sglang/python/sglang/srt/layers/quantization/__init__.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/v0.5.5/vllm/model_executor/layers/quantization/__init__.py
from __future__ import annotations

from typing import Dict, Type

from sglang.srt.layers.quantization.auto_round import AutoRoundConfig
from sglang.srt.layers.quantization.awq import (
    AWQConfig,
    AWQCPUConfig,
    AWQMarlinConfig,
    AWQXPUConfig,
)
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.quantization.bitsandbytes import BitsAndBytesConfig
from sglang.srt.layers.quantization.blockwise_int8 import BlockInt8Config
from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors import (
    CompressedTensorsConfig,
)
from sglang.srt.layers.quantization.fp8 import Fp8Config
from sglang.srt.layers.quantization.gguf import GGUFConfig
from sglang.srt.layers.quantization.gptq import (
    CPUGPTQConfig,
    GPTQAscendConfig,
    GPTQConfig,
    GPTQMarlinConfig,
    GPTQXPUConfig,
)
from sglang.srt.layers.quantization.humming import HummingConfig
from sglang.srt.layers.quantization.mlx import MlxQuantizationConfig
from sglang.srt.layers.quantization.modelopt_quant import (
    ModelOptFp4Config,
    ModelOptFp8Config,
    ModelOptMixedPrecisionConfig,
)
from sglang.srt.layers.quantization.modelslim.modelslim import ModelSlimConfig
from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
from sglang.srt.layers.quantization.mxfp4 import Mxfp4Config
from sglang.srt.layers.quantization.npu_mxfp4 import Mxfp4W4A8Config
from sglang.srt.layers.quantization.npu_mxfp4_w4a4 import Mxfp4W4A4Config
from sglang.srt.layers.quantization.nvfp4_online import NvFp4OnlineConfig
from sglang.srt.layers.quantization.petit import PetitNvFp4Config
from sglang.srt.layers.quantization.quark.quark import QuarkConfig
from sglang.srt.layers.quantization.quark_int4fp8_moe import QuarkInt4Fp8Config
from sglang.srt.layers.quantization.w4afp8 import W4AFp8Config
from sglang.srt.layers.quantization.w8a8_fp8 import W8A8Fp8Config
from sglang.srt.layers.quantization.w8a8_int8 import W8A8Int8Config
from sglang.srt.platforms import current_platform
from sglang.srt.utils import (
    cpu_has_amx_support,
    is_cpu,
    is_cuda,
    is_gfx95_supported,
    is_mps,
    is_npu,
    is_xpu,
)

_is_gfx95_supported = is_gfx95_supported()

# Base quantization methods
BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
    "fp8": Fp8Config,
    "mxfp8": Fp8Config,
    "blockwise_int8": BlockInt8Config,
    "modelopt": ModelOptFp8Config,  # Auto-detect, defaults to FP8
    "modelopt_fp8": ModelOptFp8Config,
    "modelopt_fp4": ModelOptFp4Config,
    "nvfp4_online": NvFp4OnlineConfig,
    "modelopt_mixed": ModelOptMixedPrecisionConfig,
    "w8a8_int8": W8A8Int8Config,
    "w8a8_fp8": W8A8Fp8Config,
    "awq": AWQConfig,
    "awq_marlin": AWQMarlinConfig,
    "bitsandbytes": BitsAndBytesConfig,
    "gguf": GGUFConfig,
    "gptq": GPTQConfig,
    "gptq_marlin": GPTQMarlinConfig,
    "moe_wna16": MoeWNA16Config,
    "compressed-tensors": CompressedTensorsConfig,
    "w4afp8": W4AFp8Config,
    "petit_nvfp4": PetitNvFp4Config,
    "quark": QuarkConfig,
    "quark_mxfp4": QuarkConfig,
    "auto-round": AutoRoundConfig,
    "auto-round-int8": W8A8Int8Config,
    "modelslim": ModelSlimConfig,
    "quark_int4fp8_moe": QuarkInt4Fp8Config,
    "humming": HummingConfig,
    "mxfp_w4a8": Mxfp4W4A8Config,
}


if is_cpu() or is_cuda() or _is_gfx95_supported:
    BASE_QUANTIZATION_METHODS.update(
        {
            "mxfp4": Mxfp4Config,
        }
    )


if is_npu():
    BASE_QUANTIZATION_METHODS.update(
        {
            "gptq": GPTQAscendConfig,
            # On NPU, `mxfp4` means single-level W4A4 MXFP4 for dense LLM (the
            # upstream `Mxfp4Config` OCP-MoE path is only registered on
            # cpu/cuda/hip above, so there is no collision here).
            "mxfp4": Mxfp4W4A4Config,
        }
    )


if is_xpu():
    BASE_QUANTIZATION_METHODS.update(
        {
            "gptq": GPTQXPUConfig,
            "awq": AWQXPUConfig,
        }
    )


if is_mps():
    BASE_QUANTIZATION_METHODS.update(
        {
            "mlx_q4": MlxQuantizationConfig,
            "mlx_q8": MlxQuantizationConfig,
        }
    )

# subset of above quant methods, supported on CPU
CPU_QUANTIZATION_METHODS = {
    "fp8": Fp8Config,
    "w8a8_int8": W8A8Int8Config,
    "compressed-tensors": CompressedTensorsConfig,
    "awq": AWQCPUConfig,
    "gptq": CPUGPTQConfig,
    "mxfp4": Mxfp4Config,
    "auto-round": AutoRoundConfig,
}

QUANTIZATION_METHODS = {**BASE_QUANTIZATION_METHODS}


def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
    if quantization not in QUANTIZATION_METHODS:
        raise ValueError(
            f"Invalid quantization method: {quantization}. "
            f"Available methods: {list(QUANTIZATION_METHODS.keys())}"
        )
    from sglang.srt.utils import is_cpu

    if is_cpu() and cpu_has_amx_support():
        if quantization not in CPU_QUANTIZATION_METHODS:
            raise ValueError(
                f"Invalid quantization method on CPU: {quantization}. "
                f"Available methods on CPU: {list(QUANTIZATION_METHODS.keys())}"
            )
        else:
            return CPU_QUANTIZATION_METHODS[quantization]

    if current_platform.is_out_of_tree():
        config = current_platform.get_quantization_config(quantization)

        # If the platform has a quantization config, use it else use the default
        if config is not None:
            return config

    return QUANTIZATION_METHODS[quantization]