minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: TRT-LLM MHA CUDA illegal address with EAGLE v2 + DP attention (#21649)

Browse Source 
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
Kangyan-Zhou
2026-04-05 09:41:14 -07:00
committed by GitHub
co-authored by Claude Opus 4.6 Baizhou Zhang
parent c5fa364b80
commit 5dd2c243eb
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/trtllm_mha_backend.py
+1 -1
View File
@@ -871,7 +871,7 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
max_kv_len=self.max_context_len, max_kv_len=self.max_context_len,
bmm1_scale=bmm1_scale, bmm1_scale=bmm1_scale,
bmm2_scale=bmm2_scale, bmm2_scale=bmm2_scale,
batch_size=forward_batch.batch_size, batch_size=self.forward_metadata.cu_seqlens_q.shape[0] - 1,
cum_seq_lens_q=self.forward_metadata.cu_seqlens_q, cum_seq_lens_q=self.forward_metadata.cu_seqlens_q,
cum_seq_lens_kv=self.forward_metadata.cu_seqlens_k, cum_seq_lens_kv=self.forward_metadata.cu_seqlens_k,
window_left=layer.sliding_window_size, window_left=layer.sliding_window_size,