📝 docs: add CPU offload optimization strategies guide

- Document chunk size optimization (simplest, most effective)
- Analyze CUDA Graph limitations for offload scenarios
- Cover CUDA Graph applicability for MLP/Proj layers
- Survey frontier research: InfiniGen, ShadowKV, L2 Prefetch, KVPR
- Add optimization priority recommendations

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

.claude/rules/nsys-profiling.md

@@ -0,0 +1,89 @@
+# Nsys Profiling Rule
+
+## 强制规则
+
+**所有 nsys profiling 任务必须使用 `scripts/profile_offload.sh` 脚本**，禁止直接运行 nsys 命令。
+
+| 禁止 | 原因 |
+|------|------|
+| `nsys profile python tests/test_ruler.py ...` | 参数不一致，输出路径混乱 |
+| 手动构造 nsys 命令 | 容易遗漏关键参数 |
+
+## 使用方法
+
+```bash
+# 基本用法（默认 4 slots）
+bash scripts/profile_offload.sh
+
+# 指定 GPU slots 数量
+bash scripts/profile_offload.sh --num-gpu-blocks 8
+
+# 指定 sample
+bash scripts/profile_offload.sh --sample 5
+
+# 指定 dataset
+bash scripts/profile_offload.sh --dataset niah_single_1
+
+# 禁用 offload（对比测试）
+bash scripts/profile_offload.sh --no-offload
+
+# 组合参数
+bash scripts/profile_offload.sh --num-gpu-blocks 8 --sample 0 --gpu 1
+```
+
+## 参数说明
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `--dataset` | `niah_single_1` | RULER 任务名称 |
+| `--sample` | `0` | 样本索引 |
+| `--gpu` | `0` | 使用的 GPU |
+| `--num-gpu-blocks` | `4` | GPU ring buffer slots 数量 |
+| `--no-offload` | - | 禁用 CPU offload |
+
+## 输出文件
+
+输出文件自动生成到 `results/nsys/` 目录：
+
+```
+results/nsys/ruler_<dataset>_sample<index>_offload_<slots>slots_<timestamp>.nsys-rep
+```
+
+示例：`ruler_niah_single_1_sample0_offload_8slots_20260127_031500.nsys-rep`
+
+## 查看结果
+
+```bash
+# GUI 查看
+nsight-sys results/nsys/<filename>.nsys-rep
+
+# 命令行统计
+nsys stats --report cuda_api_sum results/nsys/<filename>.nsys-rep
+nsys stats --report cuda_gpu_kern_sum results/nsys/<filename>.nsys-rep
+```
+
+## 典型工作流
+
+### 1. 对比不同 slots 数量
+
+```bash
+# 测试 4 slots（默认）
+bash scripts/profile_offload.sh --num-gpu-blocks 4
+
+# 测试 8 slots
+bash scripts/profile_offload.sh --num-gpu-blocks 8
+
+# 对比结果
+nsys stats --report cuda_gpu_kern_sum results/nsys/*4slots*.nsys-rep
+nsys stats --report cuda_gpu_kern_sum results/nsys/*8slots*.nsys-rep
+```
+
+### 2. 分析 pipeline overlap
+
+```bash
+# 生成 profile
+bash scripts/profile_offload.sh --num-gpu-blocks 8
+
+# 用 nsight-sys GUI 查看 CUDA HW timeline
+# 检查 H2D 和 flash_fwd_kernel 是否 overlap
+```

.gitignore

@@ -239,3 +239,4 @@ task_plan_*.md
 findings_*.md
 progress_*.md
 notes.md
+Snipaste*

CLAUDE.md

@@ -26,6 +26,9 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
 | [`docs/ruler_32k_chunked_offload_issue.md`](docs/ruler_32k_chunked_offload_issue.md) | ⚠️ OPEN ISSUE: 32K chunked offload accuracy problem (20% error rate in RULER) |
 | [`docs/chunked_attention_solutions.md`](docs/chunked_attention_solutions.md) | 🔧 SOLUTIONS: Chunked attention 准确性问题的代码分析和解决方案 |
 | [`docs/nsys_wrong_event_order_bug.md`](docs/nsys_wrong_event_order_bug.md) | 🐛 NSYS BUG: Ring buffer pipeline 触发 nsys 时间戳乱序问题的调试记录 |
+| [`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) |
 
 ## Rules Index
 

bench_offload.py

@@ -46,24 +46,41 @@ def main():
     from nanovllm.config import SparsePolicyType
 
     parser = argparse.ArgumentParser(description="Benchmark CPU offload performance")
-    parser.add_argument("--enable-quest", action="store_true", help="Enable Quest sparse attention for decode")
+    parser.add_argument("--model", type=str, default="~/models/Llama-3.1-8B-Instruct",
+                        help="Model path (default: ~/models/Llama-3.1-8B-Instruct)")
+    # Sparse policy selection (mutually exclusive)
+    sparse_group = parser.add_mutually_exclusive_group()
+    sparse_group.add_argument("--enable-quest", action="store_true",
+                              help="Enable Quest sparse attention (decode only, prefill uses full)")
+    sparse_group.add_argument("--enable-xattn", action="store_true",
+                              help="Enable XAttention BSA (prefill only, decode uses full)")
+    # Quest parameters
     parser.add_argument("--topk", type=int, default=16, help="Top-K blocks for Quest (default: 16)")
     parser.add_argument("--threshold", type=int, default=4, help="Apply sparse only when blocks > threshold (default: 4)")
+    # XAttention parameters
+    parser.add_argument("--xattn-threshold", type=float, default=0.95,
+                        help="XAttention cumulative attention threshold (default: 0.95)")
+    parser.add_argument("--xattn-stride", type=int, default=8,
+                        help="XAttention Q/K downsampling stride (default: 8)")
+    # General parameters
     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=6, help="Number of GPU blocks (default: 6)")
+    parser.add_argument("--num-gpu-blocks", type=int, default=4, help="Number of GPU blocks (default: 4)")
     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")
     args = parser.parse_args()
 
