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

🐛 fix: stream synchronization for XAttention estimate kernels in offload mode

Browse Source 
- Wrap all compute kernels in select_blocks with compute_stream context
  (Pass 1 historical blocks, Pass 1 current chunk, Step 2 merge,
   Pass 2 historical blocks, Pass 2 current chunk, Step 4 block selection)
- Fix K data mismatch between Pass 1 and Pass 2 by ensuring wait_slot_layer
  syncs with compute_stream where kernels actually run
- Remove STRONG SYNC code from offload_engine.py (now handled by events)
- Remove debug print statements and torch.save code
- Consolidate fallback conditions in compute_with_xattn
- Change default chunk_size from 16384 to 4096 for density alignment

The bug caused Pass 1 and Pass 2 to see different K data from the same
CPU block because compute kernels ran on default stream while
wait_slot_layer only synced compute_stream.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
Zijie Tian
2026-02-05 01:30:23 +08:00
parent dd0472aea8
commit 829b311c02
1 changed files with 153 additions and 170 deletions
Download Patch File Download Diff File Expand all files Collapse all files

323
nanovllm/kvcache/sparse/xattn_bsa.py
View File

@@ -96,7 +96,7 @@ class XAttentionBSAPolicy(SparsePolicy):
self, self,
threshold: float = 0.95, # High threshold for accuracy testing threshold: float = 0.95, # High threshold for accuracy testing
stride: int = 8, stride: int = 8,
chunk_size: int = 16384, chunk_size: int = 4096, # Match offload Q chunk size for density alignment
block_size: int = 128, block_size: int = 128,
samples_per_chunk: int = 128, samples_per_chunk: int = 128,
use_triton: bool = True, use_triton: bool = True,
@@ -289,9 +289,11 @@ class XAttentionBSAPolicy(SparsePolicy):
Returns: Returns:
Attention output [total_q, num_heads, head_dim] Attention output [total_q, num_heads, head_dim]
""" """
# When block_tables is provided (paged KV cache / prefix cache), # Fallback to flash attention when:
# fallback to flash_attn as XAttention expects contiguous K, V # 1. block_tables provided (paged KV cache / prefix cache) - XAttention expects contiguous K, V
if block_tables is not None: # 2. BSA kernel not available
# 3. xattn_estimate not available
if block_tables is not None or not BSA_AVAILABLE or not XATTN_AVAILABLE:
from flash_attn import flash_attn_varlen_func from flash_attn import flash_attn_varlen_func
return flash_attn_varlen_func( return flash_attn_varlen_func(
q, k, v, q, k, v,
@@ -304,32 +306,6 @@ class XAttentionBSAPolicy(SparsePolicy):
block_table=block_tables, block_table=block_tables,
) )
if not BSA_AVAILABLE:
# Fallback to flash attention if BSA not available
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,
)
if not XATTN_AVAILABLE:
# Fallback to flash attention if xattn not available
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,
)
from nanovllm.ops.xattn import xattn_estimate from nanovllm.ops.xattn import xattn_estimate
# Set DensityObserver mode on first layer # Set DensityObserver mode on first layer
@@ -477,8 +453,7 @@ class XAttentionBSAPolicy(SparsePolicy):
causal_total = q_bk * (q_bk + 1) // 2 * mask_trimmed.shape[0] * mask_trimmed.shape[1] causal_total = q_bk * (q_bk + 1) // 2 * mask_trimmed.shape[0] * mask_trimmed.shape[1]
causal_mask = torch.tril(torch.ones(q_bk, k_bk, device=mask_trimmed.device, dtype=torch.bool)) causal_mask = torch.tril(torch.ones(q_bk, k_bk, device=mask_trimmed.device, dtype=torch.bool))
selected = (mask_trimmed & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item() selected = (mask_trimmed & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()
logger.info(f"[DEBUG GPU-only Layer0] mask_shape={mask_trimmed.shape}, "
f"density={selected/causal_total:.6f}, selected={selected}, total={causal_total}")
DensityObserver.record(layer_id, mask_trimmed, causal=True) DensityObserver.record(layer_id, mask_trimmed, causal=True)
return output return output
@@ -633,98 +608,108 @@ class XAttentionBSAPolicy(SparsePolicy):
l_chunks = [] l_chunks = []
num_kv_chunks = num_historical_blocks + 1 # +1 for current chunk num_kv_chunks = num_historical_blocks + 1 # +1 for current chunk
# Get compute_stream for all compute kernels (like attention computation)
compute_stream = offload_engine.compute_stream
