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 15:12 CST

Browse Source
This commit is contained in:
Zijie Tian
2026-01-09 15:12:42 +08:00
parent 6378cb4c17
commit 59f8970ed3
2 changed files with 281 additions and 429 deletions
Download Patch File Download Diff File Expand all files Collapse all files

130
nanovllm/engine/model_runner.py
View File

@@ -44,7 +44,17 @@ class ModelRunner:
self.allocate_kv_cache() self.allocate_kv_cache()
if not self.enforce_eager: if not self.enforce_eager:
self.capture_cudagraph() if config.enable_cpu_offload:
# TODO: Implement capture_offload_cudagraph() for offload mode
# For now, offload mode uses eager execution
# The standard capture_cudagraph() cannot be used because:
# - It captures the PagedAttention decode path via Attention.forward()
# - In offload mode, Attention.k_cache/v_cache are empty (KV is in ring buffer)
# - The refactored offload decode now uses Attention.forward() with ring buffer
# - Need specialized graph capture that sets up ring buffer correctly
pass
else:
self.capture_cudagraph()
torch.set_default_device("cpu") torch.set_default_device("cpu")
torch.set_default_dtype(default_dtype) torch.set_default_dtype(default_dtype)
@@ -845,9 +855,9 @@ class ModelRunner:
Key design: Key design:
- Ring buffer pipeline: load layer N+k while computing layer N - Ring buffer pipeline: load layer N+k while computing layer N
- Uses standard Attention.forward() path (not bypassing)
- Per-layer decode buffer for accumulating new tokens - Per-layer decode buffer for accumulating new tokens
- Async block offload when decode buffer is full - Async block offload when decode buffer is full
- Uses OffloadEngine's ring buffer API for H2D pipeline
""" """
assert len(seqs) == 1, "Layer-wise offload only supports single sequence" assert len(seqs) == 1, "Layer-wise offload only supports single sequence"
seq = seqs[0] seq = seqs[0]
@@ -881,11 +891,15 @@ class ModelRunner:
# Current decode position info # Current decode position info
pos_in_block = (len(seq) - 1) % self.block_size pos_in_block = (len(seq) - 1) % self.block_size
decode_start_pos = self.kvcache_manager.get_decode_start_pos(seq) decode_start_pos = self.kvcache_manager.get_decode_start_pos(seq)
num_decode_tokens = pos_in_block - decode_start_pos + 1 num_prev_decode_tokens = pos_in_block - decode_start_pos # Previous decode tokens (not including current)
# Import FlashAttention once # Total context length (prefill + previous decode tokens)
from flash_attn.flash_attn_interface import flash_attn_varlen_func # New token will be stored at this position
cu_seqlens_q = torch.tensor([0, 1], dtype=torch.int32, device="cuda") context_len = total_prefill_tokens + num_prev_decode_tokens
# Context setup for Attention.forward() - contiguous mode (no block tables)
slot_mapping = torch.tensor([context_len], dtype=torch.int32, device="cuda")
context_lens = torch.tensor([context_len + 1], dtype=torch.int32, device="cuda")
# Phase 1: Preload first N layers to ring buffer (fill pipeline) # Phase 1: Preload first N layers to ring buffer (fill pipeline)
num_preload = min(num_buffers, num_layers) num_preload = min(num_buffers, num_layers)
@@ -902,94 +916,70 @@ class ModelRunner:
# Phase 2: Layer-by-layer processing with ring buffer pipeline # Phase 2: Layer-by-layer processing with ring buffer pipeline
for layer_id in range(num_layers): for layer_id in range(num_layers):
layer = self.model.model.layers[layer_id] layer = self.model.model.layers[layer_id]
attn_module = layer.self_attn.attn # The Attention module
current_buffer = layer_id % num_buffers current_buffer = layer_id % num_buffers
# 2a. Wait for current buffer's load to complete # 2a. Wait for current buffer's load to complete
offload_engine.wait_buffer_load(current_buffer) offload_engine.wait_buffer_load(current_buffer)
# 2c. Input LayerNorm # 2b. Copy previous decode KV from decode buffer to ring buffer
if residual is None: # Ring buffer already has prefill KV at [0:total_prefill_tokens]
hidden_ln, residual = layer.input_layernorm(hidden_states), hidden_states # We need to add decode KV at [total_prefill_tokens:]
else: if num_prev_decode_tokens > 0:
hidden_ln, residual = layer.input_layernorm(hidden_states, residual)
# 2d. QKV projection for new token
qkv = layer.self_attn.qkv_proj(hidden_ln)
q, k_new, v_new = qkv.split([
layer.self_attn.q_size,
layer.self_attn.kv_size,
layer.self_attn.kv_size
], dim=-1)
q = q.view(1, layer.self_attn.num_heads, layer.self_attn.head_dim)
k_new = k_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
v_new = v_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
# Q/K norms
if not layer.self_attn.qkv_bias:
q = layer.self_attn.q_norm(q.reshape(-1, layer.self_attn.head_dim))
q = q.view(1, layer.self_attn.num_heads, layer.self_attn.head_dim)
k_new = layer.self_attn.k_norm(k_new.reshape(-1, layer.self_attn.head_dim))
k_new = k_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
# RoPE
q, k_new = layer.self_attn.rotary_emb(positions, q, k_new)
# 2e. Get prefilled KV from ring buffer
k_prefill, v_prefill = offload_engine.get_buffer_kv(current_buffer, total_prefill_tokens)
# 2f. Get accumulated decode KV from decode buffer (if any previous decode tokens)
if num_decode_tokens > 1:
k_decode_prev, v_decode_prev = offload_engine.get_decode_kv( k_decode_prev, v_decode_prev = offload_engine.get_decode_kv(
layer_id, decode_start_pos, pos_in_block layer_id, decode_start_pos, pos_in_block
) )
k_full = torch.cat([k_prefill, k_decode_prev, k_new], dim=0) ring_k = offload_engine.layer_k_cache[current_buffer]
v_full = torch.cat([v_prefill, v_decode_prev, v_new], dim=0) ring_v = offload_engine.layer_v_cache[current_buffer]
else: ring_k[total_prefill_tokens:total_prefill_tokens + num_prev_decode_tokens].copy_(k_decode_prev)
k_full = torch.cat([k_prefill, k_new], dim=0) ring_v[total_prefill_tokens:total_prefill_tokens + num_prev_decode_tokens].copy_(v_decode_prev)
v_full = torch.cat([v_prefill, v_new], dim=0)
# 2g. Store new KV to decode buffer for future decode steps # 2c. Set Attention module's cache to ring buffer (contiguous format)
offload_engine.store_decode_kv(layer_id, pos_in_block, k_new, v_new) # Shape: [max_seq_len, kv_heads, head_dim] -> [1, max_seq_len, kv_heads, head_dim]
attn_module.k_cache = offload_engine.layer_k_cache[current_buffer:current_buffer+1]
attn_module.v_cache = offload_engine.layer_v_cache[current_buffer:current_buffer+1]
# 2h. Mark buffer compute done (allows next load to reuse this buffer) # 2d. Set context for Attention.forward() - contiguous mode
set_context(
is_prefill=False,
slot_mapping=slot_mapping,
context_lens=context_lens,
block_tables=None, # Contiguous mode, no block tables
)
# 2e. Forward through layer using standard path
# This calls Qwen3Attention.forward() -> Attention.forward()
# Attention.forward() will:
# - Store new K,V to ring buffer via store_kvcache
# - Compute attention via flash_attn_with_kvcache
hidden_states, residual = layer(positions, hidden_states, residual)
# 2f. Copy new token's KV from ring buffer to decode buffer (for persistence)
# The new token was stored at position context_len in ring buffer
ring_k = offload_engine.layer_k_cache[current_buffer]
ring_v = offload_engine.layer_v_cache[current_buffer]
offload_engine.decode_k_buffer[layer_id, pos_in_block].copy_(ring_k[context_len])
offload_engine.decode_v_buffer[layer_id, pos_in_block].copy_(ring_v[context_len])
# 2g. Mark buffer compute done (allows next load to reuse this buffer)
offload_engine.record_buffer_compute_done(current_buffer) offload_engine.record_buffer_compute_done(current_buffer)
# 2i. Start loading next layer to same buffer (after compute done) # 2h. Start loading next layer to same buffer (after compute done)
next_layer_to_load = layer_id + num_buffers next_layer_to_load = layer_id + num_buffers
if next_layer_to_load < num_layers: if next_layer_to_load < num_layers:
offload_engine.load_layer_kv_to_buffer( offload_engine.load_layer_kv_to_buffer(
current_buffer, next_layer_to_load, cpu_block_table, valid_tokens_per_block current_buffer, next_layer_to_load, cpu_block_table, valid_tokens_per_block
) )
# 2j. Compute attention
total_kv_tokens = k_full.shape[0]
cu_seqlens_k = torch.tensor([0, total_kv_tokens], dtype=torch.int32, device="cuda")
attn_output = flash_attn_varlen_func(
q, k_full, v_full,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=1,
max_seqlen_k=total_kv_tokens,
softmax_scale=layer.self_attn.attn.scale,
causal=False,
)
# O projection
attn_output = attn_output.view(1, -1)
hidden_states = layer.self_attn.o_proj(attn_output)
# 2k. Post-attention LayerNorm + MLP
hidden_states, residual = layer.post_attention_layernorm(hidden_states, residual)
hidden_states = layer.mlp(hidden_states)
# Step 3: Final norm # Step 3: Final norm
hidden_states, _ = self.model.model.norm(hidden_states, residual) hidden_states, _ = self.model.model.norm(hidden_states, residual)
# Step 4: Compute logits # Step 4: Compute logits
logits = self.model.compute_logits(hidden_states) logits = self.model.compute_logits(hidden_states)
# Reset context
reset_context()
# Step 5: Handle block-full offload (async) # Step 5: Handle block-full offload (async)
if pos_in_block == self.block_size - 1: if pos_in_block == self.block_size - 1:
last_cpu_block = self.kvcache_manager.get_last_cpu_block(seq) last_cpu_block = self.kvcache_manager.get_last_cpu_block(seq)

580
task_plan.md
View File

@@ -1,412 +1,274 @@
# Task Plan: Fix GPU-only Mode Performance Issue # Task Plan: Enable CUDA Graphs for CPU Offload Mode
## Goal
Eliminate the `store_kvcache` scatter overhead in GPU-only mode by using **contiguous KV cache layout** (like offload mode), avoiding PagedAttention's blocked layout for single-sequence inference.
## Problem Summary ## Problem Summary
GPU-only mode with MInference is **slower** than CPU offload mode: Running `bench_offload.py` fails with:
```
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)
```
| Mode | Prefill Speed (32K tokens, Qwen3-4B) | **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.
|------|--------------------------------------|
| GPU-only + MInference | 3383 tok/s |
| Offload + MInference | 5373 tok/s |
**Root cause**: PagedAttention's blocked layout requires expensive `index_copy_` scatter operations to convert contiguous K,V to blocked format. **User requirement**: Enable CUDA graphs in offload mode for better decode performance.
## Key Insight: Why Offload is Fast ## Deep Analysis: Why Current Design is Incompatible
Offload mode uses **contiguous layout** for KV cache: ### 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 ```python
# OffloadEngine's CPU cache layout # Before (dynamic, not graphable):
k_cache_cpu: [num_layers, num_blocks, block_size, kv_heads, head_dim] k_full = torch.cat([k_prefill, k_decode_prev, k_new], dim=0)
o = flash_attn_varlen_func(q, k_full, v_full, ...)
