[ROCm][Bugfix] Convert ModelOpt FP8 per-channel weights to e4m3fnuz on MI300/MI325 (#47201)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>

tests/quantization/test_modelopt.py

@@ -249,10 +249,11 @@ def test_modelopt_fp8_pc_pt_checkpoint_setup(default_vllm_config, vllm_runner):
             assert isinstance(gate_up_proj.quant_method, ModelOptFp8PcPtLinearMethod)
             assert isinstance(down_proj.quant_method, ModelOptFp8PcPtLinearMethod)
 
-            assert qkv_proj.weight.dtype == torch.float8_e4m3fn
-            assert o_proj.weight.dtype == torch.float8_e4m3fn
-            assert gate_up_proj.weight.dtype == torch.float8_e4m3fn
-            assert down_proj.weight.dtype == torch.float8_e4m3fn
+            fp8_dtype = current_platform.fp8_dtype()
+            assert qkv_proj.weight.dtype == fp8_dtype
+            assert o_proj.weight.dtype == fp8_dtype
+            assert gate_up_proj.weight.dtype == fp8_dtype
+            assert down_proj.weight.dtype == fp8_dtype
 
             # Per-channel scales; activations are dynamically scaled per token.
             assert hasattr(qkv_proj, "weight_scale")

vllm/model_executor/layers/quantization/modelopt.py

@@ -63,6 +63,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
 )
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     process_fp8_input_tensor_strategy_moe,
+    process_fp8_weight_channel_strategy,
     process_fp8_weight_tensor_strategy_moe,
 )
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
@@ -604,8 +605,11 @@ class ModelOptFp8PcPtLinearMethod(LinearMethodBase):
         )
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
-        layer.weight = Parameter(layer.weight.t(), requires_grad=False)
-        layer.weight_scale = Parameter(layer.weight_scale.data, requires_grad=False)
+        weight, weight_scale, _ = process_fp8_weight_channel_strategy(
+            layer.weight, layer.weight_scale.data
+        )
+        layer.weight = Parameter(weight.t(), requires_grad=False)
+        layer.weight_scale = Parameter(weight_scale, requires_grad=False)
         self.fp8_linear.process_weights_after_loading(layer)
 
     def apply(