with nvtx.range("xattn_estimate_pass1"): with nvtx.range("xattn_estimate_pass1"):
slot = 0 slot = 0
# Process historical blocks (from CPU) # Process historical blocks (from CPU)
for kv_chunk_idx, cpu_block_id in enumerate(available_blocks): for kv_chunk_idx, cpu_block_id in enumerate(available_blocks):
# Load K from CPU # Load K from CPU (on slot_transfer_stream)
offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id) offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
# wait_slot_layer makes compute_stream wait for H2D transfer
offload_engine.wait_slot_layer(slot) offload_engine.wait_slot_layer(slot)
k_block = offload_engine.get_k_for_slot(slot) # [1, block_size, num_kv_heads, head_dim] # All compute kernels run on compute_stream (like attention computation)
K_chunk = k_block.transpose(1, 2) # [1, num_kv_heads, block_size, head_dim] with torch.cuda.stream(compute_stream):
k_block = offload_engine.get_k_for_slot(slot) # [1, block_size, num_kv_heads, head_dim]
K_chunk = k_block.transpose(1, 2) # [1, num_kv_heads, block_size, head_dim]
# GQA expansion # GQA expansion
num_kv_heads = K_chunk.shape[1] num_kv_heads = K_chunk.shape[1]
if num_heads != num_kv_heads:
num_groups = num_heads // num_kv_heads
K_chunk = K_chunk.repeat_interleave(num_groups, dim=1)
# KV offset in reshaped space
kv_offset_reshaped = kv_chunk_idx * kv_chunk_reshaped
# Compute raw attention scores
attn_weights_kv = flat_group_gemm_fuse_reshape(
Q, K_chunk, self.stride,
chunk_start=chunk_start,
chunk_end=chunk_end,
is_causal=False, # K 不完整，不能在这里用 causal
)
# Compute partial stats (带 causal mask)
m_partial, l_partial = softmax_compute_partial_stats(
attn_weights_kv,
reshaped_block_size,
segment_size,
scale,
chunk_start=chunk_start,
kv_offset=kv_offset_reshaped,
is_causal=True,
)
m_chunks.append(m_partial)
l_chunks.append(l_partial)
offload_engine.record_slot_compute_done(slot)
del attn_weights_kv
# Process current chunk K (already on GPU) on compute_stream
with torch.cuda.stream(compute_stream):
# k: [seq_len, num_kv_heads, head_dim] -> [1, num_kv_heads, seq_len, head_dim]
K_current = k.unsqueeze(0).transpose(1, 2)
# GQA expansion for current chunk
num_kv_heads = K_current.shape[1]
if num_heads != num_kv_heads: if num_heads != num_kv_heads:
num_groups = num_heads // num_kv_heads num_groups = num_heads // num_kv_heads
K_chunk = K_chunk.repeat_interleave(num_groups, dim=1) K_current = K_current.repeat_interleave(num_groups, dim=1)
# KV offset in reshaped space # Pad current K to alignment
kv_offset_reshaped = kv_chunk_idx * kv_chunk_reshaped curr_k_len = K_current.shape[2]
padded_curr_k_len = ((curr_k_len + alignment - 1) // alignment) * alignment
if padded_curr_k_len != curr_k_len:
K_current = torch.nn.functional.pad(K_current, (0, 0, 0, padded_curr_k_len - curr_k_len), value=0)
# Compute raw attention scores # KV offset for current chunk
attn_weights_kv = flat_group_gemm_fuse_reshape( kv_offset_current = num_historical_blocks * kv_chunk_reshaped
Q, K_chunk, self.stride,
# Compute attention scores for current chunk
attn_weights_curr = flat_group_gemm_fuse_reshape(
Q, K_current, self.stride,
chunk_start=chunk_start, chunk_start=chunk_start,
chunk_end=chunk_end, chunk_end=chunk_end,
is_causal=False, # K 不完整，不能在这里用 causal is_causal=False,
) )
# Compute partial stats (带 causal mask) # Compute partial stats for current chunk
m_partial, l_partial = softmax_compute_partial_stats( m_partial_curr, l_partial_curr = softmax_compute_partial_stats(
attn_weights_kv, attn_weights_curr,
reshaped_block_size, reshaped_block_size,
segment_size, segment_size,
scale, scale,
chunk_start=chunk_start, chunk_start=chunk_start,
kv_offset=kv_offset_reshaped, kv_offset=kv_offset_current,
is_causal=True, is_causal=True,
) )
m_chunks.append(m_partial) m_chunks.append(m_partial_curr)
l_chunks.append(l_partial) l_chunks.append(l_partial_curr)
offload_engine.record_slot_compute_done(slot) del attn_weights_curr
del attn_weights_kv
# Process current chunk K (already on GPU)
# k: [seq_len, num_kv_heads, head_dim] -> [1, num_kv_heads, seq_len, head_dim]
K_current = k.unsqueeze(0).transpose(1, 2)
# GQA expansion for current chunk
num_kv_heads = K_current.shape[1]
if num_heads != num_kv_heads:
num_groups = num_heads // num_kv_heads
K_current = K_current.repeat_interleave(num_groups, dim=1)
# Pad current K to alignment
curr_k_len = K_current.shape[2]
padded_curr_k_len = ((curr_k_len + alignment - 1) // alignment) * alignment
if padded_curr_k_len != curr_k_len:
K_current = torch.nn.functional.pad(K_current, (0, 0, 0, padded_curr_k_len - curr_k_len), value=0)
# KV offset for current chunk
kv_offset_current = num_historical_blocks * kv_chunk_reshaped
# Compute attention scores for current chunk
attn_weights_curr = flat_group_gemm_fuse_reshape(
Q, K_current, self.stride,
chunk_start=chunk_start,
chunk_end=chunk_end,
is_causal=False,
)
# Compute partial stats for current chunk
m_partial_curr, l_partial_curr = softmax_compute_partial_stats(
attn_weights_curr,
reshaped_block_size,
segment_size,
scale,
chunk_start=chunk_start,
kv_offset=kv_offset_current,
is_causal=True,
)
m_chunks.append(m_partial_curr)
l_chunks.append(l_partial_curr)
del attn_weights_curr
# ================================================================ # ================================================================
# Step 2: Merge all partial stats # Step 2: Merge all partial stats (on compute_stream)
# ================================================================ # ================================================================
with nvtx.range("xattn_estimate_merge"): with torch.cuda.stream(compute_stream):
m_global, l_global = merge_softmax_stats(m_chunks, l_chunks) with nvtx.range("xattn_estimate_merge"):
del m_chunks, l_chunks m_global, l_global = merge_softmax_stats(m_chunks, l_chunks)
del m_chunks, l_chunks
# ================================================================ # ================================================================
# Step 3: Second pass - normalize and compute block sums # Step 3: Second pass - normalize and compute block sums
@@ -736,30 +721,61 @@ class XAttentionBSAPolicy(SparsePolicy):
# Process historical blocks again # Process historical blocks again
for kv_chunk_idx, cpu_block_id in enumerate(available_blocks): for kv_chunk_idx, cpu_block_id in enumerate(available_blocks):
# Load K from CPU (on slot_transfer_stream)
offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id) offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
# wait_slot_layer makes compute_stream wait for H2D transfer
offload_engine.wait_slot_layer(slot) offload_engine.wait_slot_layer(slot)
k_block = offload_engine.get_k_for_slot(slot) # All compute kernels run on compute_stream
K_chunk = k_block.transpose(1, 2) with torch.cuda.stream(compute_stream):
k_block = offload_engine.get_k_for_slot(slot)
K_chunk = k_block.transpose(1, 2)
num_kv_heads = K_chunk.shape[1] num_kv_heads = K_chunk.shape[1]
if num_heads != num_kv_heads: if num_heads != num_kv_heads:
num_groups = num_heads // num_kv_heads num_groups = num_heads // num_kv_heads
K_chunk = K_chunk.repeat_interleave(num_groups, dim=1) K_chunk = K_chunk.repeat_interleave(num_groups, dim=1)
kv_offset_reshaped = kv_chunk_idx * kv_chunk_reshaped kv_offset_reshaped = kv_chunk_idx * kv_chunk_reshaped
# Recompute attention scores (trade-off: compute vs memory) # Recompute attention scores (trade-off: compute vs memory)
attn_weights_kv = flat_group_gemm_fuse_reshape( attn_weights_kv = flat_group_gemm_fuse_reshape(
Q, K_chunk, self.stride, Q, K_chunk, self.stride,
chunk_start=chunk_start,
chunk_end=chunk_end,
is_causal=False,
)
# Normalize with global stats and compute block sums
block_sum_kv = softmax_normalize_and_block_sum(
attn_weights_kv,
m_global,
l_global,
reshaped_block_size,
segment_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(block_sum_kv)
offload_engine.record_slot_compute_done(slot)
del attn_weights_kv
# Process current chunk on compute_stream
with torch.cuda.stream(compute_stream):
# Recompute attention scores for current chunk
attn_weights_curr = flat_group_gemm_fuse_reshape(
Q, K_current, self.stride,
chunk_start=chunk_start, chunk_start=chunk_start,
chunk_end=chunk_end, chunk_end=chunk_end,
is_causal=False, is_causal=False,
) )
# Normalize with global stats and compute block sums block_sum_curr = softmax_normalize_and_block_sum(
block_sum_kv = softmax_normalize_and_block_sum( attn_weights_curr,
attn_weights_kv,
m_global, m_global,
l_global, l_global,
reshaped_block_size, reshaped_block_size,
@@ -767,67 +783,42 @@ class XAttentionBSAPolicy(SparsePolicy):
chunk_start=chunk_start, chunk_start=chunk_start,
real_q_len=k_reshaped_seq_len - k_reshaped_num_to_pad, real_q_len=k_reshaped_seq_len - k_reshaped_num_to_pad,
scale=scale, scale=scale,
kv_offset=kv_offset_reshaped, kv_offset=kv_offset_current,
is_causal=True, is_causal=True,
) )
attn_sum_per_kv.append(block_sum_kv) attn_sum_per_kv.append(block_sum_curr)
del attn_weights_curr, K_current
offload_engine.record_slot_compute_done(slot)
del attn_weights_kv
# Process current chunk
# Recompute attention scores for current chunk
attn_weights_curr = flat_group_gemm_fuse_reshape(
Q, K_current, self.stride,
chunk_start=chunk_start,
chunk_end=chunk_end,
is_causal=False,
)
block_sum_curr = softmax_normalize_and_block_sum(
attn_weights_curr,
m_global,
l_global,
reshaped_block_size,
segment_size,
chunk_start=chunk_start,
real_q_len=k_reshaped_seq_len - k_reshaped_num_to_pad,
scale=scale,
kv_offset=kv_offset_current,
is_causal=True,
)
attn_sum_per_kv.append(block_sum_curr)
del attn_weights_curr, K_current
# ================================================================ # ================================================================
# Step 4: Concatenate block sums and select blocks # Step 4: Concatenate block sums and select blocks (on compute_stream)
# ================================================================ # ================================================================
attn_sum_concat = torch.cat(attn_sum_per_kv, dim=-1) with torch.cuda.stream(compute_stream):
del attn_sum_per_kv, m_global, l_global attn_sum_concat = torch.cat(attn_sum_per_kv, dim=-1)
del attn_sum_per_kv, m_global, l_global
# Calculate q_block offset for find_blocks_chunked # Calculate q_block offset for find_blocks_chunked
# This is the number of BSA blocks before Q in the full sequence # This is the number of BSA blocks before Q in the full sequence
num_blocks_per_chunk = q_reshaped_len // reshaped_block_size num_blocks_per_chunk = q_reshaped_len // reshaped_block_size
current_index = k_block_num - q_block_num # Q starts at this BSA block index current_index = k_block_num - q_block_num # Q starts at this BSA block index
with nvtx.range("xattn_find_blocks"): with nvtx.range("xattn_find_blocks"):
mask = find_blocks_chunked( mask = find_blocks_chunked(
attn_sum_concat, attn_sum_concat,
current_index=current_index, current_index=current_index,
threshold=self.threshold, threshold=self.threshold,
num_to_choose=None, num_to_choose=None,
decoding=False, decoding=False,
mode="prefill", mode="prefill",
causal=True, causal=True,
)
# Apply causal mask post-processing (same as xattn.py lines 1300-1306)
mask[:, :, -q_block_num:, -q_block_num:] = torch.where(
torch.tril(torch.ones(q_block_num, q_block_num, dtype=torch.bool, device=mask.device), diagonal=0),
mask[:, :, -q_block_num:, -q_block_num:],
False,
) )
# Apply causal mask post-processing (same as xattn.py lines 1300-1306)
mask[:, :, -q_block_num:, -q_block_num:] = torch.where(
torch.tril(torch.ones(q_block_num, q_block_num, dtype=torch.bool, device=mask.device), diagonal=0),
mask[:, :, -q_block_num:, -q_block_num:],
False,
)
# ================================================================ # ================================================================
# Step 5: Record density (only on layer 0) # Step 5: Record density (only on layer 0)
# ================================================================ # ================================================================
@@ -908,14 +899,6 @@ class XAttentionBSAPolicy(SparsePolicy):
if available_blocks and available_blocks[-1] not in selected_block_ids: if available_blocks and available_blocks[-1] not in selected_block_ids:
selected_block_ids.append(available_blocks[-1]) selected_block_ids.append(available_blocks[-1])
# Record communication density
if available_blocks:
DensityObserver.record_comm_density(
layer_id,
selected_cpu_blocks=len(selected_block_ids),
total_cpu_blocks=len(available_blocks),
)
# Update statistics (only for layer 0 to avoid overcounting) # Update statistics (only for layer 0 to avoid overcounting)
if layer_id == 0 and available_blocks: if layer_id == 0 and available_blocks:
self._stats_total_available_blocks += len(available_blocks) self._stats_total_available_blocks += len(available_blocks)