xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-09-06 10:07:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(spec_decode): sync ngram draft tokens across TP ranks

Browse Source 
When using tensor parallelism with external_launcher, ngram draft
tokens could diverge across TP ranks due to non-determinism in
numba parallel execution. This caused verification failures and
crashes in speculative decoding.

The fix ensures that only TP rank 0 computes draft tokens and
broadcasts them to all other ranks using broadcast_object(),
guaranteeing identical draft tokens across all TP ranks.

Fixes #31154

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Signed-off-by: yurekami <yurekami@users.noreply.github.com>
This commit is contained in:
yurekami 2025-12-24 23:09:18 +09:00
parent 7cd288a4b3
commit 0f94a71de5
1 changed files with 38 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/v1/spec_decode/ngram_proposer.py
View File

@ -6,6 +6,7 @@ import numpy as np
from numba import get_num_threads, jit, njit, prange, set_num_threads from numba import get_num_threads, jit, njit, prange, set_num_threads
from vllm.config import VllmConfig from vllm.config import VllmConfig
from vllm.distributed import get_tp_group
class NgramProposer: class NgramProposer:
@ -137,6 +138,12 @@ class NgramProposer:
token_ids_cpu: np.ndarray, token_ids_cpu: np.ndarray,
spec_decode_unsupported_reqs: set, spec_decode_unsupported_reqs: set,
) -> list[list[int]]: ) -> list[list[int]]:
# Only compute draft tokens on TP rank 0 and broadcast to other ranks.
# This ensures all TP ranks have identical draft tokens, which is
# required because numba parallel execution can produce different
# results across ranks due to non-determinism.
tp_group = get_tp_group()
if tp_group.is_first_rank:
# find which requests need ngram proposals # find which requests need ngram proposals
valid_ngram_requests = [] valid_ngram_requests = []
for i, sampled_ids in enumerate(sampled_token_ids): for i, sampled_ids in enumerate(sampled_token_ids):
@ -164,6 +171,14 @@ class NgramProposer:
num_tokens_no_spec, num_tokens_no_spec,
token_ids_cpu, token_ids_cpu,
) )
else:
draft_token_ids = None
# Broadcast draft tokens from rank 0 to all other ranks.
# Rank 0 always computes valid draft_token_ids, so broadcast
# will never return None.
draft_token_ids = tp_group.broadcast_object(draft_token_ids, src=0)
assert draft_token_ids is not None
return draft_token_ids return draft_token_ids