fix

Author: shihaobai

lightllm/models/qwen3_vl/layer_infer/pre_layer_infer.py

@@ -46,8 +46,9 @@ def context_forward(self, input_ids, infer_state: Qwen3VLInferStateInfo, layer_w
                 # image_embed(token_num, hidden_dim) + deepstack(token_num*layer_num, hidden_dim)
                 all_img_embed_df = (
                     bytes2tensor(read_shm(get_shm_name_embed(img["uuid"])))
-                    .cuda(non_blocking=True)
+                    .view(dtype)
                     .view(-1, hidden_size)
+                    .cuda(non_blocking=True)
                 )
                 per_image_deepstack = []
 

lightllm/server/visualserver/model_infer/model_rpc.py

@@ -121,8 +121,6 @@ def exposed_encode(self, images: List[ImageItem]):
                     continue
                 uid = uuids[i]
                 start, end = valid_ids[i]
-                # cur_embed_bytes = tensor2bytes(all_img_embeds[start:end])
-                # create_shm(get_shm_name_embed(uid), cur_embed_bytes)
                 create_shm_and_dump(get_shm_name_embed(uid), all_img_embeds[start:end])
                 ids_to_set.append(uid)
             if ids_to_set: