[Model] use AutoWeightsLoader for bloom

Signed-off-by: calvin chen <120380290@qq.com>

Author: calvin0327

vllm/model_executor/models/bloom.py

@@ -43,7 +43,7 @@
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP, SupportsQuant, SupportsV0Only
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -278,6 +278,41 @@ def forward(
         hidden_states = self.ln_f(hidden_states)
         return hidden_states
 
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            if name == "lm_head.weight":
+                continue
+            if not name.startswith("transformer."):
+                name = "transformer." + name
+            if is_pp_missing_parameter(name, self):
+                continue
+            param = params_dict[name]
+
+            if "query_key_value" in name:
+                # NOTE: BLOOM's fused QKV's output_dim has the shape of
+                # (num_heads * 3 * head_size), while the
+                # required shape is (3 * num_heads * head_size).
+                # Thus, we need weight conversion.
+                output_dim = getattr(param, "output_dim", None)
+                num_heads = self.config.num_attention_heads
+                if output_dim is not None:
+                    loaded_weight_shape = loaded_weight.shape
+                    loaded_weight = loaded_weight.view(
+                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
+                        loaded_weight_shape[output_dim + 1:])
+                    loaded_weight = loaded_weight.transpose(
+                        output_dim, output_dim + 1)
+                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)
+
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class BloomForCausalLM(nn.Module, SupportsPP, SupportsV0Only, SupportsQuant):
 
@@ -325,35 +360,5 @@ def compute_logits(
 
     def load_weights(self, weights: Iterable[tuple[str,
                                                    torch.Tensor]]) -> set[str]:
-        params_dict = dict(self.named_parameters(remove_duplicate=False))
-        loaded_params: set[str] = set()
-        for name, loaded_weight in weights:
-            if name == "lm_head.weight":
-                continue
-            if not name.startswith("transformer."):
-                name = "transformer." + name
-            if is_pp_missing_parameter(name, self):
-                continue
-            param = params_dict[name]
-
-            if "query_key_value" in name:
-                # NOTE: BLOOM's fused QKV's output_dim has the shape of
-                # (num_heads * 3 * head_size), while the
-                # required shape is (3 * num_heads * head_size).
-                # Thus, we need weight conversion.
-                output_dim = getattr(param, "output_dim", None)
-                num_heads = self.config.num_attention_heads
-                if output_dim is not None:
-                    loaded_weight_shape = loaded_weight.shape
-                    loaded_weight = loaded_weight.view(
-                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
-                        loaded_weight_shape[output_dim + 1:])
-                    loaded_weight = loaded_weight.transpose(
-                        output_dim, output_dim + 1)
-                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)
-
-            weight_loader = getattr(param, "weight_loader",
-                                    default_weight_loader)
-            weight_loader(param, loaded_weight)
-            loaded_params.add(name)
-        return loaded_params
+        loader = AutoWeightsLoader(self, skip_prefixes=["lm_head.weight"])
+        return loader.load_weights(weights)