Zhou TaichangandClaude Opus 5
651b03225b
Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
413 lines
13 KiB
Python
413 lines
13 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
|
"""Serialized SVDQuant NVFP4 support for diffusion transformers.
|
|
|
|
The checkpoint stores NVFP4 weights plus a rank-R correction for each
|
|
quantized linear. The four-bit GEMM uses vLLM's existing NVFP4 kernel
|
|
registry, while the rank correction uses ordinary BF16 matrix multiplication.
|
|
Native SVDQuant fusion is a separate optimization and is not required to load
|
|
or run the checkpoint.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import functools
|
|
from typing import TYPE_CHECKING, Any
|
|
|
|
import torch
|
|
from torch.nn import Parameter
|
|
from vllm.logger import init_logger
|
|
from vllm.model_executor.kernels.linear import (
|
|
CutlassNvFp4LinearKernel,
|
|
FbgemmNvFp4LinearKernel,
|
|
FlashInferB12xNvFp4LinearKernel,
|
|
FlashInferCudnnNvFp4LinearKernel,
|
|
FlashInferCuteDslNvFp4LinearKernel,
|
|
FlashInferCutlassNvFp4LinearKernel,
|
|
FlashInferTrtllmNvFp4LinearKernel,
|
|
NvFp4LinearKernel,
|
|
init_nvfp4_linear_kernel,
|
|
)
|
|
from vllm.model_executor.layers.linear import (
|
|
LinearBase,
|
|
LinearMethodBase,
|
|
UnquantizedLinearMethod,
|
|
)
|
|
from vllm.model_executor.layers.quantization.base_config import (
|
|
QuantizationConfig,
|
|
QuantizeMethodBase,
|
|
)
|
|
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
|
is_layer_skipped,
|
|
)
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
from vllm.model_executor.utils import set_weight_attrs
|
|
from vllm.platforms import current_platform
|
|
from vllm.platforms.interface import DeviceCapability
|
|
|
|
if TYPE_CHECKING:
|
|
from vllm.model_executor.layers.quantization import QuantizationMethods
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
_SUPPORTED_CAPABILITIES = {(10, 3)}
|
|
_COMPATIBLE_NVFP4_KERNELS = (
|
|
CutlassNvFp4LinearKernel,
|
|
FbgemmNvFp4LinearKernel,
|
|
FlashInferB12xNvFp4LinearKernel,
|
|
FlashInferCudnnNvFp4LinearKernel,
|
|
FlashInferCuteDslNvFp4LinearKernel,
|
|
FlashInferCutlassNvFp4LinearKernel,
|
|
FlashInferTrtllmNvFp4LinearKernel,
|
|
)
|
|
|
|
|
|
def _supports_capability(capability: DeviceCapability | None) -> bool:
|
|
return (
|
|
capability is not None
|
|
and (
|
|
capability.major,
|
|
capability.minor,
|
|
)
|
|
in _SUPPORTED_CAPABILITIES
|
|
)
|
|
|
|
|
|
def _assert_supported() -> None:
|
|
if not current_platform.is_cuda():
|
|
raise RuntimeError("SVDQuant NVFP4 requires a CUDA device")
|
|
capability = current_platform.get_device_capability()
|
|
if not _supports_capability(capability):
|
|
device = current_platform.device_name
|
|
sm = capability.to_int() if capability is not None else "unknown"
|
|
raise RuntimeError(f"SVDQuant NVFP4 is validated on SM103 only; got {device!r} (SM{sm})")
|
|
|
|
|
|
@functools.cache
|
|
def _nvfp4_kernel() -> NvFp4LinearKernel:
|
|
kernel = init_nvfp4_linear_kernel()
|
|
if not isinstance(kernel, _COMPATIBLE_NVFP4_KERNELS):
|
|
raise RuntimeError(
|
|
"The selected vLLM NVFP4 backend is incompatible with the "
|
|
f"SVDQuant checkpoint layout: {type(kernel).__name__}. Use one of "
|
|
"the FlashInfer, CUTLASS, or FBGEMM NVFP4 backends."
|
|
)
|
|
return kernel
|
|
|
|
|
|
class DiffusionSVDQuantConfig(QuantizationConfig):
|
|
"""Configuration for serialized NVFP4 W4A4 plus low-rank correction."""
|
|
|
|
def __init__(
|
|
self,
|
|
rank: int = 32,
|
|
precision: str = "nvfp4",
|
|
act_unsigned: bool = False,
|
|
modules_to_not_convert: list[str] | None = None,
|
|
) -> None:
|
|
super().__init__()
|
|
if rank <= 0:
|
|
raise ValueError(f"SVDQuant rank must be positive, got {rank}")
|
|
if precision != "nvfp4":
|
|
raise ValueError(
|
|
f"Phase 1 SVDQuant supports serialized NVFP4 checkpoints only; got precision={precision!r}"
|
|
)
|
|
if act_unsigned:
|
|
raise ValueError("Phase 1 SVDQuant does not support unsigned activations")
|
|
self.rank = rank
|
|
self.precision = precision
|
|
self.modules_to_not_convert = modules_to_not_convert or []
|
|
|
|
def __repr__(self) -> str:
|
|
return f"DiffusionSVDQuantConfig(rank={self.rank}, precision={self.precision!r})"
|
|
|
|
@classmethod
|
|
def get_name(cls) -> QuantizationMethods:
|
|
return "svdquant"
|
|
|
|
@classmethod
|
|
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
|
|
return [torch.bfloat16]
|
|
|
|
@classmethod
|
|
def get_min_capability(cls) -> int:
|
|
return 103
|
|
|
|
@classmethod
|
|
def get_config_filenames(cls) -> list[str]:
|
|
return ["quantization_config.json"]
|
|
|
|
@classmethod
|
|
def from_config(cls, config: dict[str, Any]) -> DiffusionSVDQuantConfig:
|
|
return cls(
|
|
rank=config.get("rank", 32),
|
|
precision=config.get("precision", "nvfp4"),
|
|
act_unsigned=config.get("act_unsigned", False),
|
|
modules_to_not_convert=config.get("modules_to_not_convert"),
|
|
)
|
|
|
|
def get_quant_method(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
prefix: str,
|
|
) -> QuantizeMethodBase | None:
|
|
if not isinstance(layer, LinearBase):
|
|
return None
|
|
if is_layer_skipped(
|
|
prefix,
|
|
self.modules_to_not_convert,
|
|
self.packed_modules_mapping,
|
|
# vLLM 0.28: skip_with_substr was replaced by match_mode
|
|
# ("substring" preserves the v0.27 skip_with_substr=True
|
|
# behavior verbatim).
|
|
match_mode="substring",
|
|
):
|
|
return UnquantizedLinearMethod()
|
|
return DiffusionSVDQuantLinearMethod(self)
|
|
|
|
|
|
class DiffusionSVDQuantLinearMethod(LinearMethodBase):
|
|
"""Load and execute a serialized SVDQuant linear layer."""
|
|
|
|
def __init__(self, quant_config: DiffusionSVDQuantConfig) -> None:
|
|
_assert_supported()
|
|
self.quant_config = quant_config
|
|
|
|
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: Any,
|
|
) -> None:
|
|
del params_dtype
|
|
weight_loader = extra_weight_attrs.pop(
|
|
"weight_loader",
|
|
default_weight_loader,
|
|
)
|
|
if input_size_per_partition % 16 != 0:
|
|
raise ValueError(
|
|
"SVDQuant NVFP4 requires each input partition to be divisible "
|
|
f"by the block size 16; got {input_size_per_partition}"
|
|
)
|
|
output_size_per_partition = sum(output_partition_sizes)
|
|
rank = self.quant_config.rank
|
|
|
|
qweight = Parameter(
|
|
torch.empty(
|
|
output_size_per_partition,
|
|
input_size_per_partition // 2,
|
|
dtype=torch.int8,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
qweight,
|
|
{
|
|
"input_dim": 1,
|
|
"output_dim": 0,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
wscales = Parameter(
|
|
torch.empty(
|
|
input_size_per_partition // 16,
|
|
output_size_per_partition,
|
|
dtype=torch.float8_e4m3fn,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
wscales,
|
|
{
|
|
"input_dim": 0,
|
|
"output_dim": 1,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
proj_down = Parameter(
|
|
torch.empty(
|
|
input_size_per_partition,
|
|
rank,
|
|
dtype=torch.bfloat16,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
proj_down,
|
|
{
|
|
"input_dim": 0,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
proj_up = Parameter(
|
|
torch.empty(
|
|
output_size_per_partition,
|
|
rank,
|
|
dtype=torch.bfloat16,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
proj_up,
|
|
{
|
|
"output_dim": 0,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
smooth_factor = Parameter(
|
|
torch.empty(
|
|
input_size_per_partition,
|
|
dtype=torch.bfloat16,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
smooth_factor,
|
|
{
|
|
"input_dim": 0,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
wcscales = Parameter(
|
|
torch.ones(
|
|
output_size_per_partition,
|
|
dtype=torch.bfloat16,
|
|
),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(
|
|
wcscales,
|
|
{
|
|
"output_dim": 0,
|
|
"weight_loader": weight_loader,
|
|
},
|
|
)
|
|
|
|
wtscale = Parameter(
|
|
torch.ones(1, dtype=torch.bfloat16),
|
|
requires_grad=False,
|
|
)
|
|
set_weight_attrs(wtscale, {"weight_loader": default_weight_loader})
|
|
|
|
layer.register_parameter("qweight", qweight)
|
|
layer.register_parameter("wscales", wscales)
|
|
layer.register_parameter("proj_down", proj_down)
|
|
layer.register_parameter("proj_up", proj_up)
|
|
layer.register_parameter("smooth_factor", smooth_factor)
|
|
layer.register_parameter("wcscales", wcscales)
|
|
layer.register_parameter("wtscale", wtscale)
|
|
|
|
del input_size, output_size
|
|
layer.output_size_per_partition = output_size_per_partition
|
|
|
|
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
|
"""Adapt the canonical row-major checkpoint to vLLM's NVFP4 ABI."""
