vllm.model_executor.layers.quantization.quark.schemes.quark_w4a4_mxfp4 ¶

class QuarkW4A4MXFP4(QuarkScheme):

    def __init__(self, weight_quant_spec: dict[str, Any],
                 input_quant_spec: dict[str, Any]):
        self.out_dtype = torch.get_default_dtype()
        self.qscheme = "per_group"
        self.weight_quant_spec = weight_quant_spec
        self.input_quant_spec = input_quant_spec
        self.emulate = not current_platform.supports_mx()
        self.rocm_use_aiter_fp4_asm_gemm = is_rocm_aiter_fp4_asm_gemm_enabled()
        if not self.emulate and (dynamic_mxfp4_quant is None
                                 or gemm_afp4wfp4 is None):
            # Currently need these kernels if not emulating
            raise NotImplementedError(
                f"{self.__class__.__name__} requires AITER to be installed "
                "for non-emulation mode! Please refer to "
                "https://github.com/ROCm/aiter for installation details.")

    @classmethod
    def get_min_capability(cls) -> int:
        return 70

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        layer.weight = torch.nn.Parameter(layer.weight.data,
                                          requires_grad=False)

        if self.emulate:
            layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data,
                                                    requires_grad=False)
            try:
                from quark.torch.export.nn.modules import realquantizer
                from quark.torch.quantization.config.config import (
                    QuantizationSpec)
            except ImportError as err:
                raise ImportError(
                    "The package `amd-quark` is required to use AMD Quark "
                    "MX-FP4 models. Please install it with `pip install "
                    "amd-quark`.") from err

            weight_quant_spec = QuantizationSpec.from_dict(
                self.weight_quant_spec)

            weight_quantizer = realquantizer.get_real_quantizer(
                qspec=weight_quant_spec,
                quantizer=None,
                real_quantized=True,
                reorder=False,
                float_dtype=self.out_dtype,
                scale_shape=layer.weight_scale.shape,
                zero_point_shape=None,
            )
            weight_quantizer.scale.data = layer.weight_scale.data

            layer.weight = torch.nn.Parameter(
                weight_quantizer(layer.weight.data).to(self.out_dtype),
                requires_grad=False,
            )
            layer.weight_scale = None

            # This call is necessary to release the scales memory.
            torch.cuda.empty_cache()
        else:
            if self.rocm_use_aiter_fp4_asm_gemm:
                # shuffle weight scale
                weight_scale_shuffle = layer.weight_scale.data
                sm, sn = weight_scale_shuffle.shape
                weight_scale_shuffle = weight_scale_shuffle.view(
                    sm // 32, 2, 16, sn // 8, 2, 4, 1)
                weight_scale_shuffle = weight_scale_shuffle.permute(
                    0, 3, 5, 2, 4, 1, 6).contiguous()
                weight_scale_shuffle = weight_scale_shuffle.view(sm, sn)
                layer.weight_scale = torch.nn.Parameter(weight_scale_shuffle,
                                                        requires_grad=False)

                # shuffle weight
                weight_shuffle = layer.weight.data
                weight_shuffle = shuffle_weight(weight_shuffle,
                                                layout=(16, 16))
                layer.weight = torch.nn.Parameter(weight_shuffle,
                                                  requires_grad=False)
            else:
                layer.weight_scale = torch.nn.Parameter(
                    layer.weight_scale.data.T.contiguous(),
                    requires_grad=False)

    def create_weights(self, layer: torch.nn.Module,
                       output_partition_sizes: list[int],
                       input_size_per_partition: int,
                       params_dtype: torch.dtype, weight_loader: Callable,
                       **kwargs):
        output_size_per_partition = sum(output_partition_sizes)
        layer.logical_widths = output_partition_sizes

        # WEIGHT
        weight = PackedvLLMParameter(
            data=torch.empty(
                output_size_per_partition,
                input_size_per_partition // 2,
                dtype=torch.uint8,
            ),
            input_dim=1,
            output_dim=0,
            packed_dim=1,
            packed_factor=2,
            weight_loader=weight_loader,
        )
        layer.register_parameter("weight", weight)

        # WEIGHT SCALE
        weight_scale = GroupQuantScaleParameter(
            data=torch.empty(
                output_size_per_partition,
                input_size_per_partition // OCP_MX_BLOCK_SIZE,
                dtype=torch.uint8,
            ),
            input_dim=1,
            output_dim=0,
            weight_loader=weight_loader,
        )
        layer.register_parameter("weight_scale", weight_scale)

    def apply_weights(self,
                      layer: torch.nn.Module,
                      x: torch.Tensor,
                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:

        if self.emulate:
            dq_w = dequant_mxfp4(layer.weight, layer.weight_scale, x.dtype)
            x = quant_dequant_mxfp4(x)
            return F.linear(x, dq_w, bias)
        else:
            return torch.ops.vllm.gemm_with_dynamic_quant(
                x, layer.weight, layer.weight_scale,
                self.rocm_use_aiter_fp4_asm_gemm, self.out_dtype)