[ModelOpt] Load w13/w2_input_scale for all experts, nvfp4 (#26135)

Signed-off-by: Shu Wang <shuw@nvidia.com>
Signed-off-by: Shu Wang. <shuw@nvidia.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>

vllm/model_executor/layers/fused_moe/layer.py

@@ -49,6 +49,9 @@ from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
 )
+from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
+    is_flashinfer_supporting_global_sf,
+)
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.platforms.interface import CpuArchEnum
@@ -1289,6 +1292,7 @@ class FusedMoE(CustomOp):
             "intermediate_size_per_partition": self.intermediate_size_per_partition,
             "params_dtype": params_dtype,
             "weight_loader": self.weight_loader,
+            "global_num_experts": self.global_num_experts,
         }
         # need full intermediate size pre-sharding for WNA16 act order
         if self.quant_method.__class__.__name__ in (
@@ -1632,13 +1636,25 @@ class FusedMoE(CustomOp):
                 param.data[:, :dim1, :dim2].copy_(loaded_weight)
             return True if return_success else None
 
-        expert_id = self._map_global_expert_id_to_local_expert_id(expert_id)
-        if expert_id == -1:
+        quant_method_name = self.quant_method.__class__.__name__
+        global_expert_id = expert_id
+        expert_id = self._map_global_expert_id_to_local_expert_id(global_expert_id)
+
+        allow_flashinfer = getattr(self.quant_method, "allow_flashinfer", False)
+        moe_backend = getattr(self.quant_method, "flashinfer_moe_backend", None)
+
+        use_global_sf = (
+            allow_flashinfer
+            and is_flashinfer_supporting_global_sf(moe_backend)
+            and "input_scale" in weight_name
+            and quant_method_name == "ModelOptNvFp4FusedMoE"
+        )
+
+        if expert_id == -1 and not use_global_sf:
             # Failed to load this param since it's not local to this rank
             return False if return_success else None
         # Hereafter, `expert_id` is local physical id
 
-        quant_method_name = self.quant_method.__class__.__name__
         # compressed-tensors checkpoints with packed weights are stored flipped
         # TODO (mgoin): check self.quant_method.quant_config.quant_format
         # against known CompressionFormat enum values that have this quality
@@ -1723,7 +1739,9 @@ class FusedMoE(CustomOp):
                 )
 
             self._load_single_value(
-                param=param, loaded_weight=loaded_weight, expert_id=expert_id
+                param=param,
+                loaded_weight=loaded_weight,
+                expert_id=global_expert_id if use_global_sf else expert_id,
             )
             return True if return_success else None
 

vllm/model_executor/layers/quantization/modelopt.py

@@ -49,6 +49,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
     build_flashinfer_fp8_cutlass_moe_prepare_finalize,
     flashinfer_cutlass_moe_fp8,
     get_flashinfer_moe_backend,
+    is_flashinfer_supporting_global_sf,
     register_moe_scaling_factors,
     rotate_flashinfer_fp8_moe_weights,
     select_cutlass_fp8_gemm_impl,
@@ -1217,6 +1218,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
         weight_dtype = torch.uint8
         weight_scale_dtype = torch.float8_e4m3fn
         weight_loader = extra_weight_attrs.get("weight_loader")
+        global_num_experts = extra_weight_attrs.get("global_num_experts")
         # GEMM 1
         w13_weight = ModelWeightParameter(
             data=torch.empty(
@@ -1295,14 +1297,19 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
             {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
         )
 
+        use_global_sf = self.allow_flashinfer and is_flashinfer_supporting_global_sf(
+            self.flashinfer_moe_backend
+        )
+        global_scale_num_experts = global_num_experts if use_global_sf else num_experts
+
         w13_input_scale = PerTensorScaleParameter(
-            data=torch.empty(num_experts, 2, dtype=torch.float32),
+            data=torch.empty(global_scale_num_experts, 2, dtype=torch.float32),
             weight_loader=weight_loader,
         )
         layer.register_parameter("w13_input_scale", w13_input_scale)
 
         w2_input_scale = PerTensorScaleParameter(
-            data=torch.empty(num_experts, dtype=torch.float32),
+            data=torch.empty(global_scale_num_experts, dtype=torch.float32),
             weight_loader=weight_loader,
         )
         layer.register_parameter("w2_input_scale", w2_input_scale)
@@ -1457,7 +1464,17 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
         layer.w13_weight_scale_2 = Parameter(w13_weight_scale_2, requires_grad=False)
 
         # Common processing for input scales and alphas
-        w13_input_scale = layer.w13_input_scale.max(dim=1).values.to(torch.float32)
+        use_global_sf = self.allow_flashinfer and is_flashinfer_supporting_global_sf(
+            self.flashinfer_moe_backend
+        )
+        if use_global_sf:
+            # For backends provide by Flashinfer, the input global scales are
+            # shared across all experts.
+            w13_input_scale = (
+                layer.w13_input_scale.max().to(torch.float32).expand(layer.num_experts)
+            )
+        else:
+            w13_input_scale = layer.w13_input_scale.max(dim=1).values.to(torch.float32)
         layer.g1_alphas = Parameter(
             (w13_input_scale * w13_weight_scale_2).to(torch.float32),
             requires_grad=False,
@@ -1469,14 +1486,22 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
         )
 
         # GEMM 2 processing
+        if use_global_sf:
+            # For backends provide by Flashinfer, the input global scales are
+            # shared across all experts.
+            w2_input_scale = (
+                layer.w2_input_scale.max().to(torch.float32).expand(layer.num_experts)
+            )
+        else:
+            w2_input_scale = layer.w2_input_scale
         layer.g2_alphas = Parameter(
-            (layer.w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
+            (w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
             requires_grad=False,
         )
 
         # This is for quantization, so we need to invert it.
         layer.w2_input_scale_quant = Parameter(
-            (1 / layer.w2_input_scale).to(torch.float32), requires_grad=False
+            (1 / w2_input_scale).to(torch.float32), requires_grad=False
         )
 
         # TensorRT-LLM specific processing

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

@@ -263,3 +263,9 @@ def get_flashinfer_moe_backend() -> FlashinferMoeBackend:
         f"Unknown flashinfer moe backend: {flashinfer_moe_backend}"
         f" expected one of {allowed_backends}"
     )
+
+
+def is_flashinfer_supporting_global_sf(backend: FlashinferMoeBackend | None) -> bool:
+    # TODO(shuw@nvidia): Update when new backends are added.
+    backends_supporting_global_sf = (FlashinferMoeBackend.CUTLASS,)
+    return backend in backends_supporting_global_sf