[Quant] Consolidate GPTQ: rename gptq_marlin.py to auto_gptq.py (#38288)

Signed-off-by: Chengyi Nie <cnie@roblox.com>
Co-authored-by: Chengyi Nie <cnie@roblox.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Chengyi Nie
2026-05-15 08:25:52 +08:00
committed by GitHub
co-authored by Chengyi Nie Cursor
parent 3b6a204789
commit fa2a33b893
21 changed files with 170 additions and 601 deletions
@@ -2,7 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Mapping
from copy import deepcopy
from fractions import Fraction
from types import MappingProxyType
from typing import TYPE_CHECKING
@@ -16,16 +15,14 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
)
if TYPE_CHECKING:
from ..gptq import GPTQConfig
from ..gptq_marlin import GPTQMarlinConfig
from ..auto_gptq import AutoGPTQConfig
else:
GPTQConfig = object
GPTQMarlinConfig = object
AutoGPTQConfig = object
# Match dynamic rules with module name (prefix) and override quantize
# config if module (prefix) matches a rule
def override_config(config: GPTQConfig | GPTQMarlinConfig, prefix: str):
def override_config(config: AutoGPTQConfig, prefix: str):
weight_bits = get_dynamic_override(config, prefix, "bits", config.weight_bits)
if isinstance(weight_bits, int):
config.weight_bits = weight_bits
@@ -36,31 +33,23 @@ def override_config(config: GPTQConfig | GPTQMarlinConfig, prefix: str):
if isinstance(desc_act, bool):
config.desc_act = desc_act
config.pack_factor = Fraction(32, config.weight_bits) # packed into int32
if config.get_name() == "gptq_marlin":
assert isinstance(config, GPTQMarlinConfig)
is_sym = get_dynamic_override(config, prefix, "sym", config.is_sym)
if isinstance(is_sym, bool):
config.is_sym = is_sym
config.pack_factor = 32 // config.weight_bits # packed into int32
assert isinstance(config, AutoGPTQConfig)
is_sym = get_dynamic_override(config, prefix, "sym", config.is_sym)
if isinstance(is_sym, bool):
config.is_sym = is_sym
if (config.weight_bits, config.is_sym) not in config.TYPE_MAP:
raise ValueError(
"Unsupported quantization config: "
f"bits={config.weight_bits}, sym={config.is_sym}"
)
if (config.weight_bits, config.is_sym) not in config.TYPE_MAP:
raise ValueError(
"Unsupported quantization config: "
f"bits={config.weight_bits}, sym={config.is_sym}"
)
config.quant_type = config.TYPE_MAP[(config.weight_bits, config.is_sym)]
elif config.get_name() == "gptq":
assert isinstance(config, GPTQConfig)
if config.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 {config.weight_bits} bits."
)
config.quant_type = config.TYPE_MAP[(config.weight_bits, config.is_sym)]
def get_dynamic_override(
config: GPTQConfig | GPTQMarlinConfig,
config: AutoGPTQConfig,
layer_name: str,
key: str | None = None,
default_value: int | bool | None = None,
@@ -126,7 +115,7 @@ def is_layer_gptq_quantized(
def get_linear_quant_method(
config: GPTQConfig | GPTQMarlinConfig,
config: AutoGPTQConfig,
layer: torch.nn.Module,
prefix: str,
linear_method_cls: type,