EngineX-Cambricon/enginex-mlu370-vllm
10
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

add qwen3_moe

Browse Source
This commit is contained in:
Chranos
2026-02-10 18:18:32 +08:00
parent a26729bf7f
commit efbb06147a
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
View File

@@ -258,7 +258,11 @@ class Qwen3MoeAttention(nn.Module):
k_by_head = self.k_norm(k_by_head) k_by_head = self.k_norm(k_by_head)
k = k_by_head.reshape(k_shape) 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) attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
output, _ = self.o_proj(attn_output) output, _ = self.o_proj(attn_output)
return output return output