# Store is simple contiguous slice assignment # After (fixed addresses, graphable):
self.k_cache_cpu[layer_id, block_id, :actual_size].copy_(k[start:end]) # 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,
)
``` ```
The K,V computed during prefill `[seq_len, kv_heads, head_dim]` matches the cache layout - no format conversion needed! ## Implementation Plan
## Solution: Contiguous Layout for GPU-only Mode ### Phase 1: Modify Offload Decode for CUDA Graph Compatibility
For GPU-only single-sequence mode, use the **same contiguous layout as offload mode**, but on GPU:
```
Current GPU-only (PagedAttention):
Cache: [num_blocks, block_size, kv_heads, head_dim] (blocked)
Store: scatter via index_copy_ (SLOW)
Proposed GPU-only (Contiguous):
Cache: [num_layers, max_seq_len, kv_heads, head_dim] (contiguous)
Store: slice assignment k_cache[layer_id, :seq_len] = k (FAST)
```
This mirrors offload mode's architecture but keeps everything on GPU - no cross-device transfer, no layout conversion.
## Phases
- [x] Phase 1: Add contiguous GPU KV cache in GPUOnlyManager (for single-seq mode)
- [x] Phase 2: Implement `run_gpu_only_prefill()` using contiguous cache
- [x] Phase 3: Implement decode path for contiguous cache
- [x] Phase 4: Test and validate performance
## Results
| Mode | 32K Prefill Speed | Notes |
|------|-------------------|-------|
| GPU-only (before) | ~3383 tok/s | PagedAttention scatter overhead |
| GPU-only contiguous (after) | **5293 tok/s** | 56% improvement |
| Offload mode | 5391 tok/s | Baseline comparison |
**Test passed**: `test_needle.py --input-len 32768 --max-model-len 40960` - correct output retrieved.
## Detailed Design
### Phase 1: Contiguous GPU KV Cache
**File**: `nanovllm/kvcache/gpu_manager.py`
Add contiguous cache allocation for single-sequence mode:
```python
class GPUOnlyManager(KVCacheManager):
def __init__(self, num_blocks: int, block_size: int, max_seq_len: int = 0):
# ... existing code ...
self.max_seq_len = max_seq_len
# Contiguous cache for single-seq mode (allocated in allocate_cache)
self.contiguous_k_cache = None # [num_layers, max_seq_len, kv_heads, head_dim]
self.contiguous_v_cache = None
def allocate_cache(
self,
num_layers: int,
num_kv_heads: int,
head_dim: int,
dtype: torch.dtype,
) -> None:
# Existing PagedAttention cache for multi-seq/decode
self.kv_cache = torch.empty(
2, num_layers, self._num_blocks, self._block_size,
num_kv_heads, head_dim,
dtype=dtype, device="cuda"
)
# Contiguous cache for single-seq prefill (if max_seq_len specified)
if self.max_seq_len > 0:
self.contiguous_k_cache = torch.empty(
num_layers, self.max_seq_len, num_kv_heads, head_dim,
dtype=dtype, device="cuda"
)
self.contiguous_v_cache = torch.empty(
num_layers, self.max_seq_len, num_kv_heads, head_dim,
dtype=dtype, device="cuda"
)
```
### Phase 2: Layer-wise GPU-only Prefill
**File**: `nanovllm/engine/model_runner.py` **File**: `nanovllm/engine/model_runner.py`
Following offload pattern exactly - store K,V per-layer to contiguous cache: **Changes**:
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()`
```python ```python
@torch.inference_mode() @torch.inference_mode()
def run_gpu_only_prefill(self, seqs: list[Sequence]) -> list[int]: def capture_offload_cudagraph(self):
""" """
GPU-only prefill with contiguous KV cache layout. Capture CUDA graphs for offload decode.
