EngineX-Mthreads/enginex-mthreads-vllm
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Sync from v0.13

Browse Source
This commit is contained in:
xiezhongtao
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions
Download Patch File Download Diff File Expand all files Collapse all files

329
vllm/model_executor/layers/quantization/gptq.py
View File

@@ -1,16 +1,43 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import enum
from enum import Enum
from fractions import Fraction
from typing import Any, Dict, List, Optional
from typing import TYPE_CHECKING, Any, Union
import torch
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
from torch.nn.parameter import Parameter
from vllm import _custom_ops as ops
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe.layer import FusedMoE
from vllm.model_executor.layers.linear import LinearMethodBase
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.utils import set_weight_attrs
QuantizationConfig,
QuantizeMethodBase,
)
from vllm.model_executor.layers.quantization.utils.gptq_utils import (
get_linear_quant_method,
)
from vllm.model_executor.parameter import (
ChannelQuantScaleParameter,
GroupQuantScaleParameter,
PackedColumnParameter,
PackedvLLMParameter,
RowvLLMParameter,
)
from vllm.transformers_utils.config import get_safetensors_params_metadata
from vllm.utils.collection_utils import is_list_of
if TYPE_CHECKING:
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.model_executor.models.utils import WeightsMapper
else:
QuantizationMethods = str
logger = init_logger(__name__)
class GPTQConfig(QuantizationConfig):
@@ -24,27 +51,83 @@ class GPTQConfig(QuantizationConfig):
weight_bits: int,
group_size: int,
desc_act: bool,
lm_head_quantized: bool,
dynamic: dict[str, dict[str, int | bool]],
autoround_version: str = "",
modules_in_block_to_quantize: list[str] | None = None,
checkpoint_format: str = "",
) -> None:
# GPTQModel use `dynamic` config property to allow per module
# quantization config so each module can be individually optimized.
# Format is dict[str, dict] where key is a regex string that can
# perform both positive ("+:" prefixed) or negative ("-:" prefixed)
# matching of a module.
# Default to positive match, override base quant config mode, if no
# prefix is used. Value is in dict format of field key and override
# value.
# Negative matching will skip quantization init for this module
# entirely:
# non-quantized inference. More details and quantization examples can be
# found at: https://github.com/ModelCloud/GPTQModel
# Example:
# # last 1/2 of the layers 10-21 has 8bit vs 4bit for 0-9
# # last 1/4 of the layers 16-21 has 8bit and group_size 64
# dynamic = {
# #`.*\.` matches the layers_node prefix
# # positive match layer 10-15
# r"+:.*\.(?:1[0-5])\..*": {"bits": 8,},
# # positive match layer 16-21
# r"+:.*\.(?:1[6-9]|20|21)\..*": {"bits": 8, "group_size": 64,},
# r"-:.*\.moe\..*": {}, # negative match (skip) all `moe` layers
# }
super().__init__()
self.dynamic = dynamic
self.weight_bits = weight_bits
self.group_size = group_size
self.desc_act = desc_act
self.lm_head_quantized = lm_head_quantized
self.pack_factor = Fraction(32, self.weight_bits)
if self.weight_bits not in [2, 3, 4, 8]:
raise ValueError(
"Currently, only 2/3/4/8-bit weight quantization is "
f"supported for GPTQ, but got {self.weight_bits} bits.")
f"supported for GPTQ, but got {self.weight_bits} bits."
)
# Somehow gptq_gemm 4-bit is buggy, maybe fix it in the future.
# For now, show a warning, since gptq_marlin will be used by default.
if self.weight_bits == 4:
logger.warning_once(
"Currently, the 4-bit gptq_gemm kernel for GPTQ is buggy. "
"Please switch to gptq_marlin or gptq_bitblas."
)
self.modules_in_block_to_quantize = modules_in_block_to_quantize or []
# used to identify GPTQ model quantized by autoround
self.autoround_version = autoround_version
# GPTQ v1 and v2 format deals with zero points differently.
# Currently GPTQModel stores v1 format checkpoints by default,
# but provides the option to set `format="gptq_v2"` in `QuantizeConfig`.
self.checkpoint_format = checkpoint_format
def __repr__(self) -> str:
return (f"GPTQConfig(weight_bits={self.weight_bits}, "
f"group_size={self.group_size}, "
f"desc_act={self.desc_act})")
return (
f"GPTQConfig(weight_bits={self.weight_bits}, "
f"group_size={self.group_size}, "
f"desc_act={self.desc_act}), "
f"lm_head_quantized={self.lm_head_quantized}, "
f"dynamic={self.dynamic}, "
f"modules_in_block_to_quantize={self.modules_in_block_to_quantize}), "
f"checkpoint_format={self.checkpoint_format})"
)
@classmethod
def get_name(cls) -> str:
def get_name(cls) -> QuantizationMethods:
return "gptq"
@classmethod
def get_supported_act_dtypes(cls) -> List[torch.dtype]:
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
return [torch.half]
@classmethod
@@ -53,28 +136,87 @@ class GPTQConfig(QuantizationConfig):
return 60
@classmethod
def get_config_filenames(cls) -> List[str]:
def get_config_filenames(cls) -> list[str]:
return ["quantize_config.json"]
@classmethod
def from_config(cls, config: Dict[str, Any]) -> "GPTQConfig":
def from_config(cls, config: dict[str, Any]) -> "GPTQConfig":
dynamic = cls.get_from_keys_or(config, ["dynamic"], default={})
dynamic = {} if dynamic is None else dynamic
weight_bits = cls.get_from_keys(config, ["bits"])
group_size = cls.get_from_keys(config, ["group_size"])
desc_act = cls.get_from_keys(config, ["desc_act"])
return cls(weight_bits, group_size, desc_act)
lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
autoround_version = cls.get_from_keys_or(
config, ["autoround_version"], default=""
)
modules_in_block_to_quantize = cls.get_from_keys_or(
config, ["modules_in_block_to_quantize"], default=None
)
checkpoint_format = cls.get_from_keys_or(
config, ["checkpoint_format"], default=""
)
return cls(
weight_bits,
group_size,
desc_act,
lm_head_quantized,
dynamic,
autoround_version,
modules_in_block_to_quantize,
checkpoint_format,
)
def get_quant_method(
self, layer: torch.nn.Module) -> Optional["GPTQLinearMethod"]:
if isinstance(layer, LinearBase):
return GPTQLinearMethod(self)
return None
self, layer: torch.nn.Module, prefix: str
) -> Union["GPTQLinearMethod", "QuantizeMethodBase"] | None:
if isinstance(layer, FusedMoE):
# GPTQ MoE support: fall back to MoeWNA16 for broad compatibility
from .moe_wna16 import MoeWNA16Config
def get_scaled_act_names(self) -> List[str]:
return []
# TODO: maybe update this for GPTQv2 format checkpoints
config = {
"quant_method": "gptq",
"bits": self.weight_bits,
"group_size": self.group_size,
"sym": True, # GPTQ typically uses symmetric quantization
"lm_head": False,
}
return MoeWNA16Config.from_config(config).get_quant_method(layer, prefix)
return get_linear_quant_method(self, layer, prefix, GPTQLinearMethod)
def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):
if self.modules_in_block_to_quantize is not None:
self.modules_in_block_to_quantize = hf_to_vllm_mapper.apply_list(
self.modules_in_block_to_quantize
)
def maybe_update_config(self, model_name: str, revision: str | None = None):
if self.modules_in_block_to_quantize:
if is_list_of(self.modules_in_block_to_quantize, list):
# original modules_in_block_to_quantize: list[list[str]]
# flatten original modules_in_block_to_quantize
self.modules_in_block_to_quantize = [
item
for sublist in self.modules_in_block_to_quantize
for item in sublist
]
return
unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32]
metadata = get_safetensors_params_metadata(model_name, revision=revision)
quant_layers: set[str] = {
param_name.rsplit(".", 1)[0]
for param_name, info in metadata.items()
if (dtype := info.get("dtype", None))
and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes
}
self.modules_in_block_to_quantize = list(quant_layers)
class ExllamaState(Enum):
UNUSED = enum.auto()
UNINITIALIZED = enum.auto()
READY = enum.auto()
@@ -90,29 +232,34 @@ class GPTQLinearMethod(LinearMethodBase):
def __init__(self, quant_config: GPTQConfig):
self.quant_config = quant_config
# GPTQ v1 and v2 format deals with zero points differently
self.use_v2_format = quant_config.checkpoint_format == "gptq_v2"
def create_weights(
self,
layer: torch.nn.Module,
input_size_per_partition: int,
output_partition_sizes: List[int],
output_partition_sizes: list[int],
input_size: int,
output_size: int,
params_dtype: torch.dtype,
**extra_weight_attrs,
):
del output_size # Unused.
weight_loader = extra_weight_attrs.get("weight_loader")
if input_size_per_partition % self.quant_config.group_size != 0:
raise ValueError(
"The input size is not aligned with the quantized "
"weight shape. This can be caused by too large "
"tensor parallel size.")
"tensor parallel size."
)
output_size_per_partition = sum(output_partition_sizes)
if (output_size_per_partition % self.quant_config.pack_factor.numerator
!= 0):
if output_size_per_partition % self.quant_config.pack_factor.numerator != 0:
raise ValueError(
"The output size is not aligned with the quantized "
"weight shape. This can be caused by too large "
"tensor parallel size.")
"tensor parallel size."
