xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-18 07:05:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix][Spec Decode] Fix wrong valid_mask for padded speculation when chunked prefill occurs (#26231)

Browse Source 
Signed-off-by: seven-mile <i@7li.moe>
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>
This commit is contained in:
7mile 2025-10-07 02:24:22 +08:00 committed by GitHub
parent 824a3f403f
commit b2ea5ba677
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/v1/spec_decode/eagle.py
View File

@ -522,13 +522,9 @@ class EagleProposer:
) )
# Generate a mask for all valid tokens within those requests # Generate a mask for all valid tokens within those requests
max_gen_len = sampled_token_ids.shape[-1] valid_mask = (valid_sampled_token_ids_gpu != -1) & (
if max_gen_len == 1: valid_sampled_token_ids_gpu < gpu_input_batch.vocab_size
valid_mask = torch.ones_like(valid_sampled_token_ids_gpu, dtype=torch.bool) )
else:
valid_mask = (valid_sampled_token_ids_gpu != -1) & (
valid_sampled_token_ids_gpu < gpu_input_batch.vocab_size
)
# Count the number of valid tokens in each request # Count the number of valid tokens in each request
valid_sampled_tokens_count = valid_mask.sum(dim=1) valid_sampled_tokens_count = valid_mask.sum(dim=1)