Mirrors run_layerwise_offload_prefill() but stores to GPU instead of CPU. Key design:
No scatter operations - just contiguous slice assignment. - 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
""" """
assert len(seqs) == 1, "GPU-only layer-wise prefill only supports single sequence" offload_engine = self.kvcache_manager.offload_engine
seq = seqs[0]
num_layers = len(self.model.model.layers) num_layers = len(self.model.model.layers)
total_tokens = len(seq) num_buffers = offload_engine.num_kv_buffers
max_seq_len = offload_engine.max_seq_len
# Get contiguous GPU cache # Fixed-address tensors for graph capture
k_cache = self.kvcache_manager.contiguous_k_cache input_ids = torch.zeros(1, dtype=torch.int64, device="cuda")
v_cache = self.kvcache_manager.contiguous_v_cache positions = torch.zeros(1, dtype=torch.int64, device="cuda")
cache_seqlens = torch.zeros(1, dtype=torch.int32, device="cuda")
hidden_output = torch.zeros(1, self.config.hf_config.hidden_size, device="cuda")
# Prepare inputs # Graph capture per buffer slot (deterministic: layer_id % num_buffers)
input_ids = torch.tensor(seq[:], dtype=torch.int64, device="cuda") self.offload_graphs = {}
positions = torch.arange(total_tokens, dtype=torch.int64, device="cuda") self.offload_graph_pool = None
from flash_attn.flash_attn_interface import flash_attn_varlen_func for buffer_idx in range(num_buffers):
cu_seqlens = torch.tensor([0, total_tokens], dtype=torch.int32, device="cuda") graph = torch.cuda.CUDAGraph()
# Embedding # Get fixed-address ring buffer for this slot
hidden_states = self.model.model.embed_tokens(input_ids) k_cache = offload_engine.layer_k_cache[buffer_idx:buffer_idx+1] # [1, max_seq, heads, dim]
residual = None v_cache = offload_engine.layer_v_cache[buffer_idx:buffer_idx+1]
# Layer-by-layer processing (same as offload prefill) # Warmup
for layer_id in range(num_layers): with torch.cuda.stream(offload_engine.compute_stream):
layer = self.model.model.layers[layer_id] # ... (layer forward pass using k_cache, v_cache)
pass
# Input LayerNorm # Capture
if residual is None: with torch.cuda.graph(graph, self.offload_graph_pool):
hidden_ln, residual = layer.input_layernorm(hidden_states), hidden_states # ... (same layer forward pass)
else: pass
hidden_ln, residual = layer.input_layernorm(hidden_states, residual)
# QKV projection if self.offload_graph_pool is None:
qkv = layer.self_attn.qkv_proj(hidden_ln) self.offload_graph_pool = graph.pool()
q, k, v = qkv.split([ self.offload_graphs[buffer_idx] = graph
layer.self_attn.q_size,
layer.self_attn.kv_size,
layer.self_attn.kv_size
], dim=-1)
q = q.view(total_tokens, layer.self_attn.num_heads, layer.self_attn.head_dim) self.offload_graph_vars = dict(
k = k.view(total_tokens, layer.self_attn.num_kv_heads, layer.self_attn.head_dim) input_ids=input_ids,
v = v.view(total_tokens, layer.self_attn.num_kv_heads, layer.self_attn.head_dim) positions=positions,
cache_seqlens=cache_seqlens,
# Q/K norms (Qwen3 specific) hidden_output=hidden_output,
if not layer.self_attn.qkv_bias: )
q = layer.self_attn.q_norm(q.reshape(-1, layer.self_attn.head_dim))
q = q.view(total_tokens, layer.self_attn.num_heads, layer.self_attn.head_dim)
k = layer.self_attn.k_norm(k.reshape(-1, layer.self_attn.head_dim))
k = k.view(total_tokens, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
# RoPE
q, k = layer.self_attn.rotary_emb(positions, q, k)
# Store K,V to contiguous GPU cache (same layout - no conversion!)
# This is just slice assignment, not scatter
k_cache[layer_id, :total_tokens] = k
v_cache[layer_id, :total_tokens] = v
# Sparse or Full attention (uses k, v directly)
if self.sparse_prefill_policy is not None:
attn_output = self.sparse_prefill_policy.sparse_prefill_attention(
q, k, v, layer_id
)
else:
attn_output = flash_attn_varlen_func(
q, k, v,
cu_seqlens_q=cu_seqlens,
cu_seqlens_k=cu_seqlens,
max_seqlen_q=total_tokens,
max_seqlen_k=total_tokens,
softmax_scale=layer.self_attn.attn.scale,
causal=True,
)
# O projection
attn_output = attn_output.view(total_tokens, -1)
hidden_states = layer.self_attn.o_proj(attn_output)
# Post-attention LayerNorm + MLP
hidden_states, residual = layer.post_attention_layernorm(hidden_states, residual)
hidden_states = layer.mlp(hidden_states)
# Final norm
hidden_states, _ = self.model.model.norm(hidden_states, residual)
# Compute logits
logits = self.model.compute_logits(hidden_states[-1:])
# Record prefill length for decode
self.kvcache_manager.contiguous_seq_len = total_tokens
# Sample
temperatures = self.prepare_sample(seqs) if self.rank == 0 else None
token_ids = self.sampler(logits, temperatures).tolist() if self.rank == 0 else None
return token_ids
``` ```
### Phase 3: Decode with Contiguous Cache #### 1.2 Modified `run_layerwise_offload_decode()`
Key changes:
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
# In the layer loop, replace:
k_full = torch.cat([k_prefill, k_decode_prev, k_new], dim=0)
attn_output = flash_attn_varlen_func(q, k_full, v_full, ...)
# With:
# 1. Get ring buffer slice
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)
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
total_kv_tokens = total_prefill_tokens + num_decode_tokens
k_buffer[total_kv_tokens-1].copy_(k_new.squeeze(0))
v_buffer[total_kv_tokens-1].copy_(v_new.squeeze(0))
# 4. Flash attention with fixed-address cache
cache_seqlens = torch.tensor([total_kv_tokens], dtype=torch.int32, device="cuda")
attn_output = flash_attn_with_kvcache(
q.unsqueeze(1), # [1, 1, heads, dim]
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__`
**File**: `nanovllm/engine/model_runner.py` **File**: `nanovllm/engine/model_runner.py`
```python **Change**: Line 46-47
@torch.inference_mode()
def run_gpu_only_decode(self, seqs: list[Sequence]) -> list[int]:
"""
Decode using contiguous GPU KV cache.
Similar to offload decode but simpler - all KV already on GPU.
"""
assert len(seqs) == 1
seq = seqs[0]
num_layers = len(self.model.model.layers)
k_cache = self.kvcache_manager.contiguous_k_cache
v_cache = self.kvcache_manager.contiguous_v_cache
context_len = self.kvcache_manager.contiguous_seq_len
# Prepare inputs
input_ids = torch.tensor([seq.last_token], dtype=torch.int64, device="cuda")
positions = torch.tensor([len(seq) - 1], dtype=torch.int64, device="cuda")
from flash_attn.flash_attn_interface import flash_attn_varlen_func
cu_seqlens_q = torch.tensor([0, 1], dtype=torch.int32, device="cuda")
# Embedding
hidden_states = self.model.model.embed_tokens(input_ids)
residual = None
for layer_id in range(num_layers):
layer = self.model.model.layers[layer_id]
# Input LayerNorm
if residual is None:
hidden_ln, residual = layer.input_layernorm(hidden_states), hidden_states
else:
hidden_ln, residual = layer.input_layernorm(hidden_states, residual)
# QKV projection
qkv = layer.self_attn.qkv_proj(hidden_ln)
q, k_new, v_new = qkv.split([
layer.self_attn.q_size,
layer.self_attn.kv_size,
layer.self_attn.kv_size
], dim=-1)
q = q.view(1, layer.self_attn.num_heads, layer.self_attn.head_dim)
k_new = k_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
v_new = v_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
# Q/K norms
if not layer.self_attn.qkv_bias:
q = layer.self_attn.q_norm(q.reshape(-1, layer.self_attn.head_dim))
q = q.view(1, layer.self_attn.num_heads, layer.self_attn.head_dim)
k_new = layer.self_attn.k_norm(k_new.reshape(-1, layer.self_attn.head_dim))
k_new = k_new.view(1, layer.self_attn.num_kv_heads, layer.self_attn.head_dim)
# RoPE
q, k_new = layer.self_attn.rotary_emb(positions, q, k_new)
# Get cached K,V and append new token
k_cached = k_cache[layer_id, :context_len]
v_cached = v_cache[layer_id, :context_len]
# Store new K,V to cache
k_cache[layer_id, context_len] = k_new.squeeze(0)
v_cache[layer_id, context_len] = v_new.squeeze(0)
# Full K,V for attention
k_full = k_cache[layer_id, :context_len + 1]
v_full = v_cache[layer_id, :context_len + 1]
# Attention
cu_seqlens_k = torch.tensor([0, context_len + 1], dtype=torch.int32, device="cuda")
attn_output = flash_attn_varlen_func(
q, k_full, v_full,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=1,
max_seqlen_k=context_len + 1,
softmax_scale=layer.self_attn.attn.scale,
causal=False, # Single query, no causal needed
)
# O projection
attn_output = attn_output.view(1, -1)
hidden_states = layer.self_attn.o_proj(attn_output)
# Post-attention LayerNorm + MLP
hidden_states, residual = layer.post_attention_layernorm(hidden_states, residual)
hidden_states = layer.mlp(hidden_states)
# Update context length
self.kvcache_manager.contiguous_seq_len = context_len + 1
# Final norm
hidden_states, _ = self.model.model.norm(hidden_states, residual)
# Compute logits
logits = self.model.compute_logits(hidden_states)
# Sample
temperatures = self.prepare_sample(seqs) if self.rank == 0 else None
token_ids = self.sampler(logits, temperatures).tolist() if self.rank == 0 else None
return token_ids
```
### Phase 4: Decision Logic
```python ```python
def _should_use_contiguous_gpu_mode(self, seqs: list[Sequence], is_prefill: bool) -> bool: # Current (crashes in offload mode):
"""Check if contiguous GPU mode should be used.""" if not self.enforce_eager:
# Must have contiguous cache allocated self.capture_cudagraph()
if not hasattr(self.kvcache_manager, 'contiguous_k_cache'):
return False
if self.kvcache_manager.contiguous_k_cache is None:
return False
# Must NOT be offload mode # Fixed (conditional capture based on mode):
if hasattr(self.kvcache_manager, 'offload_engine'): if not self.enforce_eager:
return False if config.enable_cpu_offload:
self.capture_offload_cudagraph() # New method for offload mode
# Single sequence only else:
if len(seqs) != 1: self.capture_cudagraph() # Standard PagedAttention decode
return False
# For prefill: has blocks (not warmup)
if is_prefill and not seqs[0].block_table:
return False
return True
def run(self, seqs: list[Sequence], is_prefill: bool) -> list[int]:
# Check offload mode (existing)
if hasattr(self, 'kvcache_manager') and hasattr(self.kvcache_manager, 'offload_engine'):
...
# Check contiguous GPU mode
if self._should_use_contiguous_gpu_mode(seqs, is_prefill):
if is_prefill:
return self.run_gpu_only_prefill(seqs)
else:
return self.run_gpu_only_decode(seqs)
# Standard PagedAttention path
...
