[Core] Refactor model loading code (#4097)

vllm/model_executor/models/qwen2_moe.py

@@ -22,7 +22,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2MoE model compatible with HuggingFace weights."""
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 import torch.nn.functional as F
@@ -46,9 +46,8 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.model_executor.weight_utils import (default_weight_loader,
-                                              hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
 
 
@@ -366,6 +365,8 @@ class Qwen2MoeModel(nn.Module):
 
 class Qwen2MoeForCausalLM(nn.Module):
 
+    fall_back_to_pt_during_load = False
+
     def __init__(
         self,
         config: PretrainedConfig,
@@ -404,11 +405,7 @@ class Qwen2MoeForCausalLM(nn.Module):
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    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"),
@@ -419,12 +416,7 @@ class Qwen2MoeForCausalLM(nn.Module):
         ]
 
         params_dict = dict(self.named_parameters())
-        for name, loaded_weight in hf_model_weights_iterator(
-                model_name_or_path,
-                cache_dir,
-                load_format,
-                revision,
-                fall_back_to_pt=False):
+        for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping: