vllm.model_executor.layers.quantization.modelopt ¶

class ModelOptFp8Config(QuantizationConfig):
    """Config class for ModelOpt FP8."""

    def __init__(
        self,
        is_checkpoint_fp8_serialized: bool = False,
        kv_cache_quant_method: Optional[str] = None,
        exclude_modules: Optional[list[str]] = None,
    ) -> None:
        super().__init__()
        self.is_checkpoint_fp8_serialized = is_checkpoint_fp8_serialized
        self.kv_cache_quant_method = kv_cache_quant_method
        self.exclude_modules = exclude_modules or []
        if is_checkpoint_fp8_serialized:
            logger.warning(
                "Detected ModelOpt fp8 checkpoint. Please note that"
                " the format is experimental and could change."
            )

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

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

    @classmethod
    def get_min_capability(cls) -> int:
        return 89

    @classmethod
    def get_config_filenames(cls) -> list[str]:
        return ["hf_quant_config.json"]

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

    @classmethod
    def override_quantization_method(
        cls, hf_quant_cfg, user_quant
    ) -> Optional[QuantizationMethods]:
        """Detect if this ModelOpt config should be used based on
        quantization config."""

        if hf_quant_cfg is None:
            return None

        # Use the community standard 'quant_method'
        quant_method = hf_quant_cfg.get("quant_method", "").lower()

        # Only proceed if the method is explicitly "modelopt"
        if quant_method != "modelopt":
            return None

        # Look for ModelOpt-specific config structure
        if "quantization" in hf_quant_cfg:
            quant_config = hf_quant_cfg["quantization"]
            if isinstance(quant_config, dict):
                quant_algo = quant_config.get("quant_algo", "")
                if "FP8" in quant_algo:
                    return "modelopt"
        else:
            # Check for compressed-tensors style config with specific quant_algo
            quant_algo = hf_quant_cfg.get("quant_algo", "")
            if isinstance(quant_algo, str) and "FP8" in quant_algo:
                return "modelopt"

        return None

    @classmethod
    def from_config(cls, config: dict[str, Any]) -> "ModelOptFp8Config":
        # Handle both ModelOpt format and compressed-tensors style format
        if "quantization" in config:
            # ModelOpt format: {"quantization": {"quant_algo": "..."}}
            quant_config = cls.get_from_keys(config, ["quantization"])
            if not isinstance(quant_config, dict):
                raise ValueError("Expected 'quantization' to be a dictionary in config")
            quant_method = quant_config.get("quant_algo", "")
            if not quant_method:
                raise ValueError("Missing 'quant_algo' in quantization config")
            kv_cache_quant_method = quant_config.get("kv_cache_quant_algo")
            # "exclude_modules" is the key in the legacy hf_quant_config.json
            exclude_modules = quant_config.get("exclude_modules")
        else:
            # Compressed-tensors style format:
            # {"quant_algo": "...", "quant_method": "modelopt"}
            quant_method = config.get("quant_algo", "")
            kv_cache_quant_method = config.get("kv_cache_quant_algo")
            # "ignore" is the key in config.json
            exclude_modules = config.get("ignore")

        if quant_method not in QUANT_ALGOS:
            raise ValueError(
                f"ModelOpt currently only supports: {QUANT_ALGOS} "
                "quantizations in vLLM. Please check the "
                "`hf_quant_config.json` file for your model's "
                "quant configuration."
            )
        is_checkpoint_fp8_serialized = "FP8" in quant_method

        return cls(is_checkpoint_fp8_serialized, kv_cache_quant_method, exclude_modules)

    def is_layer_excluded(self, prefix: str) -> bool:
        """
        Check if a layer should be excluded from quantization.
        Handles both exact matching (for fused layers) and substring matching.

        This method handles both regular models and multimodal models that use
        the language_model prefix. For multimodal models, it checks if the
        module name (without the language_model prefix) is in the exclude list.
        """
        if self.exclude_modules is None:
            return False

        # First check exact matching with fused layer support
        if is_layer_skipped(prefix, self.exclude_modules, self.packed_modules_mapping):
            return True

        # Then check substring matching for patterns not caught by exact match
        for module in self.exclude_modules:
            # Skip exact matches already handled above
            if module != prefix and (
                module in prefix
                or (
                    prefix.startswith("language_model.")
                    and module in prefix.removeprefix("language_model.")
                )
            ):
                return True
        return False

    def get_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        from vllm.attention.layer import Attention  # Avoid circular import

        if isinstance(layer, LinearBase):
            if self.is_layer_excluded(prefix):
                return UnquantizedLinearMethod()
            # Check if this is a vision model layer that should not be quantized
            if "vision_tower" in prefix or "vision_model" in prefix:
                return UnquantizedLinearMethod()
            return ModelOptFp8LinearMethod(self)
        elif isinstance(layer, Attention):
            return ModelOptFp8KVCacheMethod(self)
        elif isinstance(layer, FusedMoE):
            return ModelOptFp8MoEMethod(self, layer)
        return None

class ModelOptFp8KVCacheMethod(BaseKVCacheMethod):
    """
    Supports loading kv-cache scaling factors from FP8 checkpoints.
    """

    def __init__(self, quant_config: Union[ModelOptFp8Config, ModelOptNvFp4Config]):
        super().__init__(quant_config)

class ModelOptFp8LinearMethod(LinearMethodBase):
    """Linear method for Model Optimizer static quantization.
    Supports loading FP8 checkpoints with static weight scale and
    activation scale. Future support might be added for dynamic
    scales.

    Limitations:
    1. Only support per-tensor quantization due to torch._scaled_mm support.
    2. Only support float8_e4m3fn datatype
        Args: quant_config: The ModelOpt quantization config.
    """

    def __init__(self, quant_config: ModelOptFp8Config) -> None:
        self.quant_config = quant_config
        self.fp8_linear = Fp8LinearOp(
            act_quant_static=True, act_quant_group_shape=GroupShape.PER_TENSOR
        )

    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,
    ):
        del input_size, output_size
        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
        weight_dtype = (
            torch.float8_e4m3fn
            if self.quant_config.is_checkpoint_fp8_serialized
            else params_dtype
        )
        weight = ModelWeightParameter(
            data=torch.empty(
                output_size_per_partition, input_size_per_partition, dtype=weight_dtype
            ),
            input_dim=1,
            output_dim=0,
            weight_loader=weight_loader,
        )
        layer.register_parameter("weight", weight)

        if self.quant_config.is_checkpoint_fp8_serialized:
            # WEIGHT SCALE
            weight_scale = PerTensorScaleParameter(
                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
                weight_loader=weight_loader,
            )
            weight_scale[:] = torch.finfo(torch.float32).min
            layer.register_parameter("weight_scale", weight_scale)
            # INPUT SCALE
            scale = PerTensorScaleParameter(
                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
                weight_loader=weight_loader,
            )

            scale[:] = torch.finfo(torch.float32).min
            layer.register_parameter("input_scale", scale)

    def process_weights_after_loading(self, layer: Module) -> None:
        weight = layer.weight
        max_w_scale = layer.weight_scale.max()
        if not (layer.weight_scale == layer.weight_scale[0]).all():
            max_w_scale, weight = requantize_with_max_scale(
                layer.weight, layer.weight_scale, layer.logical_widths
            )
        layer.weight = Parameter(weight.t(), requires_grad=False)
        layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
        layer.input_scale = Parameter(layer.input_scale.max(), requires_grad=False)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        return self.fp8_linear.apply(
            input=x,
            weight=layer.weight,
            weight_scale=layer.weight_scale,
            input_scale=layer.input_scale,
            bias=bias,
        )

class ModelOptFp8MoEMethod(FusedMoEMethodBase):
    """MoE method for ModelOpt FP8.
    Supports loading FP8 checkpoints with static weight scale and
    activation scale.
    Args:
        quant_config: The ModelOpt quantization config.
    """

    def __init__(
        self,
        quant_config: ModelOptFp8Config,
        layer: torch.nn.Module,
    ) -> None:
        super().__init__(layer.moe_config)
        self.layer = layer
        self.quant_config = quant_config
        from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
            cutlass_fp8_supported,
        )

        self.cutlass_fp8_supported = cutlass_fp8_supported()
        self.flashinfer_moe_backend: Optional[FlashinferMoeBackend] = None
        if envs.VLLM_USE_FLASHINFER_MOE_FP8 and has_flashinfer_moe():
            self.flashinfer_moe_backend = get_flashinfer_moe_backend()
            logger.info_once(
                f"Using FlashInfer {self.flashinfer_moe_backend.value} kernels"
            )