-    path = os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/")
+    path = os.path.expanduser(args.model)
     max_len = args.max_len
 
     # Setup policy configuration
     if args.enable_quest:
         sparse_policy = SparsePolicyType.QUEST
-        print(f"\n[Quest Sparse Attention] topk={args.topk}, threshold={args.threshold}")
+        print(f"\n[Quest Sparse Attention] decode: Quest (topk={args.topk}, threshold={args.threshold}), prefill: Full")
+    elif args.enable_xattn:
+        sparse_policy = SparsePolicyType.XATTN_BSA
+        print(f"\n[XAttention BSA] prefill: XAttn (tau={args.xattn_threshold}, stride={args.xattn_stride}), decode: Full")
     else:
         sparse_policy = SparsePolicyType.FULL
         print("\n[Full Attention] baseline (no sparse)")
@@ -78,8 +95,12 @@ def main():
         enable_cpu_offload=True,
         num_gpu_blocks=args.num_gpu_blocks,
         sparse_policy=sparse_policy,
+        # Quest parameters
         sparse_topk_blocks=args.topk,
         sparse_threshold_blocks=args.threshold,
+        # XAttention parameters
+        sparse_threshold=args.xattn_threshold,
+        sparse_stride=args.xattn_stride,
     )
 
     # Warmup

docs/bench_offload_results.md

@@ -0,0 +1,89 @@
+# CPU Offload Benchmark Results
+
+本文档记录 `bench_offload.py` 在不同配置下的性能测试结果。
+
+## 测试环境
+
+| 参数 | 值 |
+|------|-----|
+| 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
+
+| 策略 | Prefill | Decode | 说明 |
+|------|---------|--------|------|
+| FULL | Full Attention | Full Attention | 基线，加载所有 blocks |
+| XATTN_BSA | XAttention (tau=0.95, stride=8) | Full Attention (fallback) | 稀疏 prefill |
+
+## 测试结果
+
+### 32K 上下文
+
+| 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 |
+
+### 128K 上下文
+
+| 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 |
+
+### KV Cache 配置
+
+| 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 |
+
+## 分析
+
+### XAttention 性能特点
+
+1. **32K 上下文**: XAttention 比 Full 慢 26%
+2. **128K 上下文**: XAttention 比 Full 慢 16%
+
+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。
+
+### 原因分析
+
+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
+
+### 适用场景
+
+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
+
+## 运行命令
+
+```bash
+# Full Attention (32K)
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 32768
+
+# XAttention BSA (32K)
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --max-len 32768 --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
+
+# 调整 XAttention 参数
+CUDA_VISIBLE_DEVICES=0 python bench_offload.py --enable-xattn --xattn-threshold 0.8 --xattn-stride 16
+```
+
+## 更新记录
+
+- 2026-01-27: 初始测试，Llama-3.1-8B-Instruct, A100 80GB

