[Misc] Standardize RoPE handling for Qwen2-VL (#9250)

2024-10-16 13:56:17 +08:00
parent ed920135c8
commit 7e7eae338d
16 changed files with 102 additions and 200 deletions
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -242,7 +242,7 @@ class DeepseekV2Attention(nn.Module):
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")
-        rope_scaling['type'] = 'deepseek_yarn'
+        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,