[Bugfix][CI][V1] Work around V1 + CUDA Graph + torch._scaled_mm fallback issue (vllm-project#13425)

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

Author: tlrmchlsmth

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -9,8 +9,8 @@
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
     CompressedTensorsScheme)
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    apply_fp8_linear, cutlass_fp8_supported, normalize_e4m3fn_to_e4m3fnuz,
-    requantize_with_max_scale)
+    apply_fp8_linear, cutlass_fp8_supported, maybe_create_device_identity,
+    normalize_e4m3fn_to_e4m3fnuz, requantize_with_max_scale)
 from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
                                            ModelWeightParameter,
                                            PerTensorScaleParameter)
@@ -93,6 +93,8 @@ def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
                        params_dtype: torch.dtype, weight_loader: Callable,
                        **kwargs):
+        maybe_create_device_identity()
+
         output_size_per_partition = sum(output_partition_sizes)
         layer.logical_widths = output_partition_sizes
 

vllm/model_executor/layers/quantization/fbgemm_fp8.py

@@ -17,7 +17,8 @@
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     is_layer_skipped)
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    apply_fp8_linear, normalize_e4m3fn_to_e4m3fnuz)
+    apply_fp8_linear, maybe_create_device_identity,
+    normalize_e4m3fn_to_e4m3fnuz)
 from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
                                            ModelWeightParameter)
 from vllm.platforms import current_platform
@@ -84,6 +85,7 @@ def create_weights(
         params_dtype: torch.dtype,
         **extra_weight_attrs,
     ):
+        maybe_create_device_identity()
         weight_loader = extra_weight_attrs.get("weight_loader")
         del input_size, output_size
         output_size_per_partition = sum(output_partition_sizes)

vllm/model_executor/layers/quantization/fp8.py

@@ -24,8 +24,8 @@
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     all_close_1d, apply_fp8_linear, convert_to_channelwise,
     cutlass_block_fp8_supported, cutlass_fp8_supported,
-    normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize,
-    requantize_with_max_scale)
+    maybe_create_device_identity, normalize_e4m3fn_to_e4m3fnuz,
+    per_tensor_dequantize, requantize_with_max_scale)
 from vllm.model_executor.parameter import (BlockQuantScaleParameter,
                                            ModelWeightParameter,
                                            PerTensorScaleParameter)
@@ -162,6 +162,8 @@ def create_weights(
         params_dtype: torch.dtype,
         **extra_weight_attrs,
     ):
+        maybe_create_device_identity()
+
         output_size_per_partition = sum(output_partition_sizes)
         weight_loader = extra_weight_attrs.get("weight_loader")
 

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -9,7 +9,7 @@
 
 # Input scaling factors are no longer optional in _scaled_mm starting
 # from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
-TORCH_DEVICE_IDENTITY = torch.ones(1, dtype=torch.float32)
+TORCH_DEVICE_IDENTITY = None
 
 # The condition to determine if it is on a platform that supports
 # torch._scaled_mm rowwise feature.
@@ -113,6 +113,13 @@ def requantize_with_max_scale(
     return max_w_scale, weight
 
 
+def maybe_create_device_identity():
+    # Allocate dummy ones tensor for torch._scaled_mm
+    global TORCH_DEVICE_IDENTITY
+    if TORCH_DEVICE_IDENTITY is None:
+        TORCH_DEVICE_IDENTITY = torch.ones(1, dtype=torch.float32)
+
+
 def apply_fp8_linear(
     input: torch.Tensor,
     weight: torch.Tensor,
@@ -215,11 +222,6 @@ def apply_fp8_linear(
             # For the scaled_mm fallback case, we break this down, since it
             # does not support s_w being a vector.
 
-            # Making sure the dummy tensor is on the same device as the weight
-            global TORCH_DEVICE_IDENTITY
-            if TORCH_DEVICE_IDENTITY.device != weight.device:
-                TORCH_DEVICE_IDENTITY = TORCH_DEVICE_IDENTITY.to(weight.device)
-
             # GEMM
             # This computes C = (X * W).
             # Output in fp32 to allow subsequent ops to happen in-place