xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-16 09:06:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Attention] Optimize make_local_attention_virtual_batches for Flash Attention (#23185)

Browse Source 
Signed-off-by: linzebing <linzebing1995@gmail.com>
This commit is contained in:
Zebing Lin 2025-08-19 22:57:47 -04:00 committed by GitHub
parent 64ab3c7253
commit a634733f67
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/v1/attention/backends/utils.py
View File

@ -464,8 +464,9 @@ def make_local_attention_virtual_batches(
attn_chunk_size)[arange > 0]
# convert from q_seqlens to cu_seqlens_q
cu_seqlens_q_local = np.pad(np.cumsum(seqlens_q_local), (1, 0))\
.astype(np.int32)
cu_seqlens_q_local = np.empty(virtual_batches + 1, dtype=np.int32)
np.cumsum(seqlens_q_local, out=cu_seqlens_q_local[1:])
cu_seqlens_q_local[0] = 0
# compute the seqlens_k_local,
# basically a full local attention block for all but the last block in each
@ -508,11 +509,10 @@ def make_local_attention_virtual_batches(
# [ 22, 23 ], < local-batch 6, (batch 2, starting from k[4])
# [ 24, 25 ], < local-batch 7, (batch 2, starting from k[8])
# ]
block_indices= np.broadcast_to(
np.arange(pages_per_local_batch, dtype=np.int32),
(virtual_batches, pages_per_local_batch)) \
+ np.expand_dims(block_starts, axis=1)
block_indices = block_indices.flatten().clip(max=block_table.shape[1] - 1)
block_indices = (block_starts[:, None] +
np.arange(pages_per_local_batch, dtype=np.int32))
block_indices = block_indices.reshape(-1).clip(max=block_table.shape[1] -
1)
batch_indices = np.repeat(np.arange(actual_batch_size, dtype=np.int32),
local_blocks * pages_per_local_batch)
block_table_local = block_table[batch_indices, block_indices]\