    def maybe_make_prepare_finalize(
        self,
    ) -> Optional[mk.FusedMoEPrepareAndFinalize]:
        # TRT LLM not supported with all2all yet.
        if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
            return None
        elif self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
            prepare_finalize = build_flashinfer_fp8_cutlass_moe_prepare_finalize(
                self.moe
            )
            logger.debug_once("%s", prepare_finalize.__class__.__name__)
            return prepare_finalize
        else:
            return super().maybe_make_prepare_finalize()

    def select_gemm_impl(
        self,
        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
        layer: torch.nn.Module,
    ) -> mk.FusedMoEPermuteExpertsUnpermute:
        assert self.moe_quant_config is not None
        experts = select_cutlass_fp8_gemm_impl(
            self.moe,
            self.moe_quant_config,
        )
        logger.debug_once("Using %s", experts.__class__.__name__)
        return experts

    def create_weights(
        self,
        layer: torch.nn.Module,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        # Use FP8 dtype if checkpoint is serialized
        weight_dtype = (
            torch.float8_e4m3fn
            if self.quant_config.is_checkpoint_fp8_serialized
            else params_dtype
        )
        weight_loader = extra_weight_attrs.get("weight_loader")

        w13_weight = ModelWeightParameter(
            data=torch.empty(
                num_experts,
                2 * intermediate_size_per_partition,
                hidden_size,
                dtype=weight_dtype,
            ),
            input_dim=2,
            output_dim=1,
            weight_loader=weight_loader,
        )
        layer.register_parameter("w13_weight", w13_weight)

        w2_weight = ModelWeightParameter(
            data=torch.empty(
                num_experts,
                hidden_size,
                intermediate_size_per_partition,
                dtype=weight_dtype,
            ),
            input_dim=2,
            output_dim=1,
            weight_loader=weight_loader,
        )
        layer.register_parameter("w2_weight", w2_weight)

        if self.quant_config.is_checkpoint_fp8_serialized:
            # WEIGHT SCALES - Per-tensor scaling for ModelOpts
            # Allocate 2 scales for w1 and w3 respectively.
            # They will be combined to a single scale after weight loading.
            w13_weight_scale = PerTensorScaleParameter(
                data=torch.full(
                    (num_experts, 2),
                    1.0,
                    dtype=torch.float32,
                ),
                weight_loader=weight_loader,
            )
            w2_weight_scale = PerTensorScaleParameter(
                data=torch.full((num_experts,), 1.0, dtype=torch.float32),
                weight_loader=weight_loader,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            layer.register_parameter("w2_weight_scale", w2_weight_scale)

            # Set weight loader attributes for scales
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
            )

            # INPUT SCALES - Per-tensor scaling for ModelOpt
            w13_input_scale = PerTensorScaleParameter(
                data=torch.full((num_experts,), 1.0, dtype=torch.float32),
                weight_loader=weight_loader,
            )
            w2_input_scale = PerTensorScaleParameter(
                data=torch.full((num_experts,), 1.0, dtype=torch.float32),
                weight_loader=weight_loader,
            )
            layer.register_parameter("w13_input_scale", w13_input_scale)
            layer.register_parameter("w2_input_scale", w2_input_scale)

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        """Process FP8 MoE weights after loading from serialized checkpoint.
        Only supports pre-quantized checkpoints with FP8 weights and scales.
        """

        layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False)
        layer.w2_weight = Parameter(layer.w2_weight.data, requires_grad=False)

        from vllm._custom_ops import scaled_fp8_quant
        from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
            per_tensor_dequantize,
        )

        # Handle scale parameters
        if hasattr(layer, "w13_weight_scale") and layer.w13_weight_scale is not None:
            # Fp8 moe kernel needs single weight scale for w13 per expert.
            # We take the max of the w1 and w3 scales
            # then dequant and requant each expert.
            if layer.w13_weight_scale.dim() == 2:
                # Get the maximum scale across w1 and w3 for each expert
                max_w13_scales = layer.w13_weight_scale.max(dim=1).values

