[VLM] Add MLA with pure RoPE support for deepseek-vl2 models (#12729)

2025-02-05 12:44:26 +08:00
parent 249824c3bf
commit 98fd089fc9
3 changed files with 30 additions and 6 deletions
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -414,7 +414,8 @@ class DeepseekV2MLAAttention(nn.Module):
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

-        rope_scaling["rope_type"] = 'deepseek_yarn'
+        if rope_scaling:
+            rope_scaling["rope_type"] = 'deepseek_yarn'
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,