A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 12:30 UTC
[main][BugFix] Fixed an accuracy bug of Qwen3-next-MTP when batched inferring (#30632)
Signed-off-by: drslark <slarksblood@qq.com>
drslark committed
Dec 14, 2025 at 17:32 UTC
add1b9d3dec4a6d1b404f5793a210ff77482b7ae
1 file changed
+1
-1
vllm/v1/attention/backends/gdn_attn.py
+1
-1
@@ -211,7 +211,7 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
211
spec_token_masks = torch.repeat_interleave(
212
spec_sequence_masks, query_lens
213
)
214
- index = torch.argsort(spec_token_masks)
214
+ index = torch.argsort(spec_token_masks, stable=True)
215
num_non_spec_tokens = num_prefill_tokens + num_decode_tokens
216
non_spec_token_indx = index[:num_non_spec_tokens]
217
spec_token_indx = index[num_non_spec_tokens:]