Correct model layer aliasing for Bert style models (#43896)

vllm/model_executor/models/modernbert.py

@@ -237,7 +237,11 @@ class ModernBertEncoderLayer(nn.Module):
 @default_pooling_type(seq_pooling_type="CLS")
 class ModernBertModel(nn.Module):
     hf_to_vllm_mapper = WeightsMapper(
-        orig_to_new_prefix={"layers.": "encoder_layer.layers."}
+        orig_to_new_prefix={
+            "model.layers.": "encoder_layer.layers.",
+            "layers.": "encoder_layer.layers.",
+            "model.": "",
+        }
     )
 
     def __init__(
@@ -266,6 +270,8 @@ class ModernBertModel(nn.Module):
         for name, loaded_weight in weights:
             if name.endswith(".bias") and name not in params_dict:
                 continue
+            if name not in params_dict:
+                continue
             param = params_dict[name]
             weight_loader = getattr(param, "weight_loader", default_weight_loader)
             weight_loader(param, loaded_weight)

vllm/model_executor/warmup/deep_gemm_warmup.py

@@ -138,9 +138,6 @@ def _fp8_linear_may_use_deep_gemm(module: torch.nn.Module) -> bool:
     Return True if the input module/layer could be processed with DeepGEMM.
     """
 
-    # FIXME: this logic is brittle and incorrect - since we
-    # could use DeepGEMM with for than just Fp8LinearMethod
-    block_size = get_mk_alignment_for_contiguous_layout()[0]
     if not (
         isinstance(module, LinearBase)
         and isinstance(module.quant_method, Fp8LinearMethod)
@@ -156,6 +153,8 @@ def _fp8_linear_may_use_deep_gemm(module: torch.nn.Module) -> bool:
     ):
         return False
 
+    block_size = get_mk_alignment_for_contiguous_layout()[0]
+
     w, _, block_sizes = _extract_data_from_linear_base_module(module)
     return (
         block_sizes == get_mk_alignment_for_contiguous_layout()