)
if self.quant_config.group_size != -1:
group_size = self.quant_config.group_size
@@ -121,8 +268,10 @@ class GPTQLinearMethod(LinearMethodBase):
exllama_state = ExllamaState.UNINITIALIZED
scale_and_zero_size = input_size // group_size
scale_and_zero_input_dim = None
if (input_size != input_size_per_partition
and self.quant_config.group_size != -1):
if (
input_size != input_size_per_partition
and self.quant_config.group_size != -1
):
# For act-order models, we cannot use Exllama for row parallel layer
if self.quant_config.desc_act:
exllama_state = ExllamaState.UNUSED
@@ -131,94 +280,114 @@ class GPTQLinearMethod(LinearMethodBase):
scale_and_zero_size = input_size_per_partition // group_size
scale_and_zero_input_dim = 0
qweight = Parameter(
torch.empty(
qweight = PackedvLLMParameter(
data=torch.empty(
input_size_per_partition // self.quant_config.pack_factor,
output_size_per_partition,
dtype=torch.int32,
),
requires_grad=False,
input_dim=0,
output_dim=1,
packed_dim=0,
packed_factor=self.quant_config.pack_factor,
weight_loader=weight_loader,
)
set_weight_attrs(
qweight, {
"input_dim": 0,
"output_dim": 1,
"packed_dim": 0,
"pack_factor": self.quant_config.pack_factor,
})
g_idx = Parameter(
torch.tensor(
g_idx = RowvLLMParameter(
data=torch.tensor(
[
i // self.quant_config.group_size
for i in range(input_size_per_partition)
],
dtype=torch.int32,
),
requires_grad=False,
input_dim=0,
weight_loader=weight_loader,
)
# Ignore warning from fused linear layers such as QKVParallelLinear.
set_weight_attrs(g_idx, {"input_dim": 0, "ignore_warning": True})
qzeros = Parameter(
torch.empty(
qzeros_args = {
"data": torch.empty(
scale_and_zero_size,
output_size_per_partition // self.quant_config.pack_factor,
dtype=torch.int32,
),
requires_grad=False,
)
set_weight_attrs(
qzeros, {
"input_dim": scale_and_zero_input_dim,
"output_dim": 1,
"packed_dim": 1,
"pack_factor": self.quant_config.pack_factor,
})
scales = Parameter(
torch.empty(
"weight_loader": weight_loader,
}
weight_scale_args = {
"data": torch.empty(
scale_and_zero_size,
output_size_per_partition,
dtype=params_dtype,
),
requires_grad=False,
)
set_weight_attrs(scales, {
"input_dim": scale_and_zero_input_dim,
"output_dim": 1,
})
"weight_loader": weight_loader,
}
if scale_and_zero_input_dim is None:
scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
qzeros = PackedColumnParameter(
output_dim=1,
packed_dim=1,
packed_factor=self.quant_config.pack_factor,
**qzeros_args,
)
else:
scales = GroupQuantScaleParameter(
output_dim=1, input_dim=0, **weight_scale_args
)
qzeros = PackedvLLMParameter(
input_dim=0,
output_dim=1,
packed_dim=1,
packed_factor=self.quant_config.pack_factor,
**qzeros_args,
)
layer.register_parameter("qweight", qweight)
set_weight_attrs(qweight, extra_weight_attrs)
layer.register_parameter("g_idx", g_idx)
set_weight_attrs(g_idx, extra_weight_attrs)
layer.register_parameter("qzeros", qzeros)
set_weight_attrs(qzeros, extra_weight_attrs)
layer.register_parameter("scales", scales)
set_weight_attrs(scales, extra_weight_attrs)
layer.exllama_state = exllama_state
def apply(self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
qweight = layer.qweight
out_shape = x.shape[:-1] + (qweight.shape[-1], )
reshaped_x = x.reshape(-1, x.shape[-1])
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
# for torch.compile
layer.qzeros = Parameter(layer.qzeros.data, requires_grad=False)
layer.qweight = Parameter(layer.qweight.data, requires_grad=False)
layer.g_idx = Parameter(layer.g_idx.data, requires_grad=False)
layer.scales = Parameter(layer.scales.data, requires_grad=False)
# exllama needs to shuffle the weight after the weight is loaded
# here we do the shuffle on first forward pass
if layer.exllama_state == ExllamaState.UNINITIALIZED:
if self.quant_config.desc_act:
layer.g_idx.data = torch.argsort(layer.g_idx).to(torch.int)
else:
layer.g_idx.data = torch.empty((0, ),
device=layer.g_idx.device)
layer.g_idx.data = torch.empty(
(0,), dtype=torch.int, device=layer.g_idx.device
)
layer.exllama_state = ExllamaState.READY
ops.gptq_shuffle(layer.qweight, layer.g_idx,
self.quant_config.weight_bits)
output = ops.gptq_gemm(reshaped_x, layer.qweight, layer.qzeros,
layer.scales, layer.g_idx,
layer.exllama_state == ExllamaState.READY,
self.quant_config.weight_bits)
ops.gptq_shuffle(layer.qweight, layer.g_idx, self.quant_config.weight_bits)
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
out_shape = x.shape[:-1] + (layer.qweight.shape[-1],)
reshaped_x = x.reshape(-1, x.shape[-1])
# GPTQ v1 and v2 format checkpoints deals with zero points differently,
# and require different gemm kernels.
output = ops.gptq_gemm(
reshaped_x,
layer.qweight,
layer.qzeros,
layer.scales,
layer.g_idx,
layer.exllama_state == ExllamaState.READY,
self.use_v2_format,
self.quant_config.weight_bits,
)
if bias is not None:
output.add_(bias)
return output.reshape(out_shape)