add qwen3_moe

2026-02-10 18:18:32 +08:00
parent cf92e95688
commit 2a9f483af8
1 changed files with 5 additions and 1 deletions
--- a/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
+++ b/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
@@ -258,7 +258,11 @@ class Qwen3MoeAttention(nn.Module):
        k_by_head = self.k_norm(k_by_head)
        k = k_by_head.reshape(k_shape)

-        q, k = self.rotary_emb(positions, q, k)
+        # MLU's forward_mlu signature is (positions, x, offsets=None),
+        # so we must call separately for q and k to avoid k being
+        # treated as offsets.
+        q = self.rotary_emb(positions, q)
+        k = self.rotary_emb(positions, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.o_proj(attn_output)
        return output