``` ```
## Architecture Comparison ### Phase 3: Per-Layer Graph vs Full-Decode Graph
| Aspect | Offload Mode | GPU-only (Proposed) | GPU-only (Current) | Two approaches for graph capture:
|--------|--------------|---------------------|-------------------|
| Cache location | CPU (contiguous) | GPU (contiguous) | GPU (PagedAttention) |
| Cache layout | `[layers, blocks, block_size, heads, dim]` | `[layers, max_seq_len, heads, dim]` | `[blocks, block_size, heads, dim]` |
| Prefill store | Contiguous slice copy | **Slice assignment (no copy!)** | Scatter (index_copy_) |
| Decode read | H2D ring buffer | Direct GPU access | PagedAttention |
## Key Points #### 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
1. **No explicit copy_ needed**: Slice assignment `cache[layer, :len] = k` is direct memory write #### Option B: Full-Decode Graph (More Complex, Maximum Overhead Reduction)
2. **Same layout as computed K,V**: No format conversion required - Capture one graph for entire decode step (all layers)
3. **Mirrors offload architecture**: Same layer-wise processing pattern - Requires all H2D loads completed before graph replay
4. **GPU advantage**: No cross-device transfer, faster than offload - Better kernel fusion, less CPU overhead
## Memory Usage **Recommendation**: Start with Option A (simpler), optimize to Option B later.
Contiguous GPU cache: `2 * num_layers * max_seq_len * kv_heads * head_dim * dtype_size` ## Implementation Phases
For Qwen3-4B with 32K max_seq_len: | Phase | Description | Status |
- `2 * 28 * 32768 * 8 * 128 * 2 = 3.5GB` |-------|-------------|--------|
| Phase 1 | Modify decode to use fixed-address buffers + flash_attn_with_kvcache | [ ] |
| Phase 2 | Add `capture_offload_cudagraph()` method | [ ] |
| Phase 3 | Update `__init__` to call correct capture method | [ ] |
| Phase 4 | Test with `bench_offload.py` | [ ] |
| Phase 5 | Benchmark performance improvement | [ ] |
Same as offload mode's CPU cache, but on GPU. ## Key Code Changes Summary
## Files to Modify | File | Change |
|------|--------|
| `model_runner.py:46-47` | Conditional CUDA graph capture based on offload mode |
| `model_runner.py` (new) | Add `capture_offload_cudagraph()` method |
| `model_runner.py:850-1010` | Modify `run_layerwise_offload_decode()` to use fixed-address attention |
| File | Changes | ## Alternative: Quick Fix (Skip Graph Capture)
|------|---------|
| `nanovllm/kvcache/gpu_manager.py` | Add contiguous cache allocation |
| `nanovllm/engine/model_runner.py` | Add `run_gpu_only_prefill()`, `run_gpu_only_decode()`, modify `run()` |
## Expected Performance If CUDA graph support is not immediately needed, the simpler fix is:
| Metric | Before | After | Improvement | ```python
|--------|--------|-------|-------------| # Line 46-47 in model_runner.py
| GPU-only prefill (32K) | 3383 tok/s | ~5400+ tok/s | ~60%+ | if not self.enforce_eager and not config.enable_cpu_offload:
| Decode | Baseline | Similar | ~0% | self.capture_cudagraph()
```
## Status This skips CUDA graph capture entirely in offload mode. Offload mode will use eager execution (which already works).
**Currently in Phase 1** - Ready to implement contiguous GPU cache
## Risk Assessment
| Risk | Mitigation |
|------|------------|
| 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
| Metric | Without Graph | With Graph | Improvement |
|--------|---------------|------------|-------------|
| Decode latency per token | Baseline | ~10-20% faster | Reduced kernel launch overhead |
| GPU utilization | Medium | Higher | Better kernel fusion |