vllm.model_executor.layers.quantization.fp8 ¶

class CopyNumelCounter(TorchDispatchMode):
    """
    Tracks total number of elements modified with `copy_`. Useful for keeping
    track of weight loading where underlying weights can be arbitrarily
    transformed (such as with `narrow`) before calling copy.
    """

    def __init__(self):
        super().__init__()
        self.copied_numel = 0

    def __torch_dispatch__(self, func, types, args=(), kwargs=None):
        if kwargs is None:
            kwargs = {}
        out = func(*args, **kwargs)
        if func == torch.ops.aten.copy_.default:
            self.copied_numel += args[0].numel()
        return out

class Fp8Config(QuantizationConfig):
    """Config class for FP8."""

    def __init__(
        self,
        is_checkpoint_fp8_serialized: bool = False,
        activation_scheme: str = "dynamic",
        ignored_layers: list[str] | None = None,
        weight_block_size: list[int] | None = None,
    ) -> None:
        super().__init__()

        self.is_checkpoint_fp8_serialized = is_checkpoint_fp8_serialized

        if activation_scheme not in ACTIVATION_SCHEMES:
            raise ValueError(f"Unsupported activation scheme {activation_scheme}")
        self.activation_scheme = activation_scheme
        self.ignored_layers = ignored_layers or []
        if weight_block_size is not None:
            if not is_checkpoint_fp8_serialized:
                raise ValueError(
                    "The block-wise quantization only supports fp8-serialized "
                    "checkpoint for now."
                )
            if len(weight_block_size) != 2:
                raise ValueError(
                    "The quantization block size of weight must have 2 "
                    f"dimensions, but got {len(weight_block_size)} dimensions"
                )
            if activation_scheme != "dynamic":
                raise ValueError(
                    "The block-wise quantization only supports "
                    "dynamic activation scheme for now, but got "
                    f"{activation_scheme} activation scheme."
                )
        self.weight_block_size = weight_block_size

    @classmethod
    def get_name(cls) -> QuantizationMethods:
        return "fp8"

    @classmethod
    def get_supported_act_dtypes(cls) -> list[torch.dtype]:
        return [torch.bfloat16, torch.half]

    @classmethod
    def get_min_capability(cls) -> int:
        return 75

    @classmethod
    def get_config_filenames(cls) -> list[str]:
        return []

    def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):
        if self.ignored_layers is not None:
            self.ignored_layers = hf_to_vllm_mapper.apply_list(self.ignored_layers)

    @classmethod
    def from_config(cls, config: dict[str, Any]) -> "Fp8Config":
        quant_method = cls.get_from_keys(config, ["quant_method"])
        is_checkpoint_fp8_serialized = "fp8" in quant_method
        activation_scheme = cls.get_from_keys(config, ["activation_scheme"])
        ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
        weight_block_size = cls.get_from_keys_or(config, ["weight_block_size"], None)
        if not ignored_layers:
            ignored_layers = cls.get_from_keys_or(
                config, ["modules_to_not_convert"], None
            )
        return cls(
            is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized,
            activation_scheme=activation_scheme,
            ignored_layers=ignored_layers,
            weight_block_size=weight_block_size,
        )

    def get_xpu_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        from vllm.model_executor.layers.quantization.ipex_quant import (
            XPUFp8LinearMethod,
            XPUFp8MoEMethod,
        )

        fp8_config = Fp8Config(
            is_checkpoint_fp8_serialized=self.is_checkpoint_fp8_serialized,
            activation_scheme=self.activation_scheme,
            ignored_layers=self.ignored_layers,
            weight_block_size=self.weight_block_size,
        )

        if isinstance(layer, LinearBase):
            if is_layer_skipped(
                prefix=prefix,
                ignored_layers=self.ignored_layers,
                fused_mapping=self.packed_modules_mapping,
            ):
                return UnquantizedLinearMethod()
            return XPUFp8LinearMethod(fp8_config)
        elif isinstance(layer, FusedMoE):
            if is_layer_skipped(
                prefix=prefix,
                ignored_layers=self.ignored_layers,
                fused_mapping=self.packed_modules_mapping,
            ):
                return UnquantizedFusedMoEMethod(layer.moe_config)

            return XPUFp8MoEMethod(fp8_config, layer)
        elif isinstance(layer, Attention):
            return Fp8KVCacheMethod(self)
        return None

    def get_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        if current_platform.is_xpu():
            return self.get_xpu_quant_method(layer, prefix)
        if isinstance(layer, LinearBase):
            if is_layer_skipped(
                prefix=prefix,
                ignored_layers=self.ignored_layers,
                fused_mapping=self.packed_modules_mapping,
            ):
                return UnquantizedLinearMethod()
            if not self.is_checkpoint_fp8_serialized:
                online_method = Fp8OnlineLinearMethod(self)
                online_method.marlin_input_dtype = get_marlin_input_dtype(prefix)
                return online_method
            else:
                offline_method = Fp8LinearMethod(self)
                offline_method.marlin_input_dtype = get_marlin_input_dtype(prefix)
                return offline_method
        elif isinstance(layer, FusedMoE):
            if is_layer_skipped(
                prefix=prefix,
                ignored_layers=self.ignored_layers,
                fused_mapping=self.packed_modules_mapping,
            ):
                return UnquantizedFusedMoEMethod(layer.moe_config)
            if self.is_checkpoint_fp8_serialized:
                moe_quant_method = Fp8MoEMethod(self, layer)
            else:
                moe_quant_method = Fp8OnlineMoEMethod(self, layer)
            return moe_quant_method
        elif isinstance(layer, Attention):
            return Fp8KVCacheMethod(self)
        return None

    def get_cache_scale(self, name: str) -> str | None:
        """
        Check whether the param name matches the format for k/v cache scales
        in compressed-tensors. If this is the case, return its equivalent
        param name expected by vLLM

        :param name: param name
        :return: matching param name for KV cache scale in vLLM
        """
        if name.endswith(".output_scale") and ".k_proj" in name:
            return name.replace(".k_proj.output_scale", ".attn.k_scale")
        if name.endswith(".output_scale") and ".v_proj" in name:
            return name.replace(".v_proj.output_scale", ".attn.v_scale")
        if name.endswith(".output_scale") and ".q_proj" in name:
            return name.replace(".q_proj.output_scale", ".attn.q_scale")
        if name.endswith("self_attn.prob_output_scale"):
            return name.replace(".prob_output_scale", ".attn.prob_scale")
        # If no matches, return None
        return None

class Fp8KVCacheMethod(BaseKVCacheMethod):
    """
    Supports loading kv-cache scaling factors from FP8 checkpoints.
    """

    def __init__(self, quant_config: Fp8Config):
        super().__init__(quant_config)

class Fp8LinearMethod(LinearMethodBase):
    """Linear method for FP8.
    Supports loading FP8 checkpoints with static weight scale and
    dynamic/static activation scale.

    Limitations:
    1. Only support float8_e4m3fn data type due to the limitation of
       torch._scaled_mm (https://github.com/pytorch/pytorch/blob/2e48b39603411a41c5025efbe52f89560b827825/aten/src/ATen/native/cuda/Blas.cpp#L854-L856)

    Args:
        quant_config: The quantization config.
    """

    def __init__(self, quant_config: Fp8Config):
        self.quant_config = quant_config
        self.cutlass_block_fp8_supported = cutlass_block_fp8_supported()
        self.out_dtype = torch.get_default_dtype()

        # For GPUs that lack FP8 hardware support, we can leverage the Marlin
        # kernel for fast weight-only FP8 quantization
        self.marlin_input_dtype = None
        self.use_marlin = (
            not current_platform.has_device_capability(89)
            or envs.VLLM_TEST_FORCE_FP8_MARLIN
        )
        # Disable marlin for rocm
        if current_platform.is_rocm():
            self.use_marlin = False
        if vllm_is_batch_invariant():
            self.use_marlin = False

        self.use_aiter_and_is_supported = rocm_aiter_ops.is_linear_fp8_enabled()
        self.use_deep_gemm = is_deep_gemm_supported()

        self.weight_block_size = self.quant_config.weight_block_size
        self.block_quant = self.weight_block_size is not None
        self.act_q_static = self.quant_config.activation_scheme == "static"

        if self.block_quant:
            assert not self.act_q_static
            assert self.weight_block_size is not None
            self.w8a8_block_fp8_linear = W8A8BlockFp8LinearOp(
                weight_group_shape=GroupShape(*self.weight_block_size),
                act_quant_group_shape=GroupShape(1, self.weight_block_size[0]),
                cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
                use_aiter_and_is_supported=self.use_aiter_and_is_supported,
            )
        else:
            # Use per-token quantization for better perf if dynamic and cutlass
            if self.act_q_static:
                activation_quant_key = kFp8StaticTensorSym
            elif cutlass_fp8_supported():
                activation_quant_key = kFp8DynamicTokenSym
            else:
                activation_quant_key = kFp8DynamicTensorSym

            self.fp8_linear = init_fp8_linear_kernel(
                activation_quant_key=activation_quant_key,
                weight_quant_key=kFp8StaticTensorSym,
                out_dtype=torch.get_default_dtype(),
                module_name=self.__class__.__name__,
            )

    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,
    ):
        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

        if self.block_quant:
            assert self.weight_block_size is not None
            layer.weight_block_size = self.weight_block_size
            validate_fp8_block_shape(
                layer,
                input_size,
                output_size,
                input_size_per_partition,
                output_partition_sizes,
                self.weight_block_size,
            )

        weight = create_fp8_weight_parameter(
            output_size_per_partition, input_size_per_partition, weight_loader
        )
        layer.register_parameter("weight", weight)

