[Bugfix][MRV2] Fix speculative warmup deadlock with pipeline parallelism (#62129)

Signed-off-by: chenjr97 <chenjr97@mail2.sysu.edu.cn>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
This commit is contained in:
Junru Chen
2026-10-17 23:51:32 -07:00
committed by GitHub
co-authored by Nick Hill vllm-agent
parent da9134bc89
commit 7c9042c05a
3 changed files with 28 additions and 1 deletions
+2 -1
View File
@@ -238,7 +238,8 @@ def warmup_kernels(
rejection_sampler.enable_adaptive_verification = False
try:
_warmup_kernels(model_runner, worker_execute_model, worker_sample_tokens)
if rejection_sampler is not None:
# All PP ranks must run the same forwards to match sends and receives.
if model_runner.speculative_config is not None:
# Greedy batches skip logits processing, so rejection kernels
# also see model-dtype logits instead of FP32.
_warmup_kernels(