Merge pull request #62 from vkuzo/20250926_mxfp4_moe_quant

enable mxfp4 quant in vllm

Author: vkuzo

torchao_hf_vllm/torchao_hf_script.py

@@ -30,6 +30,7 @@
     Int8DynamicActivationInt4WeightConfig,
     CutlassInt4PackedLayout,
 )
+from torchao.quantization import ModuleFqnToConfig
 from torchao.prototype.mx_formats.inference_workflow import MXFPInferenceConfig
 from torchao.prototype.mx_formats import MXGemmKernelChoice
 from jsonargparse import CLI, Namespace
@@ -67,7 +68,6 @@ def get_quantization_config(args):
         case "fp8":
             single_config = Float8DynamicActivationFloat8WeightConfig(granularity=gran)
             if args.experts_only_qwen_1_5_moe_a_2_7b:
-                from torchao.quantization import ModuleFqnToConfig
                 expert_fqn_to_config = {}
                 # TODO(future PR): this is annoying, I should be able to use a regex here
                 for layer_idx in range(24):
@@ -101,14 +101,39 @@ def get_quantization_config(args):
         case "mxfp8":
             return TorchAoConfig(MXFPInferenceConfig())
         case "mxfp4":
-            return TorchAoConfig(
-                MXFPInferenceConfig(
-                    activation_dtype=torch.float4_e2m1fn_x2,
-                    weight_dtype=torch.float4_e2m1fn_x2,
-                    block_size=32,
-                    gemm_kernel_choice=MXGemmKernelChoice.CUTLASS,
-                )
+            single_config = MXFPInferenceConfig(
+                activation_dtype=torch.float4_e2m1fn_x2,
+                weight_dtype=torch.float4_e2m1fn_x2,
+                block_size=32,
+                # gemm_kernel_choice=MXGemmKernelChoice.CUTLASS,
+                gemm_kernel_choice=MXGemmKernelChoice.EMULATED,
             )
+            if args.experts_only_qwen_1_5_moe_a_2_7b:
+                expert_fqn_to_config = {}
+                # TODO(future PR): this is annoying, I should be able to use a regex here
+                for layer_idx in range(24):
+                    for expert_idx in range(60):
+                        expert_fqn_to_config[f"model.layers.{layer_idx}.mlp.experts.{expert_idx}.gate_proj"] = single_config
+                        expert_fqn_to_config[f"model.layers.{layer_idx}.mlp.experts.{expert_idx}.up_proj"] = single_config
+                        expert_fqn_to_config[f"model.layers.{layer_idx}.mlp.experts.{expert_idx}.down_proj"] = single_config
+                module_fqn_to_config = ModuleFqnToConfig({
+                    "_default": None,
+                    **expert_fqn_to_config,
+                })
+                return TorchAoConfig(
+                    quant_type=module_fqn_to_config,
+                )
+            else:
+                modules_to_not_convert = []
+                if args.skip_gate_qwen_1_5_moe_a_2_7b:
+                    for layer_idx in range(24):
+                        modules_to_not_convert.append(f"model.layers.{layer_idx}.mlp.gate")
+                        modules_to_not_convert.append(f"model.layers.{layer_idx}.mlp.shared_expert_gate")
+                modules_to_not_convert.append(f"lm_head")
+                return TorchAoConfig(
+                    single_config,
+                    modules_to_not_convert=modules_to_not_convert,
+                )
         case _:
             raise ValueError(f"Unsupported quantization type: {args.quant_type}")
 
@@ -165,6 +190,7 @@ def main(
     bench_tokens: int = 100,
     device_map: str = "cuda",
     experts_only_qwen_1_5_moe_a_2_7b: bool = False,
+    skip_gate_qwen_1_5_moe_a_2_7b: bool = False,
     save_model_to_disk: bool = True,
 ):
     """
@@ -182,6 +208,7 @@ def main(
         bench_tokens: Number of tokens to generate for benchmarking
         device_map: Device mapping strategy
         experts_only_qwen_1_5_moe_a_2_7b: if True, quantizes experts only for Qwen1.5-MoE-A2.7B model
+        skip_gate_qwen_1_5_moe_a_2_7b: if True, skips gate quantization for Qwen1.5-MoE-A2.7B model
         save_model_to_disk: if True, saves quantized model to local disk
     """
     # Set seed before creating the model
@@ -206,11 +233,14 @@ def main(
         device_map=device_map,
         experts_only_qwen_1_5_moe_a_2_7b=experts_only_qwen_1_5_moe_a_2_7b,
         save_model_to_disk=save_model_to_disk,
+        skip_gate_qwen_1_5_moe_a_2_7b=skip_gate_qwen_1_5_moe_a_2_7b,
     )
     print(f"{args=}")
 
     if args.experts_only_qwen_1_5_moe_a_2_7b:
-        assert args.quant_type == "fp8", "unsupported"
+        assert args.quant_type in ("fp8", "mxfp4"), "unsupported"
+
+    assert not args.skip_gate_qwen_1_5_moe_a_2_7b and args.experts_only_qwen_1_5_moe_a_2_7b, "unsupported"
 
     # Create output directory
     output_dir = Path(args.output_dir)