EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Bugfix] Fix index out of bounds in local attention with large sequences (#5173)

Browse Source
This commit is contained in:
Chang Su
2025-04-08 18:43:13 -07:00
committed by GitHub
parent 39efad4fbc
commit aac531c53b
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/attention/flashattention_backend.py
View File

@@ -236,7 +236,11 @@ def make_local_attention_virtual_batches(
np.arange(pages_per_local_batch, dtype=np.int32), np.arange(pages_per_local_batch, dtype=np.int32),
(virtual_batches, pages_per_local_batch), (virtual_batches, pages_per_local_batch),
) + np.expand_dims(block_starts, axis=1) ) + np.expand_dims(block_starts, axis=1)
block_indices = block_indices.flatten() # Ensure block_indices doesn't exceed block_table dimensions
# This is a critical safety check that prevents index out of bounds errors
# when dealing with large sequences (>8192 tokens) or when the block_table
# dimensions are smaller than what would be needed for the full attention chunk size.
block_indices = block_indices.flatten().clip(max=block_table.shape[1] - 1)
batch_indices = np.repeat( batch_indices = np.repeat(
np.arange(actual_batch_size, dtype=np.int32), np.arange(actual_batch_size, dtype=np.int32),
local_blocks * pages_per_local_batch, local_blocks * pages_per_local_batch,