docs/cpu_offload_optimization_strategies.md

@@ -0,0 +1,300 @@
+# CPU Offload 优化策略
+
+本文档记录 CPU Offload 场景下的性能优化策略分析，包括实际可行的方案和前沿研究方向。
+
+## 问题回顾
+
+根据 [CPU 调度延迟分析](cpu_scheduling_latency_analysis.md)，当前 chunked attention pipeline 的主要问题：
+
+| 指标 | 当前值 | 理论值 |
+|------|--------|--------|
+| Flash kernel 执行时间 | ~138 μs | - |
+| Flash kernel 间隔 | ~942 μs | ~211 μs (仅 H2D + merge) |
+| GPU 利用率 | **12.8%** | **39.5%** (理论上限) |
+| CPU 调度空闲占比 | **77-81%** | 0% |
+
+**瓶颈根源**：每个 block 都经过完整的 Python 循环，导致大量 CPU 调度延迟。
+
+---
+
+## 优化方案一：调大 Chunk Size（推荐）
+
+### 核心洞察
+
+**Merge 多个小 chunk 和直接使用大 chunk 是等效的**：
+
+```
+方案 A: Merge 4 个小 chunks
+[H2D 2K][H2D 2K][H2D 2K][H2D 2K] → concat → [Flash 8K] → merge
+
+方案 B: 直接用大 chunk
+[H2D 8K] → [Flash 8K] → merge
+
+计算结果完全等效！
+```
+
+### 收益分析
+
+| 指标 | 小 chunk (2K) × 4 | 大 chunk (8K) × 1 |
+|------|-------------------|-------------------|
+| H2D 次数 | 4 | 1 |
+| Flash kernel 调用 | 4 | 1 |
+| Merge 调用 | 4 | 1 |
+| Python 循环次数 | 4 | 1 |
+| CPU 调度开销 | 4 × ~300μs = 1200μs | 1 × ~300μs = 300μs |
+
+**本质**：CPU 调度延迟问题的根源是循环次数太多，调大 chunk size 直接减少循环次数。
+
+### Trade-off
+
+1. **GPU 内存增加**
+   - 2K chunk: 每 slot ~4MB (K+V)
+   - 8K chunk: 每 slot ~16MB (K+V)
+   - 4 slots = 64MB，对 80GB A100 影响很小
+
+2. **单次 H2D 时间变长**
+   - H2D 8K ≈ 350μs
+   - Flash 8K ≈ 550μs
+   - 因为 Flash > H2D，pipeline 仍然有效
+
+### 配置方法
+
+```bash
+# 测试不同 block size
+python bench_offload.py --kvcache-block-size 2048   # 基准
+python bench_offload.py --kvcache-block-size 4096   # 2x
+python bench_offload.py --kvcache-block-size 8192   # 4x
+```
+
+---
+
+## 优化方案二：CUDA Graph（适用于非 Attention 部分）
+
+### CUDA Graph 在 Offload 场景的局限性
+
+CUDA Graph 的前提：所有操作在 capture 时确定，数据地址固定。
+
+**Offload 场景的现实**：
+1. **H2D 源地址动态** - 每次从不同的 CPU block 加载
+2. **加载决策在运行时** - 哪些 block 需要加载是动态的
+3. **CPU 必须协调** - H2D 和 Compute 的同步需要 CPU 参与
+
+```
+Offload 场景：
+┌─────────────────────────────────────────┐
+│  数据在 CPU，需要动态加载                 │
+│  [H2D_i] → [Compute] → [H2D_{i+n}] → ...│
+│  ↑ 动态、CPU 必须参与调度                 │
+└─────────────────────────────────────────┘
+
+即使用 Graph：
+Python: [wait_h2d] [replay] [launch_h2d] [wait_h2d] [replay] ...
+        ↑ CPU 参与           ↑ CPU 参与   ↑ CPU 参与
+
+CPU 调度开销仍然存在，Graph 只优化了中间的 compute 部分。
+```
+
+**结论**：CUDA Graph 不是 Offload 场景的银弹。
+
+### 适用场景：MLP 和 Projection 层
+
+LLM 每层的计算流程：
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│  [LayerNorm] → [QKV Proj] → [Attention] → [O Proj] → [Add]  │
+│                                  ↑                          │
+│                             KV Offload                      │
+│  [LayerNorm] → [MLP: gate + up + down] → [Add]              │
+└─────────────────────────────────────────────────────────────┘
+```
+
+| 组件 | 涉及 Offload | 能用 CUDA Graph |
+|------|-------------|-----------------|
+| LayerNorm | ❌ | ✅ |
+| QKV Projection | ❌ | ✅ |
+| **Attention** | ✅ | ❌ |
+| Output Projection | ❌ | ✅ |
+| MLP (FFN) | ❌ | ✅ |
+
+**只有 Attention 涉及动态 KV Cache 加载，其余都是"纯计算"，可以用 CUDA Graph。**
+
+### 实现方案
+
+```python
+class OptimizedLayer:
+    def __init__(self, layer):
+        # Graph 1: Attention 之前
+        self.graph_pre_attn = capture([
+            layer.input_layernorm,
+            layer.self_attn.q_proj,
+            layer.self_attn.k_proj,
+            layer.self_attn.v_proj,
+        ])
+
+        # Graph 2: Attention 之后 + MLP
+        self.graph_post_attn = capture([
+            layer.self_attn.o_proj,
+            # residual add
+            layer.post_attention_layernorm,
+            layer.mlp.gate_proj,
+            layer.mlp.up_proj,
+            layer.mlp.down_proj,
+            # residual add
+        ])
+
+    def forward(self, hidden_states, kv_cache):
+        # Pre-attention (CUDA Graph)
+        self.graph_pre_attn.replay()
+
+        # Attention with offload (动态，不能用 graph)
+        attn_output = chunked_attention_with_offload(q, kv_cache)
+
+        # Post-attention + MLP (CUDA Graph)
+        self.graph_post_attn.replay()
+```
+
+### 收益估算
+
+MLP 每层典型操作 launch 开销：
+- `gate_proj`, `up_proj`, `act_fn`, `gate * up`, `down_proj`, `residual add`
+- 每个操作 ~30-50μs launch 开销，总计 ~200μs/层
+- 用 CUDA Graph：~30μs/层
+
+**32 层 × 170μs 节省 ≈ 5.4ms**
+
+---
+
+## 优化方案三：前沿研究方向
+
+### 1. InfiniGen - 投机预取 (OSDI'24)
+
+**核心思想**：不需要加载所有 KV，只预取"重要"的 token。
+
+```
+关键洞察：相邻层的 attention pattern 高度相似
+         ↓
+用第 L 层的 attention score 预测第 L+1 层需要哪些 token
+         ↓
+只预取 top-k 重要的 KV entries（而不是全部）
+```
+
+**技术实现**：
+- 用当前层的 Q 和下一层的部分 K 做"预演"
+- 预测下一层的 attention 分布
+- 异步预取预测的重要 token
+- **减少 PCIe 带宽浪费，而不是加速传输**
+
+**效果**：最高 **3x 加速**
+
+**参考**：[InfiniGen (OSDI'24)](https://www.usenix.org/conference/osdi24/presentation/lee)
+
+### 2. ShadowKV - 低秩压缩 + Sparse Offload (ICML'25 Spotlight)
+
+**核心思想**：Key 压缩存 GPU，Value offload 到 CPU，只加载 1.56% 的 KV。
+
+```
+Pre-filling:
+┌─────────────────────────────────────────────────┐
+│  Key Cache → SVD 低秩压缩 → 保留在 GPU          │
+│  Value Cache → Offload 到 CPU                   │
+│  计算每个 chunk 的 landmark (均值)               │
+│  识别 outlier tokens → 保留在 GPU               │
+└─────────────────────────────────────────────────┘
+
+Decoding:
+┌─────────────────────────────────────────────────┐
+│  用 landmarks 快速估计 attention score          │
+│  只加载 top-k 重要的 Value (1.56% sparse)       │
+│  结合 GPU 上的 outliers 计算最终结果            │
+└─────────────────────────────────────────────────┘
+```
+
+**效果**：6x 更大 batch size，**3.04x 吞吐提升**
+
+**参考**：[ShadowKV (ByteDance)](https://github.com/ByteDance-Seed/ShadowKV)
+
+### 3. L2 Cache 异步预取 (2025)
+
+**核心思想**：利用 GPU L2 Cache 做预取，在计算时预取下一批 KV。
+
+```
+传统：
+Compute:  [Flash_i]        [Flash_{i+1}]
+H2D:              [H2D_{i+1}]
+                  ↑ 等待
+
+L2 Prefetch：
+Compute:  [Flash_i  + Prefetch_{i+1} to L2]  [Flash_{i+1} L2 hit]
+          ↑ 计算时利用空闲 memory bandwidth 预取
+```
+
+**技术**：
+- 在 Flash Attention kernel 内部发起预取指令
+- 利用计算时的空闲 memory bandwidth
+- 下一次访问直接 L2 hit
+
+**效果**：**2.15x attention kernel 效率**，1.97x 端到端吞吐
+
+**参考**：[Asynchronous KV Cache Prefetching (2025)](https://arxiv.org/abs/2504.06319)
+
+### 4. KVPR - I/O-Aware 调度 (ACL'25)
+
+**核心思想**：计算最优的 recompute vs offload 比例。
+
+```
+权衡：
+- Recompute: 重新计算 KV（用 GPU 算力换内存）
+- Offload: 从 CPU 加载（用 PCIe 带宽换算力）
+
+KVPR: 根据当前负载动态决定最优比例
+      + 预取技术重叠数据传输和计算
+```
+
+**参考**：[KVPR (ACL'25)](https://aclanthology.org/2025.findings-acl.997.pdf)
+
+---
+
+## 优化策略总结
+
+### 推荐优先级
+
+| 优先级 | 方案 | 核心优化 | 实现复杂度 | 预期收益 |
+|--------|------|---------|-----------|---------|
+| **P0** | 调大 chunk size | 减少循环次数 | 极低（改配置） | 2-4x |
+| **P1** | MLP CUDA Graph | 减少 launch 开销 | 中 | ~5ms/request |
+| **P2** | InfiniGen 式预取 | 只加载重要 token | 中高 | 2-3x |
+| **P3** | ShadowKV 式压缩 | Key 压缩 + Sparse | 高 | 3x |
+| **P3** | C++ Extension | 消除 Python 开销 | 高 | 2-3x |
+
+### 策略分离原则
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│  Attention + Offload 部分：                                 │
+│    - 瓶颈：H2D 传输 + CPU 调度                              │
+│    - 优化：调大 chunk size / 投机预取 / Sparse              │
+│                                                             │
+│  MLP + Proj + Norm 部分：                                   │
+│    - 瓶颈：Kernel launch 开销                               │
+│    - 优化：CUDA Graph                                       │
+└─────────────────────────────────────────────────────────────┘
+
+两部分优化完全正交，可以组合使用。
+```
+
+---
+
+## 相关文件
+
+- `nanovllm/kvcache/sparse/full_policy.py`: Chunked attention pipeline
+- `nanovllm/kvcache/offload_engine.py`: H2D/D2H 传输管理
+- `docs/cpu_scheduling_latency_analysis.md`: 问题分析
+
+## 参考文献
+
+1. [InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management](https://www.usenix.org/conference/osdi24/presentation/lee) - OSDI'24
+2. [ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference](https://github.com/ByteDance-Seed/ShadowKV) - ICML'25 Spotlight
+3. [Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetching](https://arxiv.org/abs/2504.06319) - 2025
+4. [KVPR: Efficient LLM Inference with I/O-Aware KV Cache](https://aclanthology.org/2025.findings-acl.997.pdf) - ACL'25
+5. [LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference](https://lmcache.ai/tech_report.pdf) - 2025

