[Bugfix] Allow fallback to AWQ from AWQMarlin at per-layer granularity (#13119)

vllm/model_executor/layers/quantization/utils/marlin_utils.py

@@ -6,6 +6,7 @@ import numpy
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.model_executor.layers.linear import LinearBase
 from vllm.platforms import current_platform
 from vllm.scalar_type import ScalarType, scalar_types
 
@@ -135,6 +136,20 @@ def check_marlin_supports_shape(output_size_per_partition: int,
     return True, None
 
 
+def check_marlin_supports_layer(layer: LinearBase, group_size: int) \
+                                    -> bool:
+    output_size_per_partition = getattr(layer, "output_size_per_partition",
+                                        None) or layer.output_size
+    input_size_per_partition = getattr(layer, "input_size_per_partition",
+                                       None) or layer.input_size
+
+    return check_marlin_supports_shape(
+        output_size_per_partition=output_size_per_partition,
+        input_size_per_partition=input_size_per_partition,
+        input_size=layer.input_size,
+        group_size=group_size)[0]
+
+
 def marlin_make_workspace(output_size_per_partition: int,
                           device: torch.device) -> torch.Tensor:
     max_workspace_size = (output_size_per_partition //