zijie-tian/nano-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[claudesquad] update from 'fix-bug-2' on 09 Jan 26 15:16 CST

Browse Source
This commit is contained in:
Zijie Tian
2026-01-09 15:16:55 +08:00
parent 59f8970ed3
commit ccf04d3917
1 changed files with 163 additions and 214 deletions
Download Patch File Download Diff File Expand all files Collapse all files

377
task_plan.md
View File

@@ -1,274 +1,223 @@
# Task Plan: Enable CUDA Graphs for CPU Offload Mode # Task Plan: Enable CUDA Graphs for CPU Offload Mode
## Problem Summary ## Current Status
Running `bench_offload.py` fails with: ### Completed: Refactor Offload Decode to Use Standard Attention Path
**Problem solved**: The original offload decode (`run_layerwise_offload_decode`) bypassed `Attention.forward()` by manually calling attention components. This was inconsistent with the standard execution path.
**Solution implemented**: Refactored to use `layer.forward()` which goes through:
``` ```
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1) Qwen3DecoderLayer.forward()
→ Qwen3Attention.forward()
→ Attention.forward() ← Now properly used!
``` ```
**Root cause**: In offload mode, `HybridKVCacheManager.get_layer_cache()` returns empty tensors (by design), but CUDA graph capture calls `Attention.forward()` decode path which expects valid k_cache/v_cache. ### Code Changes Made
**User requirement**: Enable CUDA graphs in offload mode for better decode performance.
## Deep Analysis: Why Current Design is Incompatible
### Current Offload Decode Flow (`run_layerwise_offload_decode`)
```
1. Preload N layers to ring buffer (H2D async)
2. For each layer:
a. Wait for buffer load
b. LayerNorm → QKV proj → RoPE
c. k_full = torch.cat([k_prefill, k_decode_prev, k_new]) <-- DYNAMIC SHAPE
d. flash_attn_varlen_func(q, k_full, v_full, ...) <-- VARIABLE LENGTH
e. O_proj → MLP
f. Start next layer H2D load
3. Final norm → Logits → Sample
```
### CUDA Graph Incompatibility Points
| Issue | Location | Why Incompatible |
|-------|----------|------------------|
| Dynamic tensor creation | `torch.cat([k_prefill, ...])` | Creates new tensors with variable shapes |
| Variable-length attention | `flash_attn_varlen_func` | `max_seqlen_k` changes every step |
| Data-dependent branching | `if num_decode_tokens > 1` | Control flow varies at runtime |
| Empty k_cache/v_cache | `Attention.forward()` | Current capture uses standard decode path |
### Why Empty Tensors in Offload Mode?
`HybridKVCacheManager.get_layer_cache()` returns empty tensors because:
- Offload mode manages KV via `OffloadEngine`'s ring buffer
- The standard `Attention.forward()` is NEVER used in offload inference
- Empty tensors are intentional placeholders
## Solution: Fixed-Address CUDA Graph Capture for Offload Decode
### Key Insight
The `OffloadEngine` ring buffer already has **fixed GPU addresses** with **fixed max shape**:
```python
layer_k_cache: [num_kv_buffers, max_seq_len, kv_heads, head_dim] # Fixed!
layer_v_cache: [num_kv_buffers, max_seq_len, kv_heads, head_dim] # Fixed!
```
`flash_attn_with_kvcache` supports **cache_seqlens** parameter for variable actual lengths with fixed-shape cache. This is the key to CUDA graph compatibility!
### Solution Design
Replace `torch.cat` + `flash_attn_varlen_func` with:
1. Pre-copy decode buffer content to ring buffer at correct offset
2. Store new token KV directly to ring buffer
3. Use `flash_attn_with_kvcache` with `cache_seqlens` for variable length
```python
# Before (dynamic, not graphable):
k_full = torch.cat([k_prefill, k_decode_prev, k_new], dim=0)
o = flash_attn_varlen_func(q, k_full, v_full, ...)
# After (fixed addresses, graphable):
# Ring buffer already has k_prefill at [0:prefill_len]
# Copy decode_prev and k_new to buffer at [prefill_len:]
ring_buffer[prefill_len:prefill_len+decode_len] = decode_buffer
ring_buffer[total_len-1] = k_new
o = flash_attn_with_kvcache(
q.unsqueeze(1), # [1, 1, heads, dim] - fixed shape
ring_k.unsqueeze(0), # [1, max_seq_len, heads, dim] - FIXED ADDRESS
ring_v.unsqueeze(0), # [1, max_seq_len, heads, dim] - FIXED ADDRESS
cache_seqlens=total_tokens_tensor, # [1] - variable VALUE, fixed address
softmax_scale=scale,
causal=True,
)
```
## Implementation Plan
### Phase 1: Modify Offload Decode for CUDA Graph Compatibility
**File**: `nanovllm/engine/model_runner.py` **File**: `nanovllm/engine/model_runner.py`
**Changes**: 1. **`run_layerwise_offload_decode()` (line 841-991)** - Completely refactored:
1. Add `capture_offload_cudagraph()` method
2. Modify `run_layerwise_offload_decode()` to use fixed-address buffers
3. Replace `flash_attn_varlen_func` with `flash_attn_with_kvcache`
#### 1.1 New Method: `capture_offload_cudagraph()` Before (bypassed Attention):
```python
qkv = layer.self_attn.qkv_proj(hidden_ln)
q, k_new, v_new = qkv.split(...)
q = layer.self_attn.q_norm(...)
k = layer.self_attn.k_norm(...)
q, k = layer.self_attn.rotary_emb(...)
attn_output = flash_attn_varlen_func(q, k_full, v_full, ...) # Direct call!
hidden_states = layer.self_attn.o_proj(attn_output)
```
After (uses standard path):
```python
# Set up Attention module's cache to ring buffer
attn_module.k_cache = offload_engine.layer_k_cache[buffer_idx:buffer_idx+1]
attn_module.v_cache = offload_engine.layer_v_cache[buffer_idx:buffer_idx+1]
# Set context for contiguous mode
set_context(is_prefill=False, slot_mapping=..., context_lens=..., block_tables=None)
# Standard layer forward - goes through Attention.forward()!
hidden_states, residual = layer(positions, hidden_states, residual)
```
2. **`ModelRunner.__init__()` (line 46-57)** - Conditional CUDA graph capture:
```python
if not self.enforce_eager:
if config.enable_cpu_offload:
# TODO: Implement capture_offload_cudagraph()
pass # Temporarily use eager execution
else:
self.capture_cudagraph()
```
### Test Results
| Test | Mode | Status |
|------|------|--------|
| `test_needle.py --input-len 4096` | GPU-only | PASSED |
| `test_needle.py --input-len 4096 --enable-offload` | CPU offload | PASSED |
## Remaining Work: Implement Offload CUDA Graph
### Why Standard `capture_cudagraph()` Cannot Be Used
The standard capture function captures the PagedAttention decode path:
```python
# capture_cudagraph() sets up:
k_cache: [num_blocks, block_size, kv_heads, head_dim] # PagedAttention format
block_tables: [...] # Block indices for paged indexing
```
But offload mode uses contiguous ring buffer:
```python
# Offload decode sets up:
k_cache: [1, max_seq_len, kv_heads, head_dim] # Contiguous format
block_tables: None # No paging
```
### Implementation Plan for `capture_offload_cudagraph()`
#### Phase 1: Prepare Fixed-Address Tensors
```python ```python
@torch.inference_mode() @torch.inference_mode()
def capture_offload_cudagraph(self): def capture_offload_cudagraph(self):
""" """Capture CUDA graphs for offload decode using ring buffer."""
Capture CUDA graphs for offload decode.
Key design:
- Uses OffloadEngine's ring buffer as fixed-address k_cache/v_cache
- Captures per-layer compute (after H2D load is done)
- Uses flash_attn_with_kvcache with cache_seqlens for variable context
"""
offload_engine = self.kvcache_manager.offload_engine offload_engine = self.kvcache_manager.offload_engine
num_layers = len(self.model.model.layers)
num_buffers = offload_engine.num_kv_buffers num_buffers = offload_engine.num_kv_buffers
max_seq_len = offload_engine.max_seq_len
# Fixed-address tensors for graph capture # Fixed-address tensors for graph capture
input_ids = torch.zeros(1, dtype=torch.int64, device="cuda") input_ids = torch.zeros(1, dtype=torch.int64, device="cuda")
positions = torch.zeros(1, dtype=torch.int64, device="cuda") positions = torch.zeros(1, dtype=torch.int64, device="cuda")
cache_seqlens = torch.zeros(1, dtype=torch.int32, device="cuda") slot_mapping = torch.zeros(1, dtype=torch.int32, device="cuda")
hidden_output = torch.zeros(1, self.config.hf_config.hidden_size, device="cuda") context_lens = torch.zeros(1, dtype=torch.int32, device="cuda")
# Graph capture per buffer slot (deterministic: layer_id % num_buffers)
self.offload_graphs = {} self.offload_graphs = {}
self.offload_graph_pool = None self.offload_graph_pool = None
```
#### Phase 2: Capture Per-Buffer Graphs
Since layer processing rotates through ring buffers (`layer_id % num_buffers`), we need graphs for each buffer slot:
```python
for buffer_idx in range(num_buffers): for buffer_idx in range(num_buffers):
graph = torch.cuda.CUDAGraph() graph = torch.cuda.CUDAGraph()
# Get fixed-address ring buffer for this slot # Set Attention cache to this buffer slot (fixed address)
k_cache = offload_engine.layer_k_cache[buffer_idx:buffer_idx+1] # [1, max_seq, heads, dim] for layer in self.model.model.layers:
v_cache = offload_engine.layer_v_cache[buffer_idx:buffer_idx+1] layer.self_attn.attn.k_cache = offload_engine.layer_k_cache[buffer_idx:buffer_idx+1]
layer.self_attn.attn.v_cache = offload_engine.layer_v_cache[buffer_idx:buffer_idx+1]
# Set context
set_context(is_prefill=False, slot_mapping=slot_mapping,
context_lens=context_lens, block_tables=None)
# Warmup # Warmup
with torch.cuda.stream(offload_engine.compute_stream): hidden = self.model.model.embed_tokens(input_ids)
# ... (layer forward pass using k_cache, v_cache) residual = None
pass for layer_id, layer in enumerate(self.model.model.layers):
if layer_id % num_buffers == buffer_idx:
hidden, residual = layer(positions, hidden, residual)
# Capture # Capture
with torch.cuda.graph(graph, self.offload_graph_pool): with torch.cuda.graph(graph, self.offload_graph_pool):
# ... (same layer forward pass) # Same operations
pass ...
if self.offload_graph_pool is None:
self.offload_graph_pool = graph.pool()
self.offload_graphs[buffer_idx] = graph self.offload_graphs[buffer_idx] = graph
self.offload_graph_vars = dict(
input_ids=input_ids,
positions=positions,
cache_seqlens=cache_seqlens,
hidden_output=hidden_output,
)
``` ```
#### 1.2 Modified `run_layerwise_offload_decode()` #### Phase 3: Use Graphs in Decode
Key changes: Modify `run_layerwise_offload_decode()` to replay graphs:
1. Copy decode buffer content to ring buffer before attention
2. Store new token directly to ring buffer
3. Use `flash_attn_with_kvcache` instead of `flash_attn_varlen_func`
4. Optionally use captured CUDA graph for per-layer compute
```python ```python
# In the layer loop, replace: for layer_id in range(num_layers):
k_full = torch.cat([k_prefill, k_decode_prev, k_new], dim=0) current_buffer = layer_id % num_buffers
attn_output = flash_attn_varlen_func(q, k_full, v_full, ...)
# With: # Wait for H2D load
# 1. Get ring buffer slice offload_engine.wait_buffer_load(current_buffer)
k_buffer = offload_engine.layer_k_cache[buffer_idx] # [max_seq_len, heads, dim]
v_buffer = offload_engine.layer_v_cache[buffer_idx]
# 2. Copy decode buffer to ring buffer (after prefill content) # Copy decode buffer to ring buffer (same as current)
if num_decode_tokens > 1: ...
k_buffer[total_prefill_tokens:total_prefill_tokens+num_decode_tokens-1].copy_(k_decode_prev)
v_buffer[total_prefill_tokens:total_prefill_tokens+num_decode_tokens-1].copy_(v_decode_prev)
# 3. Store new token to ring buffer # Update graph variables
total_kv_tokens = total_prefill_tokens + num_decode_tokens self.offload_graph_vars["positions"][0] = positions[0]
k_buffer[total_kv_tokens-1].copy_(k_new.squeeze(0)) self.offload_graph_vars["slot_mapping"][0] = context_len
v_buffer[total_kv_tokens-1].copy_(v_new.squeeze(0)) self.offload_graph_vars["context_lens"][0] = context_len + 1
# 4. Flash attention with fixed-address cache # Replay graph instead of eager forward
cache_seqlens = torch.tensor([total_kv_tokens], dtype=torch.int32, device="cuda") self.offload_graphs[current_buffer].replay()
attn_output = flash_attn_with_kvcache(
q.unsqueeze(1), # [1, 1, heads, dim] # Copy new KV to decode buffer (same as current)
k_buffer.unsqueeze(0), # [1, max_seq_len, heads, dim] - FIXED ADDRESS ...
v_buffer.unsqueeze(0), # [1, max_seq_len, heads, dim] - FIXED ADDRESS
cache_seqlens=cache_seqlens,
softmax_scale=layer.self_attn.attn.scale,
causal=True,
)
attn_output = attn_output.squeeze(1) # [1, heads*dim]
``` ```
### Phase 2: Handle CUDA Graph Capture in `__init__` ### Challenges and Considerations
**File**: `nanovllm/engine/model_runner.py` | Challenge | Solution |
|-----------|----------|
| H2D transfers interleaved with compute | H2D happens outside graph, only compute is captured |
| Different layers use different buffers | Capture per-buffer graphs, replay correct one |
| Variable context length | Use `cache_seqlens` parameter (fixed address, variable value) |
| Per-layer buffer rotation | Graph captures single-layer forward, loop in Python |
**Change**: Line 46-47 ### Alternative: Full-Decode Graph (More Complex)
```python Instead of per-layer graphs, capture entire decode step:
# Current (crashes in offload mode): 1. Complete all H2D loads before graph
if not self.enforce_eager: 2. Single graph covers all layers
self.capture_cudagraph() 3. Better kernel fusion, less CPU overhead
4. More complex to implement (need to handle buffer rotation inside graph)
# Fixed (conditional capture based on mode):
if not self.enforce_eager:
if config.enable_cpu_offload:
self.capture_offload_cudagraph() # New method for offload mode
else:
self.capture_cudagraph() # Standard PagedAttention decode
```
### Phase 3: Per-Layer Graph vs Full-Decode Graph
Two approaches for graph capture:
#### Option A: Per-Layer Graphs (Simpler, Less Overhead Reduction)
- Capture N graphs (one per buffer slot)
- Each graph covers: LayerNorm → QKV → RoPE → Attention → O_proj → MLP
- H2D transfers and buffer management outside graph
#### Option B: Full-Decode Graph (More Complex, Maximum Overhead Reduction)
- Capture one graph for entire decode step (all layers)