        # WEIGHT SCALE
        if not self.block_quant:
            scale = create_fp8_scale_parameter(
                PerTensorScaleParameter,
                output_partition_sizes,
                input_size_per_partition,
                None,
                weight_loader,
            )
            set_weight_attrs(scale, {"scale_type": "weight_scale"})
            layer.register_parameter("weight_scale", scale)
        else:
            assert not self.act_q_static
            assert self.weight_block_size is not None
            scale = create_fp8_scale_parameter(
                BlockQuantScaleParameter,
                output_partition_sizes,
                input_size_per_partition,
                self.weight_block_size,
                weight_loader,
            )
            set_weight_attrs(scale, {"scale_type": "weight_scale"})
            # The weight_scale_inv name is intentional for deepseekv3
            layer.register_parameter("weight_scale_inv", scale)

        # INPUT ACTIVATION SCALE
        if self.act_q_static:
            scale = create_fp8_input_scale(output_partition_sizes, weight_loader)
            set_weight_attrs(scale, {"scale_type": "input_scale"})
            layer.register_parameter("input_scale", scale)

    def process_weights_after_loading(self, layer: Module) -> None:
        size_k_first = True
        input_scale = None
        # TODO(rob): refactor block quant into separate class.
        if self.block_quant:
            assert not self.act_q_static
            size_k_first = False

            weight, weight_scale_inv = process_fp8_weight_block_strategy(
                layer.weight, layer.weight_scale_inv
            )

            # Update layer with new values
            replace_parameter(layer, "weight", weight.data)
            replace_parameter(layer, "weight_scale_inv", weight_scale_inv.data)

        # If checkpoint not serialized fp8, quantize the weights.
        else:
            # If checkpoint is fp8 per-tensor, handle that there are N scales for N
            # shards in a fused module
            weight = layer.weight
            weight_scale = layer.weight_scale

            # If using w8a8, torch._scaled_mm needs per tensor, so
            # requantize the logical shards as a single weight.
            if not self.use_marlin:
                weight, weight_scale, input_scale = process_fp8_weight_tensor_strategy(
                    weight,
                    weight_scale,
                    layer.logical_widths,
                    getattr(layer, "input_scale", None),
                )
                if self.act_q_static:
                    assert input_scale is not None
                    input_scale = input_scale.max()
            weight = weight.t()

            # Update layer with new values.
            replace_parameter(layer, "weight", weight.data)
            replace_parameter(layer, "weight_scale", weight_scale.data)

        if input_scale is not None:
            replace_parameter(layer, "input_scale", input_scale)
        else:
            layer.input_scale = None

        if self.use_marlin:
            prepare_fp8_layer_for_marlin(
                layer, size_k_first, input_dtype=self.marlin_input_dtype
            )
            # Activations not quantized for marlin.
            del layer.input_scale
            return

        if self.block_quant:
            maybe_post_process_fp8_weight_block(layer)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        # if batch invariant mode is enabled, prefer DeepGEMM FP8 path
        # we will use BF16 dequant when DeepGEMM is not supported.
        if vllm_is_batch_invariant():
            if self.block_quant:
                assert self.weight_block_size is not None
                return self.w8a8_block_fp8_linear.apply(
                    input=x,
                    weight=layer.weight,
                    weight_scale=layer.weight_scale_inv,
                    input_scale=layer.input_scale,
                    bias=bias,
                )
            else:
                # per-tensor/channel: dequant to BF16 and run GEMM
                weight_fp8 = layer.weight.to(torch.bfloat16)
                weight_scale = layer.weight_scale.to(torch.bfloat16)
                if weight_scale.numel() == 1:
                    # Per-tensor: simple scalar multiplication
                    weight_bf16 = weight_fp8 * weight_scale
                else:
                    # Multiple scales (fused modules like QKV)
                    # Try to infer correct broadcasting
                    # weight is [K, N], scale could be [num_logical_weights]
                    # Need to figure out how to broadcast - for now just try
                    # direct multiplication
                    if (
                        weight_scale.dim() == 1
                        and weight_scale.shape[0] == weight_fp8.shape[0]
                    ):
                        # Per-row scaling
                        weight_bf16 = weight_fp8 * weight_scale.unsqueeze(1)
                    else:
                        # Fallback
                        weight_bf16 = weight_fp8 * weight_scale
                return torch.nn.functional.linear(x, weight_bf16.t(), bias)

        if self.use_marlin:
            if self.block_quant:
                weight_scale = layer.weight_scale_inv
            else:
                weight_scale = layer.weight_scale

            return apply_fp8_marlin_linear(
                input=x,
                weight=layer.weight,
                weight_scale=weight_scale,
                workspace=layer.workspace,
                size_n=layer.output_size_per_partition,
                size_k=layer.input_size_per_partition,
                input_dtype=self.marlin_input_dtype,
                bias=bias,
            )

        if self.block_quant:
            assert self.weight_block_size is not None

            return self.w8a8_block_fp8_linear.apply(
                input=x,
                weight=layer.weight,
                weight_scale=layer.weight_scale_inv,
                input_scale=layer.input_scale,
                bias=bias,
            )

        return self.fp8_linear.apply_weights(layer, x, bias)