From fe36bf5e808d96ab18b8db393380dab3efb6c3f6 Mon Sep 17 00:00:00 2001 From: Canlin Guo Date: Sun, 18 Jan 2026 16:07:28 +0800 Subject: [PATCH] [Model] Remove the unnecessary dtype conversion in MiniCPM (#32523) Signed-off-by: gcanlin --- vllm/model_executor/models/minicpm.py | 3 --- 1 file changed, 3 deletions(-) diff --git a/vllm/model_executor/models/minicpm.py b/vllm/model_executor/models/minicpm.py index a05be794a..8f7b8b21b 100644 --- a/vllm/model_executor/models/minicpm.py +++ b/vllm/model_executor/models/minicpm.py @@ -300,10 +300,7 @@ class MiniCPMAttention(nn.Module): ) -> torch.Tensor: qkv, _ = self.qkv_proj(hidden_states) q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) - orig_dtype = q.dtype - q, k = q.float(), k.float() q, k = self.rotary_emb(positions, q, k) - q, k = q.to(orig_dtype), k.to(orig_dtype) attn_output = self.attn(q, k, v) output, _ = self.o_proj(attn_output) return output