xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 19:05:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix block size in block_table with PCP (#29094)

Browse Source 
Signed-off-by: Livinfly <luojie3m@gmail.com>
This commit is contained in:
Jie Luo 2025-11-22 09:34:28 +08:00 committed by GitHub
parent ed8e6843cc
commit 5c8f2adf50
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/v1/worker/block_table.py
View File

@ -84,7 +84,7 @@ class BlockTable:
self.pcp_world_size = get_pcp_group().world_size self.pcp_world_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group().rank_in_group self.pcp_rank = get_pcp_group().rank_in_group
except AssertionError: except AssertionError:
# DCP might not be initialized in testing # PCP might not be initialized in testing
self.pcp_world_size = 1 self.pcp_world_size = 1
self.pcp_rank = 0 self.pcp_rank = 0
try: try:
@ -268,6 +268,11 @@ class MultiGroupBlockTable:
# (max_model_len//dcp_world_size) tokens in kvcache, # (max_model_len//dcp_world_size) tokens in kvcache,
# so the block_size which used for calc max_num_blocks_per_req # so the block_size which used for calc max_num_blocks_per_req
# must be multiplied by dcp_world_size. # must be multiplied by dcp_world_size.
try:
pcp_world_size = get_pcp_group().world_size
except AssertionError:
# PCP might not be initialized in testing
pcp_world_size = 1
try: try:
dcp_world_size = get_dcp_group().world_size dcp_world_size = get_dcp_group().world_size
except AssertionError: except AssertionError:
@ -280,12 +285,14 @@ class MultiGroupBlockTable:
f"must match block_sizes length ({len(block_sizes)})" f"must match block_sizes length ({len(block_sizes)})"
) )
total_cp_world_size = dcp_world_size * pcp_world_size
self.block_tables = [ self.block_tables = [
BlockTable( BlockTable(
block_size, block_size,
max_num_reqs, max_num_reqs,
max( max(
cdiv(max_model_len, block_size * dcp_world_size), cdiv(max_model_len, block_size * total_cp_world_size),
1 + num_speculative_tokens, 1 + num_speculative_tokens,
), ),
max_num_batched_tokens, max_num_batched_tokens,