- Requires all H2D loads completed before graph replay
- Better kernel fusion, less CPU overhead
**Recommendation**: Start with Option A (simpler), optimize to Option B later.
## Implementation Phases ## Implementation Phases
| Phase | Description | Status | | Phase | Description | Status |
|-------|-------------|--------| |-------|-------------|--------|
| Phase 1 | Modify decode to use fixed-address buffers + flash_attn_with_kvcache | [ ] | | Phase 0 | Refactor offload decode to use Attention.forward() | ✅ Completed |
| Phase 2 | Add `capture_offload_cudagraph()` method | [ ] | | Phase 1 | Implement `capture_offload_cudagraph()` with per-buffer graphs | ⬜ Pending |
| Phase 3 | Update `__init__` to call correct capture method | [ ] | | Phase 2 | Modify `run_layerwise_offload_decode()` to use graphs | ⬜ Pending |
| Phase 4 | Test with `bench_offload.py` | [ ] | | Phase 3 | Test and benchmark | ⬜ Pending |
| Phase 5 | Benchmark performance improvement | [ ] | | Phase 4 | (Optional) Optimize to full-decode graph | ⬜ Future |
## Key Code Changes Summary ## Architecture After Refactoring
| File | Change | ```
|------|--------| ┌─────────────────────────────────────────────────────────────────────────────┐
| `model_runner.py:46-47` | Conditional CUDA graph capture based on offload mode | │ Offload Decode Flow (After Refactoring) │
| `model_runner.py` (new) | Add `capture_offload_cudagraph()` method | ├─────────────────────────────────────────────────────────────────────────────┤
| `model_runner.py:850-1010` | Modify `run_layerwise_offload_decode()` to use fixed-address attention | │ │
│ For each layer: │
## Alternative: Quick Fix (Skip Graph Capture) │ 1. Wait for H2D load (ring buffer has prefill KV) │
│ 2. Copy decode buffer → ring buffer (at prefill_len offset) │
If CUDA graph support is not immediately needed, the simpler fix is: │ 3. Set Attention.k_cache = ring_buffer[buffer_idx] │
│ 4. Set context (slot_mapping, context_lens, block_tables=None) │
```python │ 5. layer.forward() → Qwen3Attention.forward() → Attention.forward() │
# Line 46-47 in model_runner.py │ └── store_kvcache() stores new token to ring buffer │
if not self.enforce_eager and not config.enable_cpu_offload: │ └── flash_attn_with_kvcache() computes attention │
self.capture_cudagraph() │ 6. Copy new token KV: ring buffer → decode buffer │
│ 7. Start next layer H2D load │
│ │
│ Key insight: Now uses standard Attention path, just with ring buffer │
│ as k_cache/v_cache in contiguous format (block_tables=None) │
│ │
└─────────────────────────────────────────────────────────────────────────────┘
``` ```
This skips CUDA graph capture entirely in offload mode. Offload mode will use eager execution (which already works). ## Files Modified
## Risk Assessment | File | Changes |
|------|---------|
| `model_runner.py:46-57` | Conditional CUDA graph capture (skip for offload) |
| `model_runner.py:841-991` | Refactored `run_layerwise_offload_decode()` to use standard `layer.forward()` |
| Risk | Mitigation | ## Next Steps
|------|------------|
| flash_attn_with_kvcache API differences | Test with actual flash-attn version |
| Memory overhead of fixed-size buffers | Already allocated in OffloadEngine |
| Performance regression | Benchmark before/after |
| Graph capture complexity | Start with per-layer graphs |
## Expected Performance Impact 1. Implement `capture_offload_cudagraph()` method
2. Modify `run_layerwise_offload_decode()` to optionally use captured graphs
| Metric | Without Graph | With Graph | Improvement | 3. Benchmark performance improvement from CUDA graphs
|--------|---------------|------------|-------------| 4. Consider full-decode graph optimization for maximum performance
| Decode latency per token | Baseline | ~10-20% faster | Reduced kernel launch overhead |
| GPU utilization | Medium | Higher | Better kernel fusion |