                # Requantize each expert's weights using the combined scale
                # w13_weight (num_experts, 2 * intermediate_size, hidden_size)
                # where the first intermediate_size rows are w1, the next are w3
                intermediate_size = layer.w13_weight.shape[1] // 2
                for expert_id in range(layer.w13_weight.shape[0]):
                    start = 0
                    for shard_id in range(2):  # w1 and w3
                        # Dequantize using the original scale for this shard
                        dq_weight = per_tensor_dequantize(
                            layer.w13_weight[expert_id][
                                start : start + intermediate_size, :
                            ],
                            layer.w13_weight_scale[expert_id][shard_id],
                        )
                        # Requantize using the combined max scale

                        (
                            layer.w13_weight[expert_id][
                                start : start + intermediate_size, :
                            ],
                            _,
                        ) = scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])

                        start += intermediate_size

                # Update the scale parameter to be per-expert
                layer.w13_weight_scale = Parameter(max_w13_scales, requires_grad=False)
            else:
                layer.w13_weight_scale = Parameter(
                    layer.w13_weight_scale.data, requires_grad=False
                )

        if hasattr(layer, "w2_weight_scale") and layer.w2_weight_scale is not None:
            layer.w2_weight_scale = Parameter(
                layer.w2_weight_scale.data, requires_grad=False
            )
        # Input scales must be equal for each expert in fp8 MoE layers.
        if hasattr(layer, "w13_input_scale") and layer.w13_input_scale is not None:
            layer.w13_input_scale = Parameter(
                layer.w13_input_scale.max(), requires_grad=False
            )
        if hasattr(layer, "w2_input_scale") and layer.w2_input_scale is not None:
            layer.w2_input_scale = Parameter(
                layer.w2_input_scale.max(), requires_grad=False
            )

        if self.flashinfer_moe_backend is not None:
            layer.w13_weight.data = swap_w13_to_w31(layer.w13_weight.data)
            register_moe_scaling_factors(layer)
            if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
                rotate_flashinfer_fp8_moe_weights(layer.w13_weight, layer.w2_weight)

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> Optional[FusedMoEQuantConfig]:
        if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
            return None

        return fp8_w8a8_moe_quant_config(
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            per_act_token_quant=False,
        )

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: Optional[torch.Tensor] = None,
        logical_to_physical_map: Optional[torch.Tensor] = None,
        logical_replica_count: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        if enable_eplb:
            raise NotImplementedError(
                "EPLB not supported for `ModelOptFp8MoEMethod` yet."
            )

        if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
            assert self.fused_experts is None
            assert activation == "silu", (
                f"Expected 'silu' activation but got {activation}"
            )
            assert not renormalize
            return apply_flashinfer_per_tensor_scale_fp8(
                layer=layer,
                hidden_states=x,
                router_logits=router_logits,
                routing_bias=e_score_correction_bias,
                global_num_experts=global_num_experts,
                top_k=top_k,
                num_expert_group=num_expert_group,
                topk_group=topk_group,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )

        # Expert selection
        topk_weights, topk_ids, _ = FusedMoE.select_experts(
            hidden_states=x,
            router_logits=router_logits,
            use_grouped_topk=use_grouped_topk,
            top_k=top_k,
            renormalize=renormalize,
            topk_group=topk_group,
            num_expert_group=num_expert_group,
            custom_routing_function=custom_routing_function,
            scoring_func=scoring_func,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
            indices_type=self.topk_indices_dtype,
        )

        #
        # Note: the order here is important. self.fused_experts can override
        # cutlass or fused_experts.
        #
        if self.fused_experts is not None:
            return self.fused_experts(
                x,
                layer.w13_weight,
                layer.w2_weight,
                topk_weights,
                topk_ids,
                inplace=False,
                activation=activation,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )
        elif self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
            assert not renormalize
            assert activation == "silu", (
                f"Expected 'silu' activation but got {activation}"
            )
            return flashinfer_cutlass_moe_fp8(
                x,
                layer,
                topk_weights,
                topk_ids,
                inplace=False,
                activation=activation,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )
        else:
            from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts

            assert self.moe_quant_config is not None

            return fused_experts(
                x,
                layer.w13_weight,
                layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                inplace=True,
                activation=activation,
                quant_config=self.moe_quant_config,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )