Use model loader from vllm (#459)

python/sglang/srt/models/qwen2.py

@@ -1,7 +1,7 @@
 # Adapted from llama2.py
 # Modify details for the adaptation of Qwen2 model.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple, Iterable
 
 import torch
 from torch import nn
@@ -19,11 +19,11 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.managers.router.model_runner import InputMetadata
-from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator
 
 Qwen2Config = None
 
@@ -271,13 +271,7 @@ class Qwen2ForCausalLM(nn.Module):
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
 
-    def load_weights(
-        self,
-        model_name_or_path: str,
-        cache_dir: Optional[str] = None,
-        load_format: str = "auto",
-        revision: Optional[str] = None,
-    ):
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -287,9 +281,7 @@ class Qwen2ForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
         params_dict = dict(self.named_parameters())
-        for name, loaded_weight in hf_model_weights_iterator(
-            model_name_or_path, cache_dir, load_format, revision
-        ):
+        for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:
                 continue
             if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: