A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 01:30 UTC
[Model] Remove the unnecessary dtype conversion in MiniCPM (#32523)
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Canlin Guo committed
Jan 18, 2026 at 16:07 UTC
fe36bf5e808d96ab18b8db393380dab3efb6c3f6
1 file changed
-3
vllm/model_executor/models/minicpm.py
-3
@@ -300,10 +300,7 @@ class MiniCPMAttention(nn.Module):
300
) -> torch.Tensor:
301
qkv, _ = self.qkv_proj(hidden_states)
302
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
303
- orig_dtype = q.dtype
304
- q, k = q.float(), k.float()
303
q, k = self.rotary_emb(positions, q, k)
306
- q, k = q.to(orig_dtype), k.to(orig_dtype)
304
attn_output = self.attn(q, k, v)
305
output, _ = self.o_proj(attn_output)
306
return output