[Bugfix] Fix k_proj's bias for whisper self attention (vllm-project#12342)

Signed-off-by: Isotr0py <[email protected]>

Author: Isotr0py

vllm/model_executor/models/whisper.py

@@ -729,7 +729,22 @@ def sample(
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         loader = AutoWeightsLoader(self, skip_prefixes=["proj_out."])
-        loaded_weights = [(name, loaded_weight)
-                          for name, loaded_weight in weights]
         mapper = WeightsMapper({".fc1.": ".mlp.fc1.", ".fc2.": ".mlp.fc2."})
-        return loader.load_weights(loaded_weights, mapper=mapper)
+        # add fake zeros bias for k_proj to state_dict
+        weights = _create_fake_bias_for_k_proj(weights)
+        return loader.load_weights(weights, mapper=mapper)
+
+
+def _create_fake_bias_for_k_proj(
+    weights: Iterable[Tuple[str, torch.Tensor]]
+) -> Iterable[Tuple[str, torch.Tensor]]:
+    """
+    Create full zeros bias for k_proj weight in self-attention layers.
+    So that the bias for k_proj in qkv_proj can be initialized with zeros.
+    """
+    for name, weight in weights:
+        if ".self_attn.k_proj.weight" in name:
+            bias = torch.zeros(weight.size(0))
+            bias_name = name.replace("weight", "bias")
+            yield from [(name, weight), (bias_name, bias)]
+        yield name, weight