docs/cpu_scheduling_latency_analysis.md

@@ -0,0 +1,177 @@
+# CPU 调度延迟分析
+
+## 问题概述
+
+在分析 nsys profile 时发现，chunked attention pipeline 中存在大量的 **CPU 调度延迟**，导致 GPU 利用率显著下降。
+
+## 观察数据
+
+### 测试环境
+- GPU: NVIDIA A100-SXM4-80GB
+- 模型: Llama-3.1-8B-Instruct
+- 测试: RULER niah_single_1, 64K context
+- Profile 文件: `ruler_8slots_test.nsys-rep`
+- 时间段: 92.982s - 93.038s
+
+### Kernel 执行时间
+
+| Kernel | 典型执行时间 |
+|--------|-------------|
+| flash_fwd_kernel | ~138 μs |
+| H2D memcpy (2MB) | ~87 μs |
+| merge_lse_kernel | ~3.5 μs |
+| merge_output_kernel | ~34 μs |
+
+### 操作间隙分析
+
+从 cuda_gpu_trace 观察到的间隙：
+
+```
+Start (ms)     Dur (μs)   Gap (μs)   Type
+------------------------------------------------------------
+92984.680      138.3      378.3      flash_fwd_kernel     ← GAP!
+92985.051      86.8       232.9      H2D memcpy           ← GAP!
+92985.141      86.8       2.8        H2D memcpy
+92985.587      135.9      360.0      flash_fwd_kernel     ← GAP!
+92986.026      3.4        302.4      merge_lse            ← GAP!
+92986.164      33.5       135.0      merge_output         ← GAP!
+92986.371      86.9       173.4      H2D memcpy           ← GAP!
+92986.461      86.8       2.7        H2D memcpy
+92986.816      137.9      268.2      flash_fwd_kernel     ← GAP!
+```
+
+### Flash Kernel 间隙分解
+
+| 间隙 | 总时间 | 有效工作时间 | 空闲时间 |
+|------|--------|-------------|---------|
+| Flash 1 → Flash 2 | 769 μs | ~174 μs (2x H2D) | ~595 μs (77%) |
+| Flash 2 → Flash 3 | 1092 μs | ~211 μs (merge + H2D) | ~881 μs (81%) |
+| Flash 3 → Flash 4 | 965 μs | ~211 μs (merge + H2D) | ~754 μs (78%) |
+
+**关键发现**: 每个 flash kernel 之间约 **77-81% 的时间是 CPU 调度空闲**。
+
+## 间隙来源分析
+
+### 1. CPU 调度延迟类型
+
+| 转换 | 典型延迟 | 原因 |
+|------|---------|------|
+| Kernel 结束 → 下一个 Kernel 开始 | 100-400 μs | CPU 准备参数、调用 CUDA driver |
+| Flash 结束 → H2D 开始 | ~233 μs | Python 代码执行 + CUDA launch |
+| H2D 结束 → Flash 开始 | ~360 μs | 同步等待 + kernel launch |
+| Flash 结束 → merge 开始 | ~302 μs | Python 代码执行 |
+
+### 2. 延迟产生的代码位置
+
+```python
+# full_policy.py: compute_chunked_prefill
+
+for block_idx in range(num_blocks):
+    # 1. 等待 H2D 完成 (同步点)
+    offload_engine.wait_slot_layer(current_slot)  # ← 可能引入延迟
+
+    # 2. 获取 KV 数据
+    k_block, v_block = offload_engine.get_kv_for_slot(current_slot)
+
+    # 3. 调用 flash attention (kernel launch)
+    block_out, block_lse = flash_attn_with_kvcache(...)  # ← CPU 调度延迟
+
+    # 4. merge 操作
+    merge_output(...)  # ← CPU 调度延迟
+    merge_lse(...)     # ← CPU 调度延迟
+
+    # 5. 发起下一个 H2D (异步)
+    offload_engine.load_to_slot_layer(next_slot, ...)  # ← CPU 调度延迟
+```
+
+### 3. 为什么 H2D 之间间隙小
+
+注意到连续的 H2D memcpy 之间间隙只有 ~2.7 μs，这是因为：
+- 它们在同一个 stream 上连续发起
+- CUDA driver 可以批量处理
+- 没有 Python 代码介入
+
+## GPU 利用率计算
+
+基于观察数据：
+
+| 指标 | 值 |
+|------|-----|
+| Flash kernel 平均执行时间 | 138 μs |
+| Flash kernel 平均间隔 | 942 μs |
+| Flash kernel GPU 利用率 | 138 / (138 + 942) = **12.8%** |
+
+如果消除 CPU 调度延迟（仅保留必要的 H2D + merge）：
+
+| 指标 | 值 |
+|------|-----|
+| 必要间隔 (2x H2D + merge) | ~211 μs |
+| 理论 GPU 利用率 | 138 / (138 + 211) = **39.5%** |
+
+**潜在提升**: 3x GPU 利用率
+
+## 优化方向
+
+### 1. CUDA Graph
+将整个 block 处理流程编译为 CUDA Graph，消除重复的 kernel launch 开销。
+
+```python
+# 伪代码
+graph = torch.cuda.CUDAGraph()
+with torch.cuda.graph(graph):
+    # 预录制 flash + merge 操作
+    block_out, block_lse = flash_attn_with_kvcache(...)
+    merge_output(...)
+    merge_lse(...)
+
+# 运行时只需 replay
+for block_idx in range(num_blocks):
+    graph.replay()  # 单次 launch，无 Python 介入
+```
+
+### 2. 自定义 Triton Kernel
+将 flash + merge 融合为单个 kernel，减少 kernel launch 次数。
+
+### 3. C++ Extension
+将 Python 循环移到 C++ 层，减少 Python 解释器开销。
+
+### 4. 流水线重叠优化
+确保 H2D 传输与前一个 block 的计算完全重叠：
+
+```
+Block 0: [H2D slot0] [Flash slot0] [merge]
+Block 1:            [H2D slot1]   [Flash slot1] [merge]
+Block 2:                         [H2D slot2]   [Flash slot2] [merge]
+```
+
+## 验证方法
+
+### 1. 使用 nsys 分析间隙
+
+```bash
+# 生成 profile
+bash scripts/profile_offload.sh --num-gpu-blocks 8
+
+# 查看 kernel trace
+nsys stats --report cuda_gpu_trace --format csv <file>.nsys-rep | \
+    awk -F',' 'NR>1 && $1 >= START && $1 <= END'
+```
+
+### 2. 计算间隙
+
+```python
+# 从 trace 数据计算
+prev_end = start + duration
+gap = next_start - prev_end
+```
+
+## 相关文件
+
+- `nanovllm/kvcache/sparse/full_policy.py`: Pipeline 实现
+- `nanovllm/kvcache/offload_engine.py`: H2D/D2H 传输
+- `scripts/profile_offload.sh`: Profiling 脚本
+
+## 参考
+
+- [CUDA Graph 文档](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cuda-graphs)
+- [nsys 用户指南](https://docs.nvidia.com/nsight-systems/UserGuide/index.html)

nanovllm/kvcache/offload_engine.py

@@ -9,6 +9,7 @@ Key design principles for CUDA Graph compatibility:
 
 import torch
 import torch.cuda.nvtx
+import nvtx
 from torch import Tensor
 from typing import Dict, List, Tuple, Optional
 from dataclasses import dataclass
@@ -374,7 +375,9 @@ 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) -> None:
+    def load_to_slot_layer(
+        self, slot_idx: int, layer_id: int, cpu_block_id: int, chunk_idx: int = -1
+    ) -> None:
         """
         Async load a single CPU block to a ring buffer slot for one layer.
 
@@ -389,13 +392,20 @@ class OffloadEngine:
             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)
         """
         logger.debug(f"Ring load: layer={layer_id}, CPU[{cpu_block_id}] -> GPU slot[{slot_idx}]")
 
         # Use per-slot stream for parallel transfers across different slots
         stream = self.slot_transfer_streams[slot_idx]
 
-        torch.cuda.nvtx.range_push(f"H2D: L{layer_id} CPU[{cpu_block_id}]->Slot[{slot_idx}]")
+        # Build NVTX label with optional chunk info
+        if chunk_idx >= 0:
+            nvtx_label = f"H2D: L{layer_id} Chunk{chunk_idx} CPU[{cpu_block_id}]->Slot[{slot_idx}]"
+        else:
+            nvtx_label = f"H2D: L{layer_id} CPU[{cpu_block_id}]->Slot[{slot_idx}]"
+
+        nvtx.push_range(message=nvtx_label, color="blue")
         with torch.cuda.stream(stream):
             # Wait for previous compute on this slot to complete before overwriting
             # This prevents data race: transfer must not start until attention finishes reading
@@ -413,7 +423,7 @@ class OffloadEngine:
                 self.v_cache_cpu[layer_id, cpu_block_id], non_blocking=True
             )
             self.ring_slot_ready[slot_idx].record(stream)
-        torch.cuda.nvtx.range_pop()
+        nvtx.pop_range()
 
     def wait_slot_layer(self, slot_idx: int) -> None:
         """
@@ -470,7 +480,8 @@ class OffloadEngine:
             else:
                 self.sparse_policy.on_decode_offload(cpu_block_id, layer_id, k_cache, valid_tokens)
 
-        torch.cuda.nvtx.range_push(f"D2H: Slot[{slot_idx}]->CPU[L{layer_id},B{cpu_block_id}]")
+        nvtx_label = f"D2H: Slot[{slot_idx}]->CPU[L{layer_id},B{cpu_block_id}]"
+        nvtx.push_range(message=nvtx_label, color="green")
         with torch.cuda.stream(self.transfer_stream_main):
             # Wait for both compute_stream and default stream
             # - compute_stream: for flash attention operations
@@ -486,7 +497,7 @@ class OffloadEngine:
                 self.v_cache_gpu[slot_idx], non_blocking=True
             )
             self.ring_slot_offload_done[slot_idx].record(self.transfer_stream_main)
-        torch.cuda.nvtx.range_pop()
+        nvtx.pop_range()
 
     # ----- KV access methods for ring buffer -----
 
@@ -702,6 +713,61 @@ class OffloadEngine:
         v = self.prefill_v_buffer[layer_id, :num_tokens].unsqueeze(0)
         return k, v
 
+    def write_to_prefill_buffer(
+        self,
+        layer_id: int,
+        k: Tensor,
+        v: Tensor,
+        chunk_idx: int = -1,
+    ) -> None:
+        """
+        Write KV tensors to prefill buffer (D2D copy within GPU).
+
+        This is called during chunked prefill to store current chunk's KV
+        before computing attention.
+
+        Args:
+            layer_id: Layer index
+            k: Key tensor [num_tokens, kv_heads, head_dim]
+            v: Value tensor [num_tokens, kv_heads, head_dim]
+            chunk_idx: Current chunk index for NVTX labeling (-1 = not specified)
+        """
+        num_tokens = k.shape[0]
+
+        # Build NVTX label
+        if chunk_idx >= 0:
+            nvtx_label = f"D2D: L{layer_id} Chunk{chunk_idx} WritePrefillBuffer"
+        else:
+            nvtx_label = f"D2D: L{layer_id} WritePrefillBuffer"
+
+        torch.cuda.nvtx.range_push(nvtx_label)
+        self.prefill_k_buffer[layer_id, :num_tokens].copy_(k)
+        self.prefill_v_buffer[layer_id, :num_tokens].copy_(v)
+        torch.cuda.nvtx.range_pop()
+
+    def write_to_decode_buffer(
+        self,
+        layer_id: int,
+        pos_in_block: int,
+        k: Tensor,
+        v: Tensor,
+    ) -> None:
+        """
+        Write KV tensors to decode buffer (D2D copy within GPU).
+
+        This is called during chunked decode to store current decode token's KV.
+
+        Args:
+            layer_id: Layer index
+            pos_in_block: Position within the current block
+            k: Key tensor [kv_heads, head_dim] (single token, squeezed)
+            v: Value tensor [kv_heads, head_dim] (single token, squeezed)
+        """
+        torch.cuda.nvtx.range_push(f"D2D: L{layer_id} Pos{pos_in_block} WriteDecodeBuffer")
+        self.decode_k_buffer[layer_id, pos_in_block].copy_(k)
+        self.decode_v_buffer[layer_id, pos_in_block].copy_(v)
+        torch.cuda.nvtx.range_pop()
+
     def offload_prefill_buffer_async(
         self,
         layer_id: int,
@@ -729,7 +795,8 @@ class OffloadEngine:
         # Use per-layer stream for parallel offloads
         stream = self.prefill_offload_streams[layer_id]
 
-        torch.cuda.nvtx.range_push(f"AsyncPrefillOffload: L{layer_id}->CPU[{cpu_block_id}]")
+        nvtx_label = f"D2H: PrefillBuffer L{layer_id}->CPU[{cpu_block_id}]"
+        nvtx.push_range(message=nvtx_label, color="orange")
         with torch.cuda.stream(stream):
             # Wait for compute to finish writing to prefill buffer
             stream.wait_stream(self.compute_stream)
@@ -744,7 +811,7 @@ class OffloadEngine:
 
             # Record completion event
             self.prefill_offload_events[layer_id].record(stream)
-        torch.cuda.nvtx.range_pop()
+        nvtx.pop_range()
 
     def wait_all_prefill_offloads(self) -> None:
         """Wait for all prefill buffer offloads to complete."""

nanovllm/kvcache/sparse/full_policy.py

@@ -139,7 +139,8 @@ class FullAttentionPolicy(SparsePolicy):
                 slot = load_slots[0]
                 for block_idx in range(num_blocks):
                     cpu_block_id = cpu_block_table[block_idx]
-                    offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id)
+                    # cpu_block_id is the chunk index (block N = chunk N)
+                    offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
                     offload_engine.wait_slot_layer(slot)
 
                     with torch.cuda.stream(compute_stream):
@@ -159,7 +160,8 @@ class FullAttentionPolicy(SparsePolicy):
                 num_slots = len(load_slots)
                 num_preload = min(num_slots, num_blocks)
                 for i in range(num_preload):
-                    offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_table[i])
+                    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)
 
                 for block_idx in range(num_blocks):
                     current_slot = load_slots[block_idx % num_slots]
@@ -186,7 +188,7 @@ class FullAttentionPolicy(SparsePolicy):
                     if next_block_idx < num_blocks:
                         next_slot = load_slots[next_block_idx % num_slots]
                         next_cpu_block_id = cpu_block_table[next_block_idx]
-                        offload_engine.load_to_slot_layer(next_slot, layer_id, next_cpu_block_id)
+                        offload_engine.load_to_slot_layer(next_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id)
 
         # Step 4: Compute attention to current chunk (causal mask)
         with torch.cuda.stream(compute_stream):
@@ -350,7 +352,8 @@ class FullAttentionPolicy(SparsePolicy):
         # Phase 1: Pre-load up to num_slots blocks
         num_preload = min(num_slots, num_blocks)
         for i in range(num_preload):
-            offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_table[i])
+            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)
 
         # Phase 2: Process blocks with pipeline
         for block_idx in range(num_blocks):
@@ -383,7 +386,8 @@ class FullAttentionPolicy(SparsePolicy):
             # Start loading next block (pipeline)
             next_block_idx = block_idx + num_slots
             if next_block_idx < num_blocks:
-                offload_engine.load_to_slot_layer(current_slot, layer_id, cpu_block_table[next_block_idx])
+                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)
 
             # Merge with accumulated
             with torch.cuda.stream(compute_stream):

nanovllm/kvcache/sparse/xattn_bsa.py

@@ -189,8 +189,8 @@ class XAttentionBSAPolicy(SparsePolicy):
         reshaped_block_size = block_size // self.stride  # e.g., 1024/8 = 128
 
         for cpu_block_id in available_blocks:
-            # Load K block from CPU to GPU
-            offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id)
+            # Load K block from CPU to GPU (cpu_block_id is chunk index)
+            offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
             offload_engine.wait_slot_layer(slot)
 
             # Get KV: [1, block_size, num_kv_heads, head_dim]
@@ -382,7 +382,7 @@ class XAttentionBSAPolicy(SparsePolicy):
                 slot = load_slots[0]
                 for block_idx in range(num_blocks):
                     cpu_block_id = cpu_block_table[block_idx]
-                    offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id)
+                    offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
                     offload_engine.wait_slot_layer(slot)
 
                     with torch.cuda.stream(compute_stream):
@@ -402,7 +402,8 @@ class XAttentionBSAPolicy(SparsePolicy):
                 num_slots = len(load_slots)
                 num_preload = min(num_slots, num_blocks)
                 for i in range(num_preload):
-                    offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_table[i])
+                    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)
 
                 for block_idx in range(num_blocks):
                     current_slot = load_slots[block_idx % num_slots]
@@ -428,7 +429,7 @@ class XAttentionBSAPolicy(SparsePolicy):
                     if next_block_idx < num_blocks:
                         next_slot = load_slots[next_block_idx % num_slots]
                         next_cpu_block_id = cpu_block_table[next_block_idx]
-                        offload_engine.load_to_slot_layer(next_slot, layer_id, next_cpu_block_id)
+                        offload_engine.load_to_slot_layer(next_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id)
 
