chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Deepseek V3.2] Enable flashmla_auto with MTP (#12294)

Browse Source 
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
This commit is contained in:
hlu1
2025-10-28 23:51:20 -07:00
committed by GitHub
parent 83a7c89c3f
commit 42e1a72efb
1 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/attention/nsa_backend.py
View File

@@ -1222,11 +1222,9 @@ class NativeSparseAttnBackend(AttentionBackend):
if self.enable_auto_select_prefill_impl:
if self.nsa_kv_cache_store_fp8:
if (
# TODO(hlu1): enable MTP
is_blackwell()
and forward_batch is not None
and forward_batch.forward_mode.is_extend()
and forward_batch.spec_algorithm.is_none()
and forward_batch.forward_mode == ForwardMode.EXTEND
):
total_kv_tokens = forward_batch.seq_lens_sum
total_q_tokens = forward_batch.extend_num_tokens