biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix[DP][v1]: Prevent hangs from mismatched worker configurations (#26218)

Browse Source 
Signed-off-by: Ayush Satyam <ayushsatyam146@gmail.com>
This commit is contained in:
Ayush Satyam
2025-10-08 11:25:08 +05:30
committed by GitHub
parent 0d4f48fa10
commit 5e65d6b2ad
3 changed files with 46 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/config/parallel.py
View File

@@ -336,6 +336,9 @@ class ParallelConfig:
graph from input ids/embeddings to the final hidden states,
excluding anything before input ids/embeddings and after
the final hidden states.
This hash is also used for DP worker configuration validation
to prevent hangs from mismatched collective communication patterns.
"""
factors: list[Any] = []
factors.append(self.pipeline_parallel_size)
@@ -343,6 +346,12 @@ class ParallelConfig:
factors.append(self.enable_expert_parallel)
factors.append(self.data_parallel_size)
factors.append(envs.VLLM_ALL2ALL_BACKEND)
factors.append(self.enable_eplb)
if self.enable_eplb:
factors.append(self.eplb_config.log_balancedness)
factors.append(self.eplb_config.window_size)
factors.append(self.eplb_config.step_interval)
factors.append(self.eplb_config.num_redundant_experts)
return hashlib.sha256(str(factors).encode()).hexdigest()
def __post_init__(self) -> None: