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

Fix BatchMLAPagedAttentionWrapper query/qo_inptr mismatch for EAGLE (#15601)

Browse Source
This commit is contained in:
Shu Wang
2026-02-27 11:35:45 -08:00
committed by GitHub
parent 6a1480ce45
commit 1b75d0d1a9
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/trtllm_mla_backend.py
+7
View File
@@ -616,6 +616,13 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
): ):
bs = forward_batch.batch_size bs = forward_batch.batch_size
self.forward_decode_metadata = TRTLLMMLADecodeMetadata() self.forward_decode_metadata = TRTLLMMLADecodeMetadata()
# This is necessary because the backend instance persists across forward passes,
# and forward_prefill_metadata from a previous regular extend call could still be set.
if (
forward_batch.forward_mode.is_target_verify()
or forward_batch.forward_mode.is_draft_extend(include_v2=True)
):
self.forward_prefill_metadata = None
# Get maximum sequence length. # Get maximum sequence length.
if getattr(forward_batch, "seq_lens_cpu", None) is not None: if getattr(forward_batch, "seq_lens_cpu", None) is not None:
max_seq = forward_batch.seq_lens_cpu.max().item() max_seq = forward_batch.seq_lens_cpu.max().item()