|
|
qweight = layer.qweight
|
|
wscales = layer.wscales
|
|
del layer.qweight
|
|
del layer.wscales
|
|
layer.register_parameter(
|
|
"weight",
|
|
Parameter(
|
|
qweight.detach().view(torch.uint8),
|
|
requires_grad=False,
|
|
),
|
|
)
|
|
layer.register_parameter(
|
|
"weight_scale",
|
|
Parameter(
|
|
wscales.detach().transpose(0, 1).contiguous(),
|
|
requires_grad=False,
|
|
),
|
|
)
|
|
|
|
layer.register_parameter(
|
|
"input_global_scale_inv",
|
|
Parameter(
|
|
torch.ones(
|
|
1,
|
|
dtype=torch.float32,
|
|
device=layer.weight.device,
|
|
),
|
|
requires_grad=False,
|
|
),
|
|
)
|
|
|
|
wtscale = layer.wtscale.detach().to(dtype=torch.float32)
|
|
del layer.wtscale
|
|
layer.register_parameter(
|
|
"alpha",
|
|
Parameter(wtscale, requires_grad=False),
|
|
)
|
|
|
|
channel_scale = layer.wcscales.detach()
|
|
del layer.wcscales
|
|
if torch.all(channel_scale == 1).item():
|
|
layer.output_channel_scale = None
|
|
else:
|
|
layer.register_parameter(
|
|
"output_channel_scale",
|
|
Parameter(channel_scale, requires_grad=False),
|
|
)
|
|
|
|
_nvfp4_kernel().process_weights_after_loading(layer)
|
|
logger.info_once(
|
|
"SVDQuant NVFP4 is using vLLM's compatibility path; the rank correction is not fused into the GEMM."
|
|
)
|
|
|
|
def apply(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
"""Compute the base NVFP4 GEMM plus the BF16 rank correction."""
|
|
if x.dtype != torch.bfloat16:
|
|
raise ValueError(f"SVDQuant NVFP4 requires BF16 activations; got {x.dtype}")
|
|
|
|
original_shape = x.shape
|
|
x_2d = x.reshape(-1, original_shape[-1]).contiguous()
|
|
|
|
# The residual branch consumes the original activation. Only the
|
|
# four-bit base GEMM consumes the smoothed activation.
|
|
smoothed = x_2d / layer.smooth_factor
|
|
out = _nvfp4_kernel().apply_weights(
|
|
layer=layer,
|
|
x=smoothed,
|
|
bias=None,
|
|
)
|
|
|
|
channel_scale = getattr(layer, "output_channel_scale", None)
|
|
if channel_scale is not None:
|
|
# Fused QKV can store independent Q/K/V outer scales. Apply them
|
|
# explicitly until a vector-alpha GEMM epilogue is available.
|
|
out.mul_(channel_scale)
|
|
|
|
correction_input = torch.mm(x_2d, layer.proj_down)
|
|
out = torch.addmm(
|
|
out,
|
|
correction_input,
|
|
layer.proj_up.transpose(0, 1),
|
|
)
|
|
if bias is not None:
|
|
out.add_(bias)
|
|
return out.reshape(
|
|
*original_shape[:-1],
|
|
layer.output_size_per_partition,
|
|
)
|
|
|
|
|
|
__all__ = [
|
|
"DiffusionSVDQuantConfig",
|
|
"DiffusionSVDQuantLinearMethod",
|
|
]
|