[Bugfix] Fix Qwen3/DSV3/DSV3.2 model support (#11510)

2025-10-16 15:14:09 +08:00
parent b0d20cdec7
commit 3cceaa381a
12 changed files with 102 additions and 33 deletions
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1357,6 +1357,7 @@ class DeepseekV2AttentionMLA(nn.Module):
                inner_state = self.mla_preprocess.forward(
                    positions, hidden_states, forward_batch, zero_allocator
                )
+                inner_state = (*inner_state, None)  # add a position for topk_indices
        elif attn_forward_method == AttnForwardMethod.NPU_MLA_SPARSE:
            inner_state = self.forward_npu_sparse_prepare(
                positions, hidden_states, forward_batch, zero_allocator