[XPU] fix broken fp8 online quantization for XPU platform (#30831)

Signed-off-by: Yan Ma <yan.ma@intel.com>

vllm/model_executor/layers/quantization/ipex_quant.py

@@ -27,6 +27,10 @@ from vllm.model_executor.layers.quantization.awq import AWQLinearMethod
 from vllm.model_executor.layers.quantization.fp8 import Fp8Config, Fp8LinearMethod
 from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
 from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    maybe_create_device_identity,
+)
+from vllm.model_executor.parameter import ModelWeightParameter
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 
@@ -305,6 +309,37 @@ class XPUFp8LinearMethod(Fp8LinearMethod):
     def __init__(self, quant_config: Fp8Config):
         super().__init__(quant_config)
 
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: list[int],
+        input_size: int,
+        output_size: int,
+        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")
+        layer.logical_widths = output_partition_sizes
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.orig_dtype = params_dtype
+        layer.weight_block_size = None
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition,
+                dtype=params_dtype,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
     def process_weights_after_loading(self, layer: Module) -> None:
         # If checkpoint not serialized fp8, quantize the weights.
         if not self.quant_config.is_checkpoint_fp8_serialized: