+
+验证结果: Max difference = 0.0 ✅ 完全等价
+```
+
+### 验证代码
+
+`tests/test_hierarchical_estimate.py` - 纯 torch + xattn kernels 实现
+
+### 性能提升
+
+| 指标 | 当前 (4096) | 优化后 (1024) | 提升 |
+|------|-------------|---------------|------|
+| softmax kernel | 12.07 ms | 0.29 ms | **41x** |
+| 端到端 estimate | 95 ms | ~6 ms | **15x** |
+
+## ⚠️ 选择策略变更
+
+**重要**: 分级求和方案使用新的选择策略：
+
+| 特性 | 原策略 (mask + voting) | 新策略 (score + threshold) |
+|------|------------------------|----------------------------|
+| 输入 | `[batch, heads, q_blocks, k_blocks]` | `[batch, heads, num_cpu_blocks]` |
+| 选择粒度 | Per-q-block | Per-chunk |
+| 聚合方式 | majority voting | threshold on scores |
+
+新策略更简洁，直接利用分级求和产生的 score，避免了 mask 生成和 voting 的复杂逻辑。
+
+## 实现状态 ✅ (2026-01-28)
+
+### 已实现
+
+分级求和方案已在 `xattn_bsa.py` 中实现：
+
+```python
+class XAttentionBSAPolicy:
+    def __init__(self, ..., estimate_block_size: int = 1024):
+        self.estimate_block_size = estimate_block_size  # 新参数
+
+    def select_blocks(self, ...):
+        # Step 2: Hierarchical softmax_fuse_block_sum
+        reshaped_est_bs = estimate_bs // self.stride  # 1024/8 = 128
+        block_sums_fine = softmax_fuse_block_sum(attn_scores, reshaped_est_bs, ...)
+
+        # Step 3: Aggregate to CPU block level
+        block_sums_coarse = block_sums_fine.view(..., num_cpu_blocks, ratio).sum(dim=-1)
+        cpu_block_scores = block_sums_coarse.sum(dim=2)
+
+        # Step 4: Score + threshold selection (replaces mask + voting)
+        scores_per_block = cpu_block_scores.mean(dim=(0, 1))
+        # ... cumulative threshold selection
+```
+
+### 实测结果 (Nsys Profiling)
+
+| Kernel | 优化前 | 优化后 | 改进 |
+|--------|--------|--------|------|
+| softmax_fuse_block_sum 占比 | 48.1% | **1.1%** | **44x** |
+| softmax_fuse_block_sum 平均时间 | ~2ms | 489us | **4x** |
+
+### 端到端性能 (32K context)
+
+| 指标 | FULL Policy | XATTN Policy | 改进 |
+|------|-------------|--------------|------|
+| Prefill throughput | 3511 tok/s | 3695 tok/s | +5% |
+| TTFT | 9327 ms | 8863 ms | -5% |
+
+## 结论
+
+当前 estimate 阶段使用全局 `kvcache_block_size=4096` 导致 `softmax_fuse_block_sum` kernel 性能处于最差点。通过将 estimate block_size 改为 512-1024，可以获得 **15x** 的性能提升，显著降低 estimate 阶段的开销。
+
+**⚠️ 重要变更**: 选择策略从 `mask + majority voting` 改为 `score + threshold`，更简洁且更直接。

docs/gpu_only_sparse_integration.md

@@ -0,0 +1,77 @@
+# GPU-only Sparse Policy 整合
+
+本文档记录将 sparse attention 策略整合到 GPU-only 模式的过程和性能对比。
+
+## 背景
+
+当前 sparse policy（Quest、XAttention）仅在 CPU offload 路径中实现。目标是将其扩展到 GPU-only 模式，以提升长上下文场景下的性能。
+
+## 基准性能（优化前）
+
+**测试环境**:
+- GPU: NVIDIA A100-SXM4-80GB
+- 模型: Llama-3.1-8B-Instruct
+- 上下文长度: 32K tokens
+- 日期: 2026-01-27
+
+### Prefill Benchmark (32K context)
+
+| 模式 | Throughput | Time | KV Cache 分配 |
+|------|------------|------|---------------|
+| **GPU-only (Full Attention)** | 4869.67 tok/s | 6.73s | 438 blocks (56GB GPU) |
+| CPU Offload (Full Attention) | 1500.29 tok/s | 21.84s | 4 blocks GPU + 32 blocks CPU |
+
+**性能比**: GPU-only 比 CPU Offload 快 **3.2x**
+
+### 配置详情
+
+**GPU-only 模式**:
+```bash
+CUDA_VISIBLE_DEVICES=0 python bench.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --max-len 32768
+```
+
+**CPU Offload 模式**:
+```bash
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --max-len 32768
+```
+
+### KV Cache 配置
+
+| 参数 | GPU-only | CPU Offload |
+|------|----------|-------------|
+| block_size | 1024 tokens | 1024 tokens |
+| per-token KV | 128 KB | 128 KB |
+| per-block KV | 128 MB | 128 MB |
+| GPU blocks | 438 | 4 |
+| CPU blocks | 0 | 32 |
+| Total memory | 56 GB | 4.6 GB |
+
+## 目标
+
+将以下 sparse policy 整合到 GPU-only 模式：
+
+| Policy | 阶段 | 描述 |
+|--------|------|------|
+| Quest | Decode | Top-K block selection based on query-key scores |
+| XAttention BSA | Prefill | Block sparse attention with cumulative threshold |
+
+## 实现进度
+
+- [ ] 分析现有 sparse policy 代码结构
+- [ ] 设计 GPU-only sparse policy 接口
+- [ ] 实现 GPU-only Quest decode
+- [ ] 实现 GPU-only XAttention prefill
+- [ ] 性能测试和对比
+
+## 优化后性能
+
+*待测试*
+
+| 模式 | Throughput | Speedup vs Full |
+|------|------------|-----------------|
+| GPU-only + Quest (decode) | TBD | TBD |
+| GPU-only + XAttn (prefill) | TBD | TBD |

docs/gpu_only_xattn_guide.md

@@ -0,0 +1,296 @@
+# GPU-Only XAttention 指南
+
+本文档介绍 GPU-only 模式下 XAttention BSA 的实现、内存优化和性能分析。
+
+## 概述
+
+GPU-only 模式下，所有 KV cache 存储在 GPU 上，无需 CPU offload。XAttention 通过稀疏注意力加速 prefill 阶段。
+
+### 执行路径对比
+
+| 模式 | Prefill 方法 | Decode 方法 | KV 存储 |
+|------|-------------|-------------|---------|
+| GPU-only Full | `compute_prefill()` | `compute_decode()` | GPU |
+| GPU-only XAttn | `compute_prefill()` | `compute_decode()` | GPU |
+| CPU Offload | `compute_chunked_prefill()` | `compute_chunked_decode()` | CPU + GPU |
+
+## 架构设计
+
+### SparsePolicy 接口
+
+```python
+class SparsePolicy:
+    # GPU-only 方法
+    def compute_prefill(self, q, k, v, ...) -> Tensor
+    def compute_decode(self, q, k_cache, v_cache, ...) -> Tensor
+
+    # CPU Offload 方法
+    def compute_chunked_prefill(self, q, k, v, ...) -> Tensor
+    def compute_chunked_decode(self, q, ...) -> Tensor
+
+    # 初始化方法
+    def initialize(self, num_layers, ...) -> None           # CPU offload metadata
+    def alloc_policy_metadata(self, num_heads, ...) -> None  # GPU-only buffers
+```
+
+### XAttentionBSAPolicy 实现
+
+```
+GPU-only Prefill 流程:
+┌─────────────────────────────────────────────────────────────┐
+│  1. GQA 扩展 (使用预分配 buffer)                              │
+│     K: [seq, kv_heads, dim] → K_exp: [1, heads, seq, dim]   │
+│                                                              │
+│  2. XAttention 估计                                          │
+│     flat_group_gemm_fuse_reshape_kernel (Q@K^T)             │
+│     softmax_fuse_block_sum_kernel (block 重要性)             │
+│     → sparse mask                                            │
+│                                                              │
+│  3. BSA 稀疏注意力                                           │
+│     flash_fwd_block_kernel (只计算选中的 blocks)             │
+│     → output                                                 │
+└─────────────────────────────────────────────────────────────┘
+```
+
+## 内存预分配
+
+### 问题背景
+
+XAttention 的 `compute_prefill()` 需要 GQA 扩展：
+
+```python
+# 之前: 动态分配 (~2GB for 64K)
+K_exp = K.repeat_interleave(num_groups, dim=1)  # 分配 1
+k_bsa = k.repeat_interleave(num_groups, dim=1)  # 分配 2 (重复!)
+```
+
+每次 prefill 都动态分配，导致：
+- 内存碎片
+- 分配延迟
+- 可能 OOM
+
+### 解决方案: alloc_policy_metadata()
+
+在框架初始化时预分配 buffer：
+
+```python
+class XAttentionBSAPolicy(SparsePolicy):
+    def alloc_policy_metadata(self, num_heads, num_kv_heads, head_dim,
+                               max_seq_len, dtype, device):
+        # 预分配 GQA 扩展 buffer
+        shape = (1, num_heads, max_seq_len, head_dim)
+        self._k_expanded = torch.empty(shape, dtype=dtype, device=device)
+        self._v_expanded = torch.empty(shape, dtype=dtype, device=device)
+
+    def compute_prefill(self, q, k, v, ...):
+        seq_len = k.shape[0]
+        # 使用预分配 buffer 的 slice
+        K_exp = self._k_expanded[:, :, :seq_len, :]
+        # 原地 GQA 扩展
+        K_exp.view(...).copy_(K.unsqueeze(2).expand(...))
+        # 复用同一 buffer 给 BSA
+        k_bsa = K_exp.squeeze(0).transpose(0, 1)
+```
+
+### 内存使用
+
+| 序列长度 | 预分配大小 | 说明 |
+|---------|-----------|------|
+| 32K | 512 MB | `2 * 32 * 32768 * 128 * 2 bytes` |
+| 64K | 1024 MB | `2 * 32 * 65536 * 128 * 2 bytes` |
+
+优化效果：
+- 之前: ~2GB 动态分配 (xattn_estimate + BSA 各一次)
+- 之后: ~1GB 预分配 (复用同一 buffer)
+
+### 框架集成
+
+```python
+# model_runner.py - allocate_kv_cache()
+def allocate_kv_cache(self):
+    # ... KV cache 分配 ...
+
+    # GPU-only 模式: 预分配 policy buffers
+    if not config.enable_cpu_offload:
+        self.kvcache_manager.sparse_policy.alloc_policy_metadata(
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=head_dim,
+            max_seq_len=config.max_model_len,
+            dtype=dtype,
+            device=torch.device("cuda"),
+        )
+```
+
+## 性能分析
+
+### 32K Prefill 性能
+
+| Policy | Throughput | 相对提升 |
+|--------|------------|----------|
+| Baseline | 4880 tok/s | - |
+| Full | 4892 tok/s | +0.2% |
+| **XAttention** | **5602 tok/s** | **+15%** |
+
+### 64K Prefill 性能
+
+| Policy | Throughput | 相对提升 |
+|--------|------------|----------|
+| Baseline | 3386 tok/s | - |
+| Full | 3355 tok/s | -0.9% |
+| **XAttention** | **4775 tok/s** | **+41%** |
+
+### Kernel 时间分解 (32K)
+
+**XAttention:**
+```
+FFN GEMM:           3219 ms  (54%)
+BSA Attention:      1231 ms  (21%)
+XAttn Estimation:    415 ms  (7%)
+Other:              1020 ms  (18%)
+─────────────────────────────
+Total:              5885 ms
+```
+
+**Full:**
+```
+FFN GEMM:           3244 ms  (48%)
+Dense Attention:    2861 ms  (43%)
+Other:               595 ms  (9%)
+─────────────────────────────
+Total:              6700 ms
+```
+
+### 加速来源
+
+```
+Dense Attention:    2861 ms
+BSA Attention:      1231 ms  (节省 1630 ms, -57%)
+XAttn Estimation:    415 ms  (额外开销)
+─────────────────────────────
+净节省:             1215 ms  (42% attention 时间)
+```
+
+## CUDA Graph 限制
+
+### 为什么 Prefill 不能用 CUDA Graph
+
+CUDA Graph 要求所有操作在 capture 时确定：
+
+| 必须固定 | Prefill 的情况 |
+|---------|---------------|
+| Tensor 形状 | seq_len 可变 (1 ~ max_model_len) |
+| Kernel grid | 依赖 seq_len |
+| 内存地址 | 中间 tensor 大小变化 |
+
+```python
+# 不同请求的 seq_len 不同
+request_1: prefill(seq_len=1024)   # grid=(8, 32, 1)
+request_2: prefill(seq_len=32768)  # grid=(256, 32, 1)
+```
+
+### Decode 可以用 CUDA Graph
+
+```python
+# Decode 每次只处理 1 token
+q: [batch_size, 1, heads, dim]  # 形状固定
+```
+
+nanovllm 为每个 batch_size 预先 capture 一个 graph：
+
+```python
+def capture_cudagraph(self):
+    for batch_size in [1, 2, 4, 8, ...]:
+        with torch.cuda.graph(g):
+            self.run_model(dummy_input, is_prefill=False)
+        self.graphs[batch_size] = g
+```
+
+### Nsys Profile 结果
+
+```
+XAttention 32K Prefill:
+  Total kernels: 41,904
+  Non-graph: 41,904 (100%)
+  Graph: 0
+
+Full 32K Prefill:
+  Total kernels: 35,308
+  Non-graph: 35,308 (100%)
+  Graph: 0
+```
+
+**两者都是 100% NON-GRAPH**，这是 prefill 的本质特性。
+
+## Profiling 工具
+
+### 使用 profile.sh
+
+```bash
+# XAttention 32K
+bash scripts/profile.sh --max-len 32768 --policy xattn
+
+# Full 32K
+bash scripts/profile.sh --max-len 32768 --policy full
+
+# 64K (需要降低 gpu-util)
+bash scripts/profile.sh --max-len 65536 --policy xattn --gpu-util 0.7
+```
+
+### 分析 nsys 结果
+
+```bash
+# 查看 kernel 统计
+nsys stats --report cuda_gpu_kern_sum results/nsys/<file>.nsys-rep
+
+# 用 sqlite 查询详细数据
+sqlite3 results/nsys/<file>.sqlite "
+SELECT
+    (SELECT value FROM StringIds WHERE id = shortName) as kernel,
+    COUNT(*) as count,
+    SUM(end-start)/1e6 as total_ms
+FROM CUPTI_ACTIVITY_KIND_KERNEL
+GROUP BY shortName
+ORDER BY total_ms DESC
+LIMIT 10
+"
+```
+
+## 使用指南
+
+### 启用 XAttention GPU-only
+
+```python
+from nanovllm import LLM
+from nanovllm.config import SparsePolicyType
+
+llm = LLM(
+    model_path,
+    max_model_len=32768,
+    sparse_policy=SparsePolicyType.XATTN_BSA,
+    gpu_memory_utilization=0.9,  # 64K 时可能需要降低
+)
+```
+
+### 命令行测试
+
+```bash
+# bench.py
+python bench.py --max-len 32768 --policy xattn
+
+# 64K 需要降低 gpu-util
+python bench.py --max-len 65536 --policy xattn --gpu-util 0.7
+```
+
+### 最佳实践
+
+1. **32K 及以下**: 使用默认 `gpu_memory_utilization=0.9`
+2. **64K**: 降低到 `gpu_memory_utilization=0.7`
+3. **Decode**: XAttention 自动 fallback 到 FullAttentionPolicy
+4. **Paged KV Cache**: 当 `block_tables` 存在时自动 fallback 到 flash_attn
+
+## 相关文档
+
+- [Sparse Policy 架构](sparse_policy_architecture.md)
+- [XAttention 算法详解](xattention_algorithm_guide.md)
+- [BSA 接口文档](block_sparse_attn_interface.md)

docs/gpuonly_density_alignment_test.md

@@ -0,0 +1,246 @@
+# Density Alignment Test Results
+
+验证 GPU-only 和 Offload 模式下三阶段 KV chunking 流程的正确性。
+
+## 测试配置
+
+### GPU-only 模式
+- **模型**: Qwen3-0.6B (28 layers, 16 heads, 8 KV heads, head_dim=128)
+- **Threshold**: 0.9
+- **Block Size**: 128 tokens (BSA block)
+- **Stride**: 8
+- **Chunk Size**: 16384 tokens
+
+### Offload 模式
+- **模型**: Llama-3.1-8B-Instruct (32 layers, 32 heads, 8 KV heads, head_dim=128)
+- **Threshold**: 0.9
+- **Block Size**: 128 tokens (BSA block)
+- **Stride**: 4
+- **Chunk Size**: 4096 tokens
+
+## 三阶段 KV Chunking 对齐测试 (2026-02-02)
+
+### 测试目的
+
+验证 `xattn_estimate` 高层 API 与手动实现的三阶段 KV chunking 流程是否完全一致。
+
+### 三阶段流程
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│  Stage 1: softmax_compute_partial_stats                     │
+│    └── 每个 KV chunk 独立计算 partial stats (m_i, l_i)       │
+│                                                             │
+│  Stage 2: merge_softmax_stats                               │
+│    └── Host 端合并所有 chunks: (m_global, l_global)          │
+│                                                             │
+│  Stage 3: softmax_normalize_and_block_sum                   │
+│    └── 使用全局 stats 归一化并计算 block sums                 │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### 测试结果
+
+#### CHUNK_SIZE = 16384 (默认)
+
+| Context | Tokens | Q Chunks | KV Chunks | Density | Mask 差异 | attn_sums 差异 | 结果 |
+|---------|--------|----------|-----------|---------|-----------|----------------|------|
+| 4K | 3,692 | 1 | 1 | 63.84% | 0 | 0.0 | ✅ |
+| 8K | 7,892 | 1 | 1 | 64.98% | 0 | 0.0 | ✅ |
+| 16K | 15,689 | 1 | 1 | 61.63% | 0 | 0.0 | ✅ |
+| 32K | 32,485 | 2 | 2 | 50.21% | 0 | 0.0 | ✅ |
+| **64K** | **64,891** | **4** | **4** | **37.00%** | **0** | **0.0** | ✅ |
+
+#### CHUNK_SIZE = 4096 (更多 chunks)
+
+| Context | Tokens | Q Chunks | KV Chunks | Density | xattn_estimate vs KV chunking | 结果 |
+|---------|--------|----------|-----------|---------|-------------------------------|------|
+| 4K | 3,692 | 1 | 1 | 63.84% | 0.000000 | ✅ |
+| 8K | 7,892 | 2 | 2 | 63.02% | 0.000000 | ✅ |
+| 16K | 15,689 | 4 | 4 | 60.08% | 0.000000 | ✅ |
+| 32K | 32,485 | 8 | 8 | 49.84% | 0.000000 | ✅ |
+| **64K** | **64,891** | **16** | **16** | **36.91%** | **0.000000** | ✅ |
+
+### 64K 详细验证 (CHUNK_SIZE=4096)
+
+64K 序列使用 chunk_size=4096 时产生 16×16 的 chunk 矩阵：
+
+```
+seq_len: 64891, q_chunk_num: 16, kv_chunk_num: 16
+
+Q chunk 0:  merged 16 KV chunks → attn_sum shape=[1, 32, 32, 512]
+Q chunk 1:  merged 16 KV chunks → attn_sum shape=[1, 32, 32, 512]
+...
+Q chunk 15: merged 16 KV chunks → attn_sum shape=[1, 32, 32, 512]
+```
+
+每个 Q chunk 需要合并 16 个 KV chunks 的 softmax stats，充分验证了 `merge_softmax_stats` 在大规模 chunk 合并场景下的正确性。
+
+### 验证指标
+
+| 指标 | 预期 | 所有长度实际结果 |
+|------|------|------------------|
+| attn_sums max diff | 0 | 0.000000e+00 |
+| attn_sums mean diff | 0 | 0.000000e+00 |
+| mask exact match | True | True |
+| density diff | 0% | 0.000000% |
+
+### 结论
+
+✅ **三阶段 KV chunking 与一次性处理完全等价，无任何精度损失。**
+
+- 当 seq_len < CHUNK_SIZE (16384)：单 chunk 处理
+- 当 seq_len >= CHUNK_SIZE：多 chunk 分段处理后合并，结果与一次性处理完全一致
+
+---
+
+## Offload 模式测试 (2026-02-02)
+
+使用 Offload 模式保存的真实 KV cache 数据进行测试。
+
+### 测试结果
+
+| 文件 | Tokens | Layer | Saved Density | Computed Density | Q/KV Chunks | 结果 |
+|------|--------|-------|---------------|------------------|-------------|------|
+| `qkv_3688.pt` | 3.7K | 3 | 38.34% | 38.34% | 1/1 | ✅ PASSED |
+| `qkv_7888.pt` | 7.9K | 3 | 29.06% | 27.56% | 2/2 | ✅ PASSED |
+| `qkv_15685.pt` | 15.7K | 3 | 19.77% | 18.60% | 4/4 | ✅ PASSED |
+| `qkv_32485.pt` | 32.5K | 5 | 15.71% | 15.62% | 8/8 | ✅ PASSED |
+| `qkv_64891.pt` | 64.9K | 3 | 11.09% | 11.09% | 16/16 | ✅ PASSED |
+
+### Layer 5 GPU-only 测试 (threshold=0.9)
+
+| 指标 | 结果 |
+|------|------|
+| Q/K shape | `[1, 16, 21001, 128]` (21K tokens) |
+| Density | 6.24% |
+| xattn_estimate vs KV chunking | 完全一致 (0.0000%) |
+| mask 差异 | 0 / 435600 blocks |
+| attn_sums 差异 | max=0.0, mean=0.0 |
+
+### 观察
+
+1. **Density 随 context 增长而降低**: 3.7K (38%) → 64.9K (11%)
+2. **xattn_estimate API 与三阶段 KV chunking 完全一致**: 所有长度差异均为 0.0000%
+3. **Saved density vs Computed density 略有差异**: 这是因为 saved density 可能在不同 chunk 下记录，累积计算方式略有不同
+
+---
+
+## 附录：xattn_bsa vs xattn_estimate 对齐
+
+| Context | Tokens | Layer 0 Density | Compute Density | Min Layer | 验证结果 |
+|---------|--------|-----------------|-----------------|-----------|----------|
+| 4k | 3,692 | 63.8% | 52.9% | Layer 3 (31.3%) | ✅ PASSED |
+| 8k | 7,892 | 65.0% | 52.5% | Layer 5 (27.3%) | ✅ PASSED |
+| 16k | 15,689 | 61.6% | 47.8% | Layer 5 (23.5%) | ✅ PASSED |
+| 32k | 32,485 | 50.2% | 40.1% | Layer 5 (18.5%) | ✅ PASSED |
+| 64k | 64,891 | 37.0% | 29.6% | Layer 5 (12.4%) | ✅ PASSED |
+
+## Density 计算公式
+
+### Total (分母)
+
+```python
+# Causal mask: Q block i 只能看到 K block 0 到 i
+causal_mask[i, j] = (j <= i + q_offset_blocks)
+
+# Total = causal 区域内的 block 数 × batch × heads
+total = causal_mask.sum() × batch × heads
+      = (n × (n+1) / 2) × 1 × 32  # n = valid_q_blocks
+```
+
+### Selected (分子)
+
+```python
+# 在 causal 区域内，被选中 (mask=True) 的 block 数量
+selected = (mask & causal_mask).sum()
+```
+
+### Density
+
+```python
+density = selected / total
+```
+
+## 观察
+
+1. **Density 随 context 增长而降低**: 4k (63.8%) → 64k (37.0%)，这是因为长序列中 attention 更加分散
+
+2. **Layer 5 通常是最稀疏的层**: 在所有长度测试中，Layer 5 的 density 最低
+
+3. **Layer 0 density 最高**: 第一层的 attention pattern 最密集，可能与 sink token 效应有关
+
+4. **Threshold=0.9 对应 ~50% density**: 在 32k context 下，threshold=0.9 意味着选择覆盖 90% attention 的 blocks，实际 density 约 50%
+
+## 使用方法
+
+### Step 1: 启用 debug 保存
+
+```python
+# nanovllm/kvcache/sparse/xattn_bsa.py
+_DEBUG_SAVE_MASK = True  # 改为 True
+```
+
+### Step 2: 运行 GPU-only 推理
+
+```bash
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+### Step 3: 运行 KV chunking 对齐验证
+
+```bash
+# 使用 GPU-only 保存的数据
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_xattn_estimate_alignment.py --gpuonly
+
+# 使用 Offload 模式保存的数据 (默认)
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_xattn_estimate_alignment.py
+
+# 指定自定义数据文件
+python tests/test_xattn_estimate_alignment.py --data-file /path/to/data.pt
+
+# 批量测试所有 Offload 数据
+for f in results/kvcache/qkv_*.pt; do
+    echo "Testing: $(basename $f)"
+    python tests/test_xattn_estimate_alignment.py --data-file "$f"
+done
+```
+
+### 批量测试所有长度
+
+```bash
+for ctx in 4k 8k 16k 32k 64k; do
+    case $ctx in
+        4k)  max_len=5000 ;;
+        8k)  max_len=9000 ;;
+        16k) max_len=17000 ;;
+        32k) max_len=34000 ;;
+        64k) max_len=65664 ;;
+    esac
+
+    echo "Testing $ctx..."
+    python tests/test_ruler.py \
+        --data-dir tests/data/ruler_$ctx \
+        --max-model-len $max_len \
+        --sparse-policy XATTN_BSA \
+        --num-samples 1 --quiet
+
+    python tests/test_xattn_estimate_alignment.py --gpuonly
+done
+```
+
+## 相关文件
+
+- `nanovllm/kvcache/sparse/xattn_bsa.py`: XAttention BSA Policy 实现
+- `nanovllm/ops/xattn.py`: xattn_estimate 函数及三阶段 KV chunking kernels
+- `tests/test_xattn_estimate_alignment.py`: KV chunking 对齐验证脚本

docs/issue_xattn_offload_gqa_buffer_oom.md

@@ -0,0 +1,209 @@
+# Issue: XAttention Offload Mode GQA Buffer OOM
+
+## 问题描述
+
+在使用 XAttention BSA (Block Sparse Attention) + CPU Offload 模式运行 GLM-4-9B 等大模型时，出现 CUDA OOM 错误。
+
+### 错误信息
+
+```
+torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8.00 GiB.
+GPU 0 has a total capacity of 23.57 GiB of which 4.19 GiB is free.
+```
+
+### 复现环境
+
+| 项目 | 值 |
+|------|-----|
+| 模型 | GLM-4-9B-Chat-1M |
+| GPU | RTX 3090 (24GB) |
+| Context Length | 32K |
+| sparse_policy | XATTN_BSA |
+| enable_cpu_offload | true |
+| max_model_len | 1048576 (1M) |
+
+### 错误位置
+
+```
+File "nanovllm/kvcache/sparse/xattn_bsa.py", line 246, in alloc_policy_metadata
+    self._k_expanded = torch.empty(shape, dtype=dtype, device=device)
+```
+
+---
+
+## 问题分析
+
+### 内存分配分析
+
+`alloc_policy_metadata()` 在 KV cache 初始化时分配以下 buffer：
+
+| Buffer | 用途 | 大小 (GLM-4, 1M seq) |
+|--------|------|----------------------|
+| `_prefill_mask_buffer` | BSA mask | ~32 MB |
+| `_m_partial_buffer` | KV chunking m stats | ~32 MB |
+| `_l_partial_buffer` | KV chunking l stats | ~32 MB |
+| `_block_sums_buffer` | Block sums | ~64 MB |
+| **`_k_expanded`** | GQA K 扩展 | **~8 GB** |
+| **`_v_expanded`** | GQA V 扩展 | **~8 GB** |
+
+### GQA Buffer 计算
+
+```python
+shape = (1, num_heads, max_seq_len, head_dim)
+     = (1, 32, 1048576, 128)
+
+size = 1 × 32 × 1048576 × 128 × 2 bytes (fp16)
+     = 8,589,934,592 bytes
+     = 8 GB per buffer
+```
+
+### 根本原因
+
+1. **设计意图冲突**：`_k_expanded` 和 `_v_expanded` 的文档注释明确说是 "for GPU-only mode"
+2. **条件检查不完整**：代码只检查了 `num_heads == num_kv_heads` 来跳过分配，没有检查 offload 模式
+3. **Offload 模式不需要这些 buffer**：`compute_chunked_prefill()` 使用不同的计算路径，不依赖预分配的 GQA buffer
+
+### 相关代码
+
+```python
+# xattn_bsa.py:238-247
+# Only allocate GQA expansion buffers if GQA (num_heads != num_kv_heads)
+if num_heads == num_kv_heads:
+    logger.info(f"[XAttn] No GQA expansion needed (num_heads == num_kv_heads = {num_heads})")
+    return  # <-- 只检查了 GQA，没检查 offload 模式
+
+# Shape: [1, num_heads, max_seq_len, head_dim] for xattn_estimate format
+shape = (1, num_heads, max_seq_len, head_dim)
+self._k_expanded = torch.empty(shape, dtype=dtype, device=device)  # <-- OOM here
+self._v_expanded = torch.empty(shape, dtype=dtype, device=device)
+```
+
+---
+
+## 解决思路
+
+### 方案 1: 在 Offload 模式下跳过 GQA Buffer 分配 (推荐)
+
+在 `alloc_policy_metadata()` 中添加 offload 模式检查：
+
+```python
+def alloc_policy_metadata(
+    self,
+    num_heads: int,
+    num_kv_heads: int,
+    head_dim: int,
+    max_seq_len: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    enable_cpu_offload: bool = False,  # <-- 新增参数
+) -> None:
+    # ... 分配 mask buffer 和 KV chunking buffers (offload 模式需要)
+
+    # Skip GQA buffers in offload mode
+    # Chunked prefill uses compute_chunked_prefill() which doesn't need these
+    if enable_cpu_offload:
+        logger.info("[XAttn] Offload mode: skipping GQA expansion buffers")
+        return
+
+    # GPU-only mode: pre-allocate GQA buffers for compute_prefill()
+    if num_heads == num_kv_heads:
+        logger.info(f"[XAttn] No GQA expansion needed")
+        return
+
+    shape = (1, num_heads, max_seq_len, head_dim)
+    self._k_expanded = torch.empty(shape, dtype=dtype, device=device)
+    self._v_expanded = torch.empty(shape, dtype=dtype, device=device)
+```
+
+**需要修改的文件**：
+1. `nanovllm/kvcache/sparse/xattn_bsa.py` - `alloc_policy_metadata()` 方法
+2. `nanovllm/engine/model_runner.py` - 调用 `alloc_policy_metadata()` 时传入 `enable_cpu_offload`
+
+### 方案 2: 延迟分配 (Lazy Allocation)
+
+只在 `compute_prefill()` 首次调用时分配 GQA buffer，offload 模式走 `compute_chunked_prefill()` 不会触发分配。
+
+```python
+def compute_prefill(self, ...):
+    # Lazy allocation on first use
+    if self._k_expanded is None and num_heads != num_kv_heads:
+        self._allocate_gqa_buffers(...)
+    ...
+```
+
+### 方案 3: 基于 chunk_size 限制 buffer 大小
+
+不预分配 max_seq_len 大小，而是只分配 chunk_size 大小：
+
+```python
+# 原来: max_seq_len (1M tokens) -> 8 GB
+# 修改后: chunk_size (16K tokens) -> ~130 MB
+buffer_len = self.chunk_size if enable_cpu_offload else max_seq_len
+shape = (1, num_heads, buffer_len, head_dim)
+```
+
+---
+
+## 验证方法
+
+修复后运行以下命令验证：
+
+```bash
+cd /home/zijie/Code/COMPASS
+GPULIST=0 ./scripts/run_ruler.sh glm4-9b-xattn-nanovllm synthetic xattn --task niah_single_1
+```
+
+预期结果：
+- 不再出现 8GB allocation 的 OOM 错误
+- 模型正常加载并完成推理
+
+---
+
+## 相关文档
+
+- `docs/xattn_bsa_policy_design.md` - XAttention BSA Policy 设计文档
+- `docs/gpu_only_xattn_guide.md` - GPU-Only XAttention 指南
+
+## 优先级
+
+**High** - 阻塞 9B+ 模型在 24GB 显存 GPU 上使用 XAttention + Offload 模式
+
+---
+
+## 修复状态
+
+**✅ 已修复** (2026-02-05)
+
+### 修复内容
+
+采用方案 1，在 offload 模式下跳过 GQA buffer 分配：
+
+1. `nanovllm/kvcache/sparse/policy.py`: 基类添加 `enable_cpu_offload` 参数
+2. `nanovllm/kvcache/sparse/xattn_bsa.py`: 实现 offload 模式检查，跳过 GQA buffer
+3. `nanovllm/engine/model_runner.py`: 传入 `enable_cpu_offload` 参数
+
+### 验证结果
+
+```bash
+# 64K offload 测试
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_64k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 72000 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA
+```
+
+- ✅ 日志显示: `[XAttn] Offload mode: skipping GQA expansion buffers`
+- ✅ 测试通过: 100% 准确率
+- ✅ 内存节省: ~16 GB (for 1M max_seq_len)
+
+### 内存对比
+
+| 配置 | 修复前 | 修复后 |
+|------|--------|--------|
+| max_model_len=72K | +1.1 GB | 0 GB |
+| max_model_len=1M | +16 GB | 0 GB |

docs/long_context_models_1m.md

@@ -0,0 +1,184 @@
+# 1M+ 上下文长度模型列表
+
+本文档收集了 Hugging Face 上支持 1M (1,048,576) 及以上上下文长度的开源模型。
+
+> 更新时间: 2026-01-28
+
+---
+
+## 一、纯语言模型 (≤10B 参数)
+
+### 1. 官方原版模型
+
+| 厂商 | 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|------|--------|------|--------|------|
+| **Qwen** | Qwen2.5-7B-Instruct-1M | 1M | 7B | 69.3K | [HF](https://hf.co/Qwen/Qwen2.5-7B-Instruct-1M) |
+| **THUDM** | GLM-4-9B-Chat-1M | 1M | 9B | 5.0K | [HF](https://hf.co/zai-org/glm-4-9b-chat-1m) |
+| **InternLM** | InternLM2.5-7B-Chat-1M | 1M | 7B | 322 | [HF](https://hf.co/internlm/internlm2_5-7b-chat-1m) |
+| **NVIDIA** | Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct | 1M | 8B | 2.9K | [HF](https://hf.co/nvidia/Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct) |
+| **LWM** | LWM-Text-1M | 1M | 7B | 75 | [HF](https://hf.co/LargeWorldModel/LWM-Text-1M) |
+| **LWM** | LWM-Text-Chat-1M | 1M | 7B | 3.0K | [HF](https://hf.co/LargeWorldModel/LWM-Text-Chat-1M) |
+
+### 2. Gradient AI 扩展系列 (基于 Llama 3)
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Llama-3-8B-Instruct-Gradient-1048k | **1M** | 8B | 44.8K | [HF](https://hf.co/gradientai/Llama-3-8B-Instruct-Gradient-1048k) |
+| Llama-3-8B-Instruct-Gradient-4194k | **4M** | 8B | 9 | [HF](https://hf.co/gradientai/Llama-3-8B-Instruct-Gradient-4194k) |
+
+### 3. 社区衍生版本 (Abliterated)
+
+| 模型 | 上下文 | 基础模型 | 下载量 | 链接 |
+|------|--------|----------|--------|------|
+| Qwen2.5-7B-Instruct-1M-abliterated | 1M | Qwen2.5-7B | 375 | [HF](https://hf.co/huihui-ai/Qwen2.5-7B-Instruct-1M-abliterated) |
+| Nemotron-8B-UltraLong-1M-Abliterated | 1M | Nemotron-8B | 46 | [HF](https://hf.co/SicariusSicariiStuff/Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct_Abliterated) |
+
+---
+
+## 二、视觉-语言模型 (≤10B 参数)
+
+### Qwen3 VL 系列
+
+#### Instruct 版本
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Qwen3-VL-2B-Instruct-1M-GGUF | 1M | 2B | 824 | [HF](https://hf.co/unsloth/Qwen3-VL-2B-Instruct-1M-GGUF) |
+| Qwen3-VL-4B-Instruct-1M-GGUF | 1M | 4B | 936 | [HF](https://hf.co/unsloth/Qwen3-VL-4B-Instruct-1M-GGUF) |
+| Qwen3-VL-8B-Instruct-1M-GGUF | 1M | 8B | 962 | [HF](https://hf.co/unsloth/Qwen3-VL-8B-Instruct-1M-GGUF) |
+
+#### Thinking 推理版本
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Qwen3-VL-2B-Thinking-1M-GGUF | 1M | 2B | 808 | [HF](https://hf.co/unsloth/Qwen3-VL-2B-Thinking-1M-GGUF) |
+| Qwen3-VL-4B-Thinking-1M-GGUF | 1M | 4B | 666 | [HF](https://hf.co/unsloth/Qwen3-VL-4B-Thinking-1M-GGUF) |
+| Qwen3-VL-8B-Thinking-1M-GGUF | 1M | 8B | 4.6K | [HF](https://hf.co/unsloth/Qwen3-VL-8B-Thinking-1M-GGUF) |
+
+---
+
+## 三、推荐模型 (≤10B)
+
+| 用途 | 推荐模型 | 理由 |
+|------|----------|------|
+| **通用对话** | Qwen2.5-7B-Instruct-1M | 官方支持，RULER 93.1分，Apache 2.0 |
+| **中英双语** | GLM-4-9B-Chat-1M | 清华出品，中文优化 |
+| **最长上下文** | Llama-3-8B-Gradient-4194k | 支持 4M 上下文 |
+| **多模态** | Qwen3-VL-8B-Thinking-1M | 视觉理解 + 推理能力 |
+| **无审查** | Qwen2.5-7B-Instruct-1M-abliterated | 移除安全限制 |
+
+---
+
+## 四、VRAM 需求参考
+
+| 模型规模 | 1M 上下文 VRAM | 备注 |
+|----------|----------------|------|
+| 7B (FP16) | ~120GB | 需多卡 |
+| 7B (INT4) | ~40GB | 单卡 A100 可行 |
+| 8B (FP16) | ~130GB | 需多卡 |
+| 9B (FP16) | ~140GB | 需多卡 |
+
+---
+
+## 五、技术对比
+
+| 模型系列 | 扩展技术 | RULER 得分 | 许可证 |
+|---------|---------|------------|--------|
+| Qwen2.5-1M | Dual Chunk Attention | 93.1 | Apache 2.0 |
+| GLM-4-1M | - | 89.9 | 自定义 |
+| Gradient-Llama | 渐进式扩展 | - | Llama 3 |
+| Nemotron-1M | NVIDIA 训练 | - | CC-BY-NC-4.0 |
+| LWM-1M | RingAttention | - | 开源 |
+
+---
+
+---
+
+# 附录：大参数模型 (>10B)
+
+> 以下模型参数量超过 10B，需要更多计算资源。
+
+## A. 纯语言模型 (>10B)
+
+### 官方模型
+
+| 厂商 | 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|------|--------|------|--------|------|
+| **Qwen** | Qwen2.5-14B-Instruct-1M | 1M | 14B | 4.7K | [HF](https://hf.co/Qwen/Qwen2.5-14B-Instruct-1M) |
+| **MiniMax** | MiniMax-Text-01 | 1M | 456B MoE | 721 | [HF](https://hf.co/MiniMaxAI/MiniMax-Text-01) |
+| **Gradient** | Llama-3-70B-Instruct-Gradient-1048k | 1M | 70B | 9 | [HF](https://hf.co/gradientai/Llama-3-70B-Instruct-Gradient-1048k) |
+
+### Qwen3 Coder 系列 (MoE)
+
+| 模型 | 上下文 | 总参数/激活参数 | 下载量 | 链接 |
+|------|--------|-----------------|--------|------|
+| Qwen3-Coder-30B-A3B-Instruct-1M-GGUF | 1M | 30B / 3B | 13.1K | [HF](https://hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-1M-GGUF) |
+| Qwen3-Coder-480B-A35B-Instruct-1M | 1M | 480B / 35B | 50 | [HF](https://hf.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-1M) |
+| Qwen3-Coder-480B-A35B-Instruct-1M-GGUF | 1M | 480B / 35B | 1.7K | [HF](https://hf.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-1M-GGUF) |
+| Qwen3-Coder-42B-A3B-TOTAL-RECALL-1M | 1M | 42B / 3B | - | [HF](https://hf.co/DavidAU/Qwen3-Coder-42B-A3B-Instruct-TOTAL-RECALL-MASTER-CODER-M-1million-ctx) |
+
+### 社区衍生版本
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Qwen2.5-14B-Instruct-1M-abliterated | 1M | 14B | 147 | [HF](https://hf.co/huihui-ai/Qwen2.5-14B-Instruct-1M-abliterated) |
+
+---
+
+## B. 视觉-语言模型 (>10B)
+
+### Meta Llama 4 系列 (MoE 多模态)
+
+| 模型 | 上下文 | 总参数/激活参数 | 下载量 | 链接 |
+|------|--------|-----------------|--------|------|
+| Llama-4-Scout-17B-16E-Instruct | **10M** | 109B / 17B | 180K | [HF](https://hf.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) |
+| Llama-4-Maverick-17B-128E-Instruct | **1M** | 400B / 17B | 32.6K | [HF](https://hf.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct) |
+| Llama-4-Scout-17B-16E | 10M | 109B / 17B | 8.4K | [HF](https://hf.co/meta-llama/Llama-4-Scout-17B-16E) |
+| Llama-4-Maverick-17B-128E | 1M | 400B / 17B | 368 | [HF](https://hf.co/meta-llama/Llama-4-Maverick-17B-128E) |
+| Llama-4-Maverick-17B-128E-Instruct-FP8 | 1M | 400B / 17B | 29.6K | [HF](https://hf.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8) |
+
+### Qwen3 VL 大模型系列
+
+#### Dense 模型
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Qwen3-VL-32B-Instruct-1M-GGUF | 1M | 32B | 1.2K | [HF](https://hf.co/unsloth/Qwen3-VL-32B-Instruct-1M-GGUF) |
+| Qwen3-VL-32B-Thinking-1M-GGUF | 1M | 32B | 452 | [HF](https://hf.co/unsloth/Qwen3-VL-32B-Thinking-1M-GGUF) |
+
+#### MoE 模型
+
+| 模型 | 上下文 | 总参数/激活参数 | 下载量 | 链接 |
+|------|--------|-----------------|--------|------|
+| Qwen3-VL-30B-A3B-Instruct-1M-GGUF | 1M | 30B / 3B | 821 | [HF](https://hf.co/unsloth/Qwen3-VL-30B-A3B-Instruct-1M-GGUF) |
+| Qwen3-VL-30B-A3B-Thinking-1M-GGUF | 1M | 30B / 3B | 944 | [HF](https://hf.co/unsloth/Qwen3-VL-30B-A3B-Thinking-1M-GGUF) |
+| Qwen3-VL-235B-A22B-Instruct-1M-GGUF | 1M | 235B / 22B | 581 | [HF](https://hf.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF) |
+| Qwen3-VL-235B-A22B-Thinking-1M-GGUF | 1M | 235B / 22B | 733 | [HF](https://hf.co/unsloth/Qwen3-VL-235B-A22B-Thinking-1M-GGUF) |
+
+#### MXFP4 量化版本
+
+| 模型 | 上下文 | 规模 | 下载量 | 链接 |
+|------|--------|------|--------|------|
+| Qwen3-VL-30B-A3B-Instruct-1M-MXFP4_MOE-GGUF | 1M | 30B MoE | 689 | [HF](https://hf.co/noctrex/Qwen3-VL-30B-A3B-Instruct-1M-MXFP4_MOE-GGUF) |
+| Qwen3-VL-30B-A3B-Thinking-1M-MXFP4_MOE-GGUF | 1M | 30B MoE | 565 | [HF](https://hf.co/noctrex/Qwen3-VL-30B-A3B-Thinking-1M-MXFP4_MOE-GGUF) |
+| Qwen3-VL-235B-A22B-Instruct-1M-MXFP4_MOE-GGUF | 1M | 235B MoE | 136 | [HF](https://hf.co/noctrex/Qwen3-VL-235B-A22B-Instruct-1M-MXFP4_MOE-GGUF) |
+| Qwen3-VL-235B-A22B-Thinking-1M-MXFP4_MOE-GGUF | 1M | 235B MoE | 244 | [HF](https://hf.co/noctrex/Qwen3-VL-235B-A22B-Thinking-1M-MXFP4_MOE-GGUF) |
+
+---
+
+## 统计汇总
+
+| 类别 | ≤10B 模型数 | >10B 模型数 | 最大上下文 |
+|------|-------------|-------------|-----------|
+| 纯语言模型 | 10 | 8 | 4M |
+| 视觉-语言模型 | 6 | 14 | 10M |
+| **合计** | **16** | **22** | **10M** |
+
+---
+
+## 参考资源
+
+- [Qwen2.5-1M 官方博客](https://qwenlm.github.io/blog/qwen2.5-1m/)
+- [LongRoPE 论文](https://huggingface.co/papers/2402.13753)
+- [InfiniteHiP 论文](https://huggingface.co/papers/2502.08910)
+- [Top LLMs for Long Context Windows](https://www.siliconflow.com/articles/en/top-LLMs-for-long-context-windows)

docs/memory_communication_benchmark.md

@@ -0,0 +1,120 @@
+# Memory Communication Benchmark
+
+GPU-CPU 通信量测试结果，对比 Full Policy 和 XAttention BSA Policy。
+
+## 测试环境
+
+- **模型**: Llama-3.1-8B-Instruct
+- **GPU**: RTX 3090 (24GB)
+- **配置**: `num_gpu_blocks=4`, `block_size=1024`, `enable_cpu_offload=True`
+- **XAttention 参数**: `threshold=0.95`, `stride=8`
+
+## 32K 上下文测试结果
+
+| 指标 | Full Policy | XAttention | 比率 |
+|------|-------------|------------|------|
+| **Prefill H2D** | 66.57 GB | 111.12 GB | **1.67x** |
+| Prefill D2H | 4.29 GB | 4.29 GB | 1.00x |
+| TTFT | 8473 ms | 10367 ms | 1.22x |
+
+### XAttention Block Selection (32K)
+
+| 指标 | 数值 |
+|------|------|
+| 可用 blocks | 465 |
+| 选中 blocks | 374 |
+| 选择密度 | 80.4% |
+
+## 64K 上下文测试结果
+
+| 指标 | Full Policy | XAttention | 比率 |
+|------|-------------|------------|------|
+| **Prefill H2D** | 262.13 GB | 386.62 GB | **1.48x** |
+| Prefill D2H | 8.46 GB | 8.46 GB | 1.00x |
+| Decode H2D (32 tokens) | 262.13 GB | 262.13 GB | 1.00x |
+| TTFT | 27081 ms | 33634 ms | 1.24x |
+
+## 通信量比率对比 (K-only 优化前)
+
+| 上下文长度 | XAttn/Full Prefill H2D 比率 |
+|------------|----------------------------|
+| 32K | 1.67x |
+| 64K | 1.48x |
+
+### 分析 (优化前)
+
+1. **XAttention 通信量增加原因**：
+   - Estimate 阶段：加载 **100%** 历史 blocks 的 **K+V**（用于 attention score 估计）
+   - Compute 阶段：加载 **选中的** blocks（约 70-80%）
+   - 理论比率：`1 + selection_density`
+
+2. **64K 比率更低的原因**：
+   - 更长上下文时，attention 分布更稀疏
+   - XAttention 的 block 选择更有效（选中比例更低）
+   - First/last block 强制包含的影响相对减小
+
+3. **Decode 阶段通信量相同**：
+   - XAttention 仅支持 prefill 阶段
+   - Decode 阶段 fallback 到 Full Policy
+
+---
+
+## K-only 优化 (2026-01-28)
+
+### 优化原理
+
+XAttention 的 `select_blocks` 估计阶段只需要 K 来计算 attention scores：
+```python
+# flat_group_gemm_fuse_reshape 只使用 Q 和 K
+attn_scores = flat_group_gemm_fuse_reshape(Q, K_chunk, stride, ...)
+```
+
+V 在估计阶段完全不使用，但之前代码会同时加载 K 和 V，造成 50% 通信量浪费。
+
+### 优化实现
+
+1. **新增方法**: `OffloadEngine.load_k_only_to_slot_layer()` - 只加载 K
+2. **修改 select_blocks**: 使用只加载 K 的新方法
+
+### 优化后测试结果
+
+| 上下文 | Full Policy | XAttn (优化前) | XAttn (优化后) | 优化节省 |
+|--------|-------------|---------------|---------------|---------|
+| 32K | 66.57 GB | 111.12 GB | **79.76 GB** | **28.2%** |
+| 64K | 262.13 GB | 386.62 GB | **258.78 GB** | **33.1%** |
+
+### XAttn/Full 比率变化
+
+| 上下文 | 优化前比率 | 优化后比率 |
+|--------|-----------|-----------|
+| 32K | 1.67x | **1.20x** |
+| 64K | 1.48x | **0.99x** |
+
+### 结论
+
+优化后，64K 上下文的 XAttention 通信量与 Full Policy 基本持平 (0.99x)，
+而 32K 也从 1.67x 降到 1.20x。这说明估计阶段的 K-only 优化非常有效
+
+## 测试命令
+
+```bash
+# 32K Full Policy
+python bench_offload.py --max-len 32768 --input-len 32000
+
+# 32K XAttention
+python bench_offload.py --max-len 32768 --input-len 32000 --enable-xattn
+
+# 64K Full Policy
+python bench_offload.py --max-len 65536 --input-len 64000
+
+# 64K XAttention
+python bench_offload.py --max-len 65536 --input-len 64000 --enable-xattn
+
+# 包含 decode 测试
+python bench_offload.py --max-len 65536 --input-len 64000 --bench-decode --output-len 32
+```
+
+## 相关文档
+
+- [`observer_architecture.md`](observer_architecture.md) - Observer 架构设计
+- [`xattn_bsa_policy_design.md`](xattn_bsa_policy_design.md) - XAttention BSA 算法设计

docs/new_model_integration_guide.md

@@ -0,0 +1,323 @@
+# 新模型整合指南
+
+本文档总结了将新模型（如GLM-4）整合到nanovllm的经验和常见问题。
+
+## 整合流程概览
+
+```
+1. 分析模型配置 (config.json)
+   ↓
+2. 创建模型文件 (nanovllm/models/<model>.py)
+   ↓
+3. 实现权重加载 (nanovllm/utils/loader.py)
+   ↓
+4. 处理特殊组件 (RoPE, Attention, etc.)
+   ↓
+5. 处理tokenizer差异 (EOS tokens, chat template)
+   ↓
+6. 验证输出正确性
+```
+
+---
+
+## 1. 配置字段映射
+
+不同模型使用不同的配置字段名称，需要建立映射关系：
+
+| 标准字段 | GLM-4 | Qwen | Llama | 说明 |
+|----------|-------|------|-------|------|
+| `num_key_value_heads` | `multi_query_group_num` | `num_key_value_heads` | `num_key_value_heads` | KV heads数量 |
+| `head_dim` | `kv_channels` | 计算得出 | 计算得出 | 每个head的维度 |
+| `intermediate_size` | `ffn_hidden_size` | `intermediate_size` | `intermediate_size` | FFN隐藏层大小 |
+| `max_position_embeddings` | `seq_length` | `max_position_embeddings` | `max_position_embeddings` | 最大位置 |
+| `rope_theta` | `10000 * rope_ratio` | `rope_theta` | `rope_theta` | RoPE基础频率 |
+
+### 代码示例
+
+```python
+# 在模型 __init__ 中处理配置差异
+num_kv_heads = getattr(config, 'num_key_value_heads',
+                       getattr(config, 'multi_query_group_num', num_heads))
+
+head_dim = getattr(config, 'head_dim',
+                   getattr(config, 'kv_channels', hidden_size // num_heads))
+
+intermediate_size = getattr(config, 'intermediate_size',
+                           getattr(config, 'ffn_hidden_size', None))
+
+max_position = getattr(config, 'max_position_embeddings',
+                       getattr(config, 'seq_length', 4096))
+```
+
+---
+
+## 2. RoPE实现差异
+
+RoPE是模型整合中**最容易出错**的部分。不同模型可能使用不同的RoPE变体：
+
+### 2.1 旋转方式
+
+| 类型 | 描述 | 使用模型 |
+|------|------|----------|
+| **Half rotation** | 前半和后半分别旋转 `[x0,x1,...] → [x0*cos-x_{d/2}*sin, ...]` | Llama, Qwen |
+| **Interleaved rotation** | 相邻元素配对旋转 `[x0,x1,...] → [x0*cos-x1*sin, x1*cos+x0*sin, ...]` | GLM-4 |
+
+### 2.2 旋转维度
+
+| 类型 | 描述 | 使用模型 |
+|------|------|----------|
+| **Full rotation** | 旋转整个head_dim | Llama, Qwen |
+| **Partial rotation** | 只旋转head_dim的一部分，其余pass-through | GLM-4 (rotary_dim = head_dim // 2) |
+
+### 2.3 GLM-4 RoPE实现
+
+```python
+class GLM4RotaryEmbedding(nn.Module):
+    def __init__(self, head_dim, rotary_dim, ...):
+        # GLM-4只旋转一半维度
+        self.rotary_dim = rotary_dim  # = head_dim // 2
+
+    def forward(self, positions, query, key):
+        # 分离旋转部分和pass-through部分
+        q_rot = query[..., :self.rotary_dim]
+        q_pass = query[..., self.rotary_dim:]
+
+        # 只对旋转部分应用interleaved RoPE
+        q_rot = apply_rotary_emb_interleaved(q_rot, cos, sin)
+
+        # 拼接回去
+        return torch.cat([q_rot, q_pass], dim=-1), ...
+```
+
+### 2.4 调试RoPE问题
+
+**症状**：模型输出乱码或重复无意义的内容（如 "The. The. The..."）
+
+**调试方法**：
+```python
+# 对比HuggingFace参考实现的输出
+hf_q, hf_k = hf_model.apply_rotary_pos_emb(query, key, cos, sin)
+my_q, my_k = my_rotary_emb(positions, query, key)
+
+print(f"Q max diff: {(hf_q - my_q).abs().max()}")  # 应该 < 1e-5
+print(f"K max diff: {(hf_k - my_k).abs().max()}")  # 应该 < 1e-5
+```
+
+---
+
+## 3. 权重名称映射
+
+不同模型的权重命名规范不同：
+
+### 3.1 常见映射
+
+| 组件 | Llama/Qwen | GLM-4 |
+|------|------------|-------|
+| Attention QKV | `q_proj`, `k_proj`, `v_proj` | `query_key_value` (合并) |
+| Attention Output | `o_proj` | `dense` |
+| MLP Gate | `gate_proj` | `dense_h_to_4h` (部分) |
+| MLP Up | `up_proj` | `dense_h_to_4h` (部分) |
+| MLP Down | `down_proj` | `dense_4h_to_h` |
+| LayerNorm | `input_layernorm` | `input_layernorm` |
+| Post-Attention LN | `post_attention_layernorm` | `post_attention_layernorm` |
+
+### 3.2 实现权重转换
+
+```python
+def convert_glm4_weights(name, param):
+    """将GLM-4权重名称转换为nanovllm格式"""
+    # 处理合并的QKV权重
+    if "query_key_value" in name:
+        # 拆分为q, k, v
+        q, k, v = param.split([q_size, kv_size, kv_size], dim=0)
+        return {"q_proj": q, "k_proj": k, "v_proj": v}
+
+    # 处理合并的gate+up权重
+    if "dense_h_to_4h" in name:
+        gate, up = param.chunk(2, dim=0)
+        return {"gate_proj": gate, "up_proj": up}
+
+    return {name: param}
+```
+
+---
+
+## 4. EOS Token处理
+
+### 4.1 问题
+
+某些模型使用**多个EOS tokens**：
+
+| 模型 | EOS Token(s) | 说明 |
+|------|--------------|------|
+| Llama | `128001` | 单一EOS |
+| Qwen | `151643` | 单一EOS |
+| GLM-4 | `[151329, 151336, 151338]` | 多个：endoftext, user, observation |
+
+**问题**：`tokenizer.eos_token_id` 只返回第一个，导致模型不会在其他EOS token处停止。
+
+### 4.2 解决方案
+
+```python
+# config.py - 支持多个EOS
+eos: int | list[int] = -1
+
+# llm_engine.py - 从hf_config读取完整EOS列表
+eos_from_config = getattr(config.hf_config, 'eos_token_id', None)
+if eos_from_config is not None:
+    config.eos = eos_from_config
+else:
+    config.eos = self.tokenizer.eos_token_id
+
+# scheduler.py - 使用set进行高效查找
+self.eos_set = set(eos) if isinstance(eos, list) else {eos}
+
+# 检查时使用 in 而不是 ==
+if token_id in self.eos_set:
+    # 停止生成
+```
+
+### 4.3 调试EOS问题
+
+**症状**：模型总是生成到max_tokens才停止
+
+**调试方法**：
+```python
+# 检查EOS配置
+print(f"tokenizer.eos_token_id: {tokenizer.eos_token_id}")
+print(f"hf_config.eos_token_id: {config.hf_config.eos_token_id}")
+
+# 检查输出中的EOS tokens
+output = llm.generate([prompt], params)
+for eos_id in [151329, 151336, 151338]:
+    if eos_id in output[0]['token_ids']:
+        print(f"Found EOS {eos_id} at position {output[0]['token_ids'].index(eos_id)}")
+```
+
+---
+
+## 5. Chat Template
+
+不同模型使用不同的对话模板：
+
+| 模型 | 模板格式 |
+|------|----------|
+| Llama-3 | `<\|begin_of_text\|><\|start_header_id\|>user<\|end_header_id\|>\n{content}<\|eot_id\|><\|start_header_id\|>assistant<\|end_header_id\|>\n` |
+| Qwen | `<\|im_start\|>user\n{content}<\|im_end\|>\n<\|im_start\|>assistant\n` |
+| GLM-4 | `[gMASK]<sop><\|user\|>\n{content}<\|assistant\|>\n` |
+
+### 实现模板转换
+
+```python
+def convert_to_model_prompt(prompt: str, model_type: str) -> str:
+    """将标准prompt转换为模型特定格式"""
+    if model_type == "glm4":
+        return f"[gMASK]<sop><|user|>\n{prompt}<|assistant|>\n"
+    elif model_type == "llama3":
+        return f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n"
+    # ...
+```
+
+---
+
+## 6. 验证清单
+
+整合新模型后，按以下顺序验证：
+
+### 6.1 权重加载验证
+
+```python
+# 检查所有权重是否正确加载
+for name, param in model.named_parameters():
+    if param.abs().sum() == 0:
+        print(f"WARNING: {name} is all zeros!")
+```
+
+### 6.2 单层输出验证
+
+```python
+# 对比embedding层输出
+my_emb = my_model.embed_tokens(input_ids)
+hf_emb = hf_model.model.embed_tokens(input_ids)
+print(f"Embedding diff: {(my_emb - hf_emb).abs().max()}")  # < 1e-5
+
+# 对比第一层输出
+my_out = my_model.layers[0](my_emb, ...)
+hf_out = hf_model.model.layers[0](hf_emb, ...)
+print(f"Layer 0 diff: {(my_out - hf_out).abs().max()}")  # < 1e-4
+```
+
+### 6.3 生成质量验证
+
+```python
+# 简单问答测试
+prompt = "Hello, how are you?"
+output = llm.generate([prompt], SamplingParams(max_tokens=50))
+print(output[0]['text'])  # 应该是连贯的回答
+
+# 检查是否正确停止
+print(f"Generated {len(output[0]['token_ids'])} tokens (max=50)")
+```
+
+### 6.4 RULER基准测试
+
+```bash
+# 运行1个sample快速验证
+python tests/test_ruler.py --model <path> --num-samples 1
+
+# 验证通过后运行完整测试
+python tests/test_ruler.py --model <path> --num-samples 100
+```
+
+---
+
+## 7. 常见问题速查
+
+| 症状 | 可能原因 | 解决方案 |
+|------|----------|----------|
+| 输出乱码/重复 | RoPE实现错误 | 检查旋转方式(interleaved vs half)和旋转维度(full vs partial) |
+| 数值爆炸(NaN/Inf) | 权重加载错误或dtype不匹配 | 检查权重映射，确保dtype一致 |
+| 不停止生成 | EOS token处理错误 | 从hf_config读取完整EOS列表 |
+| 输出质量差 | LayerNorm或bias缺失 | 检查add_qkv_bias等配置 |
+| 位置编码错误 | max_position_embeddings读取错误 | 检查配置字段名称(seq_length等) |
+
+---
+
+## 8. 文件结构
+
+新模型整合需要修改/创建的文件：
+
+```
+nanovllm/
+├── models/
+│   └── <model>.py          # 新建：模型定义
+├── layers/
+│   └── rotary_embedding.py # 修改：如需特殊RoPE
+├── utils/
+│   └── loader.py           # 修改：权重加载
+├── config.py               # 可能修改：新配置字段
+└── engine/
+    ├── llm_engine.py       # 可能修改：EOS处理
+    └── scheduler.py        # 可能修改：EOS检查
+tests/
+└── test_ruler.py           # 修改：chat template
+```
+
+---
+
+## 附录：GLM-4整合案例
+
+### 遇到的问题及解决
+
+1. **配置字段差异** → 添加getattr fallback链
+2. **Interleaved RoPE** → 实现`apply_rotary_emb_interleaved`
+3. **Partial rotation (head_dim//2)** → 实现`GLM4RotaryEmbedding`
+4. **多EOS tokens** → 修改config/llm_engine/scheduler支持list
+5. **合并的QKV权重** → 在loader中拆分
+
+### 关键代码位置
+
+- RoPE实现: `nanovllm/layers/rotary_embedding.py:GLM4RotaryEmbedding`
+- 模型定义: `nanovllm/models/glm4.py`
+- 权重加载: `nanovllm/utils/loader.py:load_glm4_weights`
+- EOS处理: `nanovllm/engine/scheduler.py:eos_set`

docs/observer_architecture.md

@@ -0,0 +1,194 @@
+# Observer Architecture
+
+nanovllm 的 Observer 架构用于统计推理过程中的关键指标，采用类变量（class variable）模式实现全局状态管理。
+
+## 架构概览
+
+```
+Observer (基类)
+├── InferenceObserver - 推理时间指标 (TTFT, TPOT)
+└── MemoryObserver - 内存传输统计 (H2D, D2H, D2D)
+```
+
+## 设计原则
+
+### 1. 类变量模式
+
+所有 Observer 使用类变量（而非实例变量）存储状态：
+
+```python
+class Observer:
+    """Observer 基类"""
+    _enabled: bool = True  # 类变量，控制是否启用
+
+class InferenceObserver(Observer):
+    ttft: int = 0          # 类变量，全局共享
+    tpot: int = 0
+    ttft_start: int = 0
+    tpot_start: int = 0
+```
+
+**优点**：
+- 无需实例化，任何地方都可以直接访问
+- 避免跨模块传递 observer 实例
+- 适合全局统计场景
+
+### 2. 启用/禁用控制
+
+每个 Observer 可独立启用/禁用：
+
+```python
+# 启用 MemoryObserver
+MemoryObserver._enabled = True
+
+# 禁用后，record_* 方法不会记录
+MemoryObserver._enabled = False
+```
+
+### 3. 阶段分离
+
+MemoryObserver 支持 prefill/decode 阶段分离统计：
+
+```python
+@classmethod
+def record_h2d(cls, num_bytes: int, is_prefill: bool = True) -> None:
+    if not cls._enabled:
+        return
+    cls.h2d_bytes += num_bytes
+    cls.h2d_count += 1
+    if is_prefill:
+        cls.prefill_h2d_bytes += num_bytes
+    else:
+        cls.decode_h2d_bytes += num_bytes
+```
+
+## Observer 实现
+
+### InferenceObserver
+
+**位置**: `nanovllm/utils/observer.py`
+
+**统计指标**：
+| 指标 | 说明 | 单位 |
+|------|------|------|
+| `ttft` | Time To First Token | 纳秒 |
+| `tpot` | Time Per Output Token | 纳秒 |
+| `ttft_start` | TTFT 计时开始点 | 纳秒 |
+| `tpot_start` | TPOT 计时开始点 | 纳秒 |
+
+**统计位置**：
+| 位置 | 代码 | 说明 |
+|------|------|------|
+| `scheduler.py:add()` | `InferenceObserver.ttft_start = perf_counter_ns()` | 开始计时 |
+| `llm_engine.py:step()` | `InferenceObserver.ttft = ... - ttft_start` | Prefill 完成后计算 TTFT |
+| `llm_engine.py:step()` | `InferenceObserver.tpot = ... - tpot_start` | Decode 时计算 TPOT |
+
+### MemoryObserver
+
+**位置**: `nanovllm/utils/memory_observer.py`
+
+**统计指标**：
+| 指标 | 说明 |
+|------|------|
+| `h2d_bytes` / `h2d_count` | Host to Device 传输量/次数 |
+| `d2h_bytes` / `d2h_count` | Device to Host 传输量/次数 |
+| `d2d_bytes` / `d2d_count` | Device to Device 复制量/次数 |
+| `prefill_h2d_bytes` / `prefill_d2h_bytes` | Prefill 阶段 H2D/D2H |
+| `decode_h2d_bytes` / `decode_d2h_bytes` | Decode 阶段 H2D/D2H |
+
+**统计位置** (均在 `offload_engine.py`)：
+
+| 方法 | 传输类型 | 说明 |
+|------|----------|------|
+| `load_to_slot_layer()` | H2D | 从 CPU 加载 block 到 GPU slot |
+| `load_block_sample_from_cpu()` | H2D | 采样加载（Quest） |
+| `load_block_full_from_cpu()` | H2D | 完整加载 block |
+| `offload_slot_layer_to_cpu()` | D2H | GPU slot 卸载到 CPU |
+| `offload_prefill_buffer_async()` | D2H | Prefill buffer 异步卸载 |
+| `write_to_prefill_buffer()` | D2D | 写入 prefill buffer |
+| `write_to_decode_buffer()` | D2D | 写入 decode buffer |
+
+**重置位置**：
+| 位置 | 代码 |
+|------|------|
+| `llm_engine.py:generate()` | `MemoryObserver.complete_reset()` |
+| `llm_engine.py:generate()` | `InferenceObserver.complete_reset()` |
+
+## 使用示例
+
+### 1. 启用并统计
+
+```python
+from nanovllm.utils.memory_observer import MemoryObserver
+
+# 启用统计
+MemoryObserver._enabled = True
+
+# 运行推理
+outputs = llm.generate(prompts, sampling_params)
+
+# 获取结果
+print(f"Prefill H2D: {MemoryObserver.prefill_h2d_bytes / 1e9:.2f} GB")
+print(f"Decode H2D: {MemoryObserver.decode_h2d_bytes / 1e9:.2f} GB")
+
+# 或使用 print_summary
+MemoryObserver.print_summary()
+```
+
+### 2. 在 bench_offload.py 中
+
+```python
+from nanovllm.utils.memory_observer import MemoryObserver
+
+# 启用
+MemoryObserver._enabled = True
+
+# benchmark 结束后打印
+def print_memory_stats():
+    fmt = MemoryObserver._fmt_bytes
+    print(f"[Memory] Prefill H2D: {fmt(MemoryObserver.prefill_h2d_bytes)}")
+    print(f"         Decode  H2D: {fmt(MemoryObserver.decode_h2d_bytes)}")
+```
+
+### 3. 获取结构化数据
+
+```python
+summary = MemoryObserver.get_summary()
+# {
+#     "total": {"h2d_bytes": ..., "d2h_bytes": ..., "d2d_bytes": ...},
+#     "prefill": {"h2d_bytes": ..., "d2h_bytes": ...},
+#     "decode": {"h2d_bytes": ..., "d2h_bytes": ...}
+# }
+```
+
+## 添加新 Observer
+
+1. 继承 `Observer` 基类
+2. 定义类变量存储统计数据
+3. 实现 `record_*` 方法（需检查 `_enabled`）
+4. 实现 `complete_reset()` 方法
+5. 在相关代码位置添加 `record_*` 调用
+6. 在 `llm_engine.py:generate()` 中添加 reset 调用
+
+```python
+from nanovllm.utils.observer import Observer
+
+class MyObserver(Observer):
+    _enabled: bool = False
+    my_metric: int = 0
+
+    @classmethod
+    def record_event(cls, value: int) -> None:
+        if not cls._enabled:
+            return
+        cls.my_metric += value
+
+    @classmethod
+    def complete_reset(cls) -> None:
+        cls.my_metric = 0
+```
+
+## 相关文档
+
+- [`memory_communication_benchmark.md`](memory_communication_benchmark.md) - 通信量测试结果
+- [`architecture_guide.md`](architecture_guide.md) - 整体架构指南

docs/test_ruler_usage_guide.md

@@ -0,0 +1,338 @@
+# test_ruler.py 使用指南
+
+RULER benchmark 综合测试工具，用于评估 LLM 长上下文能力。
+
+**测试日期**: 2026-02-05
+**测试 GPU**: RTX 3090 (GPU 4)
+
+---
+
+## 支持的任务
+
+| 类别 | 任务 |
+|------|------|
+| NIAH (Needle-In-A-Haystack) | `niah_single_1/2/3`, `niah_multikey_1/2/3`, `niah_multiquery`, `niah_multivalue` |
+| QA (Question Answering) | `qa_1`, `qa_2` |
+| Recall | `cwe`, `fwe`, `vt` |
+
+---
+
+## 基本命令格式
+
+```bash
+CUDA_VISIBLE_DEVICES=<GPU_ID> PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py [OPTIONS]
+```
+
+---
+
+## 参数说明
+
+### 必要参数
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--model` | `~/models/Llama-3.1-8B-Instruct` | 模型路径 |
+| `--data-dir` | `tests/data/ruler_64k` | 数据目录 |
+| `--max-model-len` | 65664 | 最大上下文长度 |
+
+### 数据选择
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--datasets` | 全部 | 逗号分隔的数据集名 |
+| `--num-samples` | 0 (全部) | 每个数据集测试样本数 |
+| `--sample-indices` | - | 指定样本索引 (如 `0,5,10`) |
+
+### Offload 配置
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--enable-offload` | False | 启用 CPU offload 模式 |
+| `--num-gpu-blocks` | 4 | GPU 上的 KV cache blocks 数量 |
+| `--block-size` | 4096 | KV cache block 大小 (tokens) |
+| `--num-kv-buffers` | 4 | Ring buffer 数量 |
+| `--gpu-utilization` | 0.9 | GPU 显存利用率 |
+
+### Sparse Attention 配置
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--sparse-policy` | - | 稀疏策略: `FULL`, `QUEST`, `XATTN_BSA` |
+| `--sparse-threshold` | 0.9 | XAttn cumulative attention 阈值 |
+| `--sparse-samples` | 128 | XAttn 每 chunk 采样数 |
+| `--sparse-stride` | 8 | XAttn Q/K 下采样步长 |
+
+### 输出控制
+
+| 参数 | 说明 |
+|------|------|
+| `--quiet` / `-q` | 安静模式 |
+| `--json-output` | JSON 格式输出 |
+| `--fresh-llm` | 每个样本重新初始化 LLM |
+
+### 其他
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--dtype` | auto | 模型数据类型 (`bfloat16`, `float16`) |
+| `--use-cuda-graph` | False | 启用 CUDA Graph |
+| `--max-new-tokens` | 16 | 最大生成 token 数 |
+
+---
+
+## 已验证的命令示例
+
+以下命令均在 RTX 3090 (24GB) 上测试通过。
+
+### 1. 基础 Offload 测试 (32K)
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload
+```
+
+**结果**: 100% 准确率, 耗时 ~16s
+
+### 2. Offload + XAttention BSA (32K)
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+**结果**: 100% 准确率, compute density ~50%, 耗时 ~19s
+
+### 3. Offload + XAttention BSA (64K)
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_64k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 72000 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+**结果**: 100% 准确率, compute density ~37%, 耗时 ~52s
+
+### 4. 多数据集多样本测试
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1,qa_1 \
+    --num-samples 2 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA
+```
+
+**结果**: 4/4 (100%), 耗时 ~71s
+
+### 5. 指定样本索引测试
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --sample-indices 0,5,10 \
+    --max-model-len 40960 \
+    --enable-offload
+```
+
+### 6. JSON 输出模式 (用于脚本)
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --json-output
+```
+
+**输出格式**:
+```json
+{
+  "total_correct": 1,
+  "total_samples": 1,
+  "overall_accuracy": 1.0,
+  "avg_score": 1.0,
+  "time": 30.44,
+  "tasks": {"niah_single_1": {"correct": 1, "total": 1, "accuracy": 1.0}},
+  "failed_samples": {}
+}
+```
+
+### 7. 安静模式
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --quiet
+```
+
+### 8. 调整 GPU blocks 数量
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --num-gpu-blocks 8 \
+    --sparse-policy XATTN_BSA
+```
+
+### 9. GLM-4 模型测试
+
+```bash
+CUDA_VISIBLE_DEVICES=4 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/GLM-4-9B-Chat-1M \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --dtype bfloat16
+```
+
+**结果**: 100% 准确率, 耗时 ~17s
+
+---
+
+## 数据目录结构
+
+```
+tests/data/
+├── ruler_4k/      # 4K context
+├── ruler_8k/      # 8K context
+├── ruler_16k/     # 16K context
+├── ruler_32k/     # 32K context (推荐测试)
+├── ruler_64k/     # 64K context
+├── ruler_128k/    # 128K context
+├── ruler_256k/    # 256K context
+├── ruler_512k/    # 512K context
+├── ruler_768k/    # 768K context
+└── ruler_1m/      # 1M context
+```
+
+每个目录包含 13 个任务子目录，每个任务有 `validation.jsonl` 文件。
+
+---
+
+## GPU 与模式选择
+
+| GPU 显存 | 推荐模式 | 说明 |
+|---------|---------|------|
+| 24GB (3090/4090) | `--enable-offload` | 必须使用 offload |
+| 40GB+ (A100) | 两种模式均可 | 可测试 GPU-only |
+
+**RTX 3090 限制**: 由于显存限制，必须使用 `--enable-offload` 参数。
+
+---
+
+## max-model-len 设置指南
+
+| 数据目录 | 推荐 max-model-len | 说明 |
+|---------|-------------------|------|
+| ruler_4k | 5000 | 留出 output 空间 |
+| ruler_8k | 9000 | |
+| ruler_16k | 17000 | |
+| ruler_32k | 40960 | |
+| ruler_64k | 72000 | |
+| ruler_128k | 135000 | |
+
+**公式**: `max_model_len >= max_input_len + max_new_tokens`
+
+---
+
+## DensityObserver 输出
+
+使用 `--sparse-policy XATTN_BSA` 时自动启用，输出示例：
+
+```
+============================================================
+Density Statistics (XAttention BSA)
+============================================================
+[DensityObserver] Mode: offload
+  Compute density: 0.3691 (min: 0.3691 @ layer 0)
+  Comm density:    1.0000 (CPU block granularity)
+  Savings ratio:   0.0% H2D transfer reduction
+  Num layers: 1
+  Layer 0 density: 0.369052
+```
+
+| 指标 | 说明 |
+|------|------|
+| Compute density | BSA block (128 tokens) 粒度的计算密度 |
+| Comm density | CPU block (4096 tokens) 粒度的通信密度 |
+| Savings ratio | H2D 传输减少比例 |
+
+---
+
+## 常见问题
+
+### 1. OOM 错误
+
+**原因**: 显存不足
+**解决**:
+- 使用 `--enable-offload`
+- 降低 `--gpu-utilization`
+- 减少 `--num-gpu-blocks`
+
+### 2. 模型加载失败
+
+**原因**: 模型配置不兼容
+**解决**:
+- 检查 `--dtype` 参数 (GLM 模型需要 `--dtype bfloat16`)
+- 确认模型路径正确
+
+### 3. 准确率异常
+
+**原因**: 状态泄漏
+**解决**: 使用 `--fresh-llm` 参数为每个样本重新初始化 LLM
+
+---
+
+## 相关文档
+
+- [`docs/xattn_density_types.md`](xattn_density_types.md) - Compute vs Comm density 解释
+- [`docs/xattn_density_alignment_verification.md`](xattn_density_alignment_verification.md) - GPU-only vs Offload 对齐验证
+- [`docs/ruler_benchmark_results_32k.md`](ruler_benchmark_results_32k.md) - RULER 32K 基准测试结果

docs/xattn_density_alignment_verification.md

@@ -0,0 +1,142 @@
+# XAttention Density Alignment Verification
+
+验证 GPU-only 和 Offload 模式的 density 对齐情况。
+
+**测试日期**: 2026-02-05
+**测试模型**: Llama-3.1-8B-Instruct
+**测试任务**: RULER niah_single_1
+
+---
+
+## 测试配置
+
+| 参数 | 值 |
+|------|-----|
+| sparse_policy | XATTN_BSA |
+| threshold | 0.9 |
+| chunk_size | 4096 (已对齐) |
+| stride | 8 |
+| BSA block_size | 128 |
+
+---
+
+## 测试结果
+
+### 32K Context
+
+| 模式 | Layer 0 Density | Overall Density | 准确率 |
+|------|-----------------|-----------------|--------|
+| GPU-only | 0.502079 | 0.4012 | 100% |
+| Offload | 0.498421 | 0.4984 | 100% |
+| **差异** | **0.37%** | - | - |
+
+### 64K Context
+
+| 模式 | Layer 0 Density | Overall Density | 准确率 |
+|------|-----------------|-----------------|--------|
+| GPU-only | 0.369972 | 0.2963 | 100% |
+| Offload | 0.369052 | 0.3691 | 100% |
+| **差异** | **0.09%** | - | - |
+
+---
+
+## 关键修复
+
+### Commit 829b311 - chunk_size 对齐 + Stream 同步修复
+
+**问题**: 之前 GPU-only 和 Offload 模式的 density 差异达 10-13%
+
+**根因**:
+1. GPU-only 使用 `chunk_size=16384`，Offload 使用 `chunk_size=4096`
+2. Stream 同步 bug 导致 Pass 1/2 K 数据不一致
+
+**修复**:
+1. 将 `XAttentionBSAPolicy.chunk_size` 默认值从 16384 改为 4096
+2. 所有 compute kernels 包装在 `compute_stream` context 中
+
+---
+
+## 测试命令
+
+### GPU-only 模式
+
+```bash
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+### Offload 模式
+
+```bash
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+---
+
+## 详细日志
+
+### 32K Offload 模式 Per-Chunk Density
+
+```
+Layer0 chunk: q_len=4096, k_len=4096,  density=0.6234
+Layer0 chunk: q_len=4096, k_len=8192,  density=0.6239
+Layer0 chunk: q_len=4096, k_len=12288, density=0.6026
+Layer0 chunk: q_len=4096, k_len=16384, density=0.5695
+Layer0 chunk: q_len=4096, k_len=20480, density=0.5285
+Layer0 chunk: q_len=4096, k_len=24576, density=0.4891
+Layer0 chunk: q_len=4096, k_len=28672, density=0.4514
+Layer0 chunk: q_len=3813, k_len=32485, density=0.4208
+```
+
+### 64K Offload 模式 Per-Chunk Density
+
+```
+Layer0 chunk: q_len=4096, k_len=4096,  density=0.6234
+Layer0 chunk: q_len=4096, k_len=8192,  density=0.6239
+Layer0 chunk: q_len=4096, k_len=12288, density=0.6026
+Layer0 chunk: q_len=4096, k_len=16384, density=0.5681
+Layer0 chunk: q_len=4096, k_len=20480, density=0.5255
+Layer0 chunk: q_len=4096, k_len=24576, density=0.4859
+Layer0 chunk: q_len=4096, k_len=28672, density=0.4485
+Layer0 chunk: q_len=4096, k_len=32768, density=0.4161
+Layer0 chunk: q_len=4096, k_len=36864, density=0.3892
+Layer0 chunk: q_len=4096, k_len=40960, density=0.3658
+Layer0 chunk: q_len=4096, k_len=45056, density=0.3464
+Layer0 chunk: q_len=4096, k_len=49152, density=0.3303
+Layer0 chunk: q_len=4096, k_len=53248, density=0.3170
+Layer0 chunk: q_len=4096, k_len=57344, density=0.3068
+Layer0 chunk: q_len=4096, k_len=61440, density=0.2988
+Layer0 chunk: q_len=3451, k_len=64891, density=0.2947
+```
+
+---
+
+## 结论
+
+1. **Density 对齐成功**: 差异从 10-13% 降到 <0.5%
+2. **准确率一致**: 两种模式都达到 100% 准确率
+3. **Density 随 context 增长下降**: 符合预期，更长的 context 稀疏性更高
+
+---
+
+## 相关文档
+
+- [`docs/xattn_offload_stream_sync_fix.md`](xattn_offload_stream_sync_fix.md) - Stream 同步修复详情
+- [`docs/xattn_density_types.md`](xattn_density_types.md) - Compute vs Comm density
+- [`docs/gpuonly_density_alignment_test.md`](gpuonly_density_alignment_test.md) - 早期对齐测试

docs/xattn_density_benchmark.md

@@ -0,0 +1,195 @@
+# XAttention Density Benchmark
+
+GPU-only 模式下 XAttention Block Sparse Attention 的 density 测试结果。
+
+## 测试配置
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| Model | Llama-3.1-8B-Instruct | 32 layers, 32 heads, 8 KV heads |
+| Block Size | 128 tokens | BSA kernel 固定要求 |
+| Threshold | 0.9 / 0.95 | 累积注意力阈值 |
+| Stride | 4 / 8 / 16 | Q/K 下采样步长 |
+| Dataset | RULER niah_single_1 | Sample 0 |
+| Mode | GPU-only | 无 CPU offload |
+
+## Density 定义
+
+```python
+# Density = selected_blocks / total_causal_blocks
+# 在 causal attention 下，只计算下三角区域的 blocks
+# Overall density = 所有层的平均值
+
+def compute_density(mask, causal=True):
+    """
+    mask: [batch, heads, q_blocks, k_blocks] boolean tensor
+    """
+    if causal:
+        causal_mask = torch.tril(torch.ones(q_blocks, k_blocks))
+        total = causal_mask.sum() * batch * heads
+        selected = (mask & causal_mask).sum()
+    return selected / total
+```
+
+## 测试结果
+
+### threshold=0.9
+
+#### Overall Density (平均)
+
+| Context | stride=4 | stride=8 | stride=16 |
+|---------|----------|----------|-----------|
+| **4K** | 0.5220 (52.2%) | 0.5292 (52.9%) | 0.5430 (54.3%) |
+| **8K** | 0.5152 (51.5%) | 0.5252 (52.5%) | 0.5396 (54.0%) |
+| **16K** | 0.4682 (46.8%) | 0.4775 (47.8%) | 0.4888 (48.9%) |
+| **32K** | 0.3700 (37.0%) | 0.4012 (40.1%) | 0.4196 (42.0%) |
+
+#### Min Density (per layer)
+
+| Context | stride=4 | stride=8 | stride=16 |
+|---------|----------|----------|-----------|
+| **4K** | 0.2805 (Layer 3) | 0.3132 (Layer 3) | 0.3376 (Layer 5) |
+| **8K** | 0.2886 (Layer 5) | 0.2725 (Layer 5) | 0.2995 (Layer 5) |
+| **16K** | 0.2247 (Layer 5) | 0.2349 (Layer 5) | 0.2451 (Layer 5) |
+| **32K** | 0.1799 (Layer 5) | 0.1846 (Layer 5) | 0.1964 (Layer 5) |
+
+### threshold=0.95
+
+#### Overall Density (平均)
+
+| Context | stride=4 | stride=8 | stride=16 |
+|---------|----------|----------|-----------|
+| **4K** | 0.6561 (65.6%) | 0.6699 (67.0%) | 0.6815 (68.2%) |
+| **8K** | 0.6462 (64.6%) | 0.6584 (65.8%) | 0.6732 (67.3%) |
+| **16K** | 0.6004 (60.0%) | 0.6114 (61.1%) | 0.6193 (61.9%) |
+| **32K** | 0.4894 (48.9%) | 0.5203 (52.0%) | 0.5385 (53.9%) |
+
+#### Min Density (per layer)
+
+| Context | stride=4 | stride=8 | stride=16 |
+|---------|----------|----------|-----------|
+| **4K** | 0.3972 (Layer 3) | 0.4348 (Layer 5) | 0.4517 (Layer 4) |
+| **8K** | 0.4004 (Layer 5) | 0.3906 (Layer 5) | 0.4239 (Layer 5) |
+| **16K** | 0.3331 (Layer 5) | 0.3453 (Layer 5) | 0.3589 (Layer 5) |
+| **32K** | 0.2656 (Layer 5) | 0.2784 (Layer 5) | 0.2917 (Layer 5) |
+
+### threshold 对比 (stride=8)
+
+| Context | threshold=0.9 | threshold=0.95 | 差异 |
+|---------|---------------|----------------|------|
+| **4K** | 0.5292 (52.9%) | 0.6699 (67.0%) | -14.1% |
+| **8K** | 0.5252 (52.5%) | 0.6584 (65.8%) | -13.3% |
+| **16K** | 0.4775 (47.8%) | 0.6114 (61.1%) | -13.4% |
+| **32K** | 0.4012 (40.1%) | 0.5203 (52.0%) | -11.9% |
+
+## 关键发现
+
+### 1. Context Length 影响最大
+
+Density 随 context length 显著下降（threshold=0.9, stride=8）：
+- 4K: 52.9% density
+- 8K: 52.5% density
+- 16K: 47.8% density
+- 32K: 40.1% density
+
+**结论**: 长序列有更多稀疏化机会，XAttention 的优势在长序列上更明显。
+
+### 2. Threshold 影响显著
+
+threshold=0.9 比 0.95 的 density 低约 12-14%：
+- 0.9 更激进，选择更少的 blocks
+- 0.95 更保守，保留更多 blocks
+- 两者准确性都不受影响（RULER NIAH 全部 PASS）
+
+### 3. Stride 影响较小
+
+同一 context 下，不同 stride 的 density 差异约 2-5%：
+- stride 越大 → density 略高（采样越粗糙，选择更保守）
+- stride=4 最激进，stride=16 最保守
+
+### 4. Min Density 集中在中间层
+
+- 大多数情况下 min density 出现在 Layer 5
+- 中间层的稀疏性最高，首尾层相对密集
+- 这符合 Transformer 注意力模式的一般规律
+
+### 5. 最佳稀疏化配置
+
+32K + stride=4 + threshold=0.9 达到最低 density：
+- Overall: **37.0%** (节省 63% 计算)
+- Min: **18.0%** (Layer 5)
+
+### 6. 准确性稳定
+
+所有配置下 RULER NIAH 测试都 PASS (score=1.0)，说明：
+- threshold=0.9 和 0.95 都足够保守，不损失准确性
+- 不同 stride 不影响最终结果
+
+## 推荐配置
+
+| 场景 | threshold | stride | 说明 |
+|------|-----------|--------|------|
+| 精度优先 | 0.95 | 8 | 保守配置，density ~52-67% |
+| 平衡 | 0.9 | 8 | 默认配置，density ~40-53% |
+| 性能优先 | 0.9 | 4 | 激进配置，density ~37-52% |
+
+## 测试命令
+
+```bash
+# 基本测试
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --sample-indices 0 \
+    --max-model-len 33792 \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9 \
+    --sparse-stride 8 \
+    --gpu-utilization 0.85
+
+# 参数说明
+# --sparse-policy XATTN_BSA    启用 XAttention Block Sparse Attention
+# --sparse-threshold 0.9       累积注意力阈值 (0.9-0.99)
+# --sparse-stride 8            Q/K 下采样步长 (4/8/16)
+```
+
+## DensityObserver 使用
+
+```python
+from nanovllm.utils.density_observer import DensityObserver
+
+# 启用并重置
+DensityObserver.enable()
+DensityObserver.complete_reset()
+
+# ... 运行 inference (compute_prefill 自动记录) ...
+
+# 获取结果
+summary = DensityObserver.get_summary()
+# {
+#     "mode": "gpu_only",
+#     "overall_density": 0.40,  # 所有层的平均值
+#     "per_layer_density": {0: 0.55, 1: 0.45, ...},
+#     "num_layers": 32
+# }
+
+# 获取最低 density
+min_layer, min_density = DensityObserver.get_min_density()
+
+# 打印摘要
+DensityObserver.print_summary()
+# [DensityObserver] Mode: gpu_only
+#   Overall density: 0.4012
+#   Min density: 0.1846 (layer 5)
+#   Num layers: 32
+```
+
+## 相关文件
+
+| 文件 | 说明 |
+|------|------|
+| `nanovllm/kvcache/sparse/xattn_bsa.py` | XAttention BSA Policy 实现 |
+| `nanovllm/utils/density_observer.py` | Density 统计 Observer |
+| `nanovllm/ops/xattn.py` | xattn_estimate 核心算法 |
+| `tests/test_ruler.py` | RULER benchmark 测试脚本 |

docs/xattn_density_types.md

@@ -0,0 +1,152 @@
+# XAttention Density Types: Compute vs Communication
+
+XAttention BSA 统计两种不同粒度的 density，它们反映不同的优化效果。
+
+## 两种 Density 的定义
+
+### 1. Compute Density（计算密度）
+
+**粒度**: BSA block (128 tokens)
+
+**公式**:
+```
+compute_density = selected_bsa_blocks / total_causal_bsa_blocks
+```
+
+**含义**: 实际需要计算 attention 的 blocks 占 causal 区域的比例。
+
+**影响**: 决定 attention 计算量的减少。
+
+### 2. Communication Density（通信密度）
+
+**粒度**: CPU block (4096 tokens = 32 BSA blocks)
+
+**公式**:
+```
+comm_density = selected_cpu_blocks / total_cpu_blocks
+```
+
+**含义**: 需要从 CPU 传输到 GPU 的 blocks 占总 blocks 的比例。
+
+**影响**: 决定 H2D 传输量的减少。
+
+## 为什么 Comm Density 通常高于 Compute Density
+
+### 聚合效应
+
+由于 CPU block 粒度是 BSA block 的 32 倍，CPU block 选择使用 `any()` 聚合：
+
+```python
+# BSA mask: [B, H, Q_bsa, K_bsa]
+# Reshape to CPU block level
+mask_per_cpu = mask.view(B, H, Q_bsa, num_cpu_blocks, bsa_per_cpu)
+# Any BSA block selected -> whole CPU block needed
+cpu_needed = mask_per_cpu.any(dim=-1).any(dim=2).any(dim=1)
+```
+
+只要 CPU block 中**任意一个**:
+- Head 选择了该 block，或
+- Q position 选择了该 block，或
+- BSA sub-block 被选中
+
+则整个 CPU block 都需要传输。
+
+### 示例
+
+| 场景 | Compute Density | Comm Density | 说明 |
+|------|-----------------|--------------|------|
+| 64K context, threshold=0.9 | 37% | 100% | 稀疏 blocks 均匀分布在所有 CPU blocks |
+| 32K context, threshold=0.9 | 50% | 100% | 同上 |
+
+## 测试结果
+
+### 测试命令
+
+```bash
+# Offload 模式测试
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=.:$PYTHONPATH python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_64k \
+    --datasets niah_single_1 \
+    --num-samples 1 \
+    --max-model-len 72000 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+### 输出示例
+
+```
+[DensityObserver] Mode: offload
+  Compute density: 0.3691 (min: 0.3691 @ layer 0)
+  Comm density:    1.0000 (CPU block granularity)
+  Savings ratio:   0.0% H2D transfer reduction
+  Num layers: 1
+  Layer 0 density: 0.369052
+```
+
+## 关键发现
+
+### 当前 XAttention 的通信优化局限
+
+1. **Compute density 有效降低**: ~37% @ 64K context（计算量减少 63%）
+2. **Comm density 没有降低**: 100%（通信量没有减少）
+
+### 原因分析
+
+Attention pattern 的特点：
+- 不同 heads 关注不同位置
+- 不同 Q positions 关注不同 K positions
+- 稀疏选择分布在整个 sequence 上
+
+这导致虽然每个 (head, Q, K) 组合只选择少量 blocks，但聚合后覆盖了所有 CPU blocks。
+
+### 潜在优化方向
+
+1. **Per-head block selection**: 每个 head 独立选择 CPU blocks
+2. **Block clustering**: 将相关 blocks 聚合到同一 CPU block
+3. **Dynamic block size**: 根据 attention pattern 动态调整 CPU block 大小
+
+## DensityObserver API
+
+### 启用和重置
+
+```python
+from nanovllm.utils.density_observer import DensityObserver
+
+DensityObserver.enable()
+DensityObserver.complete_reset()
+DensityObserver.set_mode("offload")  # or "gpu_only"
+```
+
+### 记录
+
+```python
+# Compute density (GPU-only 模式自动记录)
+DensityObserver.record(layer_id, mask, causal=True)
+
+# Comm density (Offload 模式在 select_blocks 中记录)
+DensityObserver.record_comm_density(layer_id, selected_cpu_blocks, total_cpu_blocks)
+```
+
+### 获取结果
+
+```python
+# 总体 density
+overall_compute = DensityObserver.get_overall_density()
+overall_comm = DensityObserver.get_overall_comm_density()
+
+# Per-layer density
+per_layer_compute = DensityObserver.get_per_layer_density()
+per_layer_comm = DensityObserver.get_per_layer_comm_density()
+
+# 打印摘要
+DensityObserver.print_summary()
+```
+
+## 相关文件
+
+- `nanovllm/utils/density_observer.py`: DensityObserver 实现
+- `nanovllm/kvcache/sparse/xattn_bsa.py`: XAttention BSA Policy（select_blocks 中记录 comm density）
+- `tests/test_ruler.py`: RULER benchmark 测试脚本

docs/xattn_kv_chunking_density_test.md

@@ -0,0 +1,122 @@
+# XAttention KV Chunking Density 验证测试
+
+## 背景
+
+验证 XAttention Triton kernel 是否只能沿 Q 轴分 chunk，不能沿 KV 轴分 chunk。
+
+**假设**：`softmax_fuse_block_sum` 需要完整的 K 来计算正确的归一化分母，分 chunk 后的 attention 分布与完整序列不同。
+
+## 测试方法
+
+1. **GPU-only 模式**：一次性对完整序列调用 `xattn_estimate`，记录 Layer 0 的 density
+2. **Offload DEBUG 模式**：分 chunk 调用 `xattn_estimate`，累积 selected/total counts，计算最终 density
+3. 使用相同的 `_debug_k_full` buffer 收集完整 K cache，确保输入数据一致
+
+### 关键代码逻辑
+
+```python
+# Offload DEBUG: 每个 chunk 累积 selected/total
+for each chunk:
+    K_full = _debug_k_full[:, :, :total_k_len, :]  # 累积的 K
+    _, mask_chunk = xattn_estimate(Q_chunk, K_full, threshold=threshold, causal=True)
+
+    # 裁剪到有效区域，计算正确的 causal mask (考虑 Q 偏移量)
+    q_offset_blocks = k_blocks - q_blocks
+    causal_mask = indices <= (q_indices + q_offset_blocks)
+
+    selected += (mask_valid & causal_mask).sum()
+    total += causal_mask.sum()
+
+density = selected / total
+```
+
+## 测试结果
+
+### 64K 序列 (niah_single_1, 序列长度 64891)
+
+| threshold | GPU-only selected | Offload selected | GPU-only density | Offload density | 差异 (selected) |
+|-----------|------------------|------------------|------------------|-----------------|-----------------|
+| **0.90** | 1,524,617 | 1,330,506 | **0.3700** | **0.3229** | 194,111 (12.7%) |
+| **0.95** | 1,955,015 | 1,747,585 | **0.4744** | **0.4241** | 207,430 (10.6%) |
+| **1.00** | 4,118,719 | 4,118,896 | **0.9995** | **0.9995** | -177 (~0%) |
+
+- **total**: 4,120,896 (两种模式一致)
+
+### 32K 序列 (niah_single_1, 序列长度 32485)
+
+| threshold | GPU-only selected | Offload selected | GPU-only density | Offload density | 差异 (selected) |
+|-----------|------------------|------------------|------------------|-----------------|-----------------|
+| **0.90** | 520,314 | 466,937 | **0.5021** | **0.4506** | 53,377 (10.3%) |
+| **0.95** | 647,765 | 602,953 | **0.6251** | **0.5818** | 44,812 (6.9%) |
+| **1.00** | 1,036,295 | 1,036,264 | **0.9999** | **0.9999** | 31 (~0%) |
+
+- **total**: 1,036,320 (两种模式一致)
+
+### 汇总对比
+
+| 序列长度 | threshold | GPU-only density | Offload density | density 差异 |
+|---------|-----------|------------------|-----------------|--------------|
+| 32K | 0.90 | 0.5021 | 0.4506 | 5.2% |
+| 64K | 0.90 | 0.3700 | 0.3229 | 4.7% |
+| 32K | 0.95 | 0.6251 | 0.5818 | 4.3% |
+| 64K | 0.95 | 0.4744 | 0.4241 | 5.0% |
+| 32K | 1.00 | 0.9999 | 0.9999 | ~0% |
+| 64K | 1.00 | 0.9995 | 0.9995 | ~0% |
+
+## 结论
+
+### 1. Softmax 归一化本身是正确的
+
+当 `threshold=1.0`（选择所有 blocks）时，GPU-only 和 Offload 模式的 density 几乎完全对齐（差异 < 0.01%）。
+
+这说明：
+- `_debug_k_full` 正确收集了完整的 K cache
+- 分 chunk 调用 `xattn_estimate` 时，softmax 归一化在正确的 K 序列上计算
+- causal mask 的 Q 偏移量处理正确
+
+### 2. 问题在于 threshold 的应用方式
+
+当 `threshold < 1.0` 时，差异显著（10-13%）：
+
+- **GPU-only**：对完整序列一次性应用 threshold，选择 cumulative attention >= threshold 的 blocks
+- **Offload**：每个 chunk 独立应用 threshold，累积 selected counts
+
+每个 chunk 独立应用 threshold 会导致：
+- 某些在 GPU-only 中被选中的 blocks，在分 chunk 时因 attention 分布不同而未被选中
+- 累积的 selected 比一次性计算的要少
+
+### 3. XAttention Triton kernel 的 KV chunking 限制
+
+**验证结论**：XAttention 的 `xattn_estimate` 可以正确处理 KV chunking（softmax 归一化正确），但 **threshold-based block selection 不能简单累积**。
+
+如果要在 Offload 模式下获得与 GPU-only 一致的 block selection：
+1. 需要先累积所有 chunks 的 attention scores
+2. 最后一次性应用 threshold 选择 blocks
+
+或者接受 10-13% 的 density 差异，这对实际推理准确性的影响需要进一步评估。
+
+## 测试命令
+
+```bash
+# GPU-only 模式
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py --dataset niah_single_1 --sample 0 \
+    --sparse-policy xattn_bsa --sparse-threshold 0.9
+
+# Offload 模式 (64K)
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py --dataset niah_single_1 --sample 0 \
+    --sparse-policy xattn_bsa --sparse-threshold 0.9 --enable-offload
+
+# Offload 模式 (32K)
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py --dataset niah_single_1 --sample 0 \
+    --sparse-policy xattn_bsa --sparse-threshold 0.9 --enable-offload \
+    --data-dir /home/zijie/Code/nano-vllm/tests/data/ruler_32k --max-model-len 34000
+```
+
+## 相关文件
+
+- `nanovllm/kvcache/sparse/xattn_bsa.py`: DEBUG 代码位置
+- `nanovllm/ops/xattn.py`: `xattn_estimate` 实现
+- `nanovllm/utils/density_observer.py`: DensityObserver 实现

docs/xattn_kv_chunking_kernels.md

@@ -0,0 +1,400 @@
+# XAttention KV Chunking Kernels
+
+## 概述
+
+本文档描述了支持 KV 维度分 chunk 的 softmax kernels 实现。这些 kernels 允许在 CPU offload 场景下，沿 KV 维度分块计算 sparse attention estimation，而不需要在 GPU 上保存完整的 raw attention scores。
+
+## 背景
+
+原始的 `softmax_fuse_block_sum` kernel 需要完整的 K 序列来计算正确的 softmax 归一化分母：
+
+```
+softmax(x_i) = exp(x_i) / Σ_j exp(x_j)
+```
+
+如果只有部分 K (KV chunk)，分母 `Σ_j exp(x_j)` 不完整，导致归一化错误。
+
+## 解决方案：三阶段计算
+
+通过将 softmax 计算拆分为三个阶段，实现正确的 KV chunking：
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│                        三阶段 Pipeline                           │
+├─────────────────────────────────────────────────────────────────┤
+│                                                                 │
+│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐           │
+│  │ KV Chunk 0  │   │ KV Chunk 1  │   │ KV Chunk N  │           │
+│  │ attn_scores │   │ attn_scores │   │ attn_scores │           │
+│  └──────┬──────┘   └──────┬──────┘   └──────┬──────┘           │
+│         │                 │                 │                   │
+│         ▼                 ▼                 ▼                   │
+│  ┌─────────────────────────────────────────────────┐           │
+│  │     阶段 1: softmax_compute_partial_stats       │           │
+│  │     计算每个 chunk 的 (m_partial, l_partial)    │           │
+│  └─────────────────────────────────────────────────┘           │
+│         │                 │                 │                   │
+│         ▼                 ▼                 ▼                   │
+│      (m_0, l_0)       (m_1, l_1)       (m_N, l_N)              │
+│         │                 │                 │                   │
+│         └────────────────┬┴─────────────────┘                   │
+│                          ▼                                      │
+│  ┌─────────────────────────────────────────────────┐           │
+│  │         阶段 2: merge_softmax_stats             │           │
+│  │         Host 端合并 → (m_global, l_global)      │           │
+│  └─────────────────────────────────────────────────┘           │
+│                          │                                      │
+│         ┌────────────────┼────────────────┐                     │
+│         ▼                ▼                ▼                     │
+│  ┌─────────────────────────────────────────────────┐           │
+│  │    阶段 3: softmax_normalize_and_block_sum      │           │
+│  │    使用全局 stats 归一化并计算 block sums        │           │
+│  └─────────────────────────────────────────────────┘           │
+│         │                │                │                     │
+│         ▼                ▼                ▼                     │
+│    block_sums_0    block_sums_1    block_sums_N                │
+│         │                │                │                     │
+│         └────────────────┴────────────────┘                     │
+│                          │                                      │
+│                          ▼                                      │
+│                 torch.cat → final mask                          │
+│                                                                 │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+### 阶段 1: `softmax_compute_partial_stats`
+
+计算每个 KV chunk 的 partial statistics：
+- `m_partial`: 该 chunk 内的最大值 (per query row)
+- `l_partial`: 该 chunk 内的 partial sum = Σ exp(x - m_partial)
+
+```python
+m_partial, l_partial = softmax_compute_partial_stats(
+    attn_weights_kv,      # [batch, heads, q_len, k_chunk_len]
+    reshaped_block_size,
+    segment_size,
+    scale,
+    chunk_start=chunk_start,
+    kv_offset=kv_offset,  # KV chunk 在完整序列中的偏移
+    is_causal=True,
+)
+# 输出: m_partial, l_partial 形状为 [batch, heads, q_len]
+```
+
+### 阶段 2: `merge_softmax_stats`
+
+Host 端合并所有 KV chunks 的 statistics：
+
+```python
+m_global, l_global = merge_softmax_stats(m_chunks, l_chunks)
+```
+
+合并公式 (Online Softmax):
+```
+m_new = max(m_global, m_chunk)
+l_new = l_global * exp(m_global - m_new) + l_chunk * exp(m_chunk - m_new)
+```
+
+### 阶段 3: `softmax_normalize_and_block_sum`
+
+使用全局 statistics 归一化并计算 block sums：
+
+```python
+attn_sum_kv = softmax_normalize_and_block_sum(
+    attn_weights_kv,      # [batch, heads, q_len, k_chunk_len]
+    m_global,             # [batch, heads, q_len]
+    l_global,             # [batch, heads, q_len]
+    reshaped_block_size,
+    segment_size,
+    chunk_start=chunk_start,
+    real_q_len=real_q_len,
+    scale=scale,
+    kv_offset=kv_offset,
+    is_causal=True,
+)
+# 输出: [batch, heads, q_blocks, k_chunk_blocks]
+```
+
+## 数学等价性证明
+
+原始 softmax 计算 (完整 K):
+```
+softmax(x_i) = exp(x_i - m) / Σ_j exp(x_j - m)
+```
+
+分 KV chunk 计算:
+```
+Chunk 0: m_0 = max(x[0:N/2]), l_0 = Σ exp(x[0:N/2] - m_0)
+Chunk 1: m_1 = max(x[N/2:N]), l_1 = Σ exp(x[N/2:N] - m_1)
+
+合并:
+m_global = max(m_0, m_1)
+l_global = l_0 * exp(m_0 - m_global) + l_1 * exp(m_1 - m_global)
+         = Σ exp(x[0:N] - m_global)  # 等于全局 sum
+
+归一化:
+softmax(x_i) = exp(x_i - m_global) / l_global  # 正确!
+```
+
+## Causal Mask 处理
+
+两个 kernel 都正确处理了 causal attention：
+
+1. **`softmax_partial_stats_kernel`**: 通过 `kv_offset` 参数确定当前 KV chunk 在完整序列中的位置，正确计算 causal boundary
+
+2. **`softmax_normalize_block_sum_kernel`**: 同样使用 `kv_offset`，对 causal boundary 之后的位置输出 0
+
+## 存储开销分析
+
+### 符号定义
+
+| 符号 | 含义 | 典型值 |
+|------|------|--------|
+| S | seq_len | 64K |
+| B | batch_size | 1 |
+| H | num_heads | 32 |
+| D | head_dim | 128 |
+| T | stride | 4-8 |
+| C | chunk_size | 16K |
+| n | num_kv_chunks = ceil(S/C) | 4 |
+
+### 原始方式 (无 KV chunking)
+
+**attn_weights 峰值内存**:
+```
+[B, H, S/T, S/T] × 4 bytes = B × H × (S/T)² × 4
+
+例: S=64K, T=4, B=1, H=32
+= 1 × 32 × 16384² × 4 = 32 GB
+```
+
+### KV Chunking 方式的额外存储
+
+#### 1. Partial Stats (每个 KV chunk)
+
+```
+m_partial: [B, H, C/T] × 4 bytes
+l_partial: [B, H, C/T] × 4 bytes
+
+单个 chunk = 2 × B × H × (C/T) × 4
+           = 2 × 1 × 32 × 4096 × 4 = 1 MB
+```
+
+#### 2. Global Stats
+
+```
+m_global: [B, H, S/T] × 4 bytes
+l_global: [B, H, S/T] × 4 bytes
+
+= 2 × B × H × (S/T) × 4
+= 2 × 1 × 32 × 16384 × 4 = 4 MB
+```
+
+#### 3. 总额外开销
+
+```
+total_extra = n × partial_stats + global_stats
+            = 4 × 1MB + 4MB = 8 MB
+```
+
+### 存储开销随 seqlen 变化
+
+| seqlen | num_chunks | 原始 attn_weights | 额外 stats | 比例 |
+|--------|------------|-------------------|------------|------|
+| 16K | 1 | 2 GB | 2 MB | 0.1% |
+| 32K | 2 | 8 GB | 4 MB | 0.05% |
+| 64K | 4 | 32 GB | 8 MB | 0.025% |
+| 128K | 8 | 128 GB | 16 MB | 0.012% |
+
+### 复杂度分析
+
+| 存储组件 | 复杂度 | 说明 |
+|----------|--------|------|
+| 原始 attn_weights | O(S²) | 二次增长 |
+| Partial/Global stats | O(S) | 线性增长 |
+| **相对开销** | O(1/S) | **随 seqlen 递减** |
+
+### 峰值显存优化
+
+KV chunking 的主要收益是**峰值显存**从 O(S²) 降到 O(S×C)：
+
+```
+原始: O(B × H × (S/T)²)           # 完整 attn_weights
+KV chunking: O(B × H × (S/T) × (C/T))  # 一次只处理一个 chunk
+```
+
+以 S=128K, C=16K 为例：
+- 原始峰值: ~128 GB
+- KV chunking 峰值: ~16 GB (降低 **8 倍**)
+
+## 支持不同 Q/KV Chunk Size
+
+三阶段 pipeline 支持 Q 和 KV 使用不同的 chunk size：
+
+```python
+q_chunk_size = 8192   # Q 分块大小
+kv_chunk_size = 16384 # KV 分块大小
+
+for q_chunk_idx in range(q_chunk_num):
+    Q_chunk = Q[:, :, q_start:q_end, :]  # [B, H, q_chunk_size, D]
+
+    for kv_chunk_idx in range(kv_chunk_num):
+        K_chunk = K[:, :, kv_start:kv_end, :]  # [B, H, kv_chunk_size, D]
+        # ... 三阶段处理
+```
+
+### 测试验证结果
+
+| Config | seq_len | Q chunks | KV chunks | density | 对齐 |
+|--------|---------|----------|-----------|---------|------|
+| Q=16K, KV=16K | 64891 | 4 | 4 | 0.1117 | ✓ 100% |
+| Q=8K, KV=16K | 64891 | 8 | 4 | 0.1112 | ✓ 100% |
+| Q=16K, KV=8K | 64891 | 4 | 8 | 0.1117 | ✓ 100% |
+| Q=8K, KV=8K | 64891 | 8 | 8 | 0.1112 | ✓ 100% |
+| Q=4K, KV=16K | 64891 | 16 | 4 | 0.1109 | ✓ 100% |
+
+## API 参考
+
+### `softmax_compute_partial_stats`
+
+```python
+def softmax_compute_partial_stats(
+    attn_weights_slice: torch.Tensor,  # [batch, heads, q_len, k_chunk_len]
+    reshaped_block_size: int,
+    segment_size: int,
+    scale: float,
+    chunk_start: int = 0,              # Q chunk 起始位置 (reshaped space)
+    kv_offset: int = 0,                # KV chunk 偏移 (reshaped space)
+    is_causal: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """返回 (m, l) partial stats"""
+```
+
+### `merge_softmax_stats`
+
+```python
+def merge_softmax_stats(
+    m_chunks: list,  # List of [batch, heads, q_len] tensors
+    l_chunks: list,  # List of [batch, heads, q_len] tensors
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """返回 (m_global, l_global)"""
+```
+
+### `softmax_normalize_and_block_sum`
+
+```python
+def softmax_normalize_and_block_sum(
+    attn_weights_slice: torch.Tensor,  # [batch, heads, q_len, k_chunk_len]
+    m_global: torch.Tensor,            # [batch, heads, q_len]
+    l_global: torch.Tensor,            # [batch, heads, q_len]
+    reshaped_block_size: int,
+    segment_size: int,
+    chunk_start: int,
+    real_q_len: int,
+    scale: float,
+    kv_offset: int = 0,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """返回 block sums [batch, heads, q_blocks, k_chunk_blocks]"""
+```
+
+## 使用示例
+
+```python
+from nanovllm.ops.xattn import (
+    flat_group_gemm_fuse_reshape,
+    softmax_compute_partial_stats,
+    softmax_normalize_and_block_sum,
+    merge_softmax_stats,
+    find_blocks_chunked,
+)
+
+# 对每个 Q chunk
+for q_chunk_idx in range(q_chunk_num):
+    Q_chunk = Q_padded[:, :, q_start:q_end, :]
+
+    # 阶段 1: 每个 KV chunk 计算 partial stats
+    m_chunks, l_chunks = [], []
+    attn_weights_chunks = []
+
+    for kv_chunk_idx in range(kv_chunk_num):
+        K_chunk = K_padded[:, :, kv_start:kv_end, :]
+        kv_offset = kv_chunk_idx * kv_chunk_size // STRIDE
+
+        # 计算 raw scores
+        attn_weights = flat_group_gemm_fuse_reshape(
+            Q_chunk, K_chunk, STRIDE,
+            chunk_start=chunk_start,
+            chunk_end=chunk_end,
+            is_causal=False,  # K 不完整
+        )
+        attn_weights_chunks.append(attn_weights)
+
+        # 计算 partial stats
+        m, l = softmax_compute_partial_stats(
+            attn_weights, block_size, segment_size, scale,
+            chunk_start=chunk_start,
+            kv_offset=kv_offset,
+            is_causal=True,
+        )
+        m_chunks.append(m)
+        l_chunks.append(l)
+
+    # 阶段 2: 合并 stats
+    m_global, l_global = merge_softmax_stats(m_chunks, l_chunks)
+
+    # 阶段 3: 归一化并计算 block sums
+    block_sums_list = []
+    for kv_chunk_idx, attn_weights in enumerate(attn_weights_chunks):
+        kv_offset = kv_chunk_idx * kv_chunk_size // STRIDE
+        block_sums = softmax_normalize_and_block_sum(
+            attn_weights, m_global, l_global,
+            block_size, segment_size, chunk_start, real_q_len, scale,
+            kv_offset=kv_offset,
+            is_causal=True,
+        )
+        block_sums_list.append(block_sums)
+
+    # 拼接并选择 blocks
+    attn_sum = torch.cat(block_sums_list, dim=-1)
+    mask = find_blocks_chunked(attn_sum, ...)
+```
+
+## 性能对比
+
+| 方面 | 原始实现 | KV Chunking 实现 |
+|------|---------|-----------------|
+| Kernel 数量 | 1 | 2 (stats + normalize) |
+| Raw scores 读取次数 | 2 | 2 |
+| 额外内存 | 0 | O(B × H × S/T × 2) for (m, l) |
+| Host 计算 | 无 | merge stats (轻量) |
+| **峰值显存** | O(S²) | **O(S × C)** |
+
+## 验证测试
+
+### 批量测试结果
+
+测试脚本 `tests/test_xattn_kv_chunking_batch.py` 验证了不同 seqlen 下的一致性：
+
+```
+| seq_len | stride | threshold | kv_chunks | density_api | density_kv | diff     | mask_diff | status |
+|---------|--------|-----------|-----------|-------------|------------|----------|-----------|--------|
+|    3688 |      4 |      0.90 |         1 |    0.383405 |   0.383405 | 0.000000 |   0.0000% |   PASS |
+|    7888 |      4 |      0.90 |         1 |    0.290611 |   0.290611 | 0.000000 |   0.0000% |   PASS |
+|   15685 |      4 |      0.90 |         1 |    0.197724 |   0.197724 | 0.000000 |   0.0000% |   PASS |
+|   32485 |      4 |      0.90 |         2 |    0.159023 |   0.159023 | 0.000000 |   0.0000% |   PASS |
+|   64891 |      4 |      0.90 |         4 |    0.111656 |   0.111656 | 0.000000 |   0.0000% |   PASS |
+```
+
+### 关键结论
+
+1. **数学等价性**: density_diff = 0.000000 对于所有测试
+2. **Mask 完全对齐**: mask_diff = 0.0000% 对于所有测试
+3. **支持任意 Q/KV chunk size 组合**
+
+## 相关文件
+
+- `nanovllm/ops/xattn.py`: Kernel 实现
+- `tests/test_xattn_estimate_alignment.py`: 单文件验证测试
+- `tests/test_xattn_kv_chunking_batch.py`: 批量验证测试
+- `docs/xattn_kernels_guide.md`: 原始 kernel 文档

docs/xattn_memory_benchmark.md

@@ -0,0 +1,154 @@
+# XAttention Memory Benchmark
+
+GPU-only 模式下 XAttention 的内存使用分析。
+
+## 测试配置
+
+### 硬件
+- **GPU**: NVIDIA A100 80GB (用于基准测试)
+- **目标**: 验证在 RTX 3090/4090 (24GB) 上的可行性
+
+### 模型
+- **Model**: Qwen3-0.6B (28 layers, 16 heads, 8 KV heads, head_dim=128)
+- **Context Length**: 32K (max_model_len=40960)
+
+### XAttention 配置
+- **Sparse Policy**: XATTN_BSA
+- **Threshold**: 0.9
+- **Block Size**: 128 tokens (BSA block)
+- **Stride**: 8
+
+---
+
+## 内存使用分析
+
+### 基准测试 (gpu-utilization=0.9)
+
+| 指标 | 数值 |
+|------|------|
+| KV Cache | 157 blocks × 448 MB = 70.3 GB |
+| **峰值内存** | **73,949 MiB (72.2 GB)** |
+| GPU 利用率 | 90.2% |
+
+### 24GB 显存可行性测试
+
+| gpu-utilization | KV Cache Blocks | KV Cache Size | 峰值内存 | 测试结果 |
+|-----------------|-----------------|---------------|----------|----------|
+| 0.25 | 39 blocks | 17.5 GB | **20.6 GB** | ✅ 5/5 PASSED |
+| 0.28 | 44 blocks | 19.7 GB | **22.8 GB** | ✅ 5/5 PASSED |
+
+---
+
+## 24GB 显存推荐配置
+
+适用于 **RTX 3090 / RTX 4090 (24GB)**：
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python tests/test_ruler.py \
+    --model ~/models/Qwen3-0.6B \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 5 \
+    --max-model-len 40960 \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9 \
+    --gpu-utilization 0.28
+```
+
+### 配置说明
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| `--gpu-utilization` | 0.28 | 限制 GPU 内存使用到 ~23GB |
+| `--max-model-len` | 40960 | 支持 32K+ context |
+| `--sparse-policy` | XATTN_BSA | 启用 XAttention 稀疏注意力 |
+| `--sparse-threshold` | 0.9 | 选择覆盖 90% attention 的 blocks |
+
+---
+
+## 内存分解
+
+### Qwen3-0.6B @ 32K Context
+
+| 组件 | 计算公式 | 大小 |
+|------|----------|------|
+| 模型权重 | 0.6B × 2 bytes | ~1.2 GB |
+| KV Cache (per-token) | 2 × 28 layers × 8 kv_heads × 128 head_dim × 2 bytes | 112 KB |
+| KV Cache (per-block) | 112 KB × 4096 tokens | 448 MB |
+| KV Cache (44 blocks) | 448 MB × 44 | 19.7 GB |
+| XAttention Buffers | GQA + mask + KV chunking | ~0.3 GB |
+| 中间激活 | 运行时分配 | ~1.5 GB |
+| **总计** | | **~22.8 GB** |
+
+---
+
+## 性能指标
+
+### RULER niah_single_1 (5 samples)
+
+| 指标 | gpu-util=0.25 | gpu-util=0.28 | gpu-util=0.9 |
+|------|---------------|---------------|--------------|
+| 准确率 | 100% (5/5) | 100% (5/5) | 100% (5/5) |
+| 耗时 | 11.4s | 11.5s | 11.6s |
+| Compute Density | 24.77% | 24.77% | 24.77% |
+| Min Layer Density | 4.29% (Layer 5) | 4.29% (Layer 5) | 4.29% (Layer 5) |
+
+**结论**: 降低 gpu-utilization 不影响准确率和性能，只影响可支持的最大序列长度。
+
+---
+
+## 不同模型的估算
+
+### KV Cache 公式
+
+```
+KV Cache per-token = 2 × num_layers × num_kv_heads × head_dim × dtype_size
+KV Cache per-block = per-token × block_size
+```
+
+### 常见模型估算 (32K context, block_size=4096)
+
+| 模型 | Layers | KV Heads | Head Dim | Per-Token | 32K Tokens | 24GB 可行? |
+|------|--------|----------|----------|-----------|------------|------------|
+| Qwen3-0.6B | 28 | 8 | 128 | 112 KB | 3.5 GB | ✅ 是 |
+| Qwen3-4B | 36 | 8 | 128 | 144 KB | 4.5 GB | ✅ 是 |
+| Llama-3.1-8B | 32 | 8 | 128 | 128 KB | 4.0 GB | ⚠️ 需要 offload |
+| Qwen2.5-7B | 28 | 4 | 128 | 56 KB | 1.8 GB | ✅ 是 |
+
+注: 8B 模型权重约 16GB，加上 KV cache 超过 24GB，需要 CPU offload。
+
+---
+
+## 使用建议
+
+### RTX 3090/4090 (24GB)
+
+1. **小模型 (≤4B)**：可直接使用 GPU-only + XAttention
+   ```bash
+   --gpu-utilization 0.28 --sparse-policy XATTN_BSA
+   ```
+
+2. **大模型 (7B-8B)**：需要 CPU offload
+   ```bash
+   --enable-offload --num-gpu-blocks 4 --num-cpu-blocks 32
+   ```
+
+### A100 (40GB/80GB)
+
+1. **所有模型**：可使用 GPU-only 模式
+   ```bash
+   --gpu-utilization 0.9 --sparse-policy XATTN_BSA
+   ```
+
+---
+
+## 相关文件
+
+- `tests/test_ruler.py`: RULER 测试脚本
+- `nanovllm/kvcache/sparse/xattn_bsa.py`: XAttention BSA Policy 实现
+- `docs/gpuonly_density_alignment_test.md`: Density 对齐验证
+
+---
+
+**Date**: 2026-02-02
+**Author**: Zijie Tian

docs/xattn_offload_profiling_32k.md

@@ -0,0 +1,184 @@
+# XAttention Offload Profiling - 32K Context
+
+Nsys profiling 分析 XAttention vs Full Attention 在 Offload 模式下的性能。
+
+**测试日期**: 2026-02-05
+**测试模型**: Llama-3.1-8B-Instruct
+**Context**: 32K tokens
+**GPU**: A100-80GB (GPU 0)
+
+---
+
+## 测试配置
+
+| 参数 | Full | XAttention |
+|------|------|------------|
+| Policy | FULL | XATTN_BSA |
+| Block size | 4096 | 4096 |
+| GPU blocks | 4 | 4 |
+| Threshold | - | 0.95 |
+| Density | 100% | ~50% |
+
+---
+
+## XAttention 各阶段时间统计
+
+### NVTX Markers Summary
+
+| 阶段 | 总时间(ms) | 调用次数 | 平均时间(ms) | 说明 |
+|------|------------|----------|--------------|------|
+| xattn_find_blocks | 1155.1 | 256 | 4.51 | 块选择 (threshold-based) |
+| xattn_estimate_pass1 | 588.3 | 256 | 2.30 | 第一轮: partial stats |
+| xattn_compute_historical | 512.0 | 224 | 2.29 | 历史 KV attention |
+| xattn_estimate_pass2 | 501.6 | 256 | 1.96 | 第二轮: block sums |
+| xattn_estimate_merge | 197.9 | 256 | 0.77 | 合并 softmax stats |
+| xattn_compute_merge | 93.8 | 256 | 0.37 | 计算结果合并 |
+| xattn_compute_current | 59.2 | 256 | 0.23 | 当前 chunk attention |
+
+### 时间分配
+
+```
+Total XAttention overhead: 3108 ms
+
+Estimate 阶段: 1288 ms (41.4%)
+  - pass1: 588 ms
+  - pass2: 502 ms
+  - merge: 198 ms
+
+Find blocks: 1155 ms (37.2%)
+
+Compute 阶段: 665 ms (21.4%)
+  - historical: 512 ms
+  - merge: 94 ms
+  - current: 59 ms
+```
+
+---
+
+## Chunk7 (最后一个 chunk) 对比
+
+### Per-Layer 时间
+
+| Policy | Layer 0 | Layer 1 | ... | Layer 31 | Avg |
+|--------|---------|---------|-----|----------|-----|
+| Full | 36.5 ms | 33.6 ms | ... | 32.7 ms | ~35 ms |
+| XAttn | 39.7 ms | 39.3 ms | ... | 38.5 ms | ~38 ms |
+
+### 分析
+
+Chunk7 是序列的最后 ~4K tokens (3813 tokens)，此时：
+- K 长度: 32485 tokens
+- Density: 42.08%
+
+**结论**: XAttention 在 Chunk7 比 Full 慢约 8%，原因：
+1. Estimate 开销无法被稀疏计算收益抵消
+2. 42% density 仍然较高，稀疏收益有限
+
+---
+
+## Full Attention Chunk7 详细数据
+
+```
+Layer  Time(ms)
+L0     36.5
+L1     44.3
+L2     43.7
+L3     38.7
+L4     34.2
+L5     45.2
+...
+L31    32.7
+Avg    ~35
+```
+
+---
+
+## XAttention Chunk7 详细数据
+
+```
+Layer  Time(ms)
+L0     39.7
+L1     39.3
+L2     37.1
+L3     39.1
+L4     38.7
+L5     39.4
+...
+L31    38.5
+Avg    ~38
+```
+
+---
+
+## 性能瓶颈分析
+
+### 1. xattn_find_blocks 开销过高
+
+- 平均 4.51 ms per call
+- 占总时间 37.2%
+- 原因: threshold-based 块选择涉及排序和累积求和
+
+### 2. 两轮 estimate 开销
+
+- Pass1 + Pass2 共 1090 ms
+- 需要遍历所有 KV chunks 两次
+- 可优化方向: 单轮 estimate
+
+### 3. Compute 阶段相对高效
+
+- 只占 21.4%
+- 说明 BSA 稀疏计算本身效率不错
+
+---
+
+## 优化建议
+
+### 短期
+
+1. **减少 find_blocks 开销**
+   - 使用 top-k 而不是 threshold
+   - 预分配 mask buffer 避免动态分配
+
+2. **合并 estimate 两轮**
+   - 在单轮中同时计算 stats 和 block sums
+
+### 中期
+
+1. **estimate 阶段使用更小的 block_size**
+   - 当前 block_size=4096 对 estimate 不友好
+   - 参考 `docs/estimate_block_size_performance.md`
+
+2. **Pipeline estimate 和 H2D**
+   - 将 estimate 与下一个 chunk 的 H2D 重叠
+
+### 长期
+
+1. **预测式块选择**
+   - 基于历史 pattern 预测下一个 chunk 的重要 blocks
+   - 减少 estimate 开销
+
+---
+
+## 相关文件
+
+- `results/nsys/full_offload_32k_blk4096_20260205_023257.nsys-rep`
+- `results/nsys/xattn_offload_32k_blk4096_20260205_023435.nsys-rep`
+
+---
+
+## 命令
+
+### Profile Full
+```bash
+bash scripts/profile_offload.sh --policy full --ctx-len 32k --gpu 0 --model ~/models/Llama-3.1-8B-Instruct
+```
+
+### Profile XAttention
+```bash
+bash scripts/profile_offload.sh --policy xattn --ctx-len 32k --gpu 0 --model ~/models/Llama-3.1-8B-Instruct
+```
+
+### 分析 NVTX
+```bash
+nsys stats --report nvtx_pushpop_sum <file>.nsys-rep
+```

docs/xattn_offload_stream_sync_fix.md

@@ -0,0 +1,307 @@
+# XAttention Offload Stream Synchronization Fix
+
+修复 XAttention BSA Policy 在 Offload 模式下的 CUDA stream 同步 bug。
+
+**修复日期**: 2026-02-05
+**Commit**: `829b311`
+**影响文件**: `nanovllm/kvcache/sparse/xattn_bsa.py`, `nanovllm/kvcache/offload_engine.py`
+
+---
+
+## 问题描述
+
+### 症状
+
+在 Offload 模式下运行 RULER benchmark 时，XAttention BSA 的 `select_blocks` 方法中 Pass 1 和 Pass 2 从**同一个 CPU block** 加载的 K 数据不一致：
+
+```
+Pass 1: K_chunk sum = 745472.00 (正确)
+Pass 2: K_chunk sum = 0.00      (错误，数据未加载完成)
+```
+
+这导致 attention 计算结果错误，RULER 准确率下降。
+
+### 复现条件
+
+- 模式: Offload (`--enable-offload`)
+- Context: ≥ 32K tokens
+- 稀疏策略: `--sparse-policy XATTN_BSA`
+
+---
+
+## 根因分析
+
+### Stream 配置回顾
+
+nano-vllm 的 CPU offload 使用多个 CUDA streams 实现 pipeline：
+
+| Stream | 用途 |
+|--------|------|
+| `slot_transfer_streams[i]` | H2D 传输 (CPU → GPU slot) |
+| `compute_stream` | Attention 计算 |
+| `prefill_offload_streams[i]` | D2H 传输 (GPU → CPU cache) |
+
+### 同步机制
+
+`wait_slot_layer(slot)` 使用 event 机制同步：
+
+```python
+def wait_slot_layer(self, slot_idx: int):
+    """Make compute_stream wait for H2D transfer completion."""
+    self.compute_stream.wait_event(self.ring_slot_ready[slot_idx])
+```
+
+### Bug 根因
+
+在 `select_blocks` 方法中：
+
+1. H2D 传输在 `slot_transfer_streams` 上执行
+2. `wait_slot_layer` 让 `compute_stream` 等待传输完成
+3. **但是** 后续的 compute kernels 在**默认 stream** 上执行，而不是 `compute_stream`
+
+```python
+# Bug 代码
+offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id)
+offload_engine.wait_slot_layer(slot)  # compute_stream 等待
+
+# 这些 kernel 在默认 stream 上运行，没有等待 H2D 完成！
+k_block = offload_engine.get_k_for_slot(slot)
+K_chunk = k_block.transpose(1, 2)
+# ... 后续计算 ...
+```
+
+### 时序图
+
+```
+slot_transfer_stream:  [====H2D====]
+compute_stream:             |wait|
+default_stream:        [kernel1][kernel2]  ← 没有等待！
+                            ↑
+                       数据未就绪
+```
+
+---
+
+## 修复方案
+
+### 核心修改
+
+将所有 estimate 阶段的 compute kernels 包装在 `with torch.cuda.stream(compute_stream):` 中：
+
+```python
+# 修复后代码
+compute_stream = offload_engine.compute_stream
+
+offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id)
+offload_engine.wait_slot_layer(slot)  # compute_stream 等待
+
+# 所有计算在 compute_stream 上执行
+with torch.cuda.stream(compute_stream):
+    k_block = offload_engine.get_k_for_slot(slot)
+    K_chunk = k_block.transpose(1, 2)
+    # ... 后续计算 ...
+```
+
+### 修复位置
+
+`select_blocks` 方法中共 6 处需要修复：
+
+| 位置 | 阶段 | 修复内容 |
+|------|------|----------|
+| Pass 1 历史 blocks | `xattn_estimate_pass1` | 历史 KV chunk 处理 |
+| Pass 1 当前 chunk | `xattn_estimate_pass1` | 当前 GPU 上的 K 处理 |
+| Step 2 合并 | `merge_softmax_stats` | softmax stats 合并 |
+| Pass 2 历史 blocks | `xattn_estimate_pass2` | 带全局 stats 的 block_sum |
+| Pass 2 当前 chunk | `xattn_estimate_pass2` | 当前 chunk 的 block_sum |
+| Step 4 block 选择 | `find_blocks_chunked` | 最终 block 选择 |
+
+### 时序图（修复后）
+
+```
+slot_transfer_stream:  [====H2D====]
+compute_stream:             |wait|[kernel1][kernel2]
+                                   ↑
+                              数据已就绪
+```
+
+---
+
+## 代码变更详情
+
+### 1. Pass 1 历史 blocks 处理
+
+```python
+# Before (bug)
+for kv_chunk_idx, cpu_block_id in enumerate(available_blocks):
+    offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id)
+    offload_engine.wait_slot_layer(slot)
+
+    k_block = offload_engine.get_k_for_slot(slot)  # 默认 stream
+    K_chunk = k_block.transpose(1, 2)
+    # ... compute ...
+
+# After (fixed)
+compute_stream = offload_engine.compute_stream
+
+for kv_chunk_idx, cpu_block_id in enumerate(available_blocks):
+    offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id)
+    offload_engine.wait_slot_layer(slot)
+
+    with torch.cuda.stream(compute_stream):  # 显式指定 stream
+        k_block = offload_engine.get_k_for_slot(slot)
+        K_chunk = k_block.transpose(1, 2)
+        # ... compute ...
+```
+
+### 2. 移除 STRONG SYNC
+
+`offload_engine.py` 中移除了不必要的强同步：
+
+```python
+# Removed from load_to_slot_layer() and load_k_only_to_slot_layer()
+# STRONG SYNC: Synchronize all prefill offload streams before H2D
+# for offload_stream in self.prefill_offload_streams:
+#     offload_stream.synchronize()
+```
+
+这些同步现在由 event 机制正确处理，不再需要阻塞式同步。
+
+### 3. 其他清理
+
+- 移除 DEBUG print 语句
+- 移除 `torch.save()` debug 代码
+- 合并多个 fallback 条件
+- 将 `chunk_size` 默认值从 16384 改为 4096（匹配 offload Q chunk size）
+
+---
+
+## 测试验证
+
+### 测试命令
+
+**GPU 0 - Offload 模式测试**:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Llama-3.1-8B-Instruct \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 10 \
+    --max-model-len 40960 \
+    --enable-offload \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+**GPU 1 - GPU-only 模式测试**:
+
+```bash
+CUDA_VISIBLE_DEVICES=1 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+    python tests/test_ruler.py \
+    --model ~/models/Qwen3-0.6B \
+    --data-dir tests/data/ruler_32k \
+    --datasets niah_single_1 \
+    --num-samples 10 \
+    --max-model-len 40960 \
+    --sparse-policy XATTN_BSA \
+    --sparse-threshold 0.9
+```
+
+### 测试结果
+
+| 模式 | 模型 | Context | Samples | Pass Rate | Density |
+|------|------|---------|---------|-----------|---------|
+| Offload | Llama-3.1-8B | 32K | 10/10 | **100%** | 9.53% |
+| GPU-only | Qwen3-0.6B | 32K | 10/10 | **100%** | 9.84% |
+
+### Density 对齐验证
+
+| 模式 | Layer 0 Density | 差异 |
+|------|-----------------|------|
+| GPU-only | 9.84% | - |
+| Offload | 9.53% | ~3% |
+
+~3% 的差异是预期的，因为两种模式的 KV 累积模式不同：
+- GPU-only: 一次性处理所有 KV
+- Offload: 分 chunk 处理，每个 chunk 独立计算 softmax stats 后合并
+
+---
+
+## 技术细节
+
+### 三阶段 KV Chunking 流程
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│  Stage 1: softmax_compute_partial_stats                     │
+│    └── 每个 KV chunk 独立计算 partial stats (m_i, l_i)       │
+│                                                             │
+│  Stage 2: merge_softmax_stats                               │
+│    └── Host 端合并所有 chunks: (m_global, l_global)          │
+│                                                             │
+│  Stage 3: softmax_normalize_and_block_sum                   │
+│    └── 使用全局 stats 归一化并计算 block sums                 │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### Stream 配置要求
+
+| 操作类型 | Stream | 原因 |
+|----------|--------|------|
+| H2D 传输 | `slot_transfer_streams` | 异步传输，不阻塞计算 |
+| D2H 传输 | `prefill_offload_streams` | 异步 offload，不阻塞计算 |
+| Estimate kernels | `compute_stream` | 与 attention 计算共享，确保同步 |
+| Attention kernels | `compute_stream` | 主计算流 |
+
+### Event 同步机制
+
+```python
+# H2D 传输完成后记录 event
+self.ring_slot_ready[slot_idx].record(slot_transfer_stream)
+
+# 计算前等待 H2D 完成
+self.compute_stream.wait_event(self.ring_slot_ready[slot_idx])
+
+# 计算完成后记录 event（用于下一轮 H2D）
+self.ring_slot_compute_done[slot_idx].record(compute_stream)
+```
+
+---
+
+## 相关文档
+
+- [`docs/architecture_guide.md`](architecture_guide.md): Stream 配置和 ring buffer 架构
+- [`docs/xattn_kv_chunking_kernels.md`](xattn_kv_chunking_kernels.md): 三阶段 softmax kernels
+- [`docs/gpuonly_density_alignment_test.md`](gpuonly_density_alignment_test.md): Density 对齐测试
+- [`docs/xattn_bsa_policy_design.md`](xattn_bsa_policy_design.md): XAttention BSA Policy 设计
+
+---
+
+## 经验总结
+
+### 1. Stream 同步的隐蔽性
+
+CUDA stream 同步 bug 很难发现：
+- 数据可能"大部分时间"正确（取决于时序）
+- 错误表现为随机/间歇性的结果偏差
+- 需要精确的 debug logging 才能定位
+
+### 2. Event vs Synchronize
+
+| 方法 | 优点 | 缺点 |
+|------|------|------|
+| `stream.wait_event()` | 非阻塞，保持 pipeline | 只同步指定 stream |
+| `stream.synchronize()` | 保证完成 | 阻塞整个 stream，破坏 pipeline |
+
+**最佳实践**: 使用 event 进行精确同步，避免 synchronize 阻塞。
+
+### 3. 调试方法
+
+```python
+# 打印 tensor sum 验证数据一致性
+print(f"K_chunk sum = {K_chunk.sum().item()}")
+
+# 保存中间结果进行离线比较
+torch.save({'K': K_chunk, 'layer': layer_id}, f'/tmp/debug_{pass}_{chunk}.pt')
+```

docs/xattn_performance_analysis.md

@@ -0,0 +1,170 @@
+# XAttention Performance Analysis
+
+本文档记录 XAttention 在不同配置下的性能分析结果，包括 NVTX 标记位置、block size 影响和性能瓶颈。
+
+## NVTX 标记
+
+XAttention 代码中添加了 NVTX 标记用于 nsys profiling，便于分析 estimate 和 compute 阶段的性能。
+
+### 标记位置
+
+| 模式 | 标记名称 | 文件位置 | 说明 |
+|------|---------|---------|------|
+| GPU-only | `xattn_estimate` | `xattn_bsa.py:compute_prefill` | xattn_estimate 调用 |
+| GPU-only | `xattn_bsa_compute` | `xattn_bsa.py:compute_prefill` | BSA kernel 调用 |
+| Offload | `xattn_estimate_gemm` | `xattn_bsa.py:select_blocks` | flat_group_gemm 循环 |
+| Offload | `xattn_estimate_softmax` | `xattn_bsa.py:select_blocks` | softmax_fuse_block_sum |
+| Offload | `xattn_estimate_find_blocks` | `xattn_bsa.py:select_blocks` | find_blocks_chunked |
+| Offload | `xattn_compute_historical` | `xattn_bsa.py:compute_chunked_prefill` | 历史 chunks attention |
+| Offload | `xattn_compute_current` | `xattn_bsa.py:compute_chunked_prefill` | 当前 chunk attention |
+| Offload | `xattn_compute_merge` | `xattn_bsa.py:compute_chunked_prefill` | merge 操作 |
+
+### 查看 NVTX 统计
+
+```bash
+# 生成 profile
+bash scripts/profile_offload.sh --policy xattn --ctx-len 64k --block-size 4096 --gpu 0
+
+# 查看 NVTX 统计
+nsys stats --report nvtx_pushpop_sum results/nsys/<filename>.nsys-rep
+```
+
+## Block Size 对 Offload 模式的影响
+
+### 测试配置
+
+- Model: Llama-3.1-8B-Instruct
+- Context: 64K tokens
+- Mode: xattn + offload
+- GPU: A100 40GB
+
+### 性能对比
+
+| 指标 | block_size=4096 | block_size=1024 | 变化 |
+|------|----------------|-----------------|------|
+| **总时间** | 27.7s | 55.5s | **2x 慢** |
+| **Chunks 数量** | 16 | 64 | 4x |
+| **CPU blocks** | 18 | 71 | ~4x |
+
+### 各阶段耗时分布
+
+#### block_size=4096
+
+| 阶段 | 占比 | 总时间 | 平均时间 | 调用次数 |
+|-----|------|--------|---------|---------|
+| **xattn_estimate_find_blocks** | **39.7%** | 18.0s | 37.6ms | 480 |
+| xattn_compute_historical | 4.4% | 2.0s | 4.2ms | 480 |
+| xattn_estimate_gemm | 3.4% | 1.5s | 3.2ms | 480 |
+| xattn_compute_current | 0.2% | 113ms | 0.22ms | 512 |
+| xattn_compute_merge | 0.2% | 96ms | 0.19ms | 512 |
+| xattn_estimate_softmax | 0.2% | 88ms | 0.18ms | 480 |
+
+#### block_size=1024
+
+| 阶段 | 占比 | 总时间 | 平均时间 | 调用次数 |
+|-----|------|--------|---------|---------|
+| **xattn_estimate_gemm** | **23.6%** | 22.6s | 11.4ms | 1984 |
+| **xattn_compute_historical** | **16.9%** | 16.2s | 8.0ms | 2016 |
+| xattn_estimate_find_blocks | 1.4% | 1.3s | 0.66ms | 1984 |
+| xattn_compute_current | 0.5% | 433ms | 0.21ms | 2048 |
+| xattn_compute_merge | 0.4% | 373ms | 0.18ms | 2048 |
+| xattn_estimate_softmax | 0.2% | 222ms | 0.11ms | 1984 |
+
+### 关键发现
+
+1. **Block size 对性能影响显著**
+   - block_size=1024 比 4096 慢约 2x
+   - 更小的 block size 导致更多的 chunks，增加调用次数
+
+2. **性能瓶颈随 block size 变化**
+   - **block_size=4096**: 瓶颈是 `find_blocks_chunked` (39.7%)
+   - **block_size=1024**: 瓶颈转移到 `estimate_gemm` (23.6%) 和 `compute_historical` (16.9%)
+
+3. **Amortization 效应**
+   - 大 block size 虽然单次 `find_blocks` 更慢 (37.6ms vs 0.66ms)
+   - 但调用次数少 (480 vs 1984)，总时间反而更少
+
+4. **find_blocks_chunked 的特殊性**
+   - 该函数主要在 CPU 上执行 block 选择逻辑
+   - 处理更大的数据量时开销显著增加
+   - block_size=4096 时占用 40% 时间，是主要优化目标
+
+## softmax_fuse_block_sum_kernel 性能分析
+
+`softmax_fuse_block_sum_kernel_non_causal` 是 XAttention 估计阶段的核心 Triton kernel。
+
+### Kernel 结构
+
+```python
+# 每个 thread block 处理的数据形状
+工作负载: [block_size, segment_size]  # 单个 Q block 对所有 K 的注意力
+
+# Pass 1: 计算全局 softmax 参数 (m_i, l_i)
+for iter in range(num_iters):  # num_iters = k_len / segment_size
+    X = load [block_size, segment_size]
+    compute max, sum for softmax normalization
+
+# Pass 2: Normalize + Block Sum
+for iter in range(num_iters):
+    X = load [block_size, segment_size]
+    X = softmax(X)
+    X = reshape(X, [block_size, segment_size/block_size, block_size])
+    X = sum(X, axis=2)  # → [block_size, segment_size/block_size]
+    X = sum(X, axis=0)  # → [segment_size/block_size]
+    store output
+```
+
+### 性能随 block_size 变化的因素
+
+| 因素 | 小 block_size (64) | 大 block_size (256) |
+|------|-------------------|---------------------|
+| Grid 并行度 | 高 (更多 blocks) | 低 (更少 blocks) |
+| 寄存器使用 | 低 | 高 (可能 spill) |
+| L2 Cache 复用 | 差 | 好 |
+| 输出大小 | 大 | 小 |
+
+### 典型性能曲线
+
+```
+Performance
+    │
+    │         ┌─────┐
+    │        /       \
+    │       /         \
+    │      /           \
+    │     /             \
+    └────/───────────────\────────→ block_size
+        64   128   256   512
+
+最优点通常在 128-256 之间
+```
+
+## 优化建议
+
+1. **优先使用 block_size=4096**
+   - 减少 chunk 数量，降低调度开销
+   - 更好的 amortization 效果
+
+2. **优化 find_blocks_chunked**
+   - 当前是 block_size=4096 的主要瓶颈
+   - 考虑 GPU 加速或批量处理
+
+3. **Pipeline 优化**
+   - 利用多 slot 的 ring buffer 实现计算和传输 overlap
+   - 当前已实现，但 find_blocks 是 CPU 操作，无法 overlap
+
+## 测试命令
+
+```bash
+# GPU-only 模式 (需要 40GB+ VRAM)
+bash scripts/profile_offload.sh --policy xattn --ctx-len 64k --no-offload --gpu 0
+
+# Offload 模式，block_size=4096
+bash scripts/profile_offload.sh --policy xattn --ctx-len 64k --block-size 4096 --gpu 0
+
+# Offload 模式，block_size=1024
+bash scripts/profile_offload.sh --policy xattn --ctx-len 64k --block-size 1024 --gpu 0
+
+# 128K context
+bash scripts/profile_offload.sh --policy xattn --ctx-len 128k --block-size 4096 --gpu 0
+```

nanovllm/config.py

@@ -22,7 +22,7 @@ class Config:
     tensor_parallel_size: int = 1
     enforce_eager: bool = False
     hf_config: AutoConfig | None = None
-    eos: int = -1
+    eos: int | list[int] = -1  # Single EOS token or list of EOS tokens (e.g., GLM-4)
     kvcache_block_size: int = 1024
     num_kvcache_blocks: int = -1
     dtype: str | None = None  # "float16", "bfloat16", or None (use model default)
@@ -57,8 +57,11 @@ class Config:
         assert os.path.isdir(self.model)
         assert self.kvcache_block_size % 256 == 0
         assert 1 <= self.tensor_parallel_size <= 8
-        self.hf_config = AutoConfig.from_pretrained(self.model)
-        self.max_model_len = min(self.max_model_len, self.hf_config.max_position_embeddings)
+        self.hf_config = AutoConfig.from_pretrained(self.model, trust_remote_code=True)
+        # Get max position embeddings (GLM-4 uses seq_length instead of max_position_embeddings)
+        max_pos = getattr(self.hf_config, 'max_position_embeddings',
+                         getattr(self.hf_config, 'seq_length', 4096))
+        self.max_model_len = min(self.max_model_len, max_pos)
         assert self.max_num_batched_tokens >= self.max_model_len
 
         # Override torch_dtype if user specified

nanovllm/engine/llm_engine.py

@@ -10,7 +10,8 @@ from nanovllm.sampling_params import SamplingParams
 from nanovllm.engine.sequence import Sequence
 from nanovllm.engine.scheduler import Scheduler
 from nanovllm.engine.model_runner import ModelRunner
-from nanovllm.utils.observer import Observer
+from nanovllm.utils.observer import InferenceObserver
+from nanovllm.utils.memory_observer import MemoryObserver
 
 
 class LLMEngine:
@@ -29,8 +30,14 @@ class LLMEngine:
             self.ps.append(process)
             self.events.append(event)
         self.model_runner = ModelRunner(config, 0, self.events)
-        self.tokenizer = AutoTokenizer.from_pretrained(config.model, use_fast=True)
-        config.eos = self.tokenizer.eos_token_id
+        self.tokenizer = AutoTokenizer.from_pretrained(config.model, use_fast=True, trust_remote_code=True)
+        # Get EOS token(s) from config (may be int or list, e.g., GLM-4 uses list)
+        # Prefer hf_config.eos_token_id which contains full list, fallback to tokenizer
+        eos_from_config = getattr(config.hf_config, 'eos_token_id', None)
+        if eos_from_config is not None:
+            config.eos = eos_from_config
+        else:
+            config.eos = self.tokenizer.eos_token_id
         # Set Sequence.block_size to match the KV cache block size
         Sequence.block_size = config.kvcache_block_size
         self.scheduler = Scheduler(config, self.model_runner.kvcache_manager)
@@ -58,15 +65,18 @@ class LLMEngine:
             print(f"[DEBUG LLMEngine.step] Mode={mode}, active_sequences={len(seqs)}")
 
         if not is_prefill:
-            # The end of the prefill mode. Get TTFT.
-            if Observer.ttft_start != 0:
-                Observer.ttft = perf_counter_ns() - Observer.ttft_start
-                Observer.reset_ttft()
-            # The start of the decode mode. Get TPOT.
-            if Observer.tpot_start != 0:
-                Observer.tpot = perf_counter_ns() - Observer.tpot_start
-            Observer.tpot_start = perf_counter_ns()
+            # Decode mode: calculate TPOT from previous decode step
+            if InferenceObserver.tpot_start != 0:
+                InferenceObserver.tpot = perf_counter_ns() - InferenceObserver.tpot_start
+            InferenceObserver.tpot_start = perf_counter_ns()
+
         token_ids = self.model_runner.call("run", seqs, is_prefill)
+
+        if is_prefill:
+            # Calculate TTFT after prefill completes (including chunked prefill)
+            if InferenceObserver.ttft_start != 0:
+                InferenceObserver.ttft = perf_counter_ns() - InferenceObserver.ttft_start
+                InferenceObserver.reset_ttft()
         self.scheduler.postprocess(seqs, token_ids)
         outputs = [(seq.seq_id, seq.completion_token_ids) for seq in seqs if seq.is_finished]
 
@@ -91,7 +101,8 @@ class LLMEngine:
         log_level = os.environ.get('NANOVLLM_LOG_LEVEL', 'INFO')
         debug_enabled = log_level.upper() == 'DEBUG'
 
-        Observer.complete_reset()
+        InferenceObserver.complete_reset()
+        MemoryObserver.complete_reset()
         if use_tqdm:
             pbar = tqdm(total=len(prompts), desc="Generating", dynamic_ncols=True)
         if not isinstance(sampling_params, list):
@@ -128,8 +139,8 @@ class LLMEngine:
                 pbar.set_postfix({
                     "Prefill": f"{int(prefill_throughput)}tok/s",
                     "Decode": f"{int(decode_throughput)}tok/s",
-                    "ttft": f"{float(Observer.ttft) / 1e6}ms",
-                    "tpot": f"{float(Observer.tpot) / 1e6}ms",
+                    "ttft": f"{float(InferenceObserver.ttft) / 1e6}ms",
+                    "tpot": f"{float(InferenceObserver.tpot) / 1e6}ms",
                 })
             for seq_id, token_ids in output:
                 outputs[seq_id] = token_ids

nanovllm/engine/model_runner.py

@@ -30,6 +30,18 @@ def _find_free_port() -> int:
         return s.getsockname()[1]
 
 
+def get_num_kv_heads(hf_config) -> int:
+    """Get number of KV heads from config (handles GLM-4's multi_query_group_num)."""
+    return getattr(hf_config, 'num_key_value_heads',
+                   getattr(hf_config, 'multi_query_group_num', hf_config.num_attention_heads))
+
+
+def get_head_dim(hf_config) -> int:
+    """Get head dimension from config (handles GLM-4's kv_channels)."""
+    return getattr(hf_config, "head_dim",
+                   getattr(hf_config, "kv_channels", hf_config.hidden_size // hf_config.num_attention_heads))
+
+
 class ModelRunner:
 
     def __init__(self, config: Config, rank: int, event: Event | list[Event]):
@@ -144,8 +156,8 @@ class ModelRunner:
         used = total - free
         peak = torch.cuda.memory_stats()["allocated_bytes.all.peak"]
         current = torch.cuda.memory_stats()["allocated_bytes.all.current"]
-        num_kv_heads = hf_config.num_key_value_heads // self.world_size
-        head_dim = getattr(hf_config, "head_dim", hf_config.hidden_size // hf_config.num_attention_heads)
+        num_kv_heads = get_num_kv_heads(hf_config) // self.world_size
+        head_dim = get_head_dim(hf_config)
         block_bytes = 2 * hf_config.num_hidden_layers * self.block_size * num_kv_heads * head_dim * hf_config.torch_dtype.itemsize
 
         # Calculate max GPU blocks based on available memory
@@ -202,19 +214,37 @@ class ModelRunner:
             dtype=hf_config.torch_dtype,
         )
 
-        # Initialize sparse policy if manager has one (CPU offload mode)
+        # Initialize sparse policy if manager has one (works for both CPU offload and GPU-only modes)
         if hasattr(self.kvcache_manager, 'sparse_policy') and self.kvcache_manager.sparse_policy is not None:
+            # Use CPU blocks for offload mode, GPU blocks for GPU-only mode
+            num_blocks_for_init = config.num_cpu_kvcache_blocks if config.enable_cpu_offload else config.num_kvcache_blocks
             self.kvcache_manager.sparse_policy.initialize(
                 num_layers=hf_config.num_hidden_layers,
                 num_kv_heads=num_kv_heads,
                 head_dim=head_dim,
-                num_cpu_blocks=config.num_cpu_kvcache_blocks,
+                num_cpu_blocks=num_blocks_for_init,
                 dtype=hf_config.torch_dtype,
                 device=torch.device("cuda"),
             )
 
+            # Pre-allocate policy metadata buffers
+            # - Offload mode: allocate chunked prefill buffers (mask, KV chunking stats)
+            # - GPU-only mode: additionally allocate GQA expansion buffers
+            num_heads = hf_config.num_attention_heads // self.world_size
+            self.kvcache_manager.sparse_policy.alloc_policy_metadata(
+                num_heads=num_heads,
+                num_kv_heads=num_kv_heads,
+                head_dim=head_dim,
+                max_seq_len=config.max_model_len,
+                dtype=hf_config.torch_dtype,
+                device=torch.device("cuda"),
+                enable_cpu_offload=config.enable_cpu_offload,
+            )
+
+            # Log policy info (handle both enum and None cases)
+            policy_name = config.sparse_policy.name if config.sparse_policy is not None else "FULL"
             logger.info(
-                f"Sparse policy initialized: {config.sparse_policy.name} "
+                f"Sparse policy initialized: {policy_name} "
                 f"(topk={config.sparse_topk_blocks}, threshold={config.sparse_threshold_blocks})"
             )
 
@@ -375,7 +405,16 @@ class ModelRunner:
         cu_seqlens_q = torch.tensor(cu_seqlens_q, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
         cu_seqlens_k = torch.tensor(cu_seqlens_k, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
         slot_mapping = torch.tensor(slot_mapping, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
-        set_context(True, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, slot_mapping, None, block_tables)
+        set_context(
+            is_prefill=True,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_q,
+            max_seqlen_k=max_seqlen_k,
+            slot_mapping=slot_mapping,
+            block_tables=block_tables,
+            kvcache_manager=getattr(self, 'kvcache_manager', None),
+        )
         return input_ids, positions
 
     def prepare_decode(self, seqs: list[Sequence]):
@@ -404,7 +443,13 @@ class ModelRunner:
         context_lens = torch.tensor(context_lens, dtype=torch.int32, pin_memory=True).cuda(non_blocking=True)
         # Use GPU physical block tables for attention
         block_tables = self._prepare_gpu_block_tables(gpu_block_tables)
-        set_context(False, slot_mapping=slot_mapping, context_lens=context_lens, block_tables=block_tables)
+        set_context(
+            is_prefill=False,
+            slot_mapping=slot_mapping,
+            context_lens=context_lens,
+            block_tables=block_tables,
+            kvcache_manager=self.kvcache_manager,
+        )
         return input_ids, positions
 
     def _prepare_gpu_block_tables(self, gpu_block_tables: list[list[int]]):
@@ -713,7 +758,13 @@ class ModelRunner:
 
         for bs in reversed(self.graph_bs):
             graph = torch.cuda.CUDAGraph()
-            set_context(False, slot_mapping=slot_mapping[:bs], context_lens=context_lens[:bs], block_tables=block_tables[:bs])
+            set_context(
+                is_prefill=False,
+                slot_mapping=slot_mapping[:bs],
+                context_lens=context_lens[:bs],
+                block_tables=block_tables[:bs],
+                kvcache_manager=self.kvcache_manager,
+            )
             outputs[:bs] = self.model(input_ids[:bs], positions[:bs])    # warmup
             with torch.cuda.graph(graph, self.graph_pool):
                 outputs[:bs] = self.model(input_ids[:bs], positions[:bs])    # capture
@@ -749,8 +800,8 @@ class ModelRunner:
         - LastGraph: o_proj → post_norm → mlp → final_norm
         """
         hf_config = self.config.hf_config
-        num_kv_heads = hf_config.num_key_value_heads // self.world_size
-        head_dim = getattr(hf_config, "head_dim", hf_config.hidden_size // hf_config.num_attention_heads)
+        num_kv_heads = get_num_kv_heads(hf_config) // self.world_size
+        head_dim = get_head_dim(hf_config)
 
         # Create Decode Graph Manager (seq_len=1)
         self.decode_graph_manager = OffloadGraphManager(

nanovllm/engine/scheduler.py

@@ -4,7 +4,7 @@ from typing import TYPE_CHECKING
 
 from nanovllm.config import Config
 from nanovllm.engine.sequence import Sequence, SequenceStatus
-from nanovllm.utils.observer import Observer
+from nanovllm.utils.observer import InferenceObserver
 
 if TYPE_CHECKING:
     from nanovllm.kvcache import KVCacheManager
@@ -15,7 +15,9 @@ class Scheduler:
     def __init__(self, config: Config, kvcache_manager: "KVCacheManager"):
         self.max_num_seqs = config.max_num_seqs
         self.max_num_batched_tokens = config.max_num_batched_tokens
-        self.eos = config.eos
+        # Convert EOS to set for efficient lookup (supports single int or list)
+        eos = config.eos
+        self.eos_set = set(eos) if isinstance(eos, list) else {eos}
         self.kvcache_manager = kvcache_manager
         self.waiting: deque[Sequence] = deque()
         self.running: deque[Sequence] = deque()
@@ -32,8 +34,8 @@ class Scheduler:
         num_seqs = 0
         num_batched_tokens = 0
         while self.waiting and num_seqs < self.max_num_seqs:
-            if Observer.ttft_start == 0:
-                Observer.ttft_start = perf_counter_ns()
+            if InferenceObserver.ttft_start == 0:
+                InferenceObserver.ttft_start = perf_counter_ns()
             seq = self.waiting[0]
 
             # Check if sequence is too large
@@ -94,7 +96,7 @@ class Scheduler:
     def postprocess(self, seqs: list[Sequence], token_ids: list[int]) -> list[bool]:
         for seq, token_id in zip(seqs, token_ids):
             seq.append_token(token_id)
-            if (not seq.ignore_eos and token_id == self.eos) or seq.num_completion_tokens == seq.max_tokens:
+            if (not seq.ignore_eos and token_id in self.eos_set) or seq.num_completion_tokens == seq.max_tokens:
                 seq.status = SequenceStatus.FINISHED
                 self.kvcache_manager.deallocate(seq)
                 self.running.remove(seq)

nanovllm/kvcache/__init__.py

@@ -25,7 +25,7 @@ def create_kvcache_manager(config: "Config") -> KVCacheManager:
     Factory function to create the appropriate KV cache manager.
 
     Decision logic:
-    1. If enable_cpu_offload=False: use GPUOnlyManager
+    1. If enable_cpu_offload=False: use GPUOnlyManager (optionally with sparse policy)
     2. If enable_cpu_offload=True but all blocks fit in GPU: use GPUOnlyManager
     3. If enable_cpu_offload=True and need CPU blocks: use HybridKVCacheManager
 
@@ -37,9 +37,44 @@ def create_kvcache_manager(config: "Config") -> KVCacheManager:
     """
     if not getattr(config, 'enable_cpu_offload', False):
         # Default: pure GPU mode
+        # Check if sparse policy is requested for GPU-only mode
+        from nanovllm.config import SparsePolicyType
+        sparse_policy_type = getattr(config, 'sparse_policy', None)
+        # Handle None case - use FULL as default
+        if sparse_policy_type is None:
+            sparse_policy_type = SparsePolicyType.FULL
+
+        sparse_policy = None
+        if sparse_policy_type != SparsePolicyType.FULL:
+            # Create sparse policy for GPU-only mode
+            from nanovllm.kvcache.sparse import create_sparse_policy
+
+            policy_kwargs = {}
+            if sparse_policy_type == SparsePolicyType.QUEST:
+                policy_kwargs = {
+                    'topk_blocks': getattr(config, 'sparse_topk_blocks', 8),
+                    'threshold_blocks': getattr(config, 'sparse_threshold_blocks', 4),
+                }
+            elif sparse_policy_type == SparsePolicyType.XATTN_BSA:
+                policy_kwargs = {
+                    'block_size': getattr(config, 'sparse_block_size', 128),
+                    'samples_per_chunk': getattr(config, 'sparse_samples_per_chunk', 128),
+                    'threshold': getattr(config, 'sparse_threshold', 0.9),
+                    'use_triton': getattr(config, 'sparse_use_triton', True),
+                    'stride': getattr(config, 'sparse_stride', 8),
+                    'chunk_size': getattr(config, 'sparse_chunk_size', 16384),
+                }
+
+            sparse_policy = create_sparse_policy(sparse_policy_type, **policy_kwargs)
+        else:
+            # FULL policy for GPU-only mode - always create for consistent API
+            from nanovllm.kvcache.sparse import FullAttentionPolicy
+            sparse_policy = FullAttentionPolicy()
+
         return GPUOnlyManager(
             num_blocks=config.num_kvcache_blocks,
             block_size=config.kvcache_block_size,
+            sparse_policy=sparse_policy,
         )
 
     # CPU offload is enabled

nanovllm/kvcache/gpu_manager.py

@@ -7,13 +7,16 @@ the KVCacheManager interface.
 """
 
 from collections import deque
-from typing import List, Tuple, Dict, Optional
+from typing import List, Tuple, Dict, Optional, TYPE_CHECKING
 import torch
 from torch import Tensor
 
 from nanovllm.engine.sequence import Sequence
 from nanovllm.kvcache.base_manager import KVCacheManager
 
+if TYPE_CHECKING:
+    from nanovllm.kvcache.sparse.policy import SparsePolicy
+
 
 class Block:
     """Physical block in GPU memory."""
@@ -50,17 +53,28 @@ class GPUOnlyManager(KVCacheManager):
     all data stays on GPU at fixed addresses.
     """
 
-    def __init__(self, num_blocks: int, block_size: int):
+    def __init__(
+        self,
+        num_blocks: int,
+        block_size: int,
+        sparse_policy: Optional["SparsePolicy"] = None,
+    ):
         """
         Initialize GPU-only manager.
 
         Args:
             num_blocks: Total number of blocks to manage
             block_size: Tokens per block (default 256)
+            sparse_policy: Optional sparse attention policy for GPU-only mode
         """
         self._block_size = block_size
         self._num_blocks = num_blocks
 
+        # Sparse policy for GPU-only mode (optional)
+        self.sparse_policy = sparse_policy
+        # No offload engine in GPU-only mode
+        self.offload_engine = None
+
         # Block metadata
         self.blocks: List[Block] = [Block(i) for i in range(num_blocks)]
 

nanovllm/kvcache/offload_engine.py

@@ -17,6 +17,7 @@ from dataclasses import dataclass
 from nanovllm.kvcache.kernels import gathered_copy_kv
 from nanovllm.comm import memcpy_2d_async
 from nanovllm.utils.logger import get_logger
+from nanovllm.utils.memory_observer import MemoryObserver
 
 # Import for type hints only (avoid circular import)
 from typing import TYPE_CHECKING
@@ -376,7 +377,8 @@ class OffloadEngine:
         self.ring_slot_compute_done[slot_idx].record()
 
     def load_to_slot_layer(
-        self, slot_idx: int, layer_id: int, cpu_block_id: int, chunk_idx: int = -1
+        self, slot_idx: int, layer_id: int, cpu_block_id: int, chunk_idx: int = -1,
+        is_prefill: bool = True,
     ) -> None:
         """
         Async load a single CPU block to a ring buffer slot for one layer.
@@ -393,6 +395,7 @@ class OffloadEngine:
             layer_id: Layer index to load (for CPU cache indexing)
             cpu_block_id: Source CPU block ID
             chunk_idx: Optional chunk index for NVTX labeling (-1 means not specified)
+            is_prefill: True if in prefill phase, False if in decode phase (for MemoryObserver)
         """
         logger.debug(f"Ring load: layer={layer_id}, CPU[{cpu_block_id}] -> GPU slot[{slot_idx}]")
 
@@ -425,6 +428,65 @@ class OffloadEngine:
             self.ring_slot_ready[slot_idx].record(stream)
         nvtx.pop_range()
 
+        # Record H2D transfer: K + V = 2 * block_bytes
+        MemoryObserver.record_h2d(2 * self.gpu_block_bytes, is_prefill=is_prefill)
+
+    def load_k_only_to_slot_layer(
+        self, slot_idx: int, layer_id: int, cpu_block_id: int, chunk_idx: int = -1,
+        is_prefill: bool = True,
+    ) -> None:
+        """
+        Async load only K (not V) from CPU block to GPU slot.
+
+        Used by XAttention estimate phase which only needs K for attention score
+        computation. Saves 50% communication compared to loading K+V.
+
+        Args:
+            slot_idx: Target GPU slot index
+            layer_id: Layer index to load (for CPU cache indexing)
+            cpu_block_id: Source CPU block ID
+            chunk_idx: Optional chunk index for NVTX labeling (-1 means not specified)
+            is_prefill: True if in prefill phase, False if in decode phase
+        """
+        logger.debug(f"Ring load K-only: layer={layer_id}, CPU[{cpu_block_id}] -> GPU slot[{slot_idx}]")
+
+        stream = self.slot_transfer_streams[slot_idx]
+
+        if chunk_idx >= 0:
+            nvtx_label = f"H2D K-only: L{layer_id} Chunk{chunk_idx} CPU[{cpu_block_id}]->Slot[{slot_idx}]"
+        else:
+            nvtx_label = f"H2D K-only: L{layer_id} CPU[{cpu_block_id}]->Slot[{slot_idx}]"
+
+        nvtx.push_range(message=nvtx_label, color="cyan")
+        with torch.cuda.stream(stream):
+            stream.wait_event(self.ring_slot_compute_done[slot_idx])
+            stream.wait_event(self.ring_slot_offload_done[slot_idx])
+
+            # Only copy K, not V
+            self.k_cache_gpu[slot_idx].copy_(
+                self.k_cache_cpu[layer_id, cpu_block_id], non_blocking=True
+            )
+            self.ring_slot_ready[slot_idx].record(stream)
+        nvtx.pop_range()
+
+        # Record H2D transfer: K only = 1 * block_bytes
+        MemoryObserver.record_h2d(self.gpu_block_bytes, is_prefill=is_prefill)
+
+    def get_k_for_slot(self, slot_idx: int) -> Tensor:
+        """
+        Get only K for a ring buffer slot (no V).
+
+        Used by XAttention estimate phase which only needs K for attention
+        score computation.
+
+        Args:
+            slot_idx: GPU slot index
+
+        Returns:
+            k_cache, shape: [1, block_size, kv_heads, head_dim]
+        """
+        return self.k_cache_gpu[slot_idx].unsqueeze(0)
+
     def wait_slot_layer(self, slot_idx: int) -> None:
         """
         Wait for a slot's loading to complete.
@@ -499,6 +561,9 @@ class OffloadEngine:
             self.ring_slot_offload_done[slot_idx].record(self.transfer_stream_main)
         nvtx.pop_range()
 
+        # Record D2H transfer: K + V = 2 * block_bytes
+        MemoryObserver.record_d2h(2 * self.gpu_block_bytes, is_prefill=is_prefill)
+
     # ----- KV access methods for ring buffer -----
 
     def get_kv_for_slot(self, slot_idx: int) -> Tuple[Tensor, Tensor]:
@@ -745,6 +810,10 @@ class OffloadEngine:
         self.prefill_v_buffer[layer_id, :num_tokens].copy_(v)
         torch.cuda.nvtx.range_pop()
 
+        # Record D2D transfer: K + V
+        transfer_bytes = 2 * k.numel() * k.element_size()
+        MemoryObserver.record_d2d(transfer_bytes)
+
     def write_to_decode_buffer(
         self,
         layer_id: int,
@@ -768,6 +837,10 @@ class OffloadEngine:
         self.decode_v_buffer[layer_id, pos_in_block].copy_(v)
         torch.cuda.nvtx.range_pop()
 
+        # Record D2D transfer: K + V (single token)
+        transfer_bytes = 2 * k.numel() * k.element_size()
+        MemoryObserver.record_d2d(transfer_bytes)
+
     def offload_prefill_buffer_async(
         self,
         layer_id: int,
@@ -813,6 +886,9 @@ class OffloadEngine:
             self.prefill_offload_events[layer_id].record(stream)
         nvtx.pop_range()
 
+        # Record D2H transfer: K + V = 2 * block_bytes
+        MemoryObserver.record_d2h(2 * self.gpu_block_bytes, is_prefill=True)
+
     def wait_all_prefill_offloads(self) -> None:
         """Wait for all prefill buffer offloads to complete."""
         for stream in self.prefill_offload_streams:
@@ -851,6 +927,11 @@ class OffloadEngine:
         v_sample = self.v_cache_cpu[
             layer_id, cpu_block_id, :num_samples
         ].clone().cuda()
+
+        # Record H2D transfer: K + V samples
+        transfer_bytes = 2 * k_sample.numel() * k_sample.element_size()
+        MemoryObserver.record_h2d(transfer_bytes, is_prefill=True)
+
         return k_sample, v_sample
 
     def load_block_full_from_cpu(
@@ -877,4 +958,8 @@ class OffloadEngine:
         v_full = self.v_cache_cpu[
             layer_id, cpu_block_id
         ].clone().cuda()
+
+        # Record H2D transfer: K + V full block
+        MemoryObserver.record_h2d(2 * self.gpu_block_bytes, is_prefill=True)
+
         return k_full, v_full

nanovllm/kvcache/sparse/full_policy.py

@@ -47,6 +47,8 @@ class FullAttentionPolicy(SparsePolicy):
         available_blocks: List[int],
         offload_engine: "OffloadEngine",
         ctx: PolicyContext,
+        q: torch.Tensor,
+        k: torch.Tensor,
     ) -> List[int]:
         """Return all blocks - no sparsity."""
         # Update statistics (only for layer 0 to avoid overcounting)
@@ -76,6 +78,75 @@ class FullAttentionPolicy(SparsePolicy):
         logger.info(f"[Full Policy] Density Stats: chunks={stats['num_chunks']}, "
                    f"blocks={stats['total_available_blocks']}, density=100.0%")
 
+    # =========================================================================
+    # GPU-only methods (non-chunked)
+    # =========================================================================
+
+    def compute_prefill(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        cu_seqlens_q: torch.Tensor,
+        cu_seqlens_k: torch.Tensor,
+        max_seqlen_q: int,
+        max_seqlen_k: int,
+        softmax_scale: float,
+        layer_id: int,
+        block_tables: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        GPU-only prefill attention using flash_attn_varlen_func.
+
+        This is the simplest implementation - just call flash attention directly.
+        For sparse policies, this method would implement block selection.
+        """
+        from flash_attn import flash_attn_varlen_func
+
+        return flash_attn_varlen_func(
+            q, k, v,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_q,
+            max_seqlen_k=max_seqlen_k,
+            softmax_scale=softmax_scale,
+            causal=True,
+            block_table=block_tables,
+        )
+
+    def compute_decode(
+        self,
+        q: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        cache_seqlens: torch.Tensor,
+        softmax_scale: float,
+        layer_id: int,
+        block_tables: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        GPU-only decode attention using flash_attn_with_kvcache.
+
+        This is the simplest implementation - just call flash attention directly.
+        For sparse policies, this method would implement block selection.
+        """
+        from flash_attn import flash_attn_with_kvcache
+
+        # q is [batch, num_heads, head_dim], need to add seq dim
+        return flash_attn_with_kvcache(
+            q.unsqueeze(1),  # [batch, 1, heads, dim]
+            k_cache,
+            v_cache,
+            cache_seqlens=cache_seqlens,
+            block_table=block_tables,
+            softmax_scale=softmax_scale,
+            causal=True,
+        )
+
+    # =========================================================================
+    # Chunked offload methods
+    # =========================================================================
+
     def compute_chunked_prefill(
         self,
         q: torch.Tensor,
@@ -116,7 +187,11 @@ class FullAttentionPolicy(SparsePolicy):
         Returns:
             Attention output [seq_len, num_heads, head_dim]
         """
-        from nanovllm.ops.chunked_attention import flash_attn_with_lse, merge_attention_outputs
+        # Use FlashInfer-based implementations (more optimized)
+        from nanovllm.ops.chunked_attention import (
+            flash_attn_with_lse_flashinfer as flash_attn_with_lse,
+            merge_attention_outputs_flashinfer as merge_attention_outputs,
+        )
 
         logger.debug(f"[DEBUG] FullPolicy.compute_chunked_prefill called, "
                      f"layer={layer_id}, chunk={current_chunk_idx}, num_tokens={num_tokens}, "
@@ -244,7 +319,11 @@ class FullAttentionPolicy(SparsePolicy):
         Returns:
             Attention output [batch_size, 1, num_heads, head_dim]
         """
-        from nanovllm.ops.chunked_attention import flash_attn_with_lse, merge_attention_outputs
+        # Use FlashInfer-based implementations (more optimized)
+        from nanovllm.ops.chunked_attention import (
+            flash_attn_with_lse_flashinfer as flash_attn_with_lse,
+            merge_attention_outputs_flashinfer as merge_attention_outputs,
+        )
 
         # q shape: [batch_size, num_heads, head_dim] (single decode token per sequence)
         q_batched = q.unsqueeze(1)  # [batch, 1, heads, dim]
@@ -336,7 +415,11 @@ class FullAttentionPolicy(SparsePolicy):
         Loads one block at a time, computes attention, and merges results.
         Uses load_to_slot_layer / wait_slot_layer / get_kv_for_slot methods.
         """
-        from nanovllm.ops.chunked_attention import flash_attn_with_lse, merge_attention_outputs
+        # Use FlashInfer-based implementations (more optimized)
+        from nanovllm.ops.chunked_attention import (
+            flash_attn_with_lse_flashinfer as flash_attn_with_lse,
+            merge_attention_outputs_flashinfer as merge_attention_outputs,
+        )
 
         num_blocks = len(cpu_block_table)
         if num_blocks == 0:
@@ -353,7 +436,7 @@ class FullAttentionPolicy(SparsePolicy):
         num_preload = min(num_slots, num_blocks)
         for i in range(num_preload):
             cpu_block_id = cpu_block_table[i]
-            offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_id, chunk_idx=cpu_block_id)
+            offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_id, chunk_idx=cpu_block_id, is_prefill=False)
 
         # Phase 2: Process blocks with pipeline
         for block_idx in range(num_blocks):
@@ -387,7 +470,7 @@ class FullAttentionPolicy(SparsePolicy):
             next_block_idx = block_idx + num_slots
             if next_block_idx < num_blocks:
                 next_cpu_block_id = cpu_block_table[next_block_idx]
-                offload_engine.load_to_slot_layer(current_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id)
+                offload_engine.load_to_slot_layer(current_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id, is_prefill=False)
 
             # Merge with accumulated
             with torch.cuda.stream(compute_stream):

nanovllm/kvcache/sparse/policy.py

@@ -108,12 +108,45 @@ class SparsePolicy(ABC):
         """
         pass
 
+    def alloc_policy_metadata(
+        self,
+        num_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        max_seq_len: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        enable_cpu_offload: bool = False,
+    ) -> None:
+        """
+        Pre-allocate GPU buffers for policy computation.
+
+        Called by the framework after KV cache allocation. Implementations should
+        use enable_cpu_offload to decide which buffers to allocate:
+        - Offload mode: allocate chunked prefill buffers (mask, KV chunking stats)
+        - GPU-only mode: additionally allocate GQA expansion buffers
+
+        This is separate from initialize() which is used for CPU offload metadata.
+
+        Args:
+            num_heads: Number of query heads
+            num_kv_heads: Number of KV heads (for GQA)
+            head_dim: Dimension per head
+            max_seq_len: Maximum sequence length (for buffer sizing)
+            dtype: Data type (typically float16/bfloat16)
+            device: Target device (cuda)
+            enable_cpu_offload: Whether CPU offload is enabled
+        """
+        pass
+
     @abstractmethod
     def select_blocks(
         self,
         available_blocks: List[int],
         offload_engine: "OffloadEngine",
         ctx: PolicyContext,
+        q: torch.Tensor,
+        k: torch.Tensor,
     ) -> List[int]:
         """
         Select which KV blocks to load for the current query chunk.
@@ -130,6 +163,8 @@ class SparsePolicy(ABC):
                            to load KV to make selection decisions).
             ctx: PolicyContext with information about the current query
                  chunk, layer, phase (prefill/decode), etc.
+            q: Query tensor [seq_len, num_heads, head_dim] for current chunk
+            k: Key tensor [seq_len, num_kv_heads, head_dim] for current chunk
 
         Returns:
             List of block IDs to load (must be a subset of available_blocks).
@@ -191,6 +226,87 @@ class SparsePolicy(ABC):
         """
         pass
 
+    # =========================================================================
+    # GPU-only methods (non-chunked)
+    # These methods are used when all KV cache is on GPU, no CPU offload needed.
+    # =========================================================================
+
+    def compute_prefill(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        cu_seqlens_q: torch.Tensor,
+        cu_seqlens_k: torch.Tensor,
+        max_seqlen_q: int,
+        max_seqlen_k: int,
+        softmax_scale: float,
+        layer_id: int,
+        block_tables: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Compute GPU-only prefill attention (non-chunked).
+
+        This method is used when all KV cache resides on GPU (no CPU offload).
+        Override this to implement sparse prefill attention for GPU-only mode.
+        Default implementation raises NotImplementedError.
+
+        Args:
+            q: [total_q, num_heads, head_dim] query tensor (packed variable length)
+            k: [total_kv, num_kv_heads, head_dim] key tensor
+            v: [total_kv, num_kv_heads, head_dim] value tensor
+            cu_seqlens_q: [batch+1] cumulative sequence lengths for queries
+            cu_seqlens_k: [batch+1] cumulative sequence lengths for keys
+            max_seqlen_q: maximum query sequence length
+            max_seqlen_k: maximum key sequence length
+            softmax_scale: softmax scaling factor (1/sqrt(head_dim))
+            layer_id: transformer layer index
+            block_tables: [batch, max_blocks] paged attention block tables (optional)
+
+        Returns:
+            [total_q, num_heads, head_dim] attention output
+        """
+        raise NotImplementedError(
+            f"{self.__class__.__name__} does not implement compute_prefill for GPU-only mode"
+        )
+
+    def compute_decode(
+        self,
+        q: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        cache_seqlens: torch.Tensor,
+        softmax_scale: float,
+        layer_id: int,
+        block_tables: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Compute GPU-only decode attention (non-chunked).
+
+        This method is used when all KV cache resides on GPU (no CPU offload).
+        Override this to implement sparse decode attention for GPU-only mode.
+        Default implementation raises NotImplementedError.
+
+        Args:
+            q: [batch, num_heads, head_dim] query tensor (single token per sequence)
+            k_cache: [num_blocks, block_size, num_kv_heads, head_dim] paged key cache
+            v_cache: [num_blocks, block_size, num_kv_heads, head_dim] paged value cache
+            cache_seqlens: [batch] sequence lengths in cache
+            softmax_scale: softmax scaling factor (1/sqrt(head_dim))
+            layer_id: transformer layer index
+            block_tables: [batch, max_blocks] paged attention block tables (optional)
+
+        Returns:
+            [batch, 1, num_heads, head_dim] attention output
+        """
+        raise NotImplementedError(
+            f"{self.__class__.__name__} does not implement compute_decode for GPU-only mode"
+        )
+
+    # =========================================================================
+    # Chunked offload methods (for CPU offload mode)
+    # =========================================================================
+
     @abstractmethod
     def compute_chunked_prefill(
         self,

nanovllm/kvcache/sparse/quest.py

@@ -191,13 +191,26 @@ class QuestPolicy(SparsePolicy):
     def select_blocks(
         self,
         available_blocks: List[int],
+        offload_engine: "OffloadEngine",
         ctx: PolicyContext,
+        q: torch.Tensor,
+        k: torch.Tensor,
     ) -> List[int]:
         """
         Select Top-K blocks based on query-key similarity bounds.
 
         If query is not available (some prefill scenarios), falls back
         to loading all blocks.
+
+        Args:
+            available_blocks: List of CPU block IDs
+            offload_engine: OffloadEngine for loading KV (unused in Quest)
+            ctx: PolicyContext with metadata
+            q: Query tensor [seq_len, num_heads, head_dim] or [batch, seq_len, num_heads, head_dim]
+            k: Key tensor [seq_len, num_kv_heads, head_dim] (unused in Quest, uses metadata instead)
+
+        Returns:
+            Selected block IDs
         """
         if self.metadata is None:
             raise RuntimeError(
@@ -211,7 +224,7 @@ class QuestPolicy(SparsePolicy):
         if n <= self.config.threshold_blocks:
             return available_blocks
 
-        if ctx.query is None:
+        if q is None:
             # No query available - cannot compute scores
             return available_blocks
 
@@ -221,11 +234,10 @@ class QuestPolicy(SparsePolicy):
         )
 
         # Metadata is already on GPU, same device as query
-        device = ctx.query.device
+        device = q.device
 
         # Compute upper bound scores
-        # query shape: [1, num_heads, head_dim] or [1, seq_len, num_heads, head_dim]
-        q = ctx.query
+        # query shape: [seq_len, num_heads, head_dim] or [batch, seq_len, num_heads, head_dim]
         if q.dim() == 4:
             # Prefill: use mean over sequence length
             q = q.mean(dim=1)  # [1, num_heads, head_dim]

nanovllm/layers/attention.py

@@ -124,24 +124,47 @@ class Attention(nn.Module):
             if k_cache.numel() and v_cache.numel():
                 store_kvcache(k, v, k_cache, v_cache, context.slot_mapping)
 
+        # Get sparse_policy from kvcache_manager (required, never None after warmup)
+        # During warmup, kvcache_manager is not yet allocated
+        if context.kvcache_manager is None:
+            # Warmup phase: use flash_attn directly
+            if context.is_prefill:
+                return flash_attn_varlen_func(
+                    q, k, v,
+                    max_seqlen_q=context.max_seqlen_q, cu_seqlens_q=context.cu_seqlens_q,
+                    max_seqlen_k=context.max_seqlen_k, cu_seqlens_k=context.cu_seqlens_k,
+                    softmax_scale=self.scale, causal=True,
+                )
+            else:
+                return flash_attn_with_kvcache(
+                    q.unsqueeze(1), k_cache, v_cache,
+                    cache_seqlens=context.context_lens, block_table=context.block_tables,
+                    softmax_scale=self.scale, causal=True,
+                )
+        sparse_policy = context.kvcache_manager.sparse_policy
+        assert sparse_policy is not None, "sparse_policy must not be None"
+
         if context.is_prefill:
             if context.is_chunked_prefill:
-                # Chunked prefill: merge attention from previous KV
+                # Chunked prefill: merge attention from previous KV (CPU offload mode)
                 o = self._chunked_prefill_attention(q, k, v, context)
-            elif context.block_tables is not None:    # prefix cache
-                k, v = k_cache, v_cache
-                o = flash_attn_varlen_func(q, k, v,
-                                           max_seqlen_q=context.max_seqlen_q, cu_seqlens_q=context.cu_seqlens_q,
-                                           max_seqlen_k=context.max_seqlen_k, cu_seqlens_k=context.cu_seqlens_k,
-                                           softmax_scale=self.scale, causal=True, block_table=context.block_tables)
             else:
-                o = flash_attn_varlen_func(q, k, v,
-                                           max_seqlen_q=context.max_seqlen_q, cu_seqlens_q=context.cu_seqlens_q,
-                                           max_seqlen_k=context.max_seqlen_k, cu_seqlens_k=context.cu_seqlens_k,
-                                           softmax_scale=self.scale, causal=True, block_table=context.block_tables)
+                # GPU-only mode: use policy for attention
+                # Use paged attention if block_tables provided, else use k, v directly
+                if context.block_tables is not None:
+                    k_for_attn, v_for_attn = k_cache, v_cache
+                else:
+                    k_for_attn, v_for_attn = k, v
+                o = sparse_policy.compute_prefill(
+                    q, k_for_attn, v_for_attn,
+                    context.cu_seqlens_q, context.cu_seqlens_k,
+                    context.max_seqlen_q, context.max_seqlen_k,
+                    self.scale, self.layer_id,
+                    context.block_tables,
+                )
         else:    # decode
             if context.is_chunked_prefill:
-                # Chunked decode: need to load all KV from CPU+GPU
+                # Chunked decode: need to load all KV from CPU+GPU (CPU offload mode)
                 # Store current decode token to per-layer decode buffer
                 # This is needed because GPU cache has no layer dimension,
                 # so all layers would overwrite each other in decode_slot.
@@ -152,9 +175,12 @@ class Attention(nn.Module):
                 offload_engine.write_to_decode_buffer(self.layer_id, pos_in_block, k.squeeze(0), v.squeeze(0))
                 o = self._chunked_decode_attention(q, k, v, context)
             else:
-                o = flash_attn_with_kvcache(q.unsqueeze(1), k_cache, v_cache,
-                                            cache_seqlens=context.context_lens, block_table=context.block_tables,
-                                            softmax_scale=self.scale, causal=True)
+                # GPU-only mode: use policy for attention
+                o = sparse_policy.compute_decode(
+                    q, k_cache, v_cache,
+                    context.context_lens, self.scale, self.layer_id,
+                    context.block_tables,
+                )
         return o
 
     def _chunked_prefill_attention(
@@ -195,20 +221,19 @@ class Attention(nn.Module):
         cpu_block_table = kvcache_manager.get_prefilled_cpu_blocks(seq)
 
         # Step 2: Apply select_blocks to filter blocks (before calling compute_chunked_prefill)
-        selected_blocks = []
-        if cpu_block_table:
-            num_chunks = current_chunk_idx + 1
-            policy_ctx = PolicyContext(
-                query_chunk_idx=current_chunk_idx,
-                num_query_chunks=num_chunks,
-                layer_id=self.layer_id,
-                query=q,  # Pass query for sparse policies that need it
-                is_prefill=True,
-                block_size=kvcache_manager.block_size,
-                total_kv_len=len(cpu_block_table) * kvcache_manager.block_size,
-            )
-            selected_blocks = sparse_policy.select_blocks(cpu_block_table, offload_engine, policy_ctx)
-            logger.debug(f"[DEBUG] select_blocks: {len(cpu_block_table)} -> {len(selected_blocks)} blocks")
+        # Always call select_blocks even for first chunk (cpu_block_table may be empty)
+        num_chunks = current_chunk_idx + 1
+        policy_ctx = PolicyContext(
+            query_chunk_idx=current_chunk_idx,
+            num_query_chunks=num_chunks,
+            layer_id=self.layer_id,
+            query=q,  # Pass query for sparse policies that need it
+            is_prefill=True,
+            block_size=kvcache_manager.block_size,
+            total_kv_len=len(cpu_block_table) * kvcache_manager.block_size if cpu_block_table else 0,
+        )
+        selected_blocks = sparse_policy.select_blocks(cpu_block_table, offload_engine, policy_ctx, q, k)
+        logger.debug(f"[DEBUG] select_blocks: {len(cpu_block_table)} -> {len(selected_blocks)} blocks")
 
         # [DEBUG] Verify execution path
         logger.debug(f"[DEBUG] Calling sparse_policy.compute_chunked_prefill, "
@@ -294,7 +319,7 @@ class Attention(nn.Module):
                 block_size=kvcache_manager.block_size,
                 total_kv_len=len(cpu_block_table) * kvcache_manager.block_size,
             )
-            selected_blocks = sparse_policy.select_blocks(cpu_block_table, offload_engine, policy_ctx)
+            selected_blocks = sparse_policy.select_blocks(cpu_block_table, offload_engine, policy_ctx, q, k)
             logger.debug(f"[DEBUG] decode select_blocks: {len(cpu_block_table)} -> {len(selected_blocks)} blocks")
 
         # [DEBUG] Verify execution path

nanovllm/layers/rotary_embedding.py

@@ -8,12 +8,43 @@ def apply_rotary_emb(
     cos: torch.Tensor,
     sin: torch.Tensor,
 ) -> torch.Tensor:
+    """Non-interleaved RoPE (used by Llama, Qwen, etc.)"""
     x1, x2 = torch.chunk(x.float(), 2, dim=-1)
     y1 = x1 * cos - x2 * sin
     y2 = x2 * cos + x1 * sin
     return torch.cat((y1, y2), dim=-1).to(x.dtype)
 
 
+def apply_rotary_emb_interleaved(
+    x: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+) -> torch.Tensor:
+    """Interleaved RoPE (used by GLM-4, etc.)
+
+    Args:
+        x: [seq_len, num_heads, head_dim]
+        cos: [seq_len, 1, head_dim // 2]
+        sin: [seq_len, 1, head_dim // 2]
+
+    x is reshaped to [seq_len, num_heads, head_dim // 2, 2] where:
+    - x[..., 0] are even positions
+    - x[..., 1] are odd positions
+    """
+    rot_dim = x.shape[-1]
+    # x_shaped: [seq_len, num_heads, rot_dim // 2, 2]
+    x_shaped = x.float().reshape(*x.shape[:-1], rot_dim // 2, 2)
+    # x_0, x_1: [seq_len, num_heads, rot_dim // 2]
+    x_0 = x_shaped[..., 0]
+    x_1 = x_shaped[..., 1]
+    # cos/sin: [seq_len, 1, rot_dim // 2] - broadcasts to num_heads
+    x_out = torch.stack([
+        x_0 * cos - x_1 * sin,
+        x_1 * cos + x_0 * sin,
+    ], dim=-1)
+    return x_out.flatten(-2).to(x.dtype)
+
+
 class RotaryEmbedding(nn.Module):
 
     def __init__(
@@ -140,6 +171,76 @@ class Llama3RotaryEmbedding(nn.Module):
         return query, key
 
 
+class GLM4RotaryEmbedding(nn.Module):
+    """
+    GLM-4 RoPE with interleaved rotation and partial rotation.
+
+    GLM-4 uses:
+    - Interleaved rotation (pairs adjacent elements, not first/second half)
+    - rope_ratio to scale base: base = 10000 * rope_ratio
+    - Partial rotation: only rotates first rotary_dim elements, rest pass through
+    - rotary_dim = head_dim // 2 (only half of head_dim is rotated)
+    """
+
+    def __init__(
+        self,
+        head_size: int,
+        rotary_dim: int,
+        max_position_embeddings: int,
+        base: float,
+    ) -> None:
+        super().__init__()
+        self.head_size = head_size
+        self.rotary_dim = rotary_dim  # GLM-4: rotary_dim = head_dim // 2
+        # inv_freq shape: [rotary_dim // 2]
+        inv_freq = 1.0 / (base ** (torch.arange(0, rotary_dim, 2, dtype=torch.float) / rotary_dim))
+        t = torch.arange(max_position_embeddings, dtype=torch.float)
+        freqs = torch.einsum("i,j -> ij", t, inv_freq)  # [max_pos, rotary_dim // 2]
+        cos = freqs.cos()
+        sin = freqs.sin()
+        # cache shape [max_pos, 1, rotary_dim // 2, 2]
+        cache = torch.stack((cos, sin), dim=-1).unsqueeze_(1)
+        self.register_buffer("cos_sin_cache", cache, persistent=False)
+
+    @torch.compile
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply RoPE to query and key.
+
+        Args:
+            positions: [seq_len]
+            query: [seq_len, num_heads, head_dim]
+            key: [seq_len, num_kv_heads, head_dim]
+
+        Returns:
+            Rotated query and key with same shapes as input.
+        """
+        cache = self.cos_sin_cache[positions]  # [seq_len, 1, rotary_dim // 2, 2]
+        cos = cache[..., 0]  # [seq_len, 1, rotary_dim // 2]
+        sin = cache[..., 1]  # [seq_len, 1, rotary_dim // 2]
+
+        # Split into rotated and pass-through parts
+        q_rot = query[..., :self.rotary_dim]
+        q_pass = query[..., self.rotary_dim:]
+        k_rot = key[..., :self.rotary_dim]
+        k_pass = key[..., self.rotary_dim:]
+
+        # Apply interleaved RoPE to rotated part
+        q_rot = apply_rotary_emb_interleaved(q_rot, cos, sin)
+        k_rot = apply_rotary_emb_interleaved(k_rot, cos, sin)
+
+        # Concatenate rotated and pass-through parts
+        query = torch.cat([q_rot, q_pass], dim=-1)
+        key = torch.cat([k_rot, k_pass], dim=-1)
+
+        return query, key
+
+
 # Cache for RoPE instances (keyed by hashable parameters)
 _rope_cache: dict[tuple, nn.Module] = {}
 
@@ -150,10 +251,11 @@ def get_rope(
     max_position: int,
     base: float,
     rope_scaling: dict | None = None,
+    is_interleaved: bool = False,
 ):
     # Create hashable cache key
     if rope_scaling is None:
-        cache_key = (head_size, rotary_dim, max_position, base, None)
+        cache_key = (head_size, rotary_dim, max_position, base, None, is_interleaved)
     else:
         rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
         if rope_type == "llama3":
@@ -163,15 +265,19 @@ def get_rope(
                 rope_scaling["low_freq_factor"],
                 rope_scaling["high_freq_factor"],
                 rope_scaling["original_max_position_embeddings"],
+                is_interleaved,
             )
         else:
-            cache_key = (head_size, rotary_dim, max_position, base, rope_type)
+            cache_key = (head_size, rotary_dim, max_position, base, rope_type, is_interleaved)
 
     if cache_key in _rope_cache:
         return _rope_cache[cache_key]
 
     if rope_scaling is None:
-        rope = RotaryEmbedding(head_size, rotary_dim, max_position, base)
+        if is_interleaved:
+            rope = GLM4RotaryEmbedding(head_size, rotary_dim, max_position, base)
+        else:
+            rope = RotaryEmbedding(head_size, rotary_dim, max_position, base)
     else:
         rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
         if rope_type == "llama3":

nanovllm/models/__init__.py

@@ -3,7 +3,9 @@
 from nanovllm.models.registry import register_model, get_model_class, MODEL_REGISTRY
 
 # Import models to trigger registration
+from nanovllm.models import qwen2
 from nanovllm.models import qwen3
 from nanovllm.models import llama
+from nanovllm.models import glm4
 
 __all__ = ["register_model", "get_model_class", "MODEL_REGISTRY"]

nanovllm/models/glm4.py

@@ -0,0 +1,235 @@
+"""GLM-4 model implementation for nano-vllm."""
+import torch
+from torch import nn
+import torch.distributed as dist
+
+from nanovllm.layers.activation import SiluAndMul
+from nanovllm.layers.attention import Attention
+from nanovllm.layers.layernorm import RMSNorm
+from nanovllm.layers.linear import QKVParallelLinear, MergedColumnParallelLinear, RowParallelLinear
+from nanovllm.layers.rotary_embedding import get_rope
+from nanovllm.layers.embed_head import VocabParallelEmbedding, ParallelLMHead
+from nanovllm.models.registry import register_model
+
+
+class GLM4Attention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        max_position: int = 1048576,
+        head_dim: int = 128,
+        rope_theta: float = 10000,
+        rope_scaling: dict | None = None,
+    ) -> None:
+        super().__init__()
+        tp_size = dist.get_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        assert self.total_num_kv_heads % tp_size == 0
+        self.num_kv_heads = self.total_num_kv_heads // tp_size
+        self.head_dim = head_dim
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim ** -0.5
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=True,  # GLM-4 has QKV bias
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,  # GLM-4 has no output bias
+        )
+        # GLM-4 only rotates half of head_dim
+        rotary_dim = self.head_dim // 2
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=rotary_dim,
+            max_position=max_position,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_interleaved=True,  # GLM-4 uses interleaved RoPE
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            self.num_kv_heads,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q = q.view(-1, self.num_heads, self.head_dim)
+        k = k.view(-1, self.num_kv_heads, self.head_dim)
+        v = v.view(-1, self.num_kv_heads, self.head_dim)
+        q, k = self.rotary_emb(positions, q, k)
+        o = self.attn(q, k, v)
+        output = self.o_proj(o.flatten(1, -1))
+        return output
+
+
+class GLM4MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,  # GLM-4 has no MLP bias
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+        )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x = self.down_proj(x)
+        return x
+
+
+class GLM4DecoderLayer(nn.Module):
+
+    def __init__(self, config) -> None:
+        super().__init__()
+        # GLM-4 config field mapping
+        hidden_size = config.hidden_size
+        num_heads = config.num_attention_heads
+        num_kv_heads = getattr(config, 'multi_query_group_num', num_heads)
+        head_dim = getattr(config, 'kv_channels', hidden_size // num_heads)
+        max_position = getattr(config, 'seq_length', 1048576)
+        rope_ratio = getattr(config, 'rope_ratio', 1)
+        rope_theta = 10000 * rope_ratio  # GLM-4 uses rope_ratio to scale base
+        intermediate_size = getattr(config, 'ffn_hidden_size', getattr(config, 'intermediate_size', None))
+        rms_norm_eps = getattr(config, 'layernorm_epsilon', 1e-5)
+
+        self.self_attn = GLM4Attention(
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            max_position=max_position,
+            head_dim=head_dim,
+            rope_theta=rope_theta,
+            rope_scaling=getattr(config, "rope_scaling", None),
+        )
+        self.mlp = GLM4MLP(
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+        )
+        self.input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if residual is None:
+            hidden_states, residual = self.input_layernorm(hidden_states), hidden_states
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(positions, hidden_states)
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class GLM4Model(nn.Module):
+
+    def __init__(self, config) -> None:
+        super().__init__()
+        vocab_size = getattr(config, 'padded_vocab_size', config.vocab_size)
+        num_layers = getattr(config, 'num_layers', config.num_hidden_layers)
+        rms_norm_eps = getattr(config, 'layernorm_epsilon', 1e-5)
+
+        self.embed_tokens = VocabParallelEmbedding(vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([GLM4DecoderLayer(config) for _ in range(num_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for layer in self.layers:
+            hidden_states, residual = layer(positions, hidden_states, residual)
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+@register_model("ChatGLMModel", "ChatGLMForConditionalGeneration")
+class ChatGLMForCausalLM(nn.Module):
+    """
+    GLM-4 model for causal language modeling.
+
+    Weight mapping from HuggingFace to nanovllm:
+    - transformer.embedding.word_embeddings → model.embed_tokens
+    - transformer.encoder.layers.X.input_layernorm → model.layers.X.input_layernorm
+    - transformer.encoder.layers.X.self_attention.query_key_value → model.layers.X.self_attn.qkv_proj (split q/k/v)
+    - transformer.encoder.layers.X.self_attention.dense → model.layers.X.self_attn.o_proj
+    - transformer.encoder.layers.X.post_attention_layernorm → model.layers.X.post_attention_layernorm
+    - transformer.encoder.layers.X.mlp.dense_h_to_4h → model.layers.X.mlp.gate_up_proj (split gate/up)
+    - transformer.encoder.layers.X.mlp.dense_4h_to_h → model.layers.X.mlp.down_proj
+    - transformer.encoder.final_layernorm → model.norm
+    - transformer.output_layer → lm_head
+    """
+    packed_modules_mapping = {
+        # QKV is merged in GLM-4 as query_key_value
+        "query_key_value": ("qkv_proj", None),  # Special handling needed
+        # MLP gate and up are merged as dense_h_to_4h
+        "dense_h_to_4h": ("gate_up_proj", None),  # Special handling needed
+    }
+
+    # Weight name mapping for loader
+    hf_to_nanovllm_mapping = {
+        "transformer.embedding.word_embeddings": "model.embed_tokens",
+        "transformer.encoder.final_layernorm": "model.norm",
+        "transformer.output_layer": "lm_head",
+    }
+
+    def __init__(self, config) -> None:
+        super().__init__()
+        vocab_size = getattr(config, 'padded_vocab_size', config.vocab_size)
+        self.config = config
+        self.model = GLM4Model(config)
+        self.lm_head = ParallelLMHead(vocab_size, config.hidden_size)
+        # GLM-4 does not tie embeddings
+        # if getattr(config, 'tie_word_embeddings', False):
+        #     self.lm_head.weight.data = self.model.embed_tokens.weight.data
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.model(input_ids, positions)
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.lm_head(hidden_states)

nanovllm/models/qwen2.py

@@ -0,0 +1,207 @@
+import torch
+from torch import nn
+import torch.distributed as dist
+from transformers import Qwen2Config
+
+from nanovllm.layers.activation import SiluAndMul
+from nanovllm.layers.attention import Attention
+from nanovllm.layers.layernorm import RMSNorm
+from nanovllm.layers.linear import QKVParallelLinear, MergedColumnParallelLinear, RowParallelLinear
+from nanovllm.layers.rotary_embedding import get_rope
+from nanovllm.layers.embed_head import VocabParallelEmbedding, ParallelLMHead
+from nanovllm.models.registry import register_model
+
+
+class Qwen2Attention(nn.Module):
+    """Qwen2/2.5 Attention without QK norm (unlike Qwen3)."""
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        max_position: int = 4096 * 32,
+        head_dim: int | None = None,
+        rope_theta: float = 10000,
+        rope_scaling: tuple | None = None,
+    ) -> None:
+        super().__init__()
+        tp_size = dist.get_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        assert self.total_num_kv_heads % tp_size == 0
+        self.num_kv_heads = self.total_num_kv_heads // tp_size
+        self.head_dim = head_dim or hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim ** -0.5
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=True,  # Qwen2/2.5 always uses bias
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            self.num_kv_heads,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q = q.view(-1, self.num_heads, self.head_dim)
+        k = k.view(-1, self.num_kv_heads, self.head_dim)
+        v = v.view(-1, self.num_kv_heads, self.head_dim)
+        q, k = self.rotary_emb(positions, q, k)
+        o = self.attn(q, k, v)
+        output = self.o_proj(o.flatten(1, -1))
+        return output
+
+
+class Qwen2MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+        )
+        assert hidden_act == "silu"
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x = self.down_proj(x)
+        return x
+
+
+class Qwen2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Qwen2Config,
+    ) -> None:
+        super().__init__()
+        self.self_attn = Qwen2Attention(
+            hidden_size=config.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            max_position=config.max_position_embeddings,
+            head_dim=getattr(config, 'head_dim', None),
+            rope_theta=getattr(config, "rope_theta", 1000000),
+            rope_scaling=getattr(config, "rope_scaling", None),
+        )
+        self.mlp = Qwen2MLP(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if residual is None:
+            hidden_states, residual = self.input_layernorm(hidden_states), hidden_states
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(positions, hidden_states)
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen2Model(nn.Module):
+
+    def __init__(
+        self,
+        config: Qwen2Config,
+    ) -> None:
+        super().__init__()
+        self.embed_tokens = VocabParallelEmbedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([Qwen2DecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for layer in self.layers:
+            hidden_states, residual = layer(positions, hidden_states, residual)
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+@register_model("Qwen2ForCausalLM")
+class Qwen2ForCausalLM(nn.Module):
+    packed_modules_mapping = {
+        "q_proj": ("qkv_proj", "q"),
+        "k_proj": ("qkv_proj", "k"),
+        "v_proj": ("qkv_proj", "v"),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: Qwen2Config
+    ) -> None:
+        super().__init__()
+        self.model = Qwen2Model(config)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        if config.tie_word_embeddings:
+            self.lm_head.weight.data = self.model.embed_tokens.weight.data
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.model(input_ids, positions)
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.lm_head(hidden_states)

nanovllm/models/qwen3.py

@@ -187,7 +187,7 @@ class Qwen3Model(nn.Module):
         return hidden_states
 
 
-@register_model("Qwen3ForCausalLM", "Qwen2ForCausalLM")
+@register_model("Qwen3ForCausalLM")
 class Qwen3ForCausalLM(nn.Module):
     packed_modules_mapping = {
         "q_proj": ("qkv_proj", "q"),

nanovllm/ops/chunked_attention.py

@@ -414,6 +414,90 @@ def merge_attention_outputs(
     return o_merged, lse_merged
 
 
+# ============================================================
+# FlashInfer-based implementations (recommended for merge only)
+# ============================================================
+
+# LSE conversion constants: FlashInfer uses log2, flash_attn uses ln
+_LOG2_E = 1.4426950408889634  # math.log2(math.e) - ln -> log2
+_LN_2 = 0.6931471805599453    # math.log(2) - log2 -> ln
+
+# Check FlashInfer availability (only for merge_state, not attention kernel)
+try:
+    from flashinfer.cascade import merge_state, merge_state_in_place
+    FLASHINFER_MERGE_AVAILABLE = True
+except ImportError:
+    FLASHINFER_MERGE_AVAILABLE = False
+
+
+def flash_attn_with_lse_flashinfer(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Flash attention that returns output and LSE.
+
+    Uses flash_attn library (FlashInfer attention has JIT compatibility issues).
+
+    Args:
+        q: Query tensor [batch, seqlen_q, nheads_q, headdim]
+        k: Key tensor [batch, seqlen_k, nheads_kv, headdim]
+        v: Value tensor [batch, seqlen_k, nheads_kv, headdim]
+        softmax_scale: Scaling factor (default: 1/sqrt(headdim))
+        causal: Whether to apply causal masking
+
+    Returns:
+        out: Output tensor [batch, seqlen_q, nheads_q, headdim]
+        lse: Log-sum-exp tensor [batch, nheads_q, seqlen_q] (ln format)
+    """
+    # Use flash_attn directly (FlashInfer attention JIT has CUDA version issues)
+    return flash_attn_with_lse(q, k, v, softmax_scale, causal)
+
+
+def merge_attention_outputs_flashinfer(
+    o1: torch.Tensor,
+    lse1: torch.Tensor,
+    o2: torch.Tensor,
+    lse2: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Merge two attention outputs using FlashInfer's optimized kernel.
+
+    Args:
+        o1: First output [batch, seqlen_q, nheads, headdim]
+        lse1: First LSE [batch, nheads, seqlen_q] (ln format)
+        o2: Second output [batch, seqlen_q, nheads, headdim]
+        lse2: Second LSE [batch, nheads, seqlen_q] (ln format)
+
+    Returns:
+        o_merged: Merged output [batch, seqlen_q, nheads, headdim]
+        lse_merged: Merged LSE [batch, nheads, seqlen_q] (ln format)
+    """
+    if not FLASHINFER_MERGE_AVAILABLE:
+        # Fallback to Triton implementation
+        return merge_attention_outputs(o1, lse1, o2, lse2)
+
+    # Convert to FlashInfer format
+    # o: [batch, seq, heads, dim] -> [seq, heads, dim]
+    # lse: [batch, heads, seq] -> [seq, heads] (convert ln -> log2)
+    v_a = o1.squeeze(0).contiguous()
+    s_a = (lse1.squeeze(0).transpose(0, 1).contiguous().float() * _LOG2_E)
+    v_b = o2.squeeze(0).contiguous()
+    s_b = (lse2.squeeze(0).transpose(0, 1).contiguous().float() * _LOG2_E)
+
+    # FlashInfer merge
+    v_merged, s_merged = merge_state(v_a, s_a, v_b, s_b)
+
+    # Convert back to flash_attn format
+    o_merged = v_merged.unsqueeze(0)  # [1, seq, heads, dim]
+    lse_merged = (s_merged * _LN_2).transpose(0, 1).unsqueeze(0)  # [1, heads, seq]
+
+    return o_merged, lse_merged
+
+
 def chunked_attention_varlen(
     q: torch.Tensor,
     kv_chunks: List[Tuple[torch.Tensor, torch.Tensor]],

nanovllm/ops/xattn.py

@@ -218,6 +218,209 @@ def softmax_fuse_block_sum_kernel_non_causal(
         tl.store(output_ptr + iter * segment_size // block_size, X.to(Out.type.element_ty))
 
 
+# ============================================================
+# KV Chunking Support Kernels
+# ============================================================
+
+@triton.jit
+def softmax_partial_stats_kernel(
+    In,
+    M_out,  # max per row
+    L_out,  # sum per row (normalized by M_out)
+    scale,
+    input_stride_0,
+    input_stride_1,
+    input_stride_2,
+    stats_stride_0,
+    stats_stride_1,
+    k_len,
+    chunk_start,  # Q start position (for causal)
+    kv_offset,    # KV chunk offset (for causal)
+    segment_size: tl.constexpr,
+    block_size: tl.constexpr,
+    is_causal: tl.constexpr,
+):
+    """
+    Compute partial softmax statistics for a KV chunk.
+
+    For each query row, computes:
+    - m: max value in this chunk
+    - l: sum of exp(x - m) in this chunk
+
+    These can be merged across chunks using online softmax formula.
+
+    Input shape: [batch, heads, q_len, k_chunk_len]
+    Output shapes: M[batch, heads, q_len], L[batch, heads, q_len]
+    """
+    block_id = tl.program_id(0)
+    head_id = tl.program_id(1)
+    batch_id = tl.program_id(2)
+
+    offs_q = tl.arange(0, block_size) + chunk_start + block_id * block_size
+    offs_k = tl.arange(0, segment_size)
+
+    num_iters = k_len // segment_size
+
+    # For causal: compute boundary
+    if is_causal:
+        # causal boundary: Q position where this KV chunk starts to be valid
+        # Q[i] can attend K[j] if i >= j
+        # For KV chunk at kv_offset, Q[i] can attend if i >= kv_offset
+        num_iters_before_causal = (chunk_start + (block_id + 1) * block_size - 1 - kv_offset) // segment_size
+        num_iters_before_causal = tl.minimum(num_iters_before_causal, num_iters)
+        num_iters_before_causal = tl.maximum(num_iters_before_causal, 0)
+    else:
+        num_iters_before_causal = num_iters
+
+    # Online softmax state
+    m_i = tl.zeros([block_size], dtype=tl.float32) - float("inf")
+    l_i = tl.zeros([block_size], dtype=tl.float32)
+
+    # Input pointer
+    input_ptr = In + batch_id * input_stride_0 + head_id * input_stride_1 + block_id * block_size * input_stride_2
+    input_ptr = input_ptr + tl.arange(0, segment_size) + tl.arange(0, block_size)[:, None] * input_stride_2
+
+    # Compute max and sum (before causal boundary)
+    for iter in range(0, num_iters_before_causal):
+        X = tl.load(input_ptr + iter * segment_size).to(tl.float32) * scale
+        m_local = tl.max(X, 1)
+        m_new = tl.maximum(m_i, m_local)
+        alpha = tl.math.exp2(m_i - m_new)
+
+        X = X - m_new[:, None]
+        l_local = tl.sum(tl.math.exp2(X), 1)
+        l_i = l_i * alpha + l_local
+
+        m_i = m_new
+
+    # Handle causal boundary
+    if is_causal:
+        for iter in range(num_iters_before_causal, num_iters_before_causal + 1):
+            if iter < num_iters:
+                X = tl.load(input_ptr + iter * segment_size).to(tl.float32) * scale
+                # causal mask: Q[i] >= K[j] + kv_offset
+                mask = offs_q[:, None] >= (offs_k[None, :] + iter * segment_size + kv_offset)
+                X = tl.where(mask, X, -1.0e6)
+                m_local = tl.max(X, 1)
+                m_new = tl.maximum(m_i, m_local)
+                alpha = tl.math.exp2(m_i - m_new)
+
+                X = X - m_new[:, None]
+                l_local = tl.sum(tl.math.exp2(X), 1)
+                l_i = l_i * alpha + l_local
+
+                m_i = m_new
+
+    # Output pointers
+    m_ptr = M_out + batch_id * stats_stride_0 + head_id * stats_stride_1 + block_id * block_size
+    l_ptr = L_out + batch_id * stats_stride_0 + head_id * stats_stride_1 + block_id * block_size
+
+    offs = tl.arange(0, block_size)
+    tl.store(m_ptr + offs, m_i.to(M_out.type.element_ty))
+    tl.store(l_ptr + offs, l_i.to(L_out.type.element_ty))
+
+
+@triton.jit
+def softmax_normalize_block_sum_kernel(
+    In,
+    Out,
+    M_global,  # global max per row
+    L_global,  # global sum per row
+    scale,
+    input_stride_0,
+    input_stride_1,
+    input_stride_2,
+    output_stride_0,
+    output_stride_1,
+    output_stride_2,
+    stats_stride_0,
+    stats_stride_1,
+    real_q_len,
+    k_len,
+    chunk_start,
+    kv_offset,  # KV chunk offset (for causal)
+    segment_size: tl.constexpr,
+    block_size: tl.constexpr,
+    is_causal: tl.constexpr,
+):
+    """
+    Normalize with global stats and compute block sums for a KV chunk.
+
+    Uses pre-computed global m and l to correctly normalize softmax
+    across all KV chunks.
+
+    Input shape: [batch, heads, q_len, k_chunk_len]
+    Output shape: [batch, heads, q_blocks, k_chunk_blocks]
+    """
+    block_id = tl.program_id(0)
+    head_id = tl.program_id(1)
+    batch_id = tl.program_id(2)
+
+    offs_q = tl.arange(0, block_size) + chunk_start + block_id * block_size
+    offs_k = tl.arange(0, segment_size)
+
+    num_iters = k_len // segment_size
+
+    # For causal: compute boundary
+    if is_causal:
+        num_iters_before_causal = (chunk_start + (block_id + 1) * block_size - 1 - kv_offset) // segment_size
+        num_iters_before_causal = tl.minimum(num_iters_before_causal, num_iters)
+        num_iters_before_causal = tl.maximum(num_iters_before_causal, 0)
+    else:
+        num_iters_before_causal = num_iters
+
+    # Load global stats
+    m_ptr = M_global + batch_id * stats_stride_0 + head_id * stats_stride_1 + block_id * block_size
+    l_ptr = L_global + batch_id * stats_stride_0 + head_id * stats_stride_1 + block_id * block_size
+
+    offs = tl.arange(0, block_size)
+    m_global = tl.load(m_ptr + offs).to(tl.float32)
+    l_global = tl.load(l_ptr + offs).to(tl.float32)
+    # Handle l_global = 0 (when all positions are masked)
+    l_global_safe = tl.where(l_global > 0, l_global, 1.0)
+    l_global_inv = 1.0 / l_global_safe
+
+    # Input pointer
+    input_ptr = In + batch_id * input_stride_0 + head_id * input_stride_1 + block_id * block_size * input_stride_2
+    input_ptr = input_ptr + tl.arange(0, segment_size) + tl.arange(0, block_size)[:, None] * input_stride_2
+
+    # Output pointer
+    output_ptr = Out + batch_id * output_stride_0 + head_id * output_stride_1 + block_id * output_stride_2
+    output_ptr = output_ptr + tl.arange(0, segment_size // block_size)
+
+    sum_mask = offs_q[:, None] < real_q_len
+
+    # Normalize and compute block sums (before causal boundary)
+    for iter in range(0, num_iters_before_causal):
+        X = tl.load(input_ptr + iter * segment_size).to(tl.float32) * scale
+        X = tl.exp2(X - m_global[:, None]) * l_global_inv[:, None]
+        X = tl.where(sum_mask, X, 0)
+        X = tl.reshape(X, (block_size, segment_size // block_size, block_size))
+        X = tl.sum(X, 2)
+        X = tl.sum(X, 0)
+        tl.store(output_ptr + iter * segment_size // block_size, X.to(Out.type.element_ty))
+
+    # Handle causal boundary
+    if is_causal:
+        for iter in range(num_iters_before_causal, num_iters_before_causal + 1):
+            if iter < num_iters:
+                X = tl.load(input_ptr + iter * segment_size).to(tl.float32) * scale
+                # causal mask: Q[i] >= K[j] + kv_offset
+                causal_mask = offs_q[:, None] >= (offs_k[None, :] + iter * segment_size + kv_offset)
+                X = tl.where(causal_mask, X, -1.0e6)
+                X = tl.exp2(X - m_global[:, None]) * l_global_inv[:, None]
+                X = tl.where(sum_mask, X, 0)
+                X = tl.reshape(X, (block_size, segment_size // block_size, block_size))
+                X = tl.sum(X, 2)
+                X = tl.sum(X, 0)
+                tl.store(output_ptr + iter * segment_size // block_size, X.to(Out.type.element_ty))
+
+        # Zero out future blocks
+        for iter in range(num_iters_before_causal + 1, num_iters):
+            X = tl.zeros([segment_size // block_size], dtype=tl.float32)
+            tl.store(output_ptr + iter * segment_size // block_size, X.to(Out.type.element_ty))
+
+
 @triton.jit
 def flat_group_gemm_fuse_reshape_kernel(
     Q, K, Out,
@@ -380,6 +583,194 @@ def softmax_fuse_block_sum(
     return output
 
 
+def softmax_compute_partial_stats(
+    attn_weights_slice: torch.Tensor,
+    reshaped_block_size: int,
+    segment_size: int,
+    scale: float,
+    chunk_start: int = 0,
+    kv_offset: int = 0,
+    is_causal: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Compute partial softmax statistics for a KV chunk.
+
+    This is the first step for KV-chunked softmax computation.
+    For each query row, computes:
+    - m: max value in this chunk
+    - l: sum of exp(x - m) in this chunk
+
+    These partial stats can be merged across KV chunks using
+    `merge_softmax_stats()`, then used with `softmax_normalize_and_block_sum()`.
+
+    Args:
+        attn_weights_slice: Raw attention scores [batch, heads, q_len, k_chunk_len]
+        reshaped_block_size: Block size in reshaped space
+        segment_size: Processing segment size
+        scale: Softmax scale factor
+        chunk_start: Q chunk start position (in reshaped space)
+        kv_offset: KV chunk offset (in reshaped space, for causal masking)
+        is_causal: Whether to apply causal masking
+
+    Returns:
+        Tuple of (m, l) where:
+        - m: [batch, heads, q_len] max values per row
+        - l: [batch, heads, q_len] partial sums per row
+    """
+    batch_size, num_heads, q_len, k_len = attn_weights_slice.shape
+
+    assert q_len % reshaped_block_size == 0
+    assert k_len % segment_size == 0
+    assert attn_weights_slice.stride(-1) == 1
+
+    m_out = torch.empty(
+        (batch_size, num_heads, q_len),
+        dtype=torch.float32,
+        device=attn_weights_slice.device
+    )
+    l_out = torch.empty(
+        (batch_size, num_heads, q_len),
+        dtype=torch.float32,
+        device=attn_weights_slice.device
+    )
+
+    grid = (q_len // reshaped_block_size, num_heads, batch_size)
+
+    softmax_partial_stats_kernel[grid](
+        attn_weights_slice,
+        m_out,
+        l_out,
+        scale,
+        attn_weights_slice.stride(0),
+        attn_weights_slice.stride(1),
+        attn_weights_slice.stride(2),
+        m_out.stride(0),
+        m_out.stride(1),
+        k_len,
+        chunk_start,
+        kv_offset,
+        segment_size,
+        reshaped_block_size,
+        is_causal,
+    )
+
+    return m_out, l_out
+
+
+def merge_softmax_stats(
+    m_chunks: list,
+    l_chunks: list,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Merge partial softmax statistics from multiple KV chunks.
+
+    Uses the online softmax merging formula:
+    m_new = max(m1, m2)
+    l_new = l1 * exp(m1 - m_new) + l2 * exp(m2 - m_new)
+
+    Args:
+        m_chunks: List of max tensors [batch, heads, q_len] from each chunk
+        l_chunks: List of sum tensors [batch, heads, q_len] from each chunk
+
+    Returns:
+        Tuple of (m_global, l_global) with same shape as inputs
+    """
+    assert len(m_chunks) == len(l_chunks)
+    assert len(m_chunks) > 0
+
+    # Use log2 scale to match kernel (exp2)
+    LOG2E = 1.4426950408889634
+
+    m_global = m_chunks[0].clone()
+    l_global = l_chunks[0].clone()
+
+    for i in range(1, len(m_chunks)):
+        m_chunk = m_chunks[i]
+        l_chunk = l_chunks[i]
+
+        m_new = torch.maximum(m_global, m_chunk)
+        # exp2(m - m_new) = 2^(m - m_new)
+        l_global = l_global * torch.pow(2.0, m_global - m_new) + l_chunk * torch.pow(2.0, m_chunk - m_new)
+        m_global = m_new
+
+    return m_global, l_global
+
+
+def softmax_normalize_and_block_sum(
+    attn_weights_slice: torch.Tensor,
+    m_global: torch.Tensor,
+    l_global: torch.Tensor,
+    reshaped_block_size: int,
+    segment_size: int,
+    chunk_start: int,
+    real_q_len: int,
+    scale: float,
+    kv_offset: int = 0,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """
+    Normalize with global stats and compute block sums for a KV chunk.
+
+    This is the second step for KV-chunked softmax computation.
+    Uses pre-computed global m and l (from `merge_softmax_stats()`)
+    to correctly normalize softmax values and compute block sums.
+
+    Args:
+        attn_weights_slice: Raw attention scores [batch, heads, q_len, k_chunk_len]
+        m_global: Global max values [batch, heads, q_len]
+        l_global: Global sum values [batch, heads, q_len]
+        reshaped_block_size: Block size in reshaped space
+        segment_size: Processing segment size
+        chunk_start: Start position for this chunk (for masking)
+        real_q_len: Actual Q length (before padding)
+        scale: Softmax scale factor
+        kv_offset: KV chunk offset (in reshaped space, for causal masking)
+        is_causal: Whether to apply causal masking
+
+    Returns:
+        Block-level attention sums [batch, heads, q_blocks, k_chunk_blocks]
+    """
+    batch_size, num_heads, q_len, k_len = attn_weights_slice.shape
+
+    assert q_len % reshaped_block_size == 0
+    assert k_len % segment_size == 0
+    assert segment_size % reshaped_block_size == 0
+    assert attn_weights_slice.stride(-1) == 1
+
+    output = torch.empty(
+        (batch_size, num_heads, q_len // reshaped_block_size, k_len // reshaped_block_size),
+        dtype=attn_weights_slice.dtype,
+        device=attn_weights_slice.device
+    )
+
+    grid = (q_len // reshaped_block_size, num_heads, batch_size)
+
+    softmax_normalize_block_sum_kernel[grid](
+        attn_weights_slice,
+        output,
+        m_global,
+        l_global,
+        scale,
+        attn_weights_slice.stride(0),
+        attn_weights_slice.stride(1),
+        attn_weights_slice.stride(2),
+        output.stride(0),
+        output.stride(1),
+        output.stride(2),
+        m_global.stride(0),
+        m_global.stride(1),
+        real_q_len,
+        k_len,
+        chunk_start,
+        kv_offset,
+        segment_size,
+        reshaped_block_size,
+        is_causal,
+    )
+
+    return output
+
+
 def flat_group_gemm_fuse_reshape(
     query_states: torch.Tensor,
     key_states: torch.Tensor,

nanovllm/utils/density_observer.py

@@ -0,0 +1,327 @@
+"""
+DensityObserver - Sparse Attention Density 统计 Observer。
+
+统计两种 density:
+1. Compute Density (计算密度): 基于 BSA block size (128)
+   - density = selected_bsa_blocks / total_causal_bsa_blocks
+   - GPU-only 和 Offload 模式应该一致
+
+2. Communication Density (通信密度): 基于 CPU block size (如 4096)
+   - comm_density = selected_cpu_blocks / total_cpu_blocks
+   - 仅用于 Offload 模式，由于粒度更粗，必然 >= compute density
+
+统计位置：
+- GPU-only: xattn_bsa.py compute_prefill() - 只记录 compute density
+- Offload: xattn_bsa.py select_blocks() - 记录两种 density
+
+对于 Offload 模式的 Density 计算:
+- 不是简单的 avg 或 min
+- 而是 sum(selected) / sum(total)，正确处理不同 chunk 大小的权重
+"""
+
+from typing import List, Dict, Optional, Tuple
+import torch
+from nanovllm.utils.observer import Observer
+
+
+class DensityObserver(Observer):
+    """
+    Sparse Attention Density Observer。
+
+    记录每层的 density，用于验证 GPU-only 和 Offload 模式的一致性。
+
+    使用方式:
+        DensityObserver.enable()
+        DensityObserver.complete_reset()
+        # ... run inference ...
+        DensityObserver.record(layer_id, mask, causal=True)
+        # 或者使用累积模式 (offload):
+        DensityObserver.record_counts(layer_id, selected, total)
+        # ...
+        DensityObserver.print_summary()
+    """
+
+    _enabled: bool = False  # 默认禁用
+
+    # 每层的 compute density 记录 (BSA block 粒度)
+    # key: layer_id, value: list of density values (每次 prefill chunk 一个)
+    _layer_densities: Dict[int, List[float]] = {}
+
+    # 每层的 communication density 记录 (CPU block 粒度，仅 offload 模式)
+    _layer_comm_densities: Dict[int, List[float]] = {}
+
+    # 累积模式: 记录 selected/total counts (用于 offload 模式)
+    # 这样可以在所有 chunks 完成后正确计算 density = sum(selected) / sum(total)
+    _layer_selected_counts: Dict[int, List[int]] = {}
+    _layer_total_counts: Dict[int, List[int]] = {}
+
+    # Mask shape 记录 (用于调试)
+    _last_q_blocks: int = 0
+    _last_k_blocks: int = 0
+
+    # 模式标记
+    _mode: str = "unknown"  # "gpu_only" or "offload"
+
+    @classmethod
+    def set_mode(cls, mode: str) -> None:
+        """设置当前模式 (gpu_only / offload)"""
+        cls._mode = mode
+
+    @classmethod
+    def record(
+        cls,
+        layer_id: int,
+        mask: torch.Tensor,
+        causal: bool = True,
+    ) -> float:
+        """
+        记录一层的 density (适用于 GPU-only 模式)。
+
+        Args:
+            layer_id: 层 ID
+            mask: [batch, heads, q_blocks, k_blocks] boolean tensor
+            causal: 是否考虑 causal mask (只计算下三角)
+
+        Returns:
+            density 值
+        """
+        if not cls._enabled:
+            return 0.0
+
+        density = cls._compute_density(mask, causal)
+
+        # 记录
+        if layer_id not in cls._layer_densities:
+            cls._layer_densities[layer_id] = []
+        cls._layer_densities[layer_id].append(density)
+
+        # 记录 mask shape
+        cls._last_q_blocks = mask.shape[2]
+        cls._last_k_blocks = mask.shape[3]
+
+        return density
+
+    @classmethod
+    def record_counts(
+        cls,
+        layer_id: int,
+        selected_blocks: int,
+        total_blocks: int,
+    ) -> None:
+        """
+        记录一层的 selected/total block counts (适用于 offload 累积模式)。
+
+        使用累积计数而不是直接计算 density，这样在所有 chunks 处理完后可以正确计算：
+        overall_density = sum(selected) / sum(total)
+
+        这比 avg(density) 更准确，因为不同 chunk 的 Q 和 K 长度不同。
+
+        Args:
+            layer_id: 层 ID
+            selected_blocks: 这个 chunk 选中的 blocks 数量
+            total_blocks: 这个 chunk 的 total possible blocks 数量
+        """
+        if not cls._enabled:
+            return
+
+        # 初始化列表
+        if layer_id not in cls._layer_selected_counts:
+            cls._layer_selected_counts[layer_id] = []
+        if layer_id not in cls._layer_total_counts:
+            cls._layer_total_counts[layer_id] = []
+
+        # 累积记录
+        cls._layer_selected_counts[layer_id].append(selected_blocks)
+        cls._layer_total_counts[layer_id].append(total_blocks)
+
+    @classmethod
+    def record_comm_density(
+        cls,
+        layer_id: int,
+        selected_cpu_blocks: int,
+        total_cpu_blocks: int,
+    ) -> float:
+        """
+        记录一层的 communication density (CPU block 粒度)。
+
+        Args:
+            layer_id: 层 ID
+            selected_cpu_blocks: 选中的 CPU blocks 数量
+            total_cpu_blocks: 总 CPU blocks 数量
+
+        Returns:
+            communication density 值
+        """
+        if not cls._enabled:
+            return 0.0
+
+        if total_cpu_blocks == 0:
+            return 1.0
+
+        comm_density = selected_cpu_blocks / total_cpu_blocks
+
+        # 记录
+        if layer_id not in cls._layer_comm_densities:
+            cls._layer_comm_densities[layer_id] = []
+        cls._layer_comm_densities[layer_id].append(comm_density)
+
+        return comm_density
+
+    @classmethod
+    def _compute_density(cls, mask: torch.Tensor, causal: bool) -> float:
+        """计算 mask 的 density"""
+        batch, heads, q_blocks, k_blocks = mask.shape
+
+        if causal:
+            # 只计算下三角区域
+            causal_mask = torch.tril(
+                torch.ones(q_blocks, k_blocks, device=mask.device, dtype=torch.bool)
+            )
+            total_blocks = causal_mask.sum().item() * batch * heads
+            selected_blocks = (mask & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()
+        else:
+            total_blocks = mask.numel()
+            selected_blocks = mask.sum().item()
+
+        if total_blocks == 0:
+            return 1.0
+
+        return selected_blocks / total_blocks
+
+    @classmethod
+    def complete_reset(cls) -> None:
+        """重置所有统计"""
+        cls._layer_densities = {}
+        cls._layer_comm_densities = {}
+        cls._layer_selected_counts = {}
+        cls._layer_total_counts = {}
+        cls._last_q_blocks = 0
+        cls._last_k_blocks = 0
+        cls._mode = "unknown"
+
+    @classmethod
+    def get_per_layer_density(cls) -> Dict[int, float]:
+        """
+        获取每层的 density。
+
+        对于累积模式 (offload): density = sum(selected) / sum(total)
+        对于直接记录模式 (gpu_only): density = avg(density_values)
+        """
+        result = {}
+
+        # 优先使用累积模式 (offload)
+        if cls._layer_selected_counts:
+            for layer_id in cls._layer_selected_counts:
+                selected_list = cls._layer_selected_counts.get(layer_id, [])
+                total_list = cls._layer_total_counts.get(layer_id, [])
+                total_selected = sum(selected_list)
+                total_total = sum(total_list)
+                if total_total > 0:
+                    result[layer_id] = total_selected / total_total
+        else:
+            # 直接记录模式 (gpu_only)
+            for layer_id, densities in cls._layer_densities.items():
+                if densities:
+                    result[layer_id] = sum(densities) / len(densities)
+
+        return result
+
+    @classmethod
+    def get_overall_density(cls) -> float:
+        """
+        获取所有层的总体 compute density。
+
+        对于累积模式 (offload): density = sum(all_selected) / sum(all_total)
+        对于直接记录模式 (gpu_only): density = avg(all_density_values)
+
+        注意: 总体 density 不是简单的 avg(per_layer_density)，
+        而是 sum(all_selected) / sum(all_total)，这样可以正确处理权重。
+        """
+        # 优先使用累积模式 (offload)
+        if cls._layer_selected_counts:
+            total_selected = 0
+            total_total = 0
+            for layer_id in cls._layer_selected_counts:
+                total_selected += sum(cls._layer_selected_counts[layer_id])
+                total_total += sum(cls._layer_total_counts.get(layer_id, []))
+            if total_total > 0:
+                return total_selected / total_total
+            return 0.0
+
+        # 直接记录模式 (gpu_only)
+        all_densities = []
+        for densities in cls._layer_densities.values():
+            all_densities.extend(densities)
+        if not all_densities:
+            return 0.0
+        return sum(all_densities) / len(all_densities)
+
+    @classmethod
+    def get_overall_comm_density(cls) -> float:
+        """获取所有层的平均 communication density"""
+        all_densities = []
+        for densities in cls._layer_comm_densities.values():
+            all_densities.extend(densities)
+        if not all_densities:
+            return 0.0
+        return sum(all_densities) / len(all_densities)
+
+    @classmethod
+    def get_per_layer_comm_density(cls) -> Dict[int, float]:
+        """
+        获取每层的 communication density (CPU block 粒度)。
+
+        Returns:
+            Dict[layer_id, avg_comm_density]
+        """
+        result = {}
+        for layer_id, densities in cls._layer_comm_densities.items():
+            if densities:
+                result[layer_id] = sum(densities) / len(densities)
+        return result
+
+    @classmethod
+    def get_summary(cls) -> dict:
+        """返回统计摘要"""
+        per_layer = cls.get_per_layer_density()
+        per_layer_comm = cls.get_per_layer_comm_density()
+        return {
+            "mode": cls._mode,
+            "overall_compute_density": cls.get_overall_density(),
+            "overall_comm_density": cls.get_overall_comm_density(),
+            "per_layer_compute_density": per_layer,
+            "per_layer_comm_density": per_layer_comm,
+            "num_layers": len(per_layer),
+            "last_mask_shape": {
+                "q_blocks": cls._last_q_blocks,
+                "k_blocks": cls._last_k_blocks,
+            },
+        }
+
+    @classmethod
+    def get_min_density(cls) -> Tuple[int, float]:
+        """获取最低 density 的层和值"""
+        per_layer = cls.get_per_layer_density()
+        if not per_layer:
+            return -1, 0.0
+        min_layer = min(per_layer, key=per_layer.get)
+        return min_layer, per_layer[min_layer]
+
+    @classmethod
+    def print_summary(cls) -> None:
+        """打印人类可读的摘要"""
+        per_layer = cls.get_per_layer_density()
+        overall = cls.get_overall_density()
+        min_layer, min_density = cls.get_min_density()
+        overall_comm = cls.get_overall_comm_density()
+
+        print(f"[DensityObserver] Mode: {cls._mode}")
+        print(f"  Compute density: {overall:.4f} (min: {min_density:.4f} @ layer {min_layer})")
+        if overall_comm > 0:
+            # Offload mode: show both densities with explanation
+            print(f"  Comm density:    {overall_comm:.4f} (CPU block granularity)")
+            print(f"  Savings ratio:   {1 - overall_comm:.1%} H2D transfer reduction")
+        print(f"  Num layers: {len(per_layer)}")
+        # 输出 layer 0 的 density 用于对比
+        if 0 in per_layer:
+            print(f"  Layer 0 density: {per_layer[0]:.6f}")

nanovllm/utils/loader.py

@@ -1,4 +1,5 @@
 import os
+import re
 from glob import glob
 import torch
 from torch import nn
@@ -9,20 +10,146 @@ def default_weight_loader(param: nn.Parameter, loaded_weight: torch.Tensor):
     param.data.copy_(loaded_weight)
 
 
+# GLM-4 weight name mappings
+GLM4_NAME_MAPPING = {
+    "transformer.embedding.word_embeddings": "model.embed_tokens",
+    "transformer.encoder.final_layernorm": "model.norm",
+    "transformer.output_layer": "lm_head",
+}
+
+GLM4_LAYER_MAPPING = {
+    "self_attention.query_key_value": "self_attn.qkv_proj",
+    "self_attention.dense": "self_attn.o_proj",
+    "mlp.dense_h_to_4h": "mlp.gate_up_proj",
+    "mlp.dense_4h_to_h": "mlp.down_proj",
+}
+
+
+def convert_glm4_weight_name(weight_name: str) -> tuple[str, str | None]:
+    """
+    Convert GLM-4 weight name to nanovllm format.
+
+    Returns:
+        tuple: (converted_name, shard_id) where shard_id is used for packed modules
+               Returns (None, None) for weights that should be skipped
+    """
+    # Skip rotary embedding weights (we use our own RoPE implementation)
+    if "rotary_pos_emb" in weight_name:
+        return None, None
+
+    # Check direct mappings first
+    for glm_name, nano_name in GLM4_NAME_MAPPING.items():
+        if weight_name.startswith(glm_name):
+            return weight_name.replace(glm_name, nano_name), None
+
+    # Handle layer weights: transformer.encoder.layers.X.xxx
+    layer_match = re.match(r"transformer\.encoder\.layers\.(\d+)\.(.+)", weight_name)
+    if layer_match:
+        layer_idx = layer_match.group(1)
+        remainder = layer_match.group(2)
+
+        # Handle packed modules (QKV and gate_up)
+        for glm_subname, nano_subname in GLM4_LAYER_MAPPING.items():
+            if remainder.startswith(glm_subname):
+                suffix = remainder[len(glm_subname):]  # .weight or .bias
+                new_name = f"model.layers.{layer_idx}.{nano_subname}{suffix}"
+
+                # Determine shard_id for packed modules
+                if "qkv_proj" in nano_subname:
+                    return new_name, "qkv"  # Special marker for GLM4 QKV
+                elif "gate_up_proj" in nano_subname:
+                    return new_name, "gate_up"  # Special marker for GLM4 gate_up
+                else:
+                    return new_name, None
+
+        # Handle non-packed layer weights (layernorms)
+        new_name = f"model.layers.{layer_idx}.{remainder}"
+        return new_name, None
+
+    # No mapping found, return original
+    return weight_name, None
+
+
+def load_glm4_qkv(param: nn.Parameter, loaded_weight: torch.Tensor, config):
+    """Load GLM-4 merged QKV weights by splitting into q, k, v."""
+    num_heads = config.num_attention_heads
+    num_kv_heads = getattr(config, 'multi_query_group_num', num_heads)
+    head_dim = getattr(config, 'kv_channels', config.hidden_size // num_heads)
+
+    q_size = num_heads * head_dim
+    kv_size = num_kv_heads * head_dim
+
+    # Split QKV: [q_size + kv_size + kv_size, hidden_size]
+    q, k, v = loaded_weight.split([q_size, kv_size, kv_size], dim=0)
+
+    # Load each part using the weight_loader
+    weight_loader = getattr(param, "weight_loader")
+    weight_loader(param, q, "q")
+    weight_loader(param, k, "k")
+    weight_loader(param, v, "v")
+
+
+def load_glm4_gate_up(param: nn.Parameter, loaded_weight: torch.Tensor, config):
+    """Load GLM-4 merged gate_up weights by splitting into gate, up."""
+    ffn_hidden_size = getattr(config, 'ffn_hidden_size', getattr(config, 'intermediate_size', None))
+
+    # Split gate_up: [ffn_hidden_size * 2, hidden_size]
+    gate, up = loaded_weight.split([ffn_hidden_size, ffn_hidden_size], dim=0)
+
+    # Load each part using the weight_loader
+    weight_loader = getattr(param, "weight_loader")
+    weight_loader(param, gate, 0)  # gate_proj is shard 0
+    weight_loader(param, up, 1)    # up_proj is shard 1
+
+
+def is_glm4_model(model: nn.Module) -> bool:
+    """Check if the model is a GLM-4 model."""
+    return model.__class__.__name__ in ("ChatGLMForCausalLM",)
+
+
 def load_model(model: nn.Module, path: str):
     packed_modules_mapping = getattr(model, "packed_modules_mapping", {})
+    is_glm4 = is_glm4_model(model)
+    config = getattr(model, "config", None)
+
     for file in glob(os.path.join(path, "*.safetensors")):
         with safe_open(file, "pt", "cpu") as f:
             for weight_name in f.keys():
+                loaded_weight = f.get_tensor(weight_name)
+
+                # GLM-4 specific handling
+                if is_glm4:
+                    param_name, shard_id = convert_glm4_weight_name(weight_name)
+
+                    # Skip weights that don't need to be loaded
+                    if param_name is None:
+                        continue
+
+                    if shard_id == "qkv":
+                        param = model.get_parameter(param_name)
+                        load_glm4_qkv(param, loaded_weight, config)
+                        continue
+                    elif shard_id == "gate_up":
+                        param = model.get_parameter(param_name)
+                        load_glm4_gate_up(param, loaded_weight, config)
+                        continue
+                    else:
+                        # Regular weight, use converted name
+                        param = model.get_parameter(param_name)
+                        weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                        weight_loader(param, loaded_weight)
+                        continue
+
+                # Original loading logic for other models
                 for k in packed_modules_mapping:
                     if k in weight_name:
                         v, shard_id = packed_modules_mapping[k]
                         param_name = weight_name.replace(k, v)
                         param = model.get_parameter(param_name)
                         weight_loader = getattr(param, "weight_loader")
-                        weight_loader(param, f.get_tensor(weight_name), shard_id)
+                        weight_loader(param, loaded_weight, shard_id)
                         break
                 else:
                     param = model.get_parameter(weight_name)
                     weight_loader = getattr(param, "weight_loader", default_weight_loader)
-                    weight_loader(param, f.get_tensor(weight_name))
+                    weight_loader(param, loaded_weight)

nanovllm/utils/memory_observer.py

@@ -0,0 +1,133 @@
+"""
+MemoryObserver - 内存传输统计 Observer。
+
+统计 GPU-CPU 间的数据传输量：
+- H2D (Host to Device): CPU → GPU
+- D2H (Device to Host): GPU → CPU
+- D2D (Device to Device): GPU → GPU (buffer copy)
+"""
+
+from nanovllm.utils.observer import Observer
+
+
+class MemoryObserver(Observer):
+    """
+    内存传输 Observer，统计 GPU-CPU 间的数据传输量。
+
+    统计类型：
+    - H2D (Host to Device): CPU → GPU
+    - D2H (Device to Host): GPU → CPU
+    - D2D (Device to Device): GPU → GPU (buffer copy)
+
+    统计位置（均在 offload_engine.py）：
+    - H2D: load_to_slot_layer(), load_block_sample_from_cpu(), load_block_full_from_cpu()
+    - D2H: offload_slot_layer_to_cpu(), offload_prefill_buffer_async()
+    - D2D: write_to_prefill_buffer(), write_to_decode_buffer()
+    - 重置: llm_engine.py:generate() - 与 InferenceObserver 一起重置
+    """
+
+    _enabled: bool = False  # 默认禁用，需要显式启用
+
+    # H2D 统计
+    h2d_bytes: int = 0
+    h2d_count: int = 0
+
+    # D2H 统计
+    d2h_bytes: int = 0
+    d2h_count: int = 0
+
+    # D2D 统计
+    d2d_bytes: int = 0
+    d2d_count: int = 0
+
+    # 按阶段统计
+    prefill_h2d_bytes: int = 0
+    prefill_d2h_bytes: int = 0
+    decode_h2d_bytes: int = 0
+    decode_d2h_bytes: int = 0
+
+    @classmethod
+    def record_h2d(cls, num_bytes: int, is_prefill: bool = True) -> None:
+        """记录 H2D 传输"""
+        if not cls._enabled:
+            return
+        cls.h2d_bytes += num_bytes
+        cls.h2d_count += 1
+        if is_prefill:
+            cls.prefill_h2d_bytes += num_bytes
+        else:
+            cls.decode_h2d_bytes += num_bytes
+
+    @classmethod
+    def record_d2h(cls, num_bytes: int, is_prefill: bool = True) -> None:
+        """记录 D2H 传输"""
+        if not cls._enabled:
+            return
+        cls.d2h_bytes += num_bytes
+        cls.d2h_count += 1
+        if is_prefill:
+            cls.prefill_d2h_bytes += num_bytes
+        else:
+            cls.decode_d2h_bytes += num_bytes
+
+    @classmethod
+    def record_d2d(cls, num_bytes: int) -> None:
+        """记录 D2D 传输"""
+        if not cls._enabled:
+            return
+        cls.d2d_bytes += num_bytes
+        cls.d2d_count += 1
+
+    @classmethod
+    def complete_reset(cls) -> None:
+        """重置所有统计"""
+        cls.h2d_bytes = cls.h2d_count = 0
+        cls.d2h_bytes = cls.d2h_count = 0
+        cls.d2d_bytes = cls.d2d_count = 0
+        cls.prefill_h2d_bytes = cls.prefill_d2h_bytes = 0
+        cls.decode_h2d_bytes = cls.decode_d2h_bytes = 0
+
+    @classmethod
+    def get_summary(cls) -> dict:
+        """返回统计摘要"""
+        return {
+            "total": {
+                "h2d_bytes": cls.h2d_bytes,
+                "h2d_count": cls.h2d_count,
+                "d2h_bytes": cls.d2h_bytes,
+                "d2h_count": cls.d2h_count,
+                "d2d_bytes": cls.d2d_bytes,
+                "d2d_count": cls.d2d_count,
+            },
+            "prefill": {
+                "h2d_bytes": cls.prefill_h2d_bytes,
+                "d2h_bytes": cls.prefill_d2h_bytes,
+            },
+            "decode": {
+                "h2d_bytes": cls.decode_h2d_bytes,
+                "d2h_bytes": cls.decode_d2h_bytes,
+            },
+        }
+
+    @classmethod
+    def _fmt_bytes(cls, b: int) -> str:
+        """格式化字节数"""
+        if b >= 1e9:
+            return f"{b/1e9:.2f} GB"
+        if b >= 1e6:
+            return f"{b/1e6:.2f} MB"
+        if b >= 1e3:
+            return f"{b/1e3:.2f} KB"
+        return f"{b} B"
+
+    @classmethod
+    def print_summary(cls) -> None:
+        """打印人类可读的摘要"""
+        fmt = cls._fmt_bytes
+        total = cls.h2d_bytes + cls.d2h_bytes + cls.d2d_bytes
+        print(f"[MemoryObserver] Total: {fmt(total)}")
+        print(f"  H2D: {fmt(cls.h2d_bytes)} ({cls.h2d_count} ops)")
+        print(f"  D2H: {fmt(cls.d2h_bytes)} ({cls.d2h_count} ops)")
+        print(f"  D2D: {fmt(cls.d2d_bytes)} ({cls.d2d_count} ops)")
+        print(f"  Prefill - H2D: {fmt(cls.prefill_h2d_bytes)}, D2H: {fmt(cls.prefill_d2h_bytes)}")
+        print(f"  Decode  - H2D: {fmt(cls.decode_h2d_bytes)}, D2H: {fmt(cls.decode_d2h_bytes)}")

nanovllm/utils/observer.py

@@ -1,17 +1,106 @@
-class Observer():
-    ttft_start = 0
-    tpot_start = 0
+"""
+Observer 基类和 InferenceObserver 实现。
 
-    ttft = 0
-    tpot = 0
+Observer 架构：
+- Observer: 基类，定义通用接口
+- InferenceObserver: 推理性能观测（TTFT/TPOT）
+- MemoryObserver: 内存传输观测（在 memory_observer.py 中定义）
+"""
+
+
+class Observer:
+    """
+    Observer 基类，提供通用的启用/禁用、重置、输出接口。
+
+    所有 Observer 子类应继承此类并实现：
+    - complete_reset(): 重置所有统计数据
+    - get_summary(): 返回统计摘要 dict
+    - print_summary(): 打印人类可读的摘要
+    """
+
+    _enabled: bool = True  # 默认启用
 
     @classmethod
-    def reset_ttft(cls):
+    def enable(cls) -> None:
+        """启用 observer"""
+        cls._enabled = True
+
+    @classmethod
+    def disable(cls) -> None:
+        """禁用 observer"""
+        cls._enabled = False
+
+    @classmethod
+    def is_enabled(cls) -> bool:
+        """检查是否启用"""
+        return cls._enabled
+
+    @classmethod
+    def complete_reset(cls) -> None:
+        """重置所有统计数据（子类实现）"""
+        raise NotImplementedError
+
+    @classmethod
+    def get_summary(cls) -> dict:
+        """返回统计摘要（子类实现）"""
+        raise NotImplementedError
+
+    @classmethod
+    def print_summary(cls) -> None:
+        """打印人类可读的摘要（子类可选覆盖）"""
+        import json
+        print(json.dumps(cls.get_summary(), indent=2))
+
+
+class InferenceObserver(Observer):
+    """
+    推理性能 Observer，统计 TTFT 和 TPOT。
+
+    - TTFT (Time To First Token): 首个 token 生成延迟
+    - TPOT (Time Per Output Token): 每个输出 token 的平均延迟
+
+    统计位置：
+    - TTFT 开始: scheduler.py:35-36 - 第一个 sequence 从 waiting 队列取出时
+    - TTFT 结束: llm_engine.py:69-72 - prefill 完成后（包括 chunked prefill 所有 chunks）
+    - TPOT 开始: llm_engine.py:65 - 每次 decode step 结束时
+    - TPOT 结束: llm_engine.py:62-63 - 下一次 decode step 开始时计算（测量上一次 decode 时间）
+    - 重置: llm_engine.py:97 - generate() 开始时
+
+    注意：TPOT 需要至少 2 个输出 token 才能计算（测量 decode step 间隔）。
+    """
+
+    # 时间戳 (nanoseconds)
+    ttft_start: int = 0
+    tpot_start: int = 0
+
+    # 统计结果 (nanoseconds)
+    ttft: int = 0
+    tpot: int = 0
+
+    @classmethod
+    def reset_ttft(cls) -> None:
+        """重置 TTFT 计时器"""
         cls.ttft_start = 0
 
     @classmethod
-    def complete_reset(cls):
+    def complete_reset(cls) -> None:
+        """重置所有统计数据"""
         cls.ttft_start = 0
         cls.tpot_start = 0
         cls.ttft = 0
         cls.tpot = 0
+
+    @classmethod
+    def get_summary(cls) -> dict:
+        """返回统计摘要"""
+        return {
+            "ttft_ns": cls.ttft,
+            "ttft_ms": cls.ttft / 1e6,
+            "tpot_ns": cls.tpot,
+            "tpot_ms": cls.tpot / 1e6,
+        }
+
+    @classmethod
+    def print_summary(cls) -> None:
+        """打印摘要"""
+        print(f"[InferenceObserver] TTFT: {cls.ttft / 1e6:.2f}ms, TPOT: {cls.tpot / 1e6:.2f}ms")

scripts/profile.sh

@@ -0,0 +1,158 @@
+#!/bin/bash
+
+# Profile bench.py using NVIDIA Nsight Systems (GPU-only mode)
+#
+# Usage:
+#   bash scripts/profile.sh [options]
+#
+# Options:
+#   --max-len LENGTH     Max sequence length (default: 32768)
+#   --policy POLICY      Sparse policy: full, xattn (default: xattn)
+#   --gpu GPU_ID         GPU to use (default: 0)
+#   --gpu-util UTIL      GPU memory utilization (default: 0.9)
+#   --input-len LENGTH   Input length (default: max-len - 1)
+#   --bench-decode       Run decode benchmark instead of prefill
+#
+# Output:
+#   results/nsys/bench_<policy>_<max_len>_<timestamp>.nsys-rep
+#
+# Examples:
+#   bash scripts/profile.sh
+#   bash scripts/profile.sh --max-len 65536 --gpu-util 0.7
+#   bash scripts/profile.sh --policy full --max-len 32768
+#   bash scripts/profile.sh --bench-decode
+
+set -e
+
+# Default configuration
+MAX_LEN="32768"
+POLICY="xattn"
+GPU_ID="0"
+GPU_UTIL="0.9"
+INPUT_LEN=""
+BENCH_MODE="prefill"
+
+# Parse arguments
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --max-len)
+            MAX_LEN="$2"
+            shift 2
+            ;;
+        --policy)
+            POLICY="$2"
+            shift 2
+            ;;
+        --gpu)
+            GPU_ID="$2"
+            shift 2
+            ;;
+        --gpu-util)
+            GPU_UTIL="$2"
+            shift 2
+            ;;
+        --input-len)
+            INPUT_LEN="$2"
+            shift 2
+            ;;
+        --bench-decode)
+            BENCH_MODE="decode"
+            shift
+            ;;
+        -h|--help)
+            echo "Usage: $0 [options]"
+            echo ""
+            echo "Options:"
+            echo "  --max-len LENGTH     Max sequence length (default: 32768)"
+            echo "  --policy POLICY      Sparse policy: full, xattn (default: xattn)"
+            echo "  --gpu GPU_ID         GPU to use (default: 0)"
+            echo "  --gpu-util UTIL      GPU memory utilization (default: 0.9)"
+            echo "  --input-len LENGTH   Input length (default: max-len - 1)"
+            echo "  --bench-decode       Run decode benchmark instead of prefill"
+            exit 0
+            ;;
+        *)
+            echo "Unknown option: $1"
+            exit 1
+            ;;
+    esac
+done
+
+# Path configuration
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(dirname "$SCRIPT_DIR")"
+OUTPUT_DIR="$PROJECT_ROOT/results/nsys"
+BENCH_SCRIPT="$PROJECT_ROOT/bench.py"
+
+# Create output directory if needed
+mkdir -p "$OUTPUT_DIR"
+
+# Generate timestamp for unique filename
+TIMESTAMP=$(date +%Y%m%d_%H%M%S)
+
+# Convert max_len to human-readable format (e.g., 32768 -> 32k)
+if [ "$MAX_LEN" -ge 1024 ]; then
+    MAX_LEN_SUFFIX="$((MAX_LEN / 1024))k"
+else
+    MAX_LEN_SUFFIX="${MAX_LEN}"
+fi
+
+OUTPUT_FILE="$OUTPUT_DIR/bench_${POLICY}_${MAX_LEN_SUFFIX}_${BENCH_MODE}_${TIMESTAMP}"
+
+# Build bench.py arguments
+BENCH_ARGS="--max-len $MAX_LEN --gpu-util $GPU_UTIL"
+
+if [ -n "$POLICY" ]; then
+    BENCH_ARGS="$BENCH_ARGS --policy $POLICY"
+fi
+
+if [ -n "$INPUT_LEN" ]; then
+    BENCH_ARGS="$BENCH_ARGS --input-len $INPUT_LEN"
+fi
+
+if [ "$BENCH_MODE" = "decode" ]; then
+    BENCH_ARGS="$BENCH_ARGS --bench-decode"
+fi
+
+echo "============================================================"
+echo "NVIDIA Nsight Systems Profiling (GPU-only)"
+echo "============================================================"
+echo "Bench script: $BENCH_SCRIPT"
+echo "Policy:       $POLICY"
+echo "Max length:   $MAX_LEN"
+echo "GPU:          $GPU_ID"
+echo "GPU util:     $GPU_UTIL"
+echo "Bench mode:   $BENCH_MODE"
+echo "Output file:  $OUTPUT_FILE.nsys-rep"
+echo ""
+
+# nsys profile options:
+# --trace=cuda,nvtx : Trace CUDA API and NVTX markers
+# --force-overwrite=true : Overwrite existing output file
+# --output=<path> : Output file path (without .nsys-rep extension)
+
+echo "Running nsys profile..."
+echo "Command: python bench.py $BENCH_ARGS"
+echo ""
+
+CUDA_VISIBLE_DEVICES=$GPU_ID PYTHONPATH="$PROJECT_ROOT:$PYTHONPATH" \
+nsys profile \
+    --trace=cuda,nvtx \
+    --force-overwrite=true \
+    --output="$OUTPUT_FILE" \
+    python "$BENCH_SCRIPT" $BENCH_ARGS
+
+echo ""
+echo "============================================================"
+echo "Profiling completed successfully!"
+echo "============================================================"
+echo "Output file: $OUTPUT_FILE.nsys-rep"
+echo ""
+echo "To view results in GUI:"
+echo "  nsight-sys $OUTPUT_FILE.nsys-rep"
+echo ""
+echo "To export statistics:"
+echo "  nsys stats --report cuda_api_sum $OUTPUT_FILE.nsys-rep"
+echo "  nsys stats --report cuda_gpu_kern_sum $OUTPUT_FILE.nsys-rep"
+echo "  nsys stats --report cuda_gpu_mem_size_sum $OUTPUT_FILE.nsys-rep"
+echo "============================================================"

scripts/profile_offload.sh

@@ -6,33 +6,47 @@
 #   bash scripts/profile_offload.sh [options]
 #
 # Options:
+#   --policy POLICY      Sparse policy name (default: full)
+#   --ctx-len LENGTH     Context length: 32k, 64k, 128k (default: 64k)
 #   --dataset DATASET    Task name (default: niah_single_1)
 #   --sample INDEX       Sample index (default: 0)
 #   --gpu GPU_ID         GPU to use (default: 0)
 #   --num-gpu-blocks N   Number of GPU blocks/slots (default: 4)
+#   --block-size SIZE    KV cache block size (default: 4096)
 #   --no-offload         Disable CPU offload
 #
 # Output:
-#   results/nsys/ruler_<dataset>_sample<index>_<timestamp>.nsys-rep
+#   results/nsys/<policy>_<gpuonly|offload>_<ctx-len>_blk<size>_<timestamp>.nsys-rep
 #
 # Examples:
 #   bash scripts/profile_offload.sh
-#   bash scripts/profile_offload.sh --dataset niah_single_1 --sample 5
-#   bash scripts/profile_offload.sh --gpu 1 --no-offload
-#   bash scripts/profile_offload.sh --num-gpu-blocks 8
-
-set -e
+#   bash scripts/profile_offload.sh --policy xattn --ctx-len 128k --no-offload
+#   bash scripts/profile_offload.sh --policy full --ctx-len 32k --num-gpu-blocks 8
 
 # Default configuration
+POLICY="full"
+CTX_LEN="64k"
 DATASET="niah_single_1"
 SAMPLE_INDEX="0"
 GPU_ID="0"
 NUM_GPU_BLOCKS="4"
+BLOCK_SIZE="4096"
+GPU_UTIL="0.9"
 ENABLE_OFFLOAD="--enable-offload"
+MODEL=""
+DATA_DIR_OVERRIDE=""
 
 # Parse arguments
 while [[ $# -gt 0 ]]; do
     case $1 in
+        --policy)
+            POLICY="$2"
+            shift 2
+            ;;
+        --ctx-len)
+            CTX_LEN="$2"
+            shift 2
+            ;;
         --dataset)
             DATASET="$2"
             shift 2
@@ -53,13 +67,33 @@ while [[ $# -gt 0 ]]; do
             NUM_GPU_BLOCKS="$2"
             shift 2
             ;;
+        --gpu-util)
+            GPU_UTIL="$2"
+            shift 2
+            ;;
+        --block-size)
+            BLOCK_SIZE="$2"
+            shift 2
+            ;;
+        --model)
+            MODEL="$2"
+            shift 2
+            ;;
+        --data-dir)
+            DATA_DIR_OVERRIDE="$2"
+            shift 2
+            ;;
         -h|--help)
             echo "Usage: $0 [options]"
             echo ""
             echo "Options:"
+            echo "  --policy POLICY      Sparse policy name (default: full)"
+            echo "  --ctx-len LENGTH     Context length: 32k, 64k, 128k (default: 64k)"
+            echo "  --block-size SIZE    KV cache block size (default: 4096)"
             echo "  --dataset DATASET    Task name (default: niah_single_1)"
             echo "  --sample INDEX       Sample index (default: 0)"
             echo "  --gpu GPU_ID         GPU to use (default: 0)"
+            echo "  --gpu-util UTIL      GPU memory utilization (default: 0.9)"
             echo "  --no-offload         Disable CPU offload"
             echo "  --num-gpu-blocks N   Number of GPU blocks/slots (default: 4)"
             exit 0
@@ -76,27 +110,62 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 PROJECT_ROOT="$(dirname "$SCRIPT_DIR")"
 OUTPUT_DIR="$PROJECT_ROOT/results/nsys"
 TEST_SCRIPT="$PROJECT_ROOT/tests/test_ruler.py"
+DATA_DIR="$PROJECT_ROOT/tests/data/ruler_${CTX_LEN}"
+
+# Set max-model-len based on context length
+case "$CTX_LEN" in
+    32k)
+        MAX_MODEL_LEN=36000
+        ;;
+    64k)
+        MAX_MODEL_LEN=72000
+        ;;
+    128k)
+        MAX_MODEL_LEN=144000
+        ;;
+    256k)
+        MAX_MODEL_LEN=288000
+        ;;
+    512k)
+        MAX_MODEL_LEN=576000
+        ;;
+    1m)
+        MAX_MODEL_LEN=1100000
+        ;;
+    *)
+        MAX_MODEL_LEN=72000
+        ;;
+esac
+
+# Override DATA_DIR if specified
+if [ -n "$DATA_DIR_OVERRIDE" ]; then
+    DATA_DIR="$DATA_DIR_OVERRIDE"
+fi
 
 # Create output directory if needed
 mkdir -p "$OUTPUT_DIR"
 
 # Generate timestamp for unique filename
 TIMESTAMP=$(date +%Y%m%d_%H%M%S)
-OFFLOAD_SUFFIX=""
 if [ -n "$ENABLE_OFFLOAD" ]; then
-    OFFLOAD_SUFFIX="_offload_${NUM_GPU_BLOCKS}slots"
+    OFFLOAD_TAG="offload"
+else
+    OFFLOAD_TAG="gpuonly"
 fi
-OUTPUT_FILE="$OUTPUT_DIR/ruler_${DATASET}_sample${SAMPLE_INDEX}${OFFLOAD_SUFFIX}_${TIMESTAMP}"
+OUTPUT_FILE="$OUTPUT_DIR/${POLICY}_${OFFLOAD_TAG}_${CTX_LEN}_blk${BLOCK_SIZE}_${TIMESTAMP}"
 
 echo "============================================================"
 echo "NVIDIA Nsight Systems Profiling"
 echo "============================================================"
-echo "Test script: $TEST_SCRIPT"
+echo "Policy:      $POLICY"
+echo "Offload:     $OFFLOAD_TAG"
+echo "Context:     $CTX_LEN"
+echo "Block Size:  $BLOCK_SIZE"
 echo "Dataset:     $DATASET"
 echo "Sample:      $SAMPLE_INDEX"
 echo "GPU:         $GPU_ID"
 echo "GPU Blocks:  $NUM_GPU_BLOCKS"
-echo "Offload:     ${ENABLE_OFFLOAD:-disabled}"
+echo "Data Dir:    $DATA_DIR"
 echo "Output file: $OUTPUT_FILE.nsys-rep"
 echo ""
 
@@ -110,17 +179,59 @@ echo ""
 echo "Running nsys profile..."
 echo ""
 
+# Map policy name to internal enum name
+# User-friendly name -> SparsePolicyType enum name
+case "$POLICY" in
+    xattn)
+        POLICY_ENUM="XATTN_BSA"
+        ;;
+    *)
+        POLICY_ENUM="$POLICY"
+        ;;
+esac
+
+# Build sparse policy argument
+SPARSE_POLICY_ARG=""
+if [ -n "$POLICY_ENUM" ] && [ "$POLICY_ENUM" != "full" ]; then
+    SPARSE_POLICY_ARG="--sparse-policy $POLICY_ENUM"
+fi
+
+# Build model argument
+MODEL_ARG=""
+if [ -n "$MODEL" ]; then
+    MODEL_ARG="--model $MODEL"
+fi
+
+# Run nsys profile and capture exit code
 CUDA_VISIBLE_DEVICES=$GPU_ID PYTHONPATH="$PROJECT_ROOT:$PYTHONPATH" \
 nsys profile \
     --trace=cuda,nvtx \
     --force-overwrite=true \
     --output="$OUTPUT_FILE" \
     python "$TEST_SCRIPT" \
+        --data-dir "$DATA_DIR" \
         --datasets "$DATASET" \
         --sample-indices "$SAMPLE_INDEX" \
         --num-gpu-blocks "$NUM_GPU_BLOCKS" \
+        --block-size "$BLOCK_SIZE" \
+        --max-model-len "$MAX_MODEL_LEN" \
+        --gpu-utilization "$GPU_UTIL" \
         $ENABLE_OFFLOAD \
+        $SPARSE_POLICY_ARG \
+        $MODEL_ARG \
         --quiet
+EXIT_CODE=$?
+
+# If test failed, delete the output file
+if [ $EXIT_CODE -ne 0 ]; then
+    echo ""
+    echo "============================================================"
+    echo "Test FAILED! Cleaning up..."
+    echo "============================================================"
+    rm -f "$OUTPUT_FILE.nsys-rep"
+    echo "Deleted: $OUTPUT_FILE.nsys-rep"
+    exit $EXIT_CODE
+fi
 
 echo ""
 echo "============================================================"

tests/modeling_qwen3.py

@@ -1,757 +0,0 @@
-"""
-Custom Qwen3 implementation using only torch and transformers.
-This file provides a clean reference implementation for understanding the model computation graph.
-
-Computation Graph:
-==================
-
-Input: token_ids [batch, seq_len]
-         │
-         ▼
-    ┌─────────────┐
-    │  Embedding  │  embed_tokens: [vocab_size, hidden_size]
-    └─────────────┘
-         │
-         ▼
-    hidden_states [batch, seq_len, hidden_size]
-         │
-         ▼
-    ┌─────────────────────────────────────────────────────────┐
-    │                   Decoder Layer (x N)                   │
-    │  ┌───────────────────────────────────────────────────┐  │
-    │  │              Self Attention Block                 │  │
-    │  │                                                   │  │
-    │  │  input_layernorm (RMSNorm)                        │  │
-    │  │         │                                         │  │
-    │  │         ▼                                         │  │
-    │  │  ┌─────────────────────────────────────────────┐  │  │
-    │  │  │            Qwen3Attention                   │  │  │
-    │  │  │  Q = q_proj(x) → q_norm → reshape           │  │  │
-    │  │  │  K = k_proj(x) → k_norm → reshape           │  │  │
-    │  │  │  V = v_proj(x) → reshape                    │  │  │
-    │  │  │         │                                   │  │  │
-    │  │  │         ▼                                   │  │  │
-    │  │  │  Q, K = apply_rotary_pos_emb(Q, K, cos, sin)│  │  │
-    │  │  │         │                                   │  │  │
-    │  │  │         ▼                                   │  │  │
-    │  │  │  attn_output = attention(Q, K, V)           │  │  │
-    │  │  │         │                                   │  │  │
-    │  │  │         ▼                                   │  │  │
-    │  │  │  output = o_proj(attn_output)               │  │  │
-    │  │  └─────────────────────────────────────────────┘  │  │
-    │  │         │                                         │  │
-    │  │         ▼                                         │  │
-    │  │  hidden_states = residual + attn_output           │  │
-    │  └───────────────────────────────────────────────────┘  │
-    │                        │                                │
-    │                        ▼                                │
-    │  ┌───────────────────────────────────────────────────┐  │
-    │  │                  MLP Block                        │  │
-    │  │                                                   │  │
-    │  │  post_attention_layernorm (RMSNorm)               │  │
-    │  │         │                                         │  │
-    │  │         ▼                                         │  │
-    │  │  ┌─────────────────────────────────────────────┐  │  │
-    │  │  │              Qwen3MLP                       │  │  │
-    │  │  │  gate = gate_proj(x)                        │  │  │
-    │  │  │  up = up_proj(x)                            │  │  │
-    │  │  │  output = down_proj(silu(gate) * up)        │  │  │
-    │  │  └─────────────────────────────────────────────┘  │  │
-    │  │         │                                         │  │
-    │  │         ▼                                         │  │
-    │  │  hidden_states = residual + mlp_output            │  │
-    │  └───────────────────────────────────────────────────┘  │
-    └─────────────────────────────────────────────────────────┘
-         │
-         ▼
-    ┌─────────────┐
-    │    norm     │  final RMSNorm
-    └─────────────┘
-         │
-         ▼
-    ┌─────────────┐
-    │   lm_head   │  [hidden_size, vocab_size]
-    └─────────────┘
-         │
-         ▼
-    logits [batch, seq_len, vocab_size]
-"""
-
-import math
-from typing import Optional, Tuple, List
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-
-class Qwen3RMSNorm(nn.Module):
-    """RMSNorm implementation."""
-
-    def __init__(self, hidden_size: int, eps: float = 1e-6):
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.eps = eps
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        input_dtype = x.dtype
-        x = x.float()
-        variance = x.pow(2).mean(-1, keepdim=True)
-        x = x * torch.rsqrt(variance + self.eps)
-        return self.weight * x.to(input_dtype)
-
-
-class Qwen3RotaryEmbedding(nn.Module):
-    """Rotary Position Embedding (RoPE)."""
-
-    def __init__(self, dim: int, max_position_embeddings: int = 32768, base: float = 10000.0):
-        super().__init__()
-        self.dim = dim
-        self.max_position_embeddings = max_position_embeddings
-        self.base = base
-
-        # Compute inverse frequencies
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-
-    @torch.no_grad()
-    def forward(self, x: torch.Tensor, position_ids: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        """
-        Args:
-            x: Input tensor [batch, seq_len, num_heads, head_dim] or similar
-            position_ids: Position indices [batch, seq_len]
-
-        Returns:
-            cos, sin: [batch, seq_len, head_dim]
-        """
-        # inv_freq: [dim/2]
-        # position_ids: [batch, seq_len]
-        inv_freq_expanded = self.inv_freq[None, :, None].float()  # [1, dim/2, 1]
-        position_ids_expanded = position_ids[:, None, :].float()  # [batch, 1, seq_len]
-
-        # freqs: [batch, dim/2, seq_len]
-        freqs = inv_freq_expanded @ position_ids_expanded
-        # freqs: [batch, seq_len, dim/2]
-        freqs = freqs.transpose(1, 2)
-
-        # Duplicate for full head_dim: [batch, seq_len, dim]
-        emb = torch.cat((freqs, freqs), dim=-1)
-
-        cos = emb.cos().to(x.dtype)
-        sin = emb.sin().to(x.dtype)
-
-        return cos, sin
-
-
-def rotate_half(x: torch.Tensor) -> torch.Tensor:
-    """Rotate half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Apply rotary position embeddings to Q and K.
-
-    Args:
-        q: [batch, num_heads, seq_len, head_dim]
-        k: [batch, num_kv_heads, seq_len, head_dim]
-        cos: [batch, seq_len, head_dim]
-        sin: [batch, seq_len, head_dim]
-
-    Returns:
-        q_embed, k_embed with same shapes as inputs
-    """
-    # Unsqueeze for broadcasting: [batch, 1, seq_len, head_dim]
-    cos = cos.unsqueeze(1)
-    sin = sin.unsqueeze(1)
-
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-
-    return q_embed, k_embed
-
-
-class Qwen3Attention(nn.Module):
-    """
-    Qwen3 Multi-Head Attention with Grouped Query Attention (GQA) support.
-
-    Data Flow:
-    ---------
-    hidden_states [batch, seq_len, hidden_size]
-            │
-            ├──► q_proj ──► q_norm ──► reshape ──► Q [batch, num_heads, seq_len, head_dim]
-            ├──► k_proj ──► k_norm ──► reshape ──► K [batch, num_kv_heads, seq_len, head_dim]
-            └──► v_proj ──► reshape ──► V [batch, num_kv_heads, seq_len, head_dim]
-                    │
-                    ▼
-            apply_rotary_pos_emb(Q, K)
-                    │
-                    ▼
-            attention(Q, K, V) ──► attn_output [batch, num_heads, seq_len, head_dim]
-                    │
-                    ▼
-            reshape ──► o_proj ──► output [batch, seq_len, hidden_size]
-    """
-
-    def __init__(
-        self,
-        hidden_size: int,
-        num_attention_heads: int,
-        num_key_value_heads: int,
-        head_dim: int,
-        max_position_embeddings: int = 32768,
-        rope_theta: float = 10000.0,
-        attention_bias: bool = False,
-        rms_norm_eps: float = 1e-6,
-        layer_idx: int = 0,
-    ):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.num_heads = num_attention_heads
-        self.num_kv_heads = num_key_value_heads
-        self.head_dim = head_dim
-        self.num_kv_groups = num_attention_heads // num_key_value_heads
-        self.layer_idx = layer_idx
-
-        # Scaling factor
-        self.scaling = head_dim ** -0.5
-
-        # QKV projections
-        self.q_proj = nn.Linear(hidden_size, num_attention_heads * head_dim, bias=attention_bias)
-        self.k_proj = nn.Linear(hidden_size, num_key_value_heads * head_dim, bias=attention_bias)
-        self.v_proj = nn.Linear(hidden_size, num_key_value_heads * head_dim, bias=attention_bias)
-        self.o_proj = nn.Linear(num_attention_heads * head_dim, hidden_size, bias=attention_bias)
-
-        # QK normalization (Qwen3 specific)
-        self.q_norm = Qwen3RMSNorm(head_dim, eps=rms_norm_eps)
-        self.k_norm = Qwen3RMSNorm(head_dim, eps=rms_norm_eps)
-
-        # Rotary embeddings
-        self.rotary_emb = Qwen3RotaryEmbedding(
-            head_dim,
-            max_position_embeddings=max_position_embeddings,
-            base=rope_theta,
-        )
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        position_ids: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        use_cache: bool = False,
-        output_qkv: bool = False,
-    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]], Optional[dict]]:
-        """
-        Args:
-            hidden_states: [batch, seq_len, hidden_size]
-            position_ids: [batch, seq_len]
-            attention_mask: [batch, 1, seq_len, kv_seq_len] (causal mask)
-            past_key_value: (k_cache, v_cache) from previous steps
-            use_cache: Whether to return updated cache
-            output_qkv: Whether to output Q, K, V tensors for debugging
-
-        Returns:
-            output: [batch, seq_len, hidden_size]
-            past_key_value: Updated cache (if use_cache=True)
-            qkv_dict: {"q": Q, "k": K, "v": V} (if output_qkv=True)
-        """
-        batch_size, seq_len, _ = hidden_states.shape
-
-        # === QKV Projections ===
-        q = self.q_proj(hidden_states)  # [batch, seq_len, num_heads * head_dim]
-        k = self.k_proj(hidden_states)  # [batch, seq_len, num_kv_heads * head_dim]
-        v = self.v_proj(hidden_states)  # [batch, seq_len, num_kv_heads * head_dim]
-
-        # Reshape to [batch, seq_len, num_heads, head_dim]
-        q = q.view(batch_size, seq_len, self.num_heads, self.head_dim)
-        k = k.view(batch_size, seq_len, self.num_kv_heads, self.head_dim)
-        v = v.view(batch_size, seq_len, self.num_kv_heads, self.head_dim)
-
-        # === QK Normalization (Qwen3 specific) ===
-        q = self.q_norm(q)
-        k = self.k_norm(k)
-
-        # Transpose to [batch, num_heads, seq_len, head_dim]
-        q = q.transpose(1, 2)
-        k = k.transpose(1, 2)
-        v = v.transpose(1, 2)
-
-        # === Rotary Position Embeddings ===
-        cos, sin = self.rotary_emb(v, position_ids)
-        q, k = apply_rotary_pos_emb(q, k, cos, sin)
-
-        # === KV Cache Update ===
-        if past_key_value is not None:
-            k_cache, v_cache = past_key_value
-            k = torch.cat([k_cache, k], dim=2)
-            v = torch.cat([v_cache, v], dim=2)
-
-        new_past_key_value = (k, v) if use_cache else None
-
-        # === Grouped Query Attention (expand KV heads if needed) ===
-        if self.num_kv_groups > 1:
-            # Repeat KV for each query group
-            k = k.repeat_interleave(self.num_kv_groups, dim=1)
-            v = v.repeat_interleave(self.num_kv_groups, dim=1)
-
-        # === Attention Computation (using SDPA for memory efficiency) ===
-        # Use PyTorch's scaled_dot_product_attention which can use FlashAttention backend
-        # is_causal only works when q_len == kv_len (prefill), not during decode
-        q_len, kv_len = q.shape[2], k.shape[2]
-        is_causal = (q_len == kv_len) and (q_len > 1)
-
-        attn_output = F.scaled_dot_product_attention(
-            q, k, v,
-            attn_mask=None,
-            dropout_p=0.0,
-            is_causal=is_causal,
-            scale=self.scaling,
-        )  # [batch, num_heads, seq_len, head_dim]
-
-        # === Output Projection ===
-        # Transpose back and reshape
-        attn_output = attn_output.transpose(1, 2).contiguous()  # [batch, seq_len, num_heads, head_dim]
-        attn_output = attn_output.view(batch_size, seq_len, -1)  # [batch, seq_len, hidden_size]
-        output = self.o_proj(attn_output)
-
-        # Optional QKV output for debugging
-        qkv_dict = None
-        if output_qkv:
-            qkv_dict = {
-                "q": q,  # [batch, num_heads, seq_len, head_dim] (post-RoPE)
-                "k": k,  # [batch, num_heads, kv_seq_len, head_dim] (post-RoPE, expanded)
-                "v": v,  # [batch, num_heads, kv_seq_len, head_dim] (expanded)
-            }
-
-        return output, new_past_key_value, qkv_dict
-
-
-class Qwen3MLP(nn.Module):
-    """
-    Qwen3 MLP with SwiGLU activation.
-
-    Data Flow:
-    ---------
-    hidden_states [batch, seq_len, hidden_size]
-            │
-            ├──► gate_proj ──► gate [batch, seq_len, intermediate_size]
-            │
-            └──► up_proj ──► up [batch, seq_len, intermediate_size]
-                    │
-                    ▼
-            silu(gate) * up
-                    │
-                    ▼
-            down_proj ──► output [batch, seq_len, hidden_size]
-    """
-
-    def __init__(self, hidden_size: int, intermediate_size: int, bias: bool = False):
-        super().__init__()
-        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=bias)
-        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=bias)
-        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=bias)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        gate = self.gate_proj(x)
-        up = self.up_proj(x)
-        return self.down_proj(F.silu(gate) * up)
-
-
-class Qwen3DecoderLayer(nn.Module):
-    """Single Qwen3 Decoder Layer."""
-
-    def __init__(
-        self,
-        hidden_size: int,
-        intermediate_size: int,
-        num_attention_heads: int,
-        num_key_value_heads: int,
-        head_dim: int,
-        max_position_embeddings: int = 32768,
-        rope_theta: float = 10000.0,
-        rms_norm_eps: float = 1e-6,
-        attention_bias: bool = False,
-        mlp_bias: bool = False,
-        layer_idx: int = 0,
-    ):
-        super().__init__()
-        self.layer_idx = layer_idx
-
-        # Pre-attention LayerNorm
-        self.input_layernorm = Qwen3RMSNorm(hidden_size, eps=rms_norm_eps)
-
-        # Self-attention
-        self.self_attn = Qwen3Attention(
-            hidden_size=hidden_size,
-            num_attention_heads=num_attention_heads,
-            num_key_value_heads=num_key_value_heads,
-            head_dim=head_dim,
-            max_position_embeddings=max_position_embeddings,
-            rope_theta=rope_theta,
-            attention_bias=attention_bias,
-            rms_norm_eps=rms_norm_eps,
-            layer_idx=layer_idx,
-        )
-
-        # Post-attention LayerNorm
-        self.post_attention_layernorm = Qwen3RMSNorm(hidden_size, eps=rms_norm_eps)
-
-        # MLP
-        self.mlp = Qwen3MLP(hidden_size, intermediate_size, bias=mlp_bias)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        position_ids: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        use_cache: bool = False,
-        output_qkv: bool = False,
-    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]], Optional[dict]]:
-        """
-        Args:
-            hidden_states: [batch, seq_len, hidden_size]
-            position_ids: [batch, seq_len]
-            attention_mask: Causal attention mask
-            past_key_value: KV cache for this layer
-            use_cache: Whether to return updated cache
-            output_qkv: Whether to output Q, K, V for debugging
-
-        Returns:
-            hidden_states: [batch, seq_len, hidden_size]
-            past_key_value: Updated cache
-            qkv_dict: QKV tensors (if output_qkv=True)
-        """
-        # === Self Attention Block ===
-        residual = hidden_states
-        hidden_states = self.input_layernorm(hidden_states)
-
-        attn_output, new_past_key_value, qkv_dict = self.self_attn(
-            hidden_states=hidden_states,
-            position_ids=position_ids,
-            attention_mask=attention_mask,
-            past_key_value=past_key_value,
-            use_cache=use_cache,
-            output_qkv=output_qkv,
-        )
-
-        hidden_states = residual + attn_output
-
-        # === MLP Block ===
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        return hidden_states, new_past_key_value, qkv_dict
-
-
-class Qwen3Model(nn.Module):
-    """Qwen3 Transformer Model (without LM head)."""
-
-    def __init__(
-        self,
-        vocab_size: int,
-        hidden_size: int,
-        intermediate_size: int,
-        num_hidden_layers: int,
-        num_attention_heads: int,
-        num_key_value_heads: int,
-        head_dim: int,
-        max_position_embeddings: int = 32768,
-        rope_theta: float = 10000.0,
-        rms_norm_eps: float = 1e-6,
-        attention_bias: bool = False,
-        mlp_bias: bool = False,
-    ):
-        super().__init__()
-        self.vocab_size = vocab_size
-        self.num_hidden_layers = num_hidden_layers
-
-        # Token embeddings
-        self.embed_tokens = nn.Embedding(vocab_size, hidden_size)
-
-        # Decoder layers
-        self.layers = nn.ModuleList([
-            Qwen3DecoderLayer(
-                hidden_size=hidden_size,
-                intermediate_size=intermediate_size,
-                num_attention_heads=num_attention_heads,
-                num_key_value_heads=num_key_value_heads,
-                head_dim=head_dim,
-                max_position_embeddings=max_position_embeddings,
-                rope_theta=rope_theta,
-                rms_norm_eps=rms_norm_eps,
-                attention_bias=attention_bias,
-                mlp_bias=mlp_bias,
-                layer_idx=i,
-            )
-            for i in range(num_hidden_layers)
-        ])
-
-        # Final LayerNorm
-        self.norm = Qwen3RMSNorm(hidden_size, eps=rms_norm_eps)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        position_ids: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
-        use_cache: bool = False,
-        output_qkv_layers: Optional[List[int]] = None,
-    ) -> Tuple[torch.Tensor, Optional[List], Optional[dict]]:
-        """
-        Args:
-            input_ids: [batch, seq_len]
-            position_ids: [batch, seq_len]
-            attention_mask: [batch, seq_len] or pre-computed 4D mask
-            past_key_values: List of (k, v) tuples for each layer
-            use_cache: Whether to return new cache
-            output_qkv_layers: List of layer indices to output QKV for
-
-        Returns:
-            hidden_states: [batch, seq_len, hidden_size]
-            new_past_key_values: Updated cache
-            qkv_outputs: {layer_idx: qkv_dict}
-        """
-        batch_size, seq_len = input_ids.shape
-
-        # Embedding
-        hidden_states = self.embed_tokens(input_ids)
-
-        # Position IDs
-        if position_ids is None:
-            past_len = past_key_values[0][0].shape[2] if past_key_values else 0
-            position_ids = torch.arange(past_len, past_len + seq_len, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand(batch_size, -1)
-
-        # Attention mask (create causal mask if not provided)
-        if attention_mask is None or attention_mask.dim() == 2:
-            kv_seq_len = seq_len + (past_key_values[0][0].shape[2] if past_key_values else 0)
-            causal_mask = torch.triu(
-                torch.full((seq_len, kv_seq_len), float("-inf"), device=input_ids.device),
-                diagonal=kv_seq_len - seq_len + 1,
-            )
-            attention_mask = causal_mask.unsqueeze(0).unsqueeze(0)  # [1, 1, seq_len, kv_seq_len]
-
-        # Initialize cache list
-        new_past_key_values = [] if use_cache else None
-        qkv_outputs = {} if output_qkv_layers else None
-
-        # Decoder layers
-        for i, layer in enumerate(self.layers):
-            past_kv = past_key_values[i] if past_key_values else None
-            output_qkv = output_qkv_layers is not None and i in output_qkv_layers
-
-            hidden_states, new_kv, qkv_dict = layer(
-                hidden_states=hidden_states,
-                position_ids=position_ids,
-                attention_mask=attention_mask,
-                past_key_value=past_kv,
-                use_cache=use_cache,
-                output_qkv=output_qkv,
-            )
-
-            if use_cache:
-                new_past_key_values.append(new_kv)
-            if qkv_dict is not None:
-                qkv_outputs[i] = qkv_dict
-
-        # Final norm
-        hidden_states = self.norm(hidden_states)
-
-        return hidden_states, new_past_key_values, qkv_outputs
-
-
-class Qwen3ForCausalLM(nn.Module):
-    """Qwen3 Model with Language Modeling head."""
-
-    def __init__(
-        self,
-        vocab_size: int,
-        hidden_size: int,
-        intermediate_size: int,
-        num_hidden_layers: int,
-        num_attention_heads: int,
-        num_key_value_heads: int,
-        head_dim: int,
-        max_position_embeddings: int = 32768,
-        rope_theta: float = 10000.0,
-        rms_norm_eps: float = 1e-6,
-        attention_bias: bool = False,
-        mlp_bias: bool = False,
-        tie_word_embeddings: bool = True,
-    ):
-        super().__init__()
-        self.vocab_size = vocab_size
-        self.tie_word_embeddings = tie_word_embeddings
-
-        # Transformer model
-        self.model = Qwen3Model(
-            vocab_size=vocab_size,
-            hidden_size=hidden_size,
-            intermediate_size=intermediate_size,
-            num_hidden_layers=num_hidden_layers,
-            num_attention_heads=num_attention_heads,
-            num_key_value_heads=num_key_value_heads,
-            head_dim=head_dim,
-            max_position_embeddings=max_position_embeddings,
-            rope_theta=rope_theta,
-            rms_norm_eps=rms_norm_eps,
-            attention_bias=attention_bias,
-            mlp_bias=mlp_bias,
-        )
-
-        # LM head
-        self.lm_head = nn.Linear(hidden_size, vocab_size, bias=False)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        position_ids: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
-        use_cache: bool = False,
-        output_qkv_layers: Optional[List[int]] = None,
-    ) -> Tuple[torch.Tensor, Optional[List], Optional[dict]]:
-        """
-        Args:
-            input_ids: [batch, seq_len]
-            ... (same as Qwen3Model)
-
-        Returns:
-            logits: [batch, seq_len, vocab_size]
-            past_key_values: Updated KV cache
-            qkv_outputs: QKV tensors for specified layers
-        """
-        hidden_states, new_past_key_values, qkv_outputs = self.model(
-            input_ids=input_ids,
-            position_ids=position_ids,
-            attention_mask=attention_mask,
-            past_key_values=past_key_values,
-            use_cache=use_cache,
-            output_qkv_layers=output_qkv_layers,
-        )
-
-        logits = self.lm_head(hidden_states)
-
-        return logits, new_past_key_values, qkv_outputs
-
-    @classmethod
-    def from_pretrained(cls, model_path: str, dtype: torch.dtype = torch.float16) -> "Qwen3ForCausalLM":
-        """
-        Load weights from a pretrained Qwen3 model.
-
-        Args:
-            model_path: Path to model directory containing config.json and model weights
-            dtype: Data type for model weights
-
-        Returns:
-            Initialized Qwen3ForCausalLM model
-        """
-        import json
-        import os
-        from safetensors.torch import load_file
-
-        # Load config
-        config_path = os.path.join(model_path, "config.json")
-        with open(config_path) as f:
-            config = json.load(f)
-
-        # Create model
-        model = cls(
-            vocab_size=config["vocab_size"],
-            hidden_size=config["hidden_size"],
-            intermediate_size=config["intermediate_size"],
-            num_hidden_layers=config["num_hidden_layers"],
-            num_attention_heads=config["num_attention_heads"],
-            num_key_value_heads=config.get("num_key_value_heads", config["num_attention_heads"]),
-            head_dim=config.get("head_dim", config["hidden_size"] // config["num_attention_heads"]),
-            max_position_embeddings=config.get("max_position_embeddings", 32768),
-            rope_theta=config.get("rope_theta", 10000.0),
-            rms_norm_eps=config.get("rms_norm_eps", 1e-6),
-            attention_bias=config.get("attention_bias", False),
-            mlp_bias=config.get("mlp_bias", False),
-            tie_word_embeddings=config.get("tie_word_embeddings", True),
-        )
-
-        # Load weights
-        weight_files = sorted([
-            f for f in os.listdir(model_path)
-            if f.endswith(".safetensors")
-        ])
-
-        state_dict = {}
-        for wf in weight_files:
-            state_dict.update(load_file(os.path.join(model_path, wf)))
-
-        # Load into model
-        model.load_state_dict(state_dict, strict=False)
-
-        # Tie lm_head weights to embed_tokens if configured
-        if model.tie_word_embeddings:
-            model.lm_head.weight = model.model.embed_tokens.weight
-
-        model = model.to(dtype)
-
-        return model
-
-    @torch.no_grad()
-    def generate(
-        self,
-        input_ids: torch.Tensor,
-        max_new_tokens: int = 32,
-        temperature: float = 1.0,
-        do_sample: bool = True,
-        pad_token_id: Optional[int] = None,
-        eos_token_id: Optional[int] = None,
-    ) -> torch.Tensor:
-        """Simple autoregressive generation."""
-        device = input_ids.device
-        batch_size, seq_len = input_ids.shape
-        past_key_values = None
-        generated = input_ids.clone()
-
-        for _ in range(max_new_tokens):
-            if past_key_values is None:
-                current_input = generated
-            else:
-                current_input = generated[:, -1:]
-
-            logits, past_key_values, _ = self(
-                input_ids=current_input,
-                past_key_values=past_key_values,
-                use_cache=True,
-            )
-
-            next_token_logits = logits[:, -1, :]
-            if temperature > 0 and do_sample:
-                next_token_logits = next_token_logits / temperature
-                probs = torch.softmax(next_token_logits, dim=-1)
-                next_token = torch.multinomial(probs, num_samples=1)
-            else:
-                next_token = next_token_logits.argmax(dim=-1, keepdim=True)
-
-            generated = torch.cat([generated, next_token], dim=1)
-
-            if eos_token_id is not None and (next_token == eos_token_id).all():
-                break
-
-        return generated
-
-
-def print_computation_graph():
-    """Print the computation graph for reference."""
-    print(__doc__)
-
-
-if __name__ == "__main__":
-    print_computation_graph()

tests/test_chunk_attention_graph.py

@@ -1,151 +0,0 @@
-#!/usr/bin/env python3
-"""
-Test: Pre-allocated chunk pair graphs for block sparse attention.
-
-Each (Q_chunk, K_chunk) pair has its own captured CUDA graph.
-Zero copy_() during replay - all data pre-filled.
-
-Usage:
-    CUDA_VISIBLE_DEVICES=0 python tests/test_chunk_attention_graph.py
-"""
-
-from dataclasses import dataclass
-from typing import List, Optional
-
-import torch
-
-from nanovllm.ops.chunked_attention import flash_attn_with_lse, merge_attention_outputs
-
-
-@dataclass
-class ChunkAttentionGraph:
-    """Container for a captured chunk attention graph."""
-    graph: torch.cuda.CUDAGraph
-    static_q: torch.Tensor
-    static_k: torch.Tensor
-    static_v: torch.Tensor
-    static_output: torch.Tensor
-    static_lse: torch.Tensor
-    causal: bool
-
-
-def capture_chunk_attention_graph(
-    chunk_size: int,
-    num_heads: int,
-    num_kv_heads: int,
-    head_dim: int,
-    scale: float,
-    device: torch.device,
-    dtype: torch.dtype,
-    causal: bool = False,
-) -> ChunkAttentionGraph:
-    """Capture a CUDA graph for single chunk attention."""
-    static_q = torch.zeros(1, chunk_size, num_heads, head_dim, dtype=dtype, device=device)
-    static_k = torch.zeros(1, chunk_size, num_kv_heads, head_dim, dtype=dtype, device=device)
-    static_v = torch.zeros(1, chunk_size, num_kv_heads, head_dim, dtype=dtype, device=device)
-
-    static_q.normal_()
-    static_k.normal_()
-    static_v.normal_()
-
-    # Warmup
-    with torch.inference_mode():
-        for _ in range(3):
-            _ = flash_attn_with_lse(static_q, static_k, static_v, scale, causal)
-        torch.cuda.synchronize()
-
-    # Capture
-    graph = torch.cuda.CUDAGraph()
-    with torch.inference_mode():
-        with torch.cuda.graph(graph):
-            static_output, static_lse = flash_attn_with_lse(static_q, static_k, static_v, scale, causal)
-
-    torch.cuda.synchronize()
-
-    return ChunkAttentionGraph(
-        graph=graph,
-        static_q=static_q,
-        static_k=static_k,
-        static_v=static_v,
-        static_output=static_output,
-        static_lse=static_lse,
-        causal=causal,
-    )
-
-
-def main():
-    device = torch.device("cuda")
-    dtype = torch.bfloat16
-
-    chunk_size = 64
-    num_chunks = 4
-    num_heads = 8
-    num_kv_heads = 8
-    head_dim = 64
-    scale = 1.0 / (head_dim ** 0.5)
-    seq_len = chunk_size * num_chunks
-
-    print(f"Device: {torch.cuda.get_device_name()}")
-    print(f"Chunk size: {chunk_size}, Num chunks: {num_chunks}")
-    print(f"Total graphs: {num_chunks * (num_chunks + 1) // 2}")
-
-    # Test data
-    full_q = torch.randn(1, seq_len, num_heads, head_dim, dtype=dtype, device=device)
-    full_k = torch.randn(1, seq_len, num_kv_heads, head_dim, dtype=dtype, device=device)
-    full_v = torch.randn(1, seq_len, num_kv_heads, head_dim, dtype=dtype, device=device)
-
-    # Reference
-    with torch.inference_mode():
-        full_output, _ = flash_attn_with_lse(full_q, full_k, full_v, scale, causal=True)
-
-    # Capture all graphs
-    graphs: List[List[Optional[ChunkAttentionGraph]]] = [[None] * num_chunks for _ in range(num_chunks)]
-    for q_idx in range(num_chunks):
-        for k_idx in range(q_idx + 1):
-            graphs[q_idx][k_idx] = capture_chunk_attention_graph(
-                chunk_size, num_heads, num_kv_heads, head_dim, scale, device, dtype,
-                causal=(k_idx == q_idx)
-            )
-    print("All graphs captured")
-
-    # Pre-fill static tensors
-    for q_idx in range(num_chunks):
-        for k_idx in range(q_idx + 1):
-            g = graphs[q_idx][k_idx]
-            g.static_q.copy_(full_q[:, q_idx*chunk_size:(q_idx+1)*chunk_size])
-            g.static_k.copy_(full_k[:, k_idx*chunk_size:(k_idx+1)*chunk_size])
-            g.static_v.copy_(full_v[:, k_idx*chunk_size:(k_idx+1)*chunk_size])
-    print("Static tensors pre-filled")
-
-    # Replay and merge
-    chunked_output = torch.zeros_like(full_output)
-    for q_idx in range(num_chunks):
-        acc_out, acc_lse = None, None
-        for k_idx in range(q_idx + 1):
-            g = graphs[q_idx][k_idx]
-            g.graph.replay()
-            out, lse = g.static_output.clone(), g.static_lse.clone()
-            if acc_out is None:
-                acc_out, acc_lse = out, lse
-            else:
-                with torch.inference_mode():
-                    acc_out, acc_lse = merge_attention_outputs(acc_out, acc_lse, out, lse)
-        chunked_output[:, q_idx*chunk_size:(q_idx+1)*chunk_size] = acc_out
-
-    torch.cuda.synchronize()
-
-    # Compare
-    all_pass = True
-    for q_idx in range(num_chunks):
-        s, e = q_idx * chunk_size, (q_idx + 1) * chunk_size
-        diff = (full_output[:, s:e] - chunked_output[:, s:e]).abs().max().item()
-        status = "✅" if diff < 1e-2 else "❌"
-        print(f"Q[{q_idx}]: max_diff={diff:.2e} {status}")
-        if diff >= 1e-2:
-            all_pass = False
-
-    print("✅ PASSED" if all_pass else "❌ FAILED")
-
-
-if __name__ == "__main__":
-    main()

tests/test_chunk_attention_graph_reuse.py

@@ -1,156 +0,0 @@
-#!/usr/bin/env python3
-"""
-Test: Reuse a single CUDA Graph across all layers and all chunk pairs.
-
-Key insight: LLM layers have identical computation structure.
-We only need 2 graphs (causal + non-causal), reused for all (layer, Q_i, K_j) combinations.
-
-Usage:
-    CUDA_VISIBLE_DEVICES=0 python tests/test_chunk_attention_graph_reuse.py
-"""
-
-from dataclasses import dataclass
-
-import torch
-
-from nanovllm.ops.chunked_attention import flash_attn_with_lse, merge_attention_outputs
-
-
-@dataclass
-class ReusableChunkGraph:
-    """A single graph that can be reused with copy_() updates."""
-    graph: torch.cuda.CUDAGraph
-    static_q: torch.Tensor
-    static_k: torch.Tensor
-    static_v: torch.Tensor
-    static_output: torch.Tensor
-    static_lse: torch.Tensor
-
-
-def capture_reusable_graph(
-    chunk_size: int,
-    num_heads: int,
-    num_kv_heads: int,
-    head_dim: int,
-    scale: float,
-    device: torch.device,
-    dtype: torch.dtype,
-    causal: bool,
-) -> ReusableChunkGraph:
-    """Capture ONE graph to be reused for all chunk pairs."""
-    static_q = torch.zeros(1, chunk_size, num_heads, head_dim, dtype=dtype, device=device)
-    static_k = torch.zeros(1, chunk_size, num_kv_heads, head_dim, dtype=dtype, device=device)
-    static_v = torch.zeros(1, chunk_size, num_kv_heads, head_dim, dtype=dtype, device=device)
-
-    static_q.normal_()
-    static_k.normal_()
-    static_v.normal_()
-
-    # Warmup
-    with torch.inference_mode():
-        for _ in range(3):
-            _ = flash_attn_with_lse(static_q, static_k, static_v, scale, causal)
-        torch.cuda.synchronize()
-
-    # Capture
-    graph = torch.cuda.CUDAGraph()
-    with torch.inference_mode():
-        with torch.cuda.graph(graph):
-            static_output, static_lse = flash_attn_with_lse(static_q, static_k, static_v, scale, causal)
-
-    torch.cuda.synchronize()
-
-    return ReusableChunkGraph(
-        graph=graph,
-        static_q=static_q,
-        static_k=static_k,
-        static_v=static_v,
-        static_output=static_output,
-        static_lse=static_lse,
-    )
-
-
-def replay_with_copy(graph: ReusableChunkGraph, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
-    """Replay graph after updating static tensors with copy_()."""
-    graph.static_q.copy_(q)
-    graph.static_k.copy_(k)
-    graph.static_v.copy_(v)
-    graph.graph.replay()
-    return graph.static_output.clone(), graph.static_lse.clone()
-
-
-def main():
-    device = torch.device("cuda")
-    dtype = torch.bfloat16
-
-    chunk_size = 64
-    num_chunks = 4
-    num_layers = 3  # Simulate multiple layers
-    num_heads = 8
-    num_kv_heads = 8
-    head_dim = 64
-    scale = 1.0 / (head_dim ** 0.5)
-    seq_len = chunk_size * num_chunks
-
-    print(f"Device: {torch.cuda.get_device_name()}")
-    print(f"Chunk size: {chunk_size}, Num chunks: {num_chunks}, Num layers: {num_layers}")
-    print(f"Only 2 graphs (causal + non-causal) for ALL layer × chunk combinations")
-
-    # Capture only 2 graphs
-    graph_causal = capture_reusable_graph(
-        chunk_size, num_heads, num_kv_heads, head_dim, scale, device, dtype, causal=True
-    )
-    graph_non_causal = capture_reusable_graph(
-        chunk_size, num_heads, num_kv_heads, head_dim, scale, device, dtype, causal=False
-    )
-    print("2 graphs captured (causal + non-causal)")
-
-    all_pass = True
-
-    for layer_id in range(num_layers):
-        # Different Q/K/V for each layer (simulating different layer outputs)
-        full_q = torch.randn(1, seq_len, num_heads, head_dim, dtype=dtype, device=device)
-        full_k = torch.randn(1, seq_len, num_kv_heads, head_dim, dtype=dtype, device=device)
-        full_v = torch.randn(1, seq_len, num_kv_heads, head_dim, dtype=dtype, device=device)
-
-        # Reference: full causal attention
-        with torch.inference_mode():
-            full_output, _ = flash_attn_with_lse(full_q, full_k, full_v, scale, causal=True)
-
-        # Chunked with graph reuse
-        chunked_output = torch.zeros_like(full_output)
-
-        for q_idx in range(num_chunks):
-            q_chunk = full_q[:, q_idx*chunk_size:(q_idx+1)*chunk_size]
-            acc_out, acc_lse = None, None
-
-            for k_idx in range(q_idx + 1):
-                k_chunk = full_k[:, k_idx*chunk_size:(k_idx+1)*chunk_size]
-                v_chunk = full_v[:, k_idx*chunk_size:(k_idx+1)*chunk_size]
-
-                # Reuse graph with copy_()
-                graph = graph_causal if k_idx == q_idx else graph_non_causal
-                out, lse = replay_with_copy(graph, q_chunk, k_chunk, v_chunk)
-
-                if acc_out is None:
-                    acc_out, acc_lse = out, lse
-                else:
-                    with torch.inference_mode():
-                        acc_out, acc_lse = merge_attention_outputs(acc_out, acc_lse, out, lse)
-
-            chunked_output[:, q_idx*chunk_size:(q_idx+1)*chunk_size] = acc_out
-
-        torch.cuda.synchronize()
-
-        # Compare
-        max_diff = (full_output - chunked_output).abs().max().item()
-        status = "✅" if max_diff < 1e-2 else "❌"
-        print(f"Layer {layer_id}: max_diff={max_diff:.2e} {status}")
-        if max_diff >= 1e-2:
-            all_pass = False
-
-    print("✅ PASSED - Single graph reuse across layers works!" if all_pass else "❌ FAILED")
-
-
-if __name__ == "__main__":
-    main()

tests/test_cudagraph_memory.py

@@ -1,357 +0,0 @@
-#!/usr/bin/env python3
-"""
-CUDA Graph Memory Analysis Test
-
-This script analyzes the memory overhead of CUDA Graph at each stage:
-1. Model loading
-2. StaticCache allocation
-3. Warmup runs
-4. Graph capture
-5. Graph replay
-
-Usage:
-    CUDA_VISIBLE_DEVICES=4 python tests/test_cudagraph_memory.py
-    CUDA_VISIBLE_DEVICES=4 python tests/test_cudagraph_memory.py --model ~/models/Qwen3-0.6B
-    CUDA_VISIBLE_DEVICES=4 python tests/test_cudagraph_memory.py --max-cache-len 2048
-"""
-
-import argparse
-import os
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.cache_utils import StaticCache
-
-
-def get_memory_mb():
-    """Get current allocated memory in MB."""
-    return torch.cuda.memory_allocated() / 1024**2
-
-
-def get_memory_gb():
-    """Get current allocated memory in GB."""
-    return torch.cuda.memory_allocated() / 1024**3
-
-
-def get_peak_memory_gb():
-    """Get peak allocated memory in GB."""
-    return torch.cuda.max_memory_allocated() / 1024**3
-
-
-def print_separator(title=None):
-    """Print a separator line."""
-    if title:
-        print(f"\n{'=' * 70}")
-        print(f" {title}")
-        print(f"{'=' * 70}")
-    else:
-        print("-" * 70)
-
-
-def test_memory_stages(model_path: str, max_cache_len: int, batch_size: int = 1):
-    """
-    Test memory usage at each stage of CUDA Graph setup.
-
-    Args:
-        model_path: Path to the model
-        max_cache_len: Maximum cache length for StaticCache
-        batch_size: Batch size for inference
-    """
-    print_separator("CUDA Graph Memory Analysis")
-    print(f"Model: {model_path}")
-    print(f"Max cache length: {max_cache_len}")
-    print(f"Batch size: {batch_size}")
-
-    results = {}
-
-    # Stage 0: Initial
-    torch.cuda.empty_cache()
-    torch.cuda.reset_peak_memory_stats()
-    results["initial"] = get_memory_mb()
-
-    # Stage 1: Load model
-    print_separator("Stage 1: Model Loading")
-    model = AutoModelForCausalLM.from_pretrained(
-        model_path,
-        torch_dtype=torch.bfloat16,
-        device_map="cuda",
-        trust_remote_code=True,
-    )
-    model.eval()
-
-    results["after_model"] = get_memory_mb()
-    model_size = results["after_model"] - results["initial"]
-    print(f"  Memory: {results['after_model']:.0f} MB")
-    print(f"  Model size: {model_size:.0f} MB ({model_size/1024:.2f} GB)")
-
-    config = model.config
-    device = next(model.parameters()).device
-    dtype = next(model.parameters()).dtype
-
-    # Stage 2: Allocate StaticCache
-    print_separator("Stage 2: StaticCache Allocation")
-    torch.cuda.reset_peak_memory_stats()
-    before = get_memory_mb()
-
-    static_cache = StaticCache(
-        config=config,
-        max_batch_size=batch_size,
-        max_cache_len=max_cache_len,
-        device=device,
-        dtype=dtype,
-    )
-
-    results["after_cache"] = get_memory_mb()
-    cache_size = results["after_cache"] - before
-    print(f"  Memory: {results['after_cache']:.0f} MB")
-    print(f"  StaticCache size: {cache_size:.0f} MB")
-
-    # Calculate theoretical cache size
-    num_layers = config.num_hidden_layers
-    num_kv_heads = getattr(config, "num_key_value_heads", config.num_attention_heads)
-    head_dim = config.hidden_size // config.num_attention_heads
-    dtype_size = 2  # bfloat16
-
-    theoretical_cache = (
-        num_layers * 2 * batch_size * num_kv_heads * max_cache_len * head_dim * dtype_size
-    ) / (1024**2)
-    print(f"  Theoretical: {theoretical_cache:.0f} MB")
-    print(f"  Overhead: {cache_size - theoretical_cache:.0f} MB ({(cache_size/theoretical_cache - 1)*100:.1f}%)")
-
-    # Stage 3: Prepare static tensors
-    print_separator("Stage 3: Static Tensor Allocation")
-    before = get_memory_mb()
-
-    static_input_ids = torch.zeros(batch_size, 1, dtype=torch.long, device=device)
-    static_position_ids = torch.zeros(batch_size, 1, dtype=torch.long, device=device)
-    static_cache_position = torch.tensor([0], dtype=torch.long, device=device)
-
-    results["after_tensors"] = get_memory_mb()
-    tensor_size = results["after_tensors"] - before
-    print(f"  Memory: {results['after_tensors']:.0f} MB")
-    print(f"  Static tensors: {tensor_size:.2f} MB (negligible)")
-
-    # Stage 4: Warmup runs
-    print_separator("Stage 4: Warmup Runs (3 iterations)")
-    torch.cuda.reset_peak_memory_stats()
-    before = get_memory_mb()
-
-    with torch.inference_mode():
-        for i in range(3):
-            _ = model(
-                input_ids=static_input_ids,
-                position_ids=static_position_ids,
-                past_key_values=static_cache,
-                cache_position=static_cache_position,
-                use_cache=True,
-            )
-        torch.cuda.synchronize()
-
-    results["after_warmup"] = get_memory_mb()
-    results["warmup_peak"] = get_peak_memory_gb() * 1024
-    warmup_size = results["after_warmup"] - before
-    print(f"  Memory: {results['after_warmup']:.0f} MB")
-    print(f"  Peak: {results['warmup_peak']:.0f} MB")
-    print(f"  Warmup overhead: {warmup_size:.0f} MB")
-
-    # Stage 5: CUDA Graph capture
-    print_separator("Stage 5: CUDA Graph Capture")
-    torch.cuda.reset_peak_memory_stats()
-    before = get_memory_mb()
-
-    graph = torch.cuda.CUDAGraph()
-    with torch.inference_mode():
-        with torch.cuda.graph(graph):
-            outputs = model(
-                input_ids=static_input_ids,
-                position_ids=static_position_ids,
-                past_key_values=static_cache,
-                cache_position=static_cache_position,
-                use_cache=True,
-            )
-            static_logits = outputs.logits
-    torch.cuda.synchronize()
-
-    results["after_capture"] = get_memory_mb()
-    results["capture_peak"] = get_peak_memory_gb() * 1024
-    capture_size = results["after_capture"] - before
-    print(f"  Memory: {results['after_capture']:.0f} MB")
-    print(f"  Peak: {results['capture_peak']:.0f} MB")
-    print(f"  Graph capture overhead: {capture_size:.0f} MB")
-
-    # Stage 6: Graph replay
-    print_separator("Stage 6: Graph Replay (10 iterations)")
-    torch.cuda.reset_peak_memory_stats()
-    before = get_memory_mb()
-
-    with torch.inference_mode():
-        for _ in range(10):
-            static_input_ids.fill_(1)
-            static_cache_position.fill_(0)
-            graph.replay()
-        torch.cuda.synchronize()
-
-    results["after_replay"] = get_memory_mb()
-    results["replay_peak"] = get_peak_memory_gb() * 1024
-    replay_change = results["after_replay"] - before
-    print(f"  Memory: {results['after_replay']:.0f} MB")
-    print(f"  Peak: {results['replay_peak']:.0f} MB")
-    print(f"  Replay memory change: {replay_change:.0f} MB (should be ~0)")
-
-    # Summary
-    print_separator("SUMMARY")
-    total_overhead = results["after_capture"] - results["after_model"]
-
-    print(f"{'Stage':<25} {'Memory (MB)':>12} {'Delta (MB)':>12}")
-    print("-" * 50)
-    print(f"{'Model loaded':<25} {results['after_model']:>12.0f} {model_size:>+12.0f}")
-    print(f"{'StaticCache allocated':<25} {results['after_cache']:>12.0f} {cache_size:>+12.0f}")
-    print(f"{'After warmup':<25} {results['after_warmup']:>12.0f} {warmup_size:>+12.0f}")
-    print(f"{'After graph capture':<25} {results['after_capture']:>12.0f} {capture_size:>+12.0f}")
-    print(f"{'After graph replay':<25} {results['after_replay']:>12.0f} {replay_change:>+12.0f}")
-    print("-" * 50)
-    print(f"{'Total (excl. model)':<25} {'':<12} {total_overhead:>+12.0f}")
-
-    print_separator("KEY FINDINGS")
-    print(f"  1. Model size:           {model_size/1024:.2f} GB")
-    print(f"  2. StaticCache:          {cache_size:.0f} MB (main overhead, scales with cache_len)")
-    print(f"  3. Graph capture:        {capture_size:.0f} MB (small, stores kernel sequence)")
-    print(f"  4. Graph replay:         {replay_change:.0f} MB (zero allocation, reuses memory)")
-    print(f"  5. Total CUDA Graph overhead: {total_overhead:.0f} MB")
-
-    return results
-
-
-def test_cache_length_scaling(model_path: str, cache_lengths: list):
-    """
-    Test how memory scales with different cache lengths.
-
-    Args:
-        model_path: Path to the model
-        cache_lengths: List of cache lengths to test
-    """
-    print_separator("Cache Length Scaling Test")
-    print(f"Model: {model_path}")
-    print(f"Cache lengths: {cache_lengths}")
-
-    # Load model once
-    model = AutoModelForCausalLM.from_pretrained(
-        model_path,
-        torch_dtype=torch.bfloat16,
-        device_map="cuda",
-        trust_remote_code=True,
-    )
-    model.eval()
-
-    config = model.config
-    device = next(model.parameters()).device
-    dtype = next(model.parameters()).dtype
-
-    model_mem = get_memory_mb()
-
-    results = []
-    for cache_len in cache_lengths:
-        torch.cuda.empty_cache()
-        torch.cuda.reset_peak_memory_stats()
-
-        # Create cache and capture graph
-        static_cache = StaticCache(
-            config=config,
-            max_batch_size=1,
-            max_cache_len=cache_len,
-            device=device,
-            dtype=dtype,
-        )
-
-        static_input_ids = torch.zeros(1, 1, dtype=torch.long, device=device)
-        static_position_ids = torch.zeros(1, 1, dtype=torch.long, device=device)
-        static_cache_position = torch.tensor([0], dtype=torch.long, device=device)
-
-        with torch.inference_mode():
-            # Warmup
-            for _ in range(3):
-                _ = model(
-                    input_ids=static_input_ids,
-                    position_ids=static_position_ids,
-                    past_key_values=static_cache,
-                    cache_position=static_cache_position,
-                    use_cache=True,
-                )
-            torch.cuda.synchronize()
-
-            # Capture
-            graph = torch.cuda.CUDAGraph()
-            with torch.cuda.graph(graph):
-                outputs = model(
-                    input_ids=static_input_ids,
-                    position_ids=static_position_ids,
-                    past_key_values=static_cache,
-                    cache_position=static_cache_position,
-                    use_cache=True,
-                )
-            torch.cuda.synchronize()
-
-        total_mem = get_memory_mb()
-        overhead = total_mem - model_mem
-        results.append((cache_len, total_mem, overhead))
-
-        del static_cache, graph
-        torch.cuda.empty_cache()
-
-    # Print results
-    print()
-    print(f"{'Cache Length':>12} | {'Total (MB)':>12} | {'Overhead (MB)':>14} | {'Per 1K tokens':>14}")
-    print("-" * 60)
-    for cache_len, total, overhead in results:
-        per_1k = overhead / (cache_len / 1000)
-        print(f"{cache_len:>12} | {total:>12.0f} | {overhead:>14.0f} | {per_1k:>14.1f}")
-
-    return results
-
-
-def main():
-    parser = argparse.ArgumentParser(description="CUDA Graph Memory Analysis")
-    parser.add_argument(
-        "--model",
-        type=str,
-        default="~/models/Qwen3-4B-Instruct-2507",
-        help="Model path",
-    )
-    parser.add_argument(
-        "--max-cache-len",
-        type=int,
-        default=1024,
-        help="Maximum cache length",
-    )
-    parser.add_argument(
-        "--batch-size",
-        type=int,
-        default=1,
-        help="Batch size",
-    )
-    parser.add_argument(
-        "--test-scaling",
-        action="store_true",
-        help="Test cache length scaling",
-    )
-    args = parser.parse_args()
-
-    model_path = os.path.expanduser(args.model)
-
-    if not torch.cuda.is_available():
-        print("CUDA is not available!")
-        return
-
-    print(f"Device: cuda:{torch.cuda.current_device()}")
-    print(f"GPU: {torch.cuda.get_device_name()}")
-
-    if args.test_scaling:
-        cache_lengths = [256, 512, 1024, 2048, 4096]
-        test_cache_length_scaling(model_path, cache_lengths)
-    else:
-        test_memory_stages(model_path, args.max_cache_len, args.batch_size)
-
-    print("\ntest_cudagraph_memory: PASSED")
-
-
-if __name__ == "__main__":
-    main()

tests/test_needle.py

@@ -1,254 +0,0 @@
-"""
-Needle-in-a-haystack test for LLM.
-
-Tests: Long context retrieval capability with configurable sequence length.
-
-NOTE: CPU offload mode has a known bug that causes incorrect outputs for
-sequences longer than ~200 tokens. Use --no-offload for correctness testing.
-"""
-
-import os
-os.environ["NANOVLLM_LOG_LEVEL"] = "INFO"
-
-import argparse
-from nanovllm import LLM, SamplingParams
-from nanovllm.config import SparsePolicyType
-from utils import generate_needle_prompt, check_needle_answer
-
-
-# ============================================================
-# Main Test
-# ============================================================
-
-def run_needle_test(
-    model_path: str,
-    max_model_len: int,
-    input_len: int,
-    num_gpu_blocks: int = 4,
-    block_size: int = 1024,
-    needle_position: float = 0.5,
-    needle_value: str = "7492",
-    max_new_tokens: int = 32,
-    enable_cpu_offload: bool = False,
-    enable_quest: bool = False,
-    enable_xattn_bsa: bool = False,
-    sparse_topk: int = 8,
-    sparse_threshold: int = 4,
-    sparse_samples: int = 128,
-    verbose: bool = True,
-) -> bool:
-    """
-    Run a needle-in-haystack test.
-
-    Args:
-        model_path: Path to model
-        max_model_len: Maximum model context length
-        input_len: Target input sequence length
-        num_gpu_blocks: Number of GPU blocks for offload
-        block_size: KV cache block size
-        needle_position: Where to place needle (0.0-1.0)
-        needle_value: The secret value to find
-        max_new_tokens: Maximum tokens to generate
-        enable_cpu_offload: Enable CPU offload mode
-        enable_quest: Enable Quest sparse attention (decode-only Top-K)
-        enable_xattn_bsa: Enable XAttention BSA sparse attention (prefill-only)
-        sparse_topk: Top-K blocks for Quest
-        sparse_threshold: Threshold for sparse selection (Quest/XAttention BSA)
-        sparse_samples: Samples per chunk for XAttention BSA estimation
-        verbose: Print detailed output
-
-    Returns:
-        True if test passed, False otherwise
-    """
-    # Determine sparse policy
-    if enable_xattn_bsa:
-        sparse_policy = SparsePolicyType.XATTN_BSA
-    elif enable_quest:
-        sparse_policy = SparsePolicyType.QUEST
-    else:
-        sparse_policy = SparsePolicyType.FULL
-
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Needle-in-Haystack Test")
-        print(f"{'='*60}")
-        print(f"Model: {model_path}")
-        print(f"Max model len: {max_model_len}")
-        print(f"Input length: {input_len}")
-        print(f"Block size: {block_size}")
-        print(f"Needle position: {needle_position:.0%}")
-        print(f"Needle value: {needle_value}")
-        print(f"CPU offload: {enable_cpu_offload}")
-        if enable_cpu_offload:
-            print(f"Sparse policy: {sparse_policy.name}")
-            if sparse_policy == SparsePolicyType.QUEST:
-                print(f"  Quest: topk={sparse_topk}, threshold={sparse_threshold}")
-            elif sparse_policy == SparsePolicyType.XATTN_BSA:
-                print(f"  XAttention BSA: threshold={sparse_threshold}, samples={sparse_samples}")
-        print(f"{'='*60}\n")
-
-    # 1. Initialize LLM
-    llm_kwargs = {
-        "enforce_eager": True,
-        "max_model_len": max_model_len,
-        "max_num_batched_tokens": max_model_len,
-        "enable_cpu_offload": enable_cpu_offload,
-        "kvcache_block_size": block_size,
-    }
-    if enable_cpu_offload:
-        llm_kwargs["num_gpu_blocks"] = num_gpu_blocks
-        llm_kwargs["sparse_policy"] = sparse_policy
-        if sparse_policy == SparsePolicyType.QUEST:
-            llm_kwargs["sparse_topk_blocks"] = sparse_topk
-            llm_kwargs["sparse_threshold_blocks"] = sparse_threshold
-        elif sparse_policy == SparsePolicyType.XATTN_BSA:
-            llm_kwargs["sparse_threshold"] = float(sparse_threshold) / 10.0  # Convert to 0.0-1.0 range
-            llm_kwargs["sparse_samples_per_chunk"] = sparse_samples
-
-    llm = LLM(model_path, **llm_kwargs)
-
-    # 2. Generate needle prompt
-    prompt, expected = generate_needle_prompt(
-        tokenizer=llm.tokenizer,
-        target_length=input_len,
-        needle_position=needle_position,
-        needle_value=needle_value,
-    )
-
-    # 3. Generate output
-    sampling_params = SamplingParams(
-        temperature=0.6,  # Moderate temperature
-        max_tokens=max_new_tokens,
-    )
-    outputs = llm.generate([prompt], sampling_params, use_tqdm=True)
-
-    # 4. Check result
-    output_text = outputs[0]["text"]
-    output_token_ids = outputs[0]["token_ids"]
-    passed = check_needle_answer(output_text, expected)
-
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Result")
-        print(f"{'='*60}")
-        print(f"Expected: {expected}")
-        print(f"Output tokens ({len(output_token_ids)}): {output_token_ids[:20]}")
-        print(f"Output: {output_text[:200]}...")
-        print(f"Status: {'PASSED' if passed else 'FAILED'}")
-        print(f"{'='*60}\n")
-
-    return passed
-
-
-# ============================================================
-# CLI Entry Point
-# ============================================================
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Needle-in-haystack test for long context LLM")
-    parser.add_argument(
-        "--model", "-m",
-        type=str,
-        default=os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/"),
-        help="Path to model"
-    )
-    parser.add_argument(
-        "--max-model-len",
-        type=int,
-        default=128 * 1024,
-        help="Maximum model context length"
-    )
-    parser.add_argument(
-        "--input-len",
-        type=int,
-        default=8 * 1024,
-        help="Target input sequence length"
-    )
-    parser.add_argument(
-        "--num-gpu-blocks",
-        type=int,
-        default=2,
-        help="Number of GPU blocks for CPU offload"
-    )
-    parser.add_argument(
-        "--block-size",
-        type=int,
-        default=1024,
-        help="KV cache block size"
-    )
-    parser.add_argument(
-        "--needle-position",
-        type=float,
-        default=0.5,
-        help="Needle position (0.0=start, 0.5=middle, 1.0=end)"
-    )
-    parser.add_argument(
-        "--needle-value",
-        type=str,
-        default="7492",
-        help="The secret value to hide"
-    )
-    parser.add_argument(
-        "--max-new-tokens",
-        type=int,
-        default=32,
-        help="Maximum tokens to generate"
-    )
-    parser.add_argument(
-        "--enable-offload",
-        action="store_true",
-        help="Enable CPU offload (has known bug for long sequences)"
-    )
-    parser.add_argument(
-        "--enable-quest",
-        action="store_true",
-        help="Enable Quest sparse attention (decode-only Top-K selection)"
-    )
-    parser.add_argument(
-        "--enable-xattn-bsa",
-        action="store_true",
-        help="Enable XAttention BSA sparse attention (prefill-only)"
-    )
-    parser.add_argument(
-        "--sparse-topk",
-        type=int,
-        default=8,
-        help="Top-K blocks for Quest sparse attention"
-    )
-    parser.add_argument(
-        "--sparse-threshold",
-        type=int,
-        default=4,
-        help="Apply sparse only when blocks > threshold (Quest) or attention threshold 0-9 (XAttention BSA)"
-    )
-    parser.add_argument(
-        "--sparse-samples",
-        type=int,
-        default=128,
-        help="Samples per chunk for XAttention BSA estimation"
-    )
-    args = parser.parse_args()
-
-    passed = run_needle_test(
-        model_path=args.model,
-        max_model_len=args.max_model_len,
-        input_len=args.input_len,
-        num_gpu_blocks=args.num_gpu_blocks,
-        block_size=args.block_size,
-        needle_position=args.needle_position,
-        needle_value=args.needle_value,
-        max_new_tokens=args.max_new_tokens,
-        enable_cpu_offload=args.enable_offload,
-        enable_quest=args.enable_quest,
-        enable_xattn_bsa=args.enable_xattn_bsa,
-        sparse_topk=args.sparse_topk,
-        sparse_threshold=args.sparse_threshold,
-        sparse_samples=args.sparse_samples,
-        verbose=True,
-    )
-
-    if passed:
-        print("test_needle: PASSED")
-    else:
-        print("test_needle: FAILED")
-        exit(1)

tests/test_needle_ref.py

@@ -1,176 +0,0 @@
-"""
-Needle-in-a-haystack reference test using pure torch + transformers.
-
-This is a reference implementation for comparison with nanovllm.
-Uses standard HuggingFace inference (no custom KV cache, no offload).
-"""
-
-import os
-import argparse
-import torch
-from transformers import AutoTokenizer
-from modeling_qwen3 import Qwen3ForCausalLM
-from utils import generate_needle_prompt, check_needle_answer
-
-
-# ============================================================
-# Main Test
-# ============================================================
-
-def run_needle_test(
-    model_path: str,
-    input_len: int,
-    needle_position: float = 0.5,
-    needle_value: str = "7492",
-    max_new_tokens: int = 32,
-    dtype: str = "auto",
-    verbose: bool = True,
-) -> bool:
-    """
-    Run a needle-in-haystack test using standard transformers inference.
-
-    Args:
-        model_path: Path to model
-        input_len: Target input sequence length
-        needle_position: Where to place needle (0.0-1.0)
-        needle_value: The secret value to find
-        max_new_tokens: Maximum tokens to generate
-        dtype: Model dtype ("auto", "float16", "bfloat16")
-        verbose: Print detailed output
-
-    Returns:
-        True if test passed, False otherwise
-    """
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Needle-in-Haystack Reference Test (torch + transformers)")
-        print(f"{'='*60}")
-        print(f"Model: {model_path}")
-        print(f"Input length: {input_len}")
-        print(f"Needle position: {needle_position:.0%}")
-        print(f"Needle value: {needle_value}")
-        print(f"Dtype: {dtype}")
-        print(f"{'='*60}\n")
-
-    # 1. Load tokenizer
-    print("[1/4] Loading tokenizer...")
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-
-    # 2. Generate needle prompt
-    print("[2/4] Generating needle prompt...")
-    prompt, expected = generate_needle_prompt(
-        tokenizer=tokenizer,
-        target_length=input_len,
-        needle_position=needle_position,
-        needle_value=needle_value,
-    )
-
-    # 3. Load model
-    print("[3/4] Loading model...")
-    torch_dtype = {
-        "auto": torch.float16,  # default to float16 for custom model
-        "float16": torch.float16,
-        "bfloat16": torch.bfloat16,
-    }.get(dtype, torch.float16)
-
-    model = Qwen3ForCausalLM.from_pretrained(model_path, dtype=torch_dtype)
-    model = model.to("cuda" if torch.cuda.is_available() else "cpu")
-    model.eval()
-
-    # 4. Generate output
-    print("[4/4] Running inference...")
-    device = next(model.parameters()).device
-    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
-    print(f"  Input shape: {input_ids.shape}")
-
-    with torch.no_grad():
-        output_ids = model.generate(
-            input_ids,
-            max_new_tokens=max_new_tokens,
-            temperature=0.6,
-            do_sample=True,
-            pad_token_id=tokenizer.eos_token_id,
-        )
-
-    # Decode only the new tokens
-    new_token_ids = output_ids[0, input_ids.shape[1]:]
-    output_text = tokenizer.decode(new_token_ids, skip_special_tokens=False)
-
-    # 5. Check result
-    passed = check_needle_answer(output_text, expected)
-
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Result")
-        print(f"{'='*60}")
-        print(f"Expected: {expected}")
-        print(f"Output tokens ({len(new_token_ids)}): {new_token_ids[:20].tolist()}")
-        print(f"Output: {output_text[:200]}...")
-        print(f"Status: {'PASSED' if passed else 'FAILED'}")
-        print(f"{'='*60}\n")
-
-    return passed
-
-
-# ============================================================
-# CLI Entry Point
-# ============================================================
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Needle-in-haystack reference test (torch + transformers)"
-    )
-    parser.add_argument(
-        "--model", "-m",
-        type=str,
-        default=os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/"),
-        help="Path to model"
-    )
-    parser.add_argument(
-        "--input-len",
-        type=int,
-        default=8 * 1024,
-        help="Target input sequence length"
-    )
-    parser.add_argument(
-        "--needle-position",
-        type=float,
-        default=0.5,
-        help="Needle position (0.0=start, 0.5=middle, 1.0=end)"
-    )
-    parser.add_argument(
-        "--needle-value",
-        type=str,
-        default="7492",
-        help="The secret value to hide"
-    )
-    parser.add_argument(
-        "--max-new-tokens",
-        type=int,
-        default=32,
-        help="Maximum tokens to generate"
-    )
-    parser.add_argument(
-        "--dtype",
-        type=str,
-        default="auto",
-        choices=["auto", "float16", "bfloat16"],
-        help="Model dtype"
-    )
-    args = parser.parse_args()
-
-    passed = run_needle_test(
-        model_path=args.model,
-        input_len=args.input_len,
-        needle_position=args.needle_position,
-        needle_value=args.needle_value,
-        max_new_tokens=args.max_new_tokens,
-        dtype=args.dtype,
-        verbose=True,
-    )
-
-    if passed:
-        print("test_needle_ref: PASSED")
-    else:
-        print("test_needle_ref: FAILED")
-        exit(1)

tests/test_quest_policy.py

@@ -1,136 +0,0 @@
-"""
-Test for QuestPolicy block selection with GQA (Grouped Query Attention).
-
-Demonstrates the key limitation: scores are AVERAGED across heads,
-so blocks strongly needed by one head but not others may be dropped.
-
-This is the expected Quest behavior - not a bug.
-"""
-
-import torch
-from nanovllm.kvcache.sparse import (
-    create_sparse_policy,
-    SparsePolicyType,
-    PolicyContext,
-)
-
-# ============================================================
-# Test: Per-Head Score Averaging in GQA
-# ============================================================
-
-# Determine device (GPU if available, else CPU)
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print(f"Running test on device: {device}")
-
-# Setup: 2 KV heads, 4 query heads (GQA group_size=2)
-# topk=2 to make selection competitive
-
-quest = create_sparse_policy(SparsePolicyType.QUEST, topk_blocks=2, threshold_blocks=0)
-quest.initialize(
-    num_layers=1,
-    num_kv_heads=2,
-    head_dim=4,
-    num_cpu_blocks=6,
-    dtype=torch.float32,
-    device=device,  # Metadata stored on GPU
-)
-
-metadata = quest.metadata
-
-def set_key(block_id, head_id, values):
-    """Set both key_min and key_max to same values for deterministic scoring."""
-    # Values need to be on the same device as metadata
-    tensor = torch.tensor(values, device=device)
-    metadata.key_min[block_id, 0, head_id, :] = tensor
-    metadata.key_max[block_id, 0, head_id, :] = tensor
-
-# ============================================================
-# Design: Different heads want different blocks
-# ============================================================
-#
-# Query = [1,1,1,1] for all heads, so score = sum(key values)
-#
-# Block | Head 0 | Head 1 | Average | Result
-# ------|--------|--------|---------|--------
-#   0   |   +4   |   -4   |    0    | Head0 wants, Head1 doesn't → DROPPED
-#   1   |   -4   |   +4   |    0    | Head1 wants, Head0 doesn't → DROPPED
-#   2   |   +4   |   +4   |   +4    | Both want → SELECTED (rank 1)
-#   3   |   +3   |   +3   |   +3    | Both want → SELECTED (rank 2)
-#   4   |   +4   |    0   |   +2    | Head0 strongly wants, Head1 neutral → rank 3
-#   5   |    0   |   +4   |   +2    | Head1 strongly wants, Head0 neutral → rank 3
-
-# Block 0: Head 0 strongly wants, Head 1 strongly rejects
-set_key(0, 0, [1.0, 1.0, 1.0, 1.0])      # head0: +4
-set_key(0, 1, [-1.0, -1.0, -1.0, -1.0])  # head1: -4
-
-# Block 1: Head 1 strongly wants, Head 0 strongly rejects
-set_key(1, 0, [-1.0, -1.0, -1.0, -1.0])  # head0: -4
-set_key(1, 1, [1.0, 1.0, 1.0, 1.0])      # head1: +4
-
-# Block 2: Both heads want equally (highest average)
-set_key(2, 0, [1.0, 1.0, 1.0, 1.0])      # head0: +4
-set_key(2, 1, [1.0, 1.0, 1.0, 1.0])      # head1: +4
-
-# Block 3: Both heads want moderately
-set_key(3, 0, [0.75, 0.75, 0.75, 0.75])  # head0: +3
-set_key(3, 1, [0.75, 0.75, 0.75, 0.75])  # head1: +3
-
-# Block 4: Head 0 strongly wants, Head 1 neutral
-set_key(4, 0, [1.0, 1.0, 1.0, 1.0])      # head0: +4
-set_key(4, 1, [0.0, 0.0, 0.0, 0.0])      # head1: 0
-
-# Block 5: Head 1 strongly wants, Head 0 neutral
-set_key(5, 0, [0.0, 0.0, 0.0, 0.0])      # head0: 0
-set_key(5, 1, [1.0, 1.0, 1.0, 1.0])      # head1: +4
-
-# ============================================================
-# Run selection
-# ============================================================
-
-# Query on same device as metadata
-query = torch.ones(1, 4, 4, device=device)  # GQA: 4 query heads → 2 KV heads
-
-ctx = PolicyContext(
-    query_chunk_idx=0,
-    num_query_chunks=1,
-    layer_id=0,
-    query=query,
-    is_prefill=False,
-    block_size=1024,
-    total_kv_len=6144,
-)
-
-available = list(range(6))
-selected = quest.select_blocks(available, ctx)
-
-# ============================================================
-# Verify: Averaging behavior
-# ============================================================
-
-# topk=2, so only blocks 2 (+4 avg) and 3 (+3 avg) should be selected
-assert len(selected) == 2, f"Expected 2 blocks, got {len(selected)}"
-assert selected == [2, 3], f"Expected [2, 3], got {selected}"
-
-# Key insight: blocks 0 and 1 have score +4 for ONE head,
-# but they cancel out due to averaging with the other head's -4
-assert 0 not in selected, "Block 0 should NOT be selected (head scores cancel out)"
-assert 1 not in selected, "Block 1 should NOT be selected (head scores cancel out)"
-
-# Blocks 4 and 5 have +4 for one head, 0 for other → avg=+2
-# But +2 < +3 (block 3), so they don't make the top-2
-assert 4 not in selected, "Block 4 avg=+2 < block 3 avg=+3"
-assert 5 not in selected, "Block 5 avg=+2 < block 3 avg=+3"
-
-print("✓ Block 2 selected: both heads want it (+4, +4) → avg=+4")
-print("✓ Block 3 selected: both heads want it (+3, +3) → avg=+3")
-print("✓ Block 0 NOT selected: head0=+4, head1=-4 → avg=0 (cancel out)")
-print("✓ Block 1 NOT selected: head0=-4, head1=+4 → avg=0 (cancel out)")
-print("✓ Block 4 NOT selected: head0=+4, head1=0 → avg=+2 (lower rank)")
-print("✓ Block 5 NOT selected: head0=0, head1=+4 → avg=+2 (lower rank)")
-
-# Verify metadata is on correct device
-assert metadata.key_min.device.type == device.type, f"key_min on wrong device: {metadata.key_min.device}"
-assert metadata.key_max.device.type == device.type, f"key_max on wrong device: {metadata.key_max.device}"
-print(f"✓ Metadata stored on {device.type.upper()}")
-
-print("\ntest_quest_policy: PASSED")

tests/test_ruler.py

@@ -41,6 +41,7 @@ from pathlib import Path
 from typing import List, Dict, Tuple, Optional
 
 from nanovllm import LLM, SamplingParams
+from nanovllm.utils.density_observer import DensityObserver
 
 
 # ============================================================
@@ -48,11 +49,67 @@ from nanovllm import LLM, SamplingParams
 # ============================================================
 
 DEFAULT_DATA_DIR = Path(__file__).parent / "data/ruler_64k"
+
+
+# ============================================================
+# Chat Template Conversion
+# ============================================================
+
+def convert_llama_to_glm4_format(prompt: str) -> str:
+    """
+    Convert Llama 3 chat template format to GLM-4 format.
+
+    Llama 3 format:
+        <|begin_of_text|><|start_header_id|>user<|end_header_id|>
+
+        {user_content}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+        {assistant_prefix}
+
+    GLM-4 format:
+        [gMASK]<sop><|user|>
+        {user_content}<|assistant|>
+        {assistant_prefix}
+    """
+    # Split into user content and assistant prefix
+    parts = prompt.split("<|eot_id|><|start_header_id|>assistant<|end_header_id|>")
+
+    # Extract user content (remove Llama header tokens)
+    user_content = parts[0]
+    user_content = user_content.replace("<|begin_of_text|>", "")
+    user_content = user_content.replace("<|start_header_id|>user<|end_header_id|>", "")
+    user_content = user_content.strip()
+
+    # Extract assistant prefix (if exists)
+    assistant_prefix = ""
+    if len(parts) > 1:
+        assistant_prefix = parts[1].replace("<|eot_id|>", "").strip()
+
+    # Apply GLM-4 format
+    glm_prompt = f"[gMASK]<sop><|user|>\n{user_content}<|assistant|>"
+    if assistant_prefix:
+        glm_prompt += f"\n{assistant_prefix}"
+
+    return glm_prompt
+
+
+def is_glm_model(model_path: str) -> bool:
+    """Check if the model is a GLM model based on config."""
+    from transformers import AutoConfig
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    return getattr(config, 'model_type', '') == 'chatglm'
+
+
+def convert_prompt_for_model(prompt: str, model_path: str) -> str:
+    """Convert prompt format based on model type."""
+    if is_glm_model(model_path):
+        return convert_llama_to_glm4_format(prompt)
+    return prompt  # Keep original format for Llama and other models
 DEFAULT_MODEL = os.path.expanduser("~/models/Llama-3.1-8B-Instruct")
 # Note: max_model_len must be > max_input_len to leave room for output tokens
 # 64k benchmark has inputs up to 65536 tokens, so we need 65536 + 128 = 65664
 DEFAULT_MAX_MODEL_LEN = 65664
-DEFAULT_MAX_NEW_TOKENS = 128  # Larger for multi-value tasks
+DEFAULT_MAX_NEW_TOKENS = 16  # Sufficient for NIAH single-value answers
 
 # Task categories for evaluation
 NIAH_TASKS = ["niah_single_1", "niah_single_2", "niah_single_3",
@@ -161,6 +218,7 @@ def run_task_test(
     verbose: bool = True,
     llm_factory: Optional[callable] = None,
     fresh_llm: bool = False,
+    model_path: Optional[str] = None,
 ) -> Dict:
     """
     Run test for a single RULER task.
@@ -198,6 +256,9 @@ def run_task_test(
     for sample in samples:
         idx = sample.get("index", sample["_local_idx"])
         prompt = sample["input"]
+        # Convert prompt format for GLM models
+        if model_path:
+            prompt = convert_prompt_for_model(prompt, model_path)
         expected = sample["outputs"]
 
         # Fresh LLM mode: reinitialize for each sample
@@ -263,7 +324,7 @@ def run_ruler_benchmark(
     max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
     enable_cpu_offload: bool = False,
     num_gpu_blocks: int = 4,
-    block_size: int = 1024,
+    block_size: int = 4096,
     num_kv_buffers: int = 4,
     gpu_utilization: float = 0.9,
     enforce_eager: bool = True,
@@ -275,6 +336,7 @@ def run_ruler_benchmark(
     sparse_samples: int = 128,
     sparse_block_size: int = 128,
     sparse_stride: int = 8,
+    dtype: Optional[str] = None,
 ) -> Dict:
     """
     Run RULER benchmark on multiple tasks.
@@ -320,6 +382,16 @@ def run_ruler_benchmark(
         print(f"Fresh LLM mode: {fresh_llm}")
         print(f"{'='*60}")
 
+    # Enable DensityObserver for XAttention BSA
+    if sparse_policy and sparse_policy.upper() == "XATTN_BSA":
+        DensityObserver.enable()
+        DensityObserver.complete_reset()
+        # Set mode for correct density interpretation
+        DensityObserver.set_mode("offload" if enable_cpu_offload else "gpu_only")
+        if not json_output:
+            mode_str = "offload" if enable_cpu_offload else "gpu_only"
+            print(f"[DensityObserver] Enabled for XAttention BSA (mode: {mode_str})")
+
     # LLM initialization kwargs
     llm_kwargs = {
         "max_model_len": max_model_len,
@@ -329,6 +401,8 @@ def run_ruler_benchmark(
         "kvcache_block_size": block_size,
         "enable_cpu_offload": enable_cpu_offload,
     }
+    if dtype:
+        llm_kwargs["dtype"] = dtype
     if enable_cpu_offload:
         llm_kwargs["num_gpu_blocks"] = num_gpu_blocks
         llm_kwargs["num_kv_buffers"] = num_kv_buffers
@@ -367,6 +441,7 @@ def run_ruler_benchmark(
             verbose=verbose and not json_output,
             llm_factory=create_llm,
             fresh_llm=fresh_llm,
+            model_path=model_path,
         )
         task_results.append(result)
 
@@ -407,6 +482,14 @@ def run_ruler_benchmark(
         print(f"{'-'*54}")
         print(f"{'TOTAL':<20} {total_correct}/{total_samples:<7} {overall_accuracy*100:>6.1f}%      {avg_score:.3f}")
         print(f"\nTime: {total_time:.1f}s")
+
+        # Print DensityObserver summary if enabled
+        if sparse_policy and sparse_policy.upper() == "XATTN_BSA" and DensityObserver.is_enabled():
+            print(f"\n{'='*60}")
+            print("Density Statistics (XAttention BSA)")
+            print(f"{'='*60}")
+            DensityObserver.print_summary()
+
         print(f"{'='*60}\n")
 
     results = {
@@ -464,8 +547,8 @@ if __name__ == "__main__":
                         help="Enable CPU offload mode")
     parser.add_argument("--num-gpu-blocks", type=int, default=4,
                         help="Number of GPU blocks for CPU offload (default: 4)")
-    parser.add_argument("--block-size", type=int, default=1024,
-                        help="KV cache block size (default: 1024)")
+    parser.add_argument("--block-size", type=int, default=4096,
+                        help="KV cache block size (default: 4096)")
     parser.add_argument("--num-kv-buffers", type=int, default=4,
                         help="Number of KV buffers for ring buffer (default: 4)")
     parser.add_argument("--gpu-utilization", type=float, default=0.9,
@@ -489,6 +572,8 @@ if __name__ == "__main__":
                         help="XAttention BSA: block size for estimation")
     parser.add_argument("--sparse-stride", type=int, default=8,
                         help="XAttention BSA: stride for Q/K downsampling")
+    parser.add_argument("--dtype", type=str, default=None,
+                        help="Model dtype (bfloat16, float16). Required for models with float32 default.")
 
     args = parser.parse_args()
 
@@ -526,6 +611,7 @@ if __name__ == "__main__":
         sparse_samples=args.sparse_samples,
         sparse_block_size=args.sparse_block_size,
         sparse_stride=args.sparse_stride,
+        dtype=args.dtype,
     )
 
     # Exit code (skip for json output mode)

tests/test_sequential.py

@@ -1,199 +0,0 @@
-"""
-Sequential inference test for LLM.
-
-Tests: After completing one prompt, the system can correctly handle
-a second prompt with a clean state (first prompt's KV cache deallocated).
-"""
-
-import os
-os.environ["NANOVLLM_LOG_LEVEL"] = "INFO"
-
-import argparse
-from nanovllm import LLM, SamplingParams
-from utils import generate_needle_prompt, check_needle_answer
-
-
-def run_sequential_test(
-    model_path: str,
-    max_model_len: int,
-    input_len: int,
-    num_gpu_blocks: int = 4,
-    block_size: int = 1024,
-    enable_cpu_offload: bool = False,
-    verbose: bool = True,
-) -> bool:
-    """
-    Run sequential inference test with two different prompts.
-
-    Each prompt has a different needle value. Both must be retrieved correctly.
-    """
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Sequential Inference Test")
-        print(f"{'='*60}")
-        print(f"Model: {model_path}")
-        print(f"Max model len: {max_model_len}")
-        print(f"Input length: {input_len}")
-        print(f"Block size: {block_size}")
-        print(f"CPU offload: {enable_cpu_offload}")
-        print(f"{'='*60}\n")
-
-    # Initialize LLM once
-    llm_kwargs = {
-        "enforce_eager": True,
-        "max_model_len": max_model_len,
-        "max_num_batched_tokens": max_model_len,
-        "enable_cpu_offload": enable_cpu_offload,
-        "kvcache_block_size": block_size,
-    }
-    if enable_cpu_offload:
-        llm_kwargs["num_gpu_blocks"] = num_gpu_blocks
-
-    llm = LLM(model_path, **llm_kwargs)
-
-    sampling_params = SamplingParams(
-        temperature=0.6,
-        max_tokens=32,
-    )
-
-    # ============================================================
-    # Test 1: First prompt with needle value "1234"
-    # ============================================================
-    needle_value_1 = "1234"
-    if verbose:
-        print(f"\n[Test 1] Generating prompt with needle value: {needle_value_1}")
-
-    prompt_1, expected_1 = generate_needle_prompt(
-        tokenizer=llm.tokenizer,
-        target_length=input_len,
-        needle_position=0.5,
-        needle_value=needle_value_1,
-    )
-
-    outputs_1 = llm.generate([prompt_1], sampling_params, use_tqdm=True)
-    output_text_1 = outputs_1[0]["text"]
-    passed_1 = check_needle_answer(output_text_1, expected_1)
-
-    if verbose:
-        print(f"  Expected: {expected_1}")
-        print(f"  Output: {output_text_1[:100]}...")
-        print(f"  Status: {'PASSED' if passed_1 else 'FAILED'}")
-
-    # ============================================================
-    # Test 2: Second prompt with needle value "5678"
-    # ============================================================
-    needle_value_2 = "5678"
-    if verbose:
-        print(f"\n[Test 2] Generating prompt with needle value: {needle_value_2}")
-
-    prompt_2, expected_2 = generate_needle_prompt(
-        tokenizer=llm.tokenizer,
-        target_length=input_len,
-        needle_position=0.5,
-        needle_value=needle_value_2,
-    )
-
-    outputs_2 = llm.generate([prompt_2], sampling_params, use_tqdm=True)
-    output_text_2 = outputs_2[0]["text"]
-    passed_2 = check_needle_answer(output_text_2, expected_2)
-
-    if verbose:
-        print(f"  Expected: {expected_2}")
-        print(f"  Output: {output_text_2[:100]}...")
-        print(f"  Status: {'PASSED' if passed_2 else 'FAILED'}")
-
-    # ============================================================
-    # Test 3: Third prompt - repeat first needle to ensure no cross-contamination
-    # ============================================================
-    needle_value_3 = "9999"
-    if verbose:
-        print(f"\n[Test 3] Generating prompt with needle value: {needle_value_3}")
-
-    prompt_3, expected_3 = generate_needle_prompt(
-        tokenizer=llm.tokenizer,
-        target_length=input_len,
-        needle_position=0.5,
-        needle_value=needle_value_3,
-    )
-
-    outputs_3 = llm.generate([prompt_3], sampling_params, use_tqdm=True)
-    output_text_3 = outputs_3[0]["text"]
-    passed_3 = check_needle_answer(output_text_3, expected_3)
-
-    if verbose:
-        print(f"  Expected: {expected_3}")
-        print(f"  Output: {output_text_3[:100]}...")
-        print(f"  Status: {'PASSED' if passed_3 else 'FAILED'}")
-
-    # ============================================================
-    # Summary
-    # ============================================================
-    all_passed = passed_1 and passed_2 and passed_3
-
-    if verbose:
-        print(f"\n{'='*60}")
-        print(f"Summary")
-        print(f"{'='*60}")
-        print(f"Test 1 (needle={needle_value_1}): {'PASSED' if passed_1 else 'FAILED'}")
-        print(f"Test 2 (needle={needle_value_2}): {'PASSED' if passed_2 else 'FAILED'}")
-        print(f"Test 3 (needle={needle_value_3}): {'PASSED' if passed_3 else 'FAILED'}")
-        print(f"Overall: {'PASSED' if all_passed else 'FAILED'}")
-        print(f"{'='*60}\n")
-
-    return all_passed
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Sequential inference test")
-    parser.add_argument(
-        "--model", "-m",
-        type=str,
-        default=os.path.expanduser("~/models/Qwen3-0.6B/"),
-        help="Path to model"
-    )
-    parser.add_argument(
-        "--max-model-len",
-        type=int,
-        default=36 * 1024,
-        help="Maximum model context length"
-    )
-    parser.add_argument(
-        "--input-len",
-        type=int,
-        default=8 * 1024,
-        help="Target input sequence length"
-    )
-    parser.add_argument(
-        "--num-gpu-blocks",
-        type=int,
-        default=2,
-        help="Number of GPU blocks for CPU offload"
-    )
-    parser.add_argument(
-        "--block-size",
-        type=int,
-        default=1024,
-        help="KV cache block size"
-    )
-    parser.add_argument(
-        "--enable-offload",
-        action="store_true",
-        help="Enable CPU offload"
-    )
-    args = parser.parse_args()
-
-    passed = run_sequential_test(
-        model_path=args.model,
-        max_model_len=args.max_model_len,
-        input_len=args.input_len,
-        num_gpu_blocks=args.num_gpu_blocks,
-        block_size=args.block_size,
-        enable_cpu_offload=args.enable_offload,
-        verbose=True,
-    )
-
-    if passed:
-        print("test_sequential: PASSED")
-    else:
-        print("test_sequential: FAILED")
-        exit(1)

tests/test_xattn_bsa.py

@@ -1,334 +0,0 @@
-"""
-Test XAttention + BSA with RULER benchmark data.
-
-Tests XAttention sparse attention correctness using RULER NIAH task.
-
-Attention methods:
-  - Prefill: XAttention + BSA (sparse) or FlashAttention (dense)
-  - Decode:  FlashAttention (always, since q_len=1)
-
-Usage (in compass conda env with BSA available):
-    CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/path/to/nano-vllm:$PYTHONPATH \
-        python tests/test_xattn_bsa.py --model ~/models/Llama-3.1-8B-Instruct
-
-    # Test with XAttention + BSA for prefill (default)
-    python tests/test_xattn_bsa.py --prefill-method xattn
-
-    # Test with FlashAttention for prefill (baseline)
-    python tests/test_xattn_bsa.py --prefill-method flash
-
-    # Test specific sample(s)
-    python tests/test_xattn_bsa.py --sample-id 0
-    python tests/test_xattn_bsa.py --sample-ids 0,1,2
-
-Note: Compatible with transformers 4.53+ (handles both old `past_key_value`
-      and new `past_key_values` API).
-"""
-
-import argparse
-import json
-import sys
-import torch
-from pathlib import Path
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.cache_utils import DynamicCache
-
-from nanovllm.ops.xattn import xattn_estimate
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
-
-
-# ============================================================
-# XAttention + BSA Functions
-# ============================================================
-
-def expand_kv_for_gqa(key_states, value_states, num_heads):
-    """Expand KV for Grouped Query Attention."""
-    num_kv_heads = key_states.shape[1]
-    if num_heads == num_kv_heads:
-        return key_states, value_states
-    num_groups = num_heads // num_kv_heads
-    return key_states.repeat_interleave(num_groups, dim=1), value_states.repeat_interleave(num_groups, dim=1)
-
-
-def flash_attention_forward(query_states, key_states, value_states, is_causal=True):
-    """Standard FlashAttention."""
-    from flash_attn import flash_attn_func
-    q = query_states.transpose(1, 2)
-    k = key_states.transpose(1, 2)
-    v = value_states.transpose(1, 2)
-    return flash_attn_func(q, k, v, causal=is_causal).transpose(1, 2)
-
-
-def xattn_bsa_forward(query_states, key_states, value_states, threshold=0.9):
-    """XAttention + BSA sparse attention."""
-    from block_sparse_attn import block_sparse_attn_func
-
-    batch_size, num_heads, q_len, head_dim = query_states.shape
-    k_len = key_states.shape[2]
-
-    _, mask = xattn_estimate(
-        query_states, key_states,
-        chunk_size=16384, block_size=128, threshold=threshold,
-        use_triton=True, causal=True,
-    )
-
-    q_block_num = (q_len + 127) // 128
-    k_block_num = (k_len + 127) // 128
-
-    q = query_states.transpose(1, 2).reshape(q_len, num_heads, head_dim)
-    k = key_states.transpose(1, 2).reshape(k_len, num_heads, head_dim)
-    v = value_states.transpose(1, 2).reshape(k_len, num_heads, head_dim)
-    
-    __import__('pdb').set_trace()
-
-    output = block_sparse_attn_func(
-        q, k, v,
-        torch.tensor([0, q_len], dtype=torch.int32, device=q.device),
-        torch.tensor([0, k_len], dtype=torch.int32, device=k.device),
-        torch.ones(num_heads, dtype=torch.int32, device=q.device),
-        None,
-        mask[:, :, :q_block_num, :k_block_num].contiguous(),
-        q_len, k_len,
-        p_dropout=0.0, deterministic=True, is_causal=True,
-    )
-    return output.reshape(batch_size, q_len, num_heads, head_dim).transpose(1, 2)
-
-
-DEBUG = False  # Set to True to enable debugging
-
-def create_patched_forward(prefill_method="xattn", threshold=0.9):
-    """Create patched forward with configurable prefill method.
-
-    Args:
-        prefill_method: "xattn" for XAttention + BSA (sparse), "flash" for FlashAttention (dense)
-        threshold: XAttention threshold for block selection (only used when prefill_method="xattn")
-
-    Note:
-        - Prefill (q_len > 1): Uses specified prefill_method
-        - Decode (q_len = 1): Always uses FlashAttention (no sparse needed for single query)
-    """
-    call_count = [0]  # Mutable to track calls across layers
-
-    def patched_forward(
-        self,
-        hidden_states,
-        position_embeddings=None,
-        attention_mask=None,
-        past_key_value=None,   # Old API (transformers < 4.57)
-        past_key_values=None,  # New API (transformers >= 4.57)
-        cache_position=None,
-        **kwargs
-    ):
-        # Handle both old and new transformers API
-        kv_cache = past_key_values if past_key_values is not None else past_key_value
-
-        bsz, q_len, _ = hidden_states.size()
-        num_heads = self.config.num_attention_heads
-        num_kv_heads = self.config.num_key_value_heads
-        head_dim = self.head_dim
-
-        # Compute Q, K, V projections
-        query_states = self.q_proj(hidden_states).view(bsz, q_len, num_heads, head_dim).transpose(1, 2)
-        key_states = self.k_proj(hidden_states).view(bsz, q_len, num_kv_heads, head_dim).transpose(1, 2)
-        value_states = self.v_proj(hidden_states).view(bsz, q_len, num_kv_heads, head_dim).transpose(1, 2)
-
-        # Apply rotary position embedding
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        # Handle KV cache
-        if kv_cache is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = kv_cache.update(
-                key_states, value_states, self.layer_idx, cache_kwargs
-            )
-
-        # Expand KV for GQA
-        key_states_exp, value_states_exp = expand_kv_for_gqa(key_states, value_states, num_heads)
-
-        # Debug output
-        if DEBUG and self.layer_idx == 0:
-            call_count[0] += 1
-            if call_count[0] <= 5:
-                phase = "prefill" if q_len > 1 else "decode"
-                print(f"\n[DEBUG] Layer {self.layer_idx}, call {call_count[0]} ({phase}): q_len={q_len}, k_len={key_states_exp.shape[2]}")
-                print(f"  kv_cache is None: {kv_cache is None}")
-
-        # Choose attention method:
-        # - Prefill (q_len > 1): Use prefill_method (xattn or flash)
-        # - Decode (q_len = 1): Always use FlashAttention
-        is_prefill = q_len > 1
-
-        if is_prefill and prefill_method == "xattn":
-            # Prefill with XAttention + BSA (sparse)
-            attn_output = xattn_bsa_forward(query_states, key_states_exp, value_states_exp, threshold)
-        else:
-            # Prefill with FlashAttention (dense) OR Decode (always FlashAttention)
-            # Note: For decode (q_len=1), causal=False since single query attends to all KV
-            attn_output = flash_attention_forward(query_states, key_states_exp, value_states_exp, is_causal=is_prefill)
-
-        attn_output = self.o_proj(attn_output.transpose(1, 2).reshape(bsz, q_len, -1))
-        return attn_output, None
-
-    return patched_forward
-
-
-# ============================================================
-# Data & Evaluation
-# ============================================================
-
-def load_samples(filepath, indices=None):
-    """Load samples from JSONL file."""
-    samples = []
-    with open(filepath) as f:
-        for i, line in enumerate(f):
-            if indices is None or i in indices:
-                sample = json.loads(line)
-                sample["_idx"] = i
-                samples.append(sample)
-    return samples
-
-
-def string_match_all(output_text, expected_list):
-    """RULER metric: fraction of expected values found in output."""
-    output_lower = output_text.lower().replace('\n', ' ')
-    if not expected_list:
-        return 1.0
-    return sum(1.0 if exp.strip().lower() in output_lower else 0.0 for exp in expected_list) / len(expected_list)
-
-
-# ============================================================
-# Test
-# ============================================================
-
-def test_with_ruler_data(model_path, data_file, sample_ids, prefill_method="xattn", threshold=0.9, max_new_tokens=50):
-    """Test attention methods using RULER data.
-
-    Args:
-        prefill_method: "xattn" for XAttention + BSA, "flash" for FlashAttention
-    """
-    prefill_desc = "XAttention + BSA (sparse)" if prefill_method == "xattn" else "FlashAttention (dense)"
-
-    print("=" * 60)
-    print("RULER NIAH Attention Test")
-    print("=" * 60)
-    print(f"Data: {data_file}")
-    print(f"Samples: {sample_ids}")
-    print(f"Prefill method: {prefill_desc}")
-    print(f"Decode method:  FlashAttention (always)")
-    if prefill_method == "xattn":
-        print(f"XAttention threshold: {threshold}")
-
-    samples = load_samples(Path(data_file), set(sample_ids) if sample_ids else None)
-    if not samples:
-        print("No samples found!")
-        return False
-    print(f"Loaded {len(samples)} samples")
-
-    # Load model
-    print(f"\nLoading model: {model_path}")
-    tokenizer = AutoTokenizer.from_pretrained(model_path)
-    model = AutoModelForCausalLM.from_pretrained(
-        model_path, torch_dtype=torch.float16, device_map="cuda",
-        attn_implementation="eager",  # Will be patched
-    )
-    model.eval()
-
-    # Patch all layers
-    print(f"Patching attention layers...")
-    print(f"  - Prefill: {prefill_desc}")
-    print(f"  - Decode:  FlashAttention")
-    for idx, layer in enumerate(model.model.layers):
-        layer.self_attn.layer_idx = idx  # Ensure layer_idx is set
-        layer.self_attn.forward = create_patched_forward(prefill_method, threshold).__get__(
-            layer.self_attn, type(layer.self_attn)
-        )
-
-    total_score = 0.0
-    results = []
-
-    for sample in samples:
-        idx = sample["_idx"]
-        prompt = sample["input"]
-        expected = sample["outputs"]
-
-        inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
-        num_tokens = inputs["input_ids"].shape[1]
-        print(f"\n--- Sample {idx} ({num_tokens} tokens) ---")
-        print(f"Expected: {expected}")
-
-        with torch.no_grad():
-            output = model.generate(
-                inputs["input_ids"],
-                max_new_tokens=max_new_tokens,
-                do_sample=False,
-                pad_token_id=tokenizer.eos_token_id,
-            )
-        output_text = tokenizer.decode(output[0][num_tokens:], skip_special_tokens=True)
-        score = string_match_all(output_text, expected)
-        total_score += score
-
-        status = "✓ PASS" if score >= 0.5 else "✗ FAIL"
-        print(f"Output: '{output_text[:100]}...'")
-        print(f"Result: {status} (score={score:.2f})")
-        results.append({"idx": idx, "score": score, "passed": score >= 0.5})
-
-    avg_score = total_score / len(samples)
-    passed = sum(1 for r in results if r["passed"])
-
-    print(f"\n{'='*60}")
-    print(f"Results: {passed}/{len(samples)} passed, avg_score={avg_score:.3f}")
-    print(f"{'='*60}")
-
-    return avg_score >= 0.5
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="Test XAttention + BSA vs FlashAttention for prefill using RULER NIAH benchmark"
-    )
-    parser.add_argument("--model", default="~/models/Llama-3.1-8B-Instruct")
-    parser.add_argument("--data-file", default="tests/data/ruler_32k/niah_single_1/validation.jsonl")
-    parser.add_argument("--sample-id", type=int, default=None, help="Test single sample by index")
-    parser.add_argument("--sample-ids", type=str, default="", help="Test multiple samples (comma-separated)")
-    parser.add_argument("--prefill-method", choices=["xattn", "flash"], default="xattn",
-                        help="Prefill attention method: xattn (XAttention+BSA sparse) or flash (FlashAttention dense)")
-    parser.add_argument("--threshold", type=float, default=0.9, help="XAttention threshold (only for --prefill-method xattn)")
-    parser.add_argument("--max-new-tokens", type=int, default=50)
-    # Keep old option for backwards compatibility
-    parser.add_argument("--no-xattn", action="store_true", help="[Deprecated] Use --prefill-method flash instead")
-    args = parser.parse_args()
-
-    model_path = args.model.replace("~", "/home/zijie")
-
-    # Handle deprecated --no-xattn option
-    prefill_method = args.prefill_method
-    if args.no_xattn:
-        prefill_method = "flash"
-        print("Warning: --no-xattn is deprecated, use --prefill-method flash instead")
-
-    if args.sample_id is not None:
-        sample_ids = [args.sample_id]
-    elif args.sample_ids:
-        sample_ids = [int(x) for x in args.sample_ids.split(",")]
-    else:
-        sample_ids = [0]
-
-    # Check BSA availability if using xattn
-    if prefill_method == "xattn":
-        try:
-            from block_sparse_attn import block_sparse_attn_func
-            print("✓ BSA (Block Sparse Attention) available")
-        except ImportError:
-            print("✗ BSA not found. Install block_sparse_attn or use --prefill-method flash")
-            sys.exit(1)
-
-    if test_with_ruler_data(model_path, args.data_file, sample_ids, prefill_method, args.threshold, args.max_new_tokens):
-        print("\ntest_xattn_bsa: PASSED")
-    else:
-        print("\ntest_xattn_bsa: FAILED")
-        sys.exit(1)
-
-
-if __name__ == "__main__":
-    main()

tests/test_xattn_chunked.py

@@ -1,259 +0,0 @@
-"""
-Test: Compare xattn_estimate vs xattn_estimate_chunked
-Verify that chunked estimation with EXTERNAL chunking produces the same mask as standard estimation.
-
-Uses real QKV data captured from model inference.
-"""
-
-import sys
-import os
-import torch
-import warnings
-
-from nanovllm.ops.xattn import xattn_estimate, xattn_estimate_chunked
-
-# ============================================================
-# Configuration
-# ============================================================
-
-BLOCK_SIZE = 64
-STRIDE = 4
-THRESHOLD = 0.9
-CHUNK_SIZE = 4096
-
-# Default QKV data directory (relative to project root)
-DEFAULT_QKV_DIR = os.path.join(os.path.dirname(os.path.dirname(__file__)), "results", "kvcache")
-
-# ============================================================
-# Utility Functions
-# ============================================================
-
-def load_qkv(path):
-    """Load saved QKV data."""
-    data = torch.load(path, map_location="cpu", weights_only=False)
-    print(f"Loaded: {path}")
-    print(f"  Query shape: {data['query'].shape}")
-    print(f"  Key shape: {data['key'].shape}")
-    print(f"  Layer: {data['layer_id']}, Density: {data['density']:.2%}")
-    return data
-
-
-def compare_masks(mask1, mask2, name1="standard", name2="chunked"):
-    """Compare two masks and report differences."""
-    if mask1.shape != mask2.shape:
-        print(f"Shape mismatch: {name1}={mask1.shape}, {name2}={mask2.shape}")
-        return False
-
-    diff = (mask1 != mask2).sum().item()
-    total = mask1.numel()
-    match_rate = (total - diff) / total * 100
-
-    print(f"  Match rate: {match_rate:.4f}% ({total - diff}/{total})")
-
-    if diff > 0:
-        diff_indices = torch.where(mask1 != mask2)
-        print(f"  First 5 diff positions: {list(zip(*[idx[:5].tolist() for idx in diff_indices]))}")
-
-    return diff == 0
-
-
-def run_chunked_externally(query, key, q_start_pos, block_size, stride, threshold, chunk_size):
-    """
-    Run xattn_estimate_chunked with EXTERNAL chunking.
-    This simulates how chunked prefill should be used in practice.
-    """
-    batch_size, num_heads, q_len, head_dim = query.shape
-    _, _, k_len, _ = key.shape
-
-    q_block_num = (q_len + block_size - 1) // block_size
-    k_block_num = (k_len + block_size - 1) // block_size
-
-    # If Q fits in one chunk, call directly
-    if q_len <= chunk_size:
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            return xattn_estimate_chunked(
-                query, key,
-                q_start_pos=q_start_pos,
-                block_size=block_size,
-                stride=stride,
-                threshold=threshold,
-                use_triton=True,
-                chunk_size=chunk_size,
-            )
-
-    # External chunking: split Q and call for each chunk
-    num_q_chunks = (q_len + chunk_size - 1) // chunk_size
-    print(f"    External chunking: {num_q_chunks} chunks")
-
-    combined_attn_sum = torch.zeros(
-        batch_size, num_heads, q_block_num, k_block_num,
-        dtype=query.dtype, device=query.device
-    )
-    combined_mask = torch.zeros(
-        batch_size, num_heads, q_block_num, k_block_num,
-        dtype=torch.bool, device=query.device
-    )
-
-    q_block_offset = 0
-    for q_chunk_idx in range(num_q_chunks):
-        q_chunk_start = q_chunk_idx * chunk_size
-        q_chunk_end = min((q_chunk_idx + 1) * chunk_size, q_len)
-
-        q_chunk = query[:, :, q_chunk_start:q_chunk_end, :]
-
-        # For causal attention, K accumulates up to current Q position
-        k_end = q_start_pos + q_chunk_end
-        k_chunk = key[:, :, :k_end, :]
-
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            attn_sum_chunk, mask_chunk = xattn_estimate_chunked(
-                q_chunk, k_chunk,
-                q_start_pos=q_start_pos + q_chunk_start,
-                block_size=block_size,
-                stride=stride,
-                threshold=threshold,
-                use_triton=True,
-                chunk_size=chunk_size,
-            )
-
-        # Place chunk results into combined output
-        chunk_q_blocks = mask_chunk.shape[2]
-        chunk_k_blocks = mask_chunk.shape[3]
-        combined_attn_sum[:, :, q_block_offset:q_block_offset+chunk_q_blocks, :chunk_k_blocks] = attn_sum_chunk
-        combined_mask[:, :, q_block_offset:q_block_offset+chunk_q_blocks, :chunk_k_blocks] = mask_chunk
-        q_block_offset += chunk_q_blocks
-
-    return combined_attn_sum, combined_mask
-
-
-def test_single_qkv(qkv_path):
-    """Test a single QKV file."""
-    data = load_qkv(qkv_path)
-    query = data["query"].cuda().to(torch.bfloat16)
-    key = data["key"].cuda().to(torch.bfloat16)
-
-    seq_len = query.shape[2]
-    print(f"\nTesting with seq_len={seq_len}")
-    print("=" * 60)
-
-    # Run standard xattn_estimate
-    print("[1] Running standard xattn_estimate...")
-    try:
-        attn_sum_std, mask_std = xattn_estimate(
-            query, key,
-            block_size=BLOCK_SIZE,
-            stride=STRIDE,
-            threshold=THRESHOLD,
-            chunk_size=CHUNK_SIZE,
-            use_triton=True,
-        )
-        print(f"  mask shape: {mask_std.shape}, density: {mask_std.float().mean().item():.4f}")
-    except Exception as e:
-        print(f"  ERROR: {e}")
-        import traceback
-        traceback.print_exc()
-        return False
-
-    # Run chunked xattn_estimate with EXTERNAL chunking
-    print("[2] Running chunked xattn_estimate (external chunking)...")
-    try:
-        attn_sum_chunked, mask_chunked = run_chunked_externally(
-            query, key,
-            q_start_pos=0,
-            block_size=BLOCK_SIZE,
-            stride=STRIDE,
-            threshold=THRESHOLD,
-            chunk_size=CHUNK_SIZE,
-        )
-        print(f"  mask shape: {mask_chunked.shape}, density: {mask_chunked.float().mean().item():.4f}")
-    except Exception as e:
-        print(f"  ERROR: {e}")
-        import traceback
-        traceback.print_exc()
-        return False
-
-    # Compare results
-    print("[3] Comparing results...")
-    chunked_q_blocks = mask_chunked.shape[2]
-    chunked_k_blocks = mask_chunked.shape[3]
-
-    # Extract comparable region from standard mask
-    mask_std_comparable = mask_std[:, :, :chunked_q_blocks, :chunked_k_blocks]
-
-    # Compare masks
-    masks_match = compare_masks(mask_std_comparable, mask_chunked, "standard", "chunked")
-
-    # Compare attn_sums
-    attn_sum_std_comparable = attn_sum_std[:, :, :chunked_q_blocks, :chunked_k_blocks]
-    if attn_sum_std_comparable.shape == attn_sum_chunked.shape:
-        attn_diff = (attn_sum_std_comparable - attn_sum_chunked).abs().max().item()
-        print(f"  Attn sum max diff: {attn_diff:.6f}")
-    else:
-        print(f"  Attn sum shape mismatch")
-
-    # Clean up GPU memory
-    del query, key, attn_sum_std, mask_std, attn_sum_chunked, mask_chunked
-    torch.cuda.empty_cache()
-
-    return masks_match
-
-
-# ============================================================
-# Main Test
-# ============================================================
-
-if __name__ == "__main__":
-    import argparse
-    parser = argparse.ArgumentParser(description="Test xattn_estimate vs xattn_estimate_chunked")
-    parser.add_argument("--qkv-dir", type=str, default=DEFAULT_QKV_DIR,
-                        help="Directory containing QKV files")
-    args = parser.parse_args()
-
-    # QKV files to test
-    qkv_files = [
-        os.path.join(args.qkv_dir, "qkv_3688.pt"),   # ~4K
-        os.path.join(args.qkv_dir, "qkv_7888.pt"),   # ~8K
-        os.path.join(args.qkv_dir, "qkv_15685.pt"),  # ~16K
-        os.path.join(args.qkv_dir, "qkv_32485.pt"),  # ~32K
-        os.path.join(args.qkv_dir, "qkv_64891.pt"),  # ~64K
-    ]
-
-    available_files = [p for p in qkv_files if os.path.exists(p)]
-
-    if not available_files:
-        print(f"No QKV file found in {args.qkv_dir}.")
-        print(f"Expected files: qkv_3688.pt, qkv_7888.pt, qkv_15685.pt, qkv_32485.pt, qkv_64891.pt")
-        sys.exit(1)
-
-    print(f"Found {len(available_files)} QKV files to test")
-    print(f"Testing EXTERNAL chunking (chunk_size={CHUNK_SIZE})")
-    print(f"Using Triton kernels")
-
-    all_passed = True
-    results = []
-
-    for qkv_path in available_files:
-        passed = test_single_qkv(qkv_path)
-        seq_len = int(os.path.basename(qkv_path).replace("qkv_", "").replace(".pt", ""))
-        results.append((seq_len, passed))
-        if not passed:
-            all_passed = False
-
-    # Summary
-    print("\n" + "=" * 60)
-    print("SUMMARY")
-    print("=" * 60)
-    for seq_len, passed in results:
-        status = "PASSED" if passed else "FAILED"
-        chunks = (seq_len + CHUNK_SIZE - 1) // CHUNK_SIZE
-        print(f"  seq_len={seq_len} ({chunks} chunk{'s' if chunks > 1 else ''}): {status}")
-
-    print("=" * 60)
-    if all_passed:
-        print("test_xattn_chunked: PASSED")
-        sys.exit(0)
-    else:
-        print("test_xattn_chunked: FAILED")
-        sys.exit(1)

tests/test_xattn_estimate_alignment.py

@@ -0,0 +1,365 @@
+"""
+Test: 验证 xattn_estimate 与 KV chunking kernels 的一致性
+
+使用真实 KV cache 数据，对比:
+1. xattn_estimate (高层 API)
+2. 三阶段 KV chunking (softmax_compute_partial_stats + merge + normalize)
+
+三阶段 KV chunking 流程:
+  1. softmax_compute_partial_stats: 计算每个 KV chunk 的 (m, l)
+  2. merge_softmax_stats: Host 端合并所有 chunks 的 stats
+  3. softmax_normalize_and_block_sum: 使用全局 stats 归一化
+
+支持两种数据格式:
+  1. offload 模式保存: {"query", "key", "stride", "threshold", "density", "layer_id"}
+  2. GPU-only 模式保存: {"Q", "K", "chunk_size", "block_size", "stride", "threshold", "mask", "attn_sums", ...}
+
+Usage:
+    # 使用 offload 模式数据
+    CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+        python tests/test_xattn_estimate_alignment.py
+
+    # 使用 GPU-only 模式数据
+    CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
+        python tests/test_xattn_estimate_alignment.py --gpuonly
+"""
+import sys
+sys.path.insert(0, "/home/zijie/Code/nano-vllm")
+
+import argparse
+import torch
+import math
+from nanovllm.ops.xattn import (
+    xattn_estimate,
+    flat_group_gemm_fuse_reshape,
+    softmax_compute_partial_stats,
+    softmax_normalize_and_block_sum,
+    merge_softmax_stats,
+    find_blocks_chunked,
+)
+
+# ============================================================
+# 命令行参数
+# ============================================================
+parser = argparse.ArgumentParser()
+parser.add_argument("--gpuonly", action="store_true", help="使用 GPU-only 模式保存的数据")
+parser.add_argument("--data-file", type=str, default=None, help="数据文件路径")
+parser.add_argument("--chunk-size", type=int, default=None, help="覆盖 CHUNK_SIZE (用于测试不同分块大小)")
+args = parser.parse_args()
+
+# ============================================================
+# 参数配置
+# ============================================================
+if args.gpuonly:
+    DATA_FILE = args.data_file or "/home/zijie/Code/nano-vllm/results/mask_alignment/gpuonly_layer0.pt"
+else:
+    DATA_FILE = args.data_file or "/home/zijie/Code/nano-vllm/results/kvcache/qkv_32485.pt"
+
+device = "cuda"
+
+# ============================================================
+# Step 1: 加载真实数据
+# ============================================================
+print("=" * 60)
+print("Step 1: 加载真实 KV cache 数据")
+print("=" * 60)
+
+data = torch.load(DATA_FILE, map_location="cpu")
+
+# 检测数据格式并加载
+if "Q" in data:
+    # GPU-only 模式保存的格式
+    print(f"[INFO] 检测到 GPU-only 模式数据格式")
+    Q = data["Q"].to(device)
+    K = data["K"].to(device)
+    BSA_BLOCK_SIZE = data.get("block_size", 128)
+    CHUNK_SIZE = data.get("chunk_size", 4096)
+    STRIDE = data.get("stride", 8)
+    THRESHOLD = data.get("threshold", 0.9)
+    if isinstance(THRESHOLD, torch.Tensor):
+        THRESHOLD = THRESHOLD.item()
+    # GPU-only 模式保存了 mask 和 attn_sums，可以用于验证
+    saved_mask = data.get("mask", None)
+    saved_attn_sums = data.get("attn_sums", None)
+    saved_density = None  # GPU-only 模式没有保存 density
+    layer_id = 0  # GPU-only 只保存 layer 0
+else:
+    # offload 模式保存的格式
+    print(f"[INFO] 检测到 offload 模式数据格式")
+    Q = data["query"].to(device)
+    K = data["key"].to(device)
+    BSA_BLOCK_SIZE = 128
+    CHUNK_SIZE = 4096
+    STRIDE = data["stride"]
+    THRESHOLD = data["threshold"]
+    if isinstance(THRESHOLD, torch.Tensor):
+        THRESHOLD = THRESHOLD[0].item()
+    saved_mask = None
+    saved_attn_sums = None
+    saved_density = data.get("density", None)
+    layer_id = data.get("layer_id", 0)
+
+batch_size, num_heads, seq_len, head_dim = Q.shape
+
+# 命令行覆盖 CHUNK_SIZE
+if args.chunk_size is not None:
+    CHUNK_SIZE = args.chunk_size
+    print(f"[INFO] 使用命令行指定的 CHUNK_SIZE={CHUNK_SIZE}")
+
+print(f"Q shape: {Q.shape}")
+print(f"K shape: {K.shape}")
+if saved_density is not None:
+    print(f"Data layer_id: {layer_id}, saved density: {saved_density:.4f}")
+else:
+    print(f"Data layer_id: {layer_id}")
+print(f"使用参数: STRIDE={STRIDE}, THRESHOLD={THRESHOLD}, CHUNK_SIZE={CHUNK_SIZE}, BSA_BLOCK_SIZE={BSA_BLOCK_SIZE}")
+print()
+
+# ============================================================
+# Step 2: 使用 xattn_estimate 高层 API
+# ============================================================
+print("=" * 60)
+print("Step 2: 调用 xattn_estimate (高层 API)")
+print("=" * 60)
+
+attn_sums_api, mask_api = xattn_estimate(
+    Q, K,
+    block_size=BSA_BLOCK_SIZE,
+    stride=STRIDE,
+    threshold=THRESHOLD,
+    chunk_size=CHUNK_SIZE,
+    causal=True,
+)
+
+# 裁剪到有效区域
+q_blocks = (seq_len + BSA_BLOCK_SIZE - 1) // BSA_BLOCK_SIZE
+k_blocks = (seq_len + BSA_BLOCK_SIZE - 1) // BSA_BLOCK_SIZE
+mask_api_valid = mask_api[:, :, :q_blocks, :k_blocks]
+
+# 计算 density (causal)
+causal_mask = torch.tril(torch.ones(q_blocks, k_blocks, device=device, dtype=torch.bool))
+total_api = causal_mask.sum().item() * batch_size * num_heads
+selected_api = (mask_api_valid & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()
+density_api = selected_api / total_api
+
+print(f"mask_api shape (padded): {mask_api.shape}")
+print(f"mask_api_valid shape:    {mask_api_valid.shape}")
+print(f"[xattn_estimate] density: {density_api:.6f} (selected={selected_api}, total={total_api})")
+print()
+
+# ============================================================
+# Step 3: 三阶段 KV Chunking
+# ============================================================
+print("=" * 60)
+print("Step 3: 三阶段 KV Chunking")
+print("=" * 60)
+print("  1) 每个 KV chunk 计算 partial stats")
+print("  2) Host 端合并 stats")
+print("  3) 使用全局 stats 归一化并计算 block sums")
+print()
+
+# 计算 padding 参数
+k_num_to_pad = ((seq_len + CHUNK_SIZE - 1) // CHUNK_SIZE) * CHUNK_SIZE - seq_len
+q_num_to_pad = ((seq_len + CHUNK_SIZE - 1) // CHUNK_SIZE) * CHUNK_SIZE - seq_len
+q_chunk_num = (seq_len + q_num_to_pad) // CHUNK_SIZE
+kv_chunk_num = (seq_len + k_num_to_pad) // CHUNK_SIZE
+
+k_block_num = (seq_len + k_num_to_pad) // BSA_BLOCK_SIZE
+q_block_num = (seq_len + q_num_to_pad) // BSA_BLOCK_SIZE
+
+reshaped_chunk_size = CHUNK_SIZE // STRIDE
+reshaped_block_size = BSA_BLOCK_SIZE // STRIDE
+k_reshaped_seq_len = (seq_len + k_num_to_pad) // STRIDE
+k_reshaped_num_to_pad = k_num_to_pad // STRIDE
+num_blocks_per_chunk = reshaped_chunk_size // reshaped_block_size
+kv_reshaped_chunk_size = CHUNK_SIZE // STRIDE
+
+print(f"seq_len: {seq_len}, q_chunk_num: {q_chunk_num}, kv_chunk_num: {kv_chunk_num}")
+print()
+
+# Padding
+if k_num_to_pad > 0:
+    K_padded = torch.nn.functional.pad(K, (0, 0, 0, k_num_to_pad), value=0)
+else:
+    K_padded = K
+
+if q_num_to_pad > 0:
+    Q_padded = torch.nn.functional.pad(Q, (0, 0, 0, q_num_to_pad), value=0)
+else:
+    Q_padded = Q
+
+# Softmax scale
+norm = 1.0
+scale = 1.4426950408889634 / math.sqrt(head_dim) / STRIDE / norm
+
+simple_mask_list = []
+
+for q_chunk_idx in range(q_chunk_num):
+    q_start = q_chunk_idx * reshaped_chunk_size * STRIDE
+    q_end = q_start + reshaped_chunk_size * STRIDE
+    Q_chunk = Q_padded[:, :, q_start:q_end, :]
+
+    chunk_start = (k_block_num - q_block_num) * reshaped_block_size + q_chunk_idx * reshaped_chunk_size
+    chunk_end = chunk_start + reshaped_chunk_size
+
+    # 阶段 1: 每个 KV chunk 计算 partial stats 和 raw scores
+    m_chunks = []
+    l_chunks = []
+    attn_weights_chunks = []
+
+    for kv_chunk_idx in range(kv_chunk_num):
+        kv_start = kv_chunk_idx * CHUNK_SIZE
+        kv_end = kv_start + CHUNK_SIZE
+        K_chunk = K_padded[:, :, kv_start:kv_end, :]
+
+        # KV offset in reshaped space
+        kv_offset_reshaped = kv_chunk_idx * kv_reshaped_chunk_size
+
+        # 计算 raw attention scores
+        attn_weights_kv = flat_group_gemm_fuse_reshape(
+            Q_chunk, K_chunk, STRIDE,
+            chunk_start=chunk_start,
+            chunk_end=chunk_end,
+            is_causal=False,  # K 不完整，不能在这里用 causal
+        )
+        attn_weights_chunks.append(attn_weights_kv)
+
+        # 计算 partial stats (带 causal mask)
+        m_partial, l_partial = softmax_compute_partial_stats(
+            attn_weights_kv,
+            reshaped_block_size,
+            min(4096, reshaped_block_size),
+            scale,
+            chunk_start=chunk_start,
+            kv_offset=kv_offset_reshaped,
+            is_causal=True,
+        )
+        m_chunks.append(m_partial)
+        l_chunks.append(l_partial)
+
+    # 阶段 2: Host 端合并 stats
+    m_global, l_global = merge_softmax_stats(m_chunks, l_chunks)
+
+    # 阶段 3: 使用全局 stats 归一化并计算 block sums
+    attn_sum_per_kv = []
+    for kv_chunk_idx, attn_weights_kv in enumerate(attn_weights_chunks):
+        kv_offset_reshaped = kv_chunk_idx * kv_reshaped_chunk_size
+        attn_sum_kv = softmax_normalize_and_block_sum(
+            attn_weights_kv,
+            m_global,
+            l_global,
+            reshaped_block_size,
+            min(4096, reshaped_block_size),
+            chunk_start=chunk_start,
+            real_q_len=k_reshaped_seq_len - k_reshaped_num_to_pad,
+            scale=scale,
+            kv_offset=kv_offset_reshaped,
+            is_causal=True,
+        )
+        attn_sum_per_kv.append(attn_sum_kv)
+
+    # 拼接各 KV chunk 的 block sums
+    attn_sum_concat = torch.cat(attn_sum_per_kv, dim=-1)
+
+    # 选择 blocks
+    simple_mask = find_blocks_chunked(
+        attn_sum_concat,
+        current_index=k_block_num - q_block_num + q_chunk_idx * num_blocks_per_chunk,
+        threshold=THRESHOLD,
+        num_to_choose=None,
+        decoding=False,
+        mode="prefill",
+        causal=True,
+    )
+    simple_mask_list.append(simple_mask)
+
+    print(f"  Q chunk {q_chunk_idx}: merged {kv_chunk_num} KV chunks, attn_sum shape={attn_sum_concat.shape}")
+
+mask_kv_chunking = torch.cat(simple_mask_list, dim=2)
+
+# 应用与 xattn_estimate 相同的 causal mask 后处理 (xattn.py 第 1300-1306 行)
+mask_kv_chunking[:, :, -q_block_num:, -q_block_num:] = torch.where(
+    torch.tril(torch.ones(q_block_num, q_block_num, dtype=bool, device=device), diagonal=0),
+    mask_kv_chunking[:, :, -q_block_num:, -q_block_num:],
+    False,
+)
+
+mask_kv_chunking_valid = mask_kv_chunking[:, :, :q_blocks, :k_blocks]
+selected_kv = (mask_kv_chunking_valid & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()
+density_kv = selected_kv / total_api
+
+print()
+print(f"[KV chunking] density: {density_kv:.6f} (selected={selected_kv}, total={total_api})")
+print()
+
+# ============================================================
+# Step 4: 对比结果
+# ============================================================
+print("=" * 60)
+print("Step 4: 对比结果")
+print("=" * 60)
+print()
+
+mask_total = mask_api_valid.numel()
+mask_diff = (mask_api_valid != mask_kv_chunking_valid).sum().item()
+
+print("| 方法 | density | 与 API 差异 | Mask 差异 |")
+print("|------|---------|-------------|-----------|")
+print(f"| xattn_estimate API | {density_api:.6f} | - | - |")
+print(f"| KV chunking | {density_kv:.6f} | {abs(density_api - density_kv):.6f} | {100*mask_diff/mask_total:.4f}% |")
+print()
+
+passed = abs(density_api - density_kv) < 1e-6 and mask_diff / mask_total < 0.001
+
+# ============================================================
+# Step 5: 与 GPU-only 保存的数据对比 (如果有)
+# ============================================================
+if saved_mask is not None or saved_attn_sums is not None:
+    print("=" * 60)
+    print("Step 5: 与 GPU-only 保存的数据对比")
+    print("=" * 60)
+    print()
+
+    if saved_mask is not None:
+        saved_mask_gpu = saved_mask.to(device)
+        # 比较 mask
+        mask_saved_diff = (mask_api_valid != saved_mask_gpu).sum().item()
+        mask_saved_total = saved_mask_gpu.numel()
+        print(f"| xattn_estimate vs GPU-only saved mask | 差异 blocks: {mask_saved_diff} / {mask_saved_total} ({100*mask_saved_diff/mask_saved_total:.4f}%) |")
+
+        if mask_saved_diff == 0:
+            print("✅ mask 与 GPU-only 保存完全一致")
+        else:
+            print("❌ mask 与 GPU-only 保存存在差异")
+            passed = False
+
+    if saved_attn_sums is not None:
+        saved_attn_sums_gpu = saved_attn_sums.to(device)
+        # 需要从 xattn_estimate 获取 attn_sums
+        # 重新调用一次获取 attn_sums
+        attn_sums_check, _ = xattn_estimate(
+            Q, K,
+            block_size=BSA_BLOCK_SIZE,
+            stride=STRIDE,
+            threshold=THRESHOLD,
+            chunk_size=CHUNK_SIZE,
+            causal=True,
+        )
+        attn_sums_check_valid = attn_sums_check[:, :, :q_blocks, :k_blocks]
+
+        max_diff = (attn_sums_check_valid - saved_attn_sums_gpu).abs().max().item()
+        mean_diff = (attn_sums_check_valid - saved_attn_sums_gpu).abs().mean().item()
+        print(f"| xattn_estimate vs GPU-only saved attn_sums | max diff: {max_diff:.6e}, mean diff: {mean_diff:.6e} |")
+
+        if max_diff < 1e-5:
+            print("✅ attn_sums 与 GPU-only 保存一致")
+        else:
+            print("❌ attn_sums 与 GPU-only 保存存在差异")
+            passed = False
+
+    print()
+
+if passed:
+    print("test_xattn_estimate_alignment: PASSED")
+else:
+    print("test_xattn_estimate_alignment: FAILED")

tests/test_xattn_estimate_chunked.py

@@ -1,244 +0,0 @@
-"""
-Test: Compare xattn_estimate vs xattn_estimate_chunked
-
-Verify that chunked estimation with EXTERNAL chunking produces the same mask
-as standard estimation. This ensures the chunked version can be used in
-chunked prefill scenarios without accuracy loss.
-
-Usage:
-    CUDA_VISIBLE_DEVICES=0 PYTHONPATH=/home/zijie/Code/nano-vllm:$PYTHONPATH \
-        python tests/test_xattn_estimate_chunked.py
-"""
-
-import sys
-import traceback
-import torch
-from nanovllm.ops.xattn import xattn_estimate, xattn_estimate_chunked
-
-# ============================================================
-# Configuration
-# ============================================================
-
-# Configuration for xattn_estimate_chunked consistency test.
-# Key requirements for 100% match:
-# 1. Use matching chunk_size for both standard and chunked versions
-# 2. Use same random seed for reproducibility
-# Note: Tiny differences (~0.000001) may occur at boundary cases due to
-# floating point precision in cumulative sum calculations.
-BLOCK_SIZE = 64
-STRIDE = 4
-THRESHOLD = 0.9
-CHUNK_SIZE = 4096  # External chunking size
-
-# Test sequence lengths
-TEST_SEQ_LENS = [4096, 8192, 16384, 32768]
-
-# ============================================================
-# Utility Functions
-# ============================================================
-
-def compare_masks(mask1, mask2, name1="standard", name2="chunked"):
-    """Compare two masks and report differences."""
-    if mask1.shape != mask2.shape:
-        print(f"  Shape mismatch: {name1}={mask1.shape}, {name2}={mask2.shape}")
-        return False
-
-    diff = (mask1 != mask2).sum().item()
-    total = mask1.numel()
-    match_rate = (total - diff) / total * 100
-
-    print(f"  Match rate: {match_rate:.4f}% ({total - diff}/{total})")
-
-    if diff > 0:
-        diff_indices = torch.where(mask1 != mask2)
-        print(f"  First 5 diff positions: {list(zip(*[idx[:5].tolist() for idx in diff_indices]))}")
-
-    return diff == 0
-
-
-def run_chunked_externally(query, key, block_size, stride, threshold, chunk_size):
-    """
-    Run xattn_estimate_chunked with EXTERNAL chunking.
-    This simulates how chunked prefill should be used in practice.
-    """
-    batch_size, num_heads, q_len, head_dim = query.shape
-    _, _, k_len, _ = key.shape
-
-    q_block_num = (q_len + block_size - 1) // block_size
-    k_block_num = (k_len + block_size - 1) // block_size
-
-    # If Q fits in one chunk, call directly
-    if q_len <= chunk_size:
-        return xattn_estimate_chunked(
-            query, key,
-            q_start_pos=0,
-            block_size=block_size,
-            stride=stride,
-            threshold=threshold,
-            use_triton=True,
-            chunk_size=chunk_size,
-        )
-
-    # External chunking: split Q and call for each chunk
-    num_q_chunks = (q_len + chunk_size - 1) // chunk_size
-    print(f"    External chunking: {num_q_chunks} chunks")
-
-    combined_attn_sum = torch.zeros(
-        batch_size, num_heads, q_block_num, k_block_num,
-        dtype=query.dtype, device=query.device
-    )
-    combined_mask = torch.zeros(
-        batch_size, num_heads, q_block_num, k_block_num,
-        dtype=torch.bool, device=query.device
-    )
-
-    q_block_offset = 0
-    for q_chunk_idx in range(num_q_chunks):
-        q_chunk_start = q_chunk_idx * chunk_size
-        q_chunk_end = min((q_chunk_idx + 1) * chunk_size, q_len)
-
-        q_chunk = query[:, :, q_chunk_start:q_chunk_end, :]
-
-        # For causal attention, K accumulates up to current Q position
-        # q_start_pos=0 means Q starts at position 0 in the full sequence
-        # K is [0, q_chunk_end) for causal attention
-        k_end = q_chunk_end
-        k_chunk = key[:, :, :k_end, :]
-
-        attn_sum_chunk, mask_chunk = xattn_estimate_chunked(
-            q_chunk, k_chunk,
-            q_start_pos=q_chunk_start,
-            block_size=block_size,
-            stride=stride,
-            threshold=threshold,
-            use_triton=True,
-            chunk_size=chunk_size,
-        )
-
-        # Place chunk results into combined output
-        chunk_q_blocks = mask_chunk.shape[2]
-        chunk_k_blocks = mask_chunk.shape[3]
-        combined_attn_sum[:, :, q_block_offset:q_block_offset+chunk_q_blocks, :chunk_k_blocks] = attn_sum_chunk
-        combined_mask[:, :, q_block_offset:q_block_offset+chunk_q_blocks, :chunk_k_blocks] = mask_chunk
-        q_block_offset += chunk_q_blocks
-
-    return combined_attn_sum, combined_mask
-
-
-def test_single_seq_len(seq_len, num_heads=32, head_dim=128):
-    """Test a single sequence length."""
-    print(f"\nTesting seq_len={seq_len}")
-    print("=" * 60)
-
-    # Generate random Q/K
-    query = torch.randn(1, num_heads, seq_len, head_dim, device="cuda", dtype=torch.bfloat16)
-    key = torch.randn(1, num_heads, seq_len, head_dim, device="cuda", dtype=torch.bfloat16)
-
-    # Run standard xattn_estimate
-    print("[1] Running standard xattn_estimate...")
-    try:
-        attn_sum_std, mask_std = xattn_estimate(
-            query, key,
-            block_size=BLOCK_SIZE,
-            stride=STRIDE,
-            threshold=THRESHOLD,
-            chunk_size=CHUNK_SIZE,
-            use_triton=True,
-            causal=True,
-        )
-        density_std = mask_std.float().mean().item()
-        print(f"  mask shape: {mask_std.shape}, density: {density_std:.4f}")
-    except Exception as e:
-        print(f"  ERROR: {e}")
-        traceback.print_exc()
-        return False
-
-    # Run chunked xattn_estimate with EXTERNAL chunking
-    print("[2] Running chunked xattn_estimate (external chunking)...")
-    try:
-        attn_sum_chunked, mask_chunked = run_chunked_externally(
-            query, key,
-            block_size=BLOCK_SIZE,
-            stride=STRIDE,
-            threshold=THRESHOLD,
-            chunk_size=CHUNK_SIZE,
-        )
-        density_chunked = mask_chunked.float().mean().item()
-        print(f"  mask shape: {mask_chunked.shape}, density: {density_chunked:.4f}")
-    except Exception as e:
-        print(f"  ERROR: {e}")
-        traceback.print_exc()
-        return False
-
-    # Compare results
-    print("[3] Comparing results...")
-    chunked_q_blocks = mask_chunked.shape[2]
-    chunked_k_blocks = mask_chunked.shape[3]
-
-    # Extract comparable region from standard mask
-    mask_std_comparable = mask_std[:, :, :chunked_q_blocks, :chunked_k_blocks]
-
-    # Compare masks
-    masks_match = compare_masks(mask_std_comparable, mask_chunked, "standard", "chunked")
-
-    # Compare attn_sums
-    attn_sum_std_comparable = attn_sum_std[:, :, :chunked_q_blocks, :chunked_k_blocks]
-    if attn_sum_std_comparable.shape == attn_sum_chunked.shape:
-        attn_diff = (attn_sum_std_comparable - attn_sum_chunked).abs().max().item()
-        print(f"  Attn sum max diff: {attn_diff:.6f}")
-    else:
-        print(f"  Attn sum shape mismatch: std={attn_sum_std_comparable.shape}, chunked={attn_sum_chunked.shape}")
-
-    # Clean up GPU memory
-    del query, key, attn_sum_std, mask_std, attn_sum_chunked, mask_chunked
-    torch.cuda.empty_cache()
-
-    return masks_match
-
-
-# ============================================================
-# Main Test
-# ============================================================
-
-if __name__ == "__main__":
-    print("XAttention Chunked vs Standard Test")
-    print("=" * 60)
-    print(f"Config: block_size={BLOCK_SIZE}, stride={STRIDE}, threshold={THRESHOLD}")
-    print(f"External chunk_size={CHUNK_SIZE}")
-    print()
-
-    # Check CUDA availability
-    if not torch.cuda.is_available():
-        print("CUDA not available!")
-        sys.exit(1)
-
-    print(f"Using GPU: {torch.cuda.get_device_name(0)}")
-    print("✓ xattn_estimate imported")
-    print("✓ xattn_estimate_chunked imported")
-
-    # Run tests
-    all_passed = True
-    results = []
-
-    for seq_len in TEST_SEQ_LENS:
-        passed = test_single_seq_len(seq_len)
-        chunks = (seq_len + CHUNK_SIZE - 1) // CHUNK_SIZE
-        results.append((seq_len, chunks, passed))
-        if not passed:
-            all_passed = False
-
-    # Summary
-    print("\n" + "=" * 60)
-    print("SUMMARY")
-    print("=" * 60)
-    for seq_len, chunks, passed in results:
-        status = "PASSED" if passed else "FAILED"
-        print(f"  seq_len={seq_len:5d} ({chunks} chunk{'s' if chunks > 1 else ' '}): {status}")
-
-    print("=" * 60)
-    if all_passed:
-        print("ALL TESTS PASSED!")
-        sys.exit(0)
-    else:
-        print("SOME TESTS FAILED!")
-        sys.exit(1)

tests/test_xattn_kernels.py

@@ -1,129 +0,0 @@
-"""
-Test: XAttention Triton kernels
-
-演示 XAttention 的两个核心 Triton kernel:
-1. flat_group_gemm_fuse_reshape: 计算 stride reshape 后的 attention scores (反对角线求和)
-2. softmax_fuse_block_sum: 对 attention scores 做 softmax 后按 block 求和
-
-数据流:
-  Q [batch, heads, q_len, head_dim]
-  K [batch, heads, kv_len, head_dim]
-    ↓ flat_group_gemm_fuse_reshape
-  attn_scores [batch, heads, q_len/stride, kv_len/stride]
-    ↓ softmax_fuse_block_sum
-  block_sums [batch, heads, q_blocks, k_blocks]
-"""
-import torch
-import sys
-sys.path.insert(0, "/home/zijie/Code/nano-vllm")
-from nanovllm.ops.xattn import flat_group_gemm_fuse_reshape, softmax_fuse_block_sum
-
-# ============================================================
-# 参数配置
-# ============================================================
-
-# Triton 约束: q_len >= stride * BLOCK_M, kv_len >= stride * BLOCK_N
-# A100: BLOCK_M = BLOCK_N = 128, 所以 min = 4 * 128 = 512
-# RTX 3090: BLOCK_M = BLOCK_N = 64, 所以 min = 4 * 64 = 256
-q_len = 512
-kv_len = 2048
-head_dim = 128
-stride = 4
-block_size = 128    # softmax block size (in reshaped space)
-segment_size = 128  # Triton kernel 要求 segment_size >= block_size
-
-# ============================================================
-# 构造输入: 偶数位置=1, 奇数位置=2
-# ============================================================
-
-Q = torch.zeros(1, 1, q_len, head_dim, dtype=torch.bfloat16).cuda()
-K = torch.zeros(1, 1, kv_len, head_dim, dtype=torch.bfloat16).cuda()
-
-for i in range(q_len):
-    if i % 2 == 0:
-        Q[0, 0, i, :] = 1
-    else:
-        Q[0, 0, i, :] = 2
-
-for i in range(kv_len):
-    if i % 2 == 0:
-        K[0, 0, i, :] = 1
-    else:
-        K[0, 0, i, :] = 2
-
-# ============================================================
-# Step 1: flat_group_gemm_fuse_reshape (chunked along K)
-# ============================================================
-
-q_reshaped_len = q_len // stride   # 128
-kv_reshaped_len = kv_len // stride  # 512
-
-# 将 K 沿着长度维度分成多个 chunk
-k_chunk_size = 512  # 每个 chunk 512 tokens
-num_k_chunks = kv_len // k_chunk_size  # 4 chunks
-
-attn_scores_list = []
-for k_chunk_idx in range(num_k_chunks):
-    k_start = k_chunk_idx * k_chunk_size
-    k_end = k_start + k_chunk_size
-    K_chunk = K[:, :, k_start:k_end, :]  # [1, 1, k_chunk_size, head_dim]
-
-    # 对每个 K chunk 调用 flat_group_gemm_fuse_reshape
-    # 输出: [batch, heads, q_len/stride, k_chunk_size/stride]
-    attn_chunk = flat_group_gemm_fuse_reshape(
-        Q, K_chunk, stride,
-        chunk_start=0,
-        chunk_end=q_reshaped_len,
-        is_causal=False
-    )
-    attn_scores_list.append(attn_chunk)
-
-# 拼接所有 K chunks 的结果
-# 每个 chunk: [1, 1, q_reshaped_len, k_chunk_size/stride]
-# 拼接后: [1, 1, q_reshaped_len, kv_reshaped_len]
-attn_scores = torch.cat(attn_scores_list, dim=-1)
-
-# 验证 shape: [batch, heads, q_len/stride, kv_len/stride]
-assert attn_scores.shape == (1, 1, q_reshaped_len, kv_reshaped_len), \
-    f"shape mismatch: {attn_scores.shape} != (1, 1, {q_reshaped_len}, {kv_reshaped_len})"
-
-# 验证: 反对角线求和
-# 每个 stride x stride 块的反对角线: Q[奇]*K[偶] + Q[偶]*K[奇] = 2*1 + 1*2 = 4
-# 反对角线有 stride/2 对，再乘以 head_dim
-expected_gemm = (2*1 + 1*2) * (stride // 2) * head_dim
-actual_gemm = attn_scores[0, 0, 0, 0].item()
-assert actual_gemm == expected_gemm, f"flat_group_gemm: {actual_gemm} != {expected_gemm}"
-
-# ============================================================
-# Step 2: softmax_fuse_block_sum
-# ============================================================
-
-scale = 1.4426950408889634  # log2(e) for exp2
-
-block_sums = softmax_fuse_block_sum(
-    attn_scores,
-    block_size,
-    segment_size,
-    chunk_start=0,
-    chunk_end=q_reshaped_len,
-    real_q_len=q_reshaped_len,
-    scale=scale,
-    is_causal=False
-)
-
-# 验证 shape: [batch, heads, q_blocks, k_blocks]
-q_blocks = q_reshaped_len // block_size  # 128 / 128 = 1
-k_blocks = kv_reshaped_len // block_size  # 512 / 128 = 4
-assert block_sums.shape == (1, 1, q_blocks, k_blocks), \
-    f"shape mismatch: {block_sums.shape} != (1, 1, {q_blocks}, {k_blocks})"
-
-# 验证: 每个 block 的 softmax 结果求和
-# 所有 attn_scores 相同 → softmax 均匀分布
-# 每行对一个 K block 的贡献 = block_size / kv_reshaped_len
-# 每个 Q block 有 block_size 行
-# block_sum = block_size * (block_size / kv_reshaped_len)
-expected_sum = block_size * block_size / kv_reshaped_len
-actual_sum = block_sums[0, 0, 0, 0].item()
-assert actual_sum == expected_sum, f"softmax_fuse_block_sum: {actual_sum} != {expected_sum}"
-
-print("test_xattn_kernels: PASSED")