         # Compute attention to current chunk (causal mask)
         with torch.cuda.stream(compute_stream):

nanovllm/layers/attention.py

@@ -104,27 +104,21 @@ class Attention(nn.Module):
             # This enables fully async offloads since each layer has its own buffer.
             offload_engine = context.kvcache_manager.offload_engine
             compute_stream = offload_engine.compute_stream
+            chunk_idx = context.current_chunk_idx if hasattr(context, 'current_chunk_idx') else -1
 
             # Wait for default stream to ensure slot_mapping tensor transfer is complete
             compute_stream.wait_stream(torch.cuda.default_stream())
 
             with torch.cuda.stream(compute_stream):
-                # Write KV to per-layer prefill buffer (contiguous write, no slot_mapping)
+                # Write KV to per-layer prefill buffer via offload_engine
                 # k, v shape: [num_tokens, kv_heads, head_dim]
-                num_tokens = k.shape[0]
-                offload_engine.prefill_k_buffer[self.layer_id, :num_tokens].copy_(k)
-                offload_engine.prefill_v_buffer[self.layer_id, :num_tokens].copy_(v)
+                #! GPU 2 GPU
+                offload_engine.write_to_prefill_buffer(self.layer_id, k, v, chunk_idx=chunk_idx)
         elif is_chunked_offload:
-            # Chunked decode mode: use compute_stream for store_kvcache
-            # This ensures proper synchronization with per-layer offload
-            compute_stream = context.kvcache_manager.offload_engine.compute_stream
-            if k_cache.numel() and v_cache.numel():
-                # CRITICAL: Wait for default stream to ensure slot_mapping tensor transfer is complete
-                # slot_mapping is created with non_blocking=True on default stream, but we use it
-                # on compute_stream. Without this sync, index_copy_ can get corrupted indices.
-                compute_stream.wait_stream(torch.cuda.default_stream())
-                with torch.cuda.stream(compute_stream):
-                    store_kvcache(k, v, k_cache, v_cache, context.slot_mapping)
+            # Chunked decode mode: write KV to per-layer decode buffer via offload_engine
+            # KV will be written to decode buffer in the decode branch below
+            # No store_kvcache needed - all KV management goes through offload_engine
+            pass
         else:
             # Normal mode: store on default stream
             if k_cache.numel() and v_cache.numel():
@@ -155,8 +149,7 @@ class Attention(nn.Module):
                 offload_engine = kvcache_manager.offload_engine
                 pos_in_block = context.decode_pos_in_block
                 # k, v shape: [1, kv_heads, head_dim]
-                offload_engine.decode_k_buffer[self.layer_id, pos_in_block].copy_(k.squeeze(0))
-                offload_engine.decode_v_buffer[self.layer_id, pos_in_block].copy_(v.squeeze(0))
+                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,

scripts/profile_offload.sh

@@ -9,6 +9,7 @@
 #   --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)
 #   --no-offload         Disable CPU offload
 #
 # Output:
@@ -18,6 +19,7 @@
 #   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
 
@@ -25,6 +27,7 @@ set -e
 DATASET="niah_single_1"
 SAMPLE_INDEX="0"
 GPU_ID="0"
+NUM_GPU_BLOCKS="4"
 ENABLE_OFFLOAD="--enable-offload"
 
 # Parse arguments
@@ -46,6 +49,10 @@ while [[ $# -gt 0 ]]; do
             ENABLE_OFFLOAD=""
             shift
             ;;
+        --num-gpu-blocks)
+            NUM_GPU_BLOCKS="$2"
+            shift 2
+            ;;
         -h|--help)
             echo "Usage: $0 [options]"
             echo ""
@@ -54,6 +61,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --sample INDEX       Sample index (default: 0)"
             echo "  --gpu GPU_ID         GPU to use (default: 0)"
             echo "  --no-offload         Disable CPU offload"
+            echo "  --num-gpu-blocks N   Number of GPU blocks/slots (default: 4)"
             exit 0
             ;;
         *)
@@ -76,7 +84,7 @@ mkdir -p "$OUTPUT_DIR"
 TIMESTAMP=$(date +%Y%m%d_%H%M%S)
 OFFLOAD_SUFFIX=""
 if [ -n "$ENABLE_OFFLOAD" ]; then
-    OFFLOAD_SUFFIX="_offload"
+    OFFLOAD_SUFFIX="_offload_${NUM_GPU_BLOCKS}slots"
 fi
 OUTPUT_FILE="$OUTPUT_DIR/ruler_${DATASET}_sample${SAMPLE_INDEX}${OFFLOAD_SUFFIX}_${TIMESTAMP}"
 
@@ -87,6 +95,7 @@ echo "Test script: $TEST_SCRIPT"
 echo "Dataset:     $DATASET"
 echo "Sample:      $SAMPLE_INDEX"
 echo "GPU:         $GPU_ID"
+echo "GPU Blocks:  $NUM_GPU_BLOCKS"
 echo "Offload:     ${ENABLE_OFFLOAD:-disabled}"
 echo "Output file: $OUTPUT_FILE.nsys-rep"
 echo ""
@@ -109,6 +118,7 @@ nsys profile \
     python "$TEST_SCRIPT" \
         --datasets "$DATASET" \
         --sample-indices "$SAMPLE_INDEX" \
+        --num-gpu-blocks "$NUM_GPU_BLOCKS" \
         $ENABLE_OFFLOAD \
         --quiet