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

[Bugfix] Fix potential EAGLE spec decode segfault during graph capture (#32818)

Browse Source 
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
Matt
2026-01-21 21:11:55 -06:00
committed by GitHub
parent 6437ff1fb9
commit c5487e2b96
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/v1/spec_decode/eagle.py
View File

@@ -1222,10 +1222,14 @@ class SpecDecodeBaseProposer:
num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp(
num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens
)
cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch(
num_tokens_dp_padded
)
num_input_tokens = batch_desc.num_tokens
if use_cudagraphs:
cudagraph_runtime_mode, batch_desc = (
self.cudagraph_dispatcher.dispatch(num_tokens_dp_padded)
)
num_input_tokens = batch_desc.num_tokens
else:
cudagraph_runtime_mode = CUDAGraphMode.NONE
num_input_tokens = num_tokens_dp_padded
if num_tokens_across_dp is not None:
num_tokens_across_dp[self.dp_rank] = num_input_tokens