[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
+12 -21
View File
@@ -356,13 +356,13 @@ class INCConfig(QuantizationConfig):
else:
use_marlin = False
if use_marlin:
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQMarlinConfig,
GPTQMarlinLinearMethod,
GPTQMarlinMoEMethod,
from vllm.model_executor.layers.quantization.auto_gptq import (
AutoGPTQConfig,
AutoGPTQLinearMethod,
AutoGPTQMoEMethod,
)
quant_args_marlin = GPTQMarlinConfig(
quant_args_marlin = AutoGPTQConfig(
weight_bits=weight_bits,
group_size=group_size,
is_sym=sym,
@@ -371,23 +371,10 @@ class INCConfig(QuantizationConfig):
dynamic={},
full_config={},
)
else:
from vllm.model_executor.layers.quantization.gptq import (
GPTQConfig,
GPTQLinearMethod,
)
quant_args = GPTQConfig(
weight_bits=weight_bits,
group_size=group_size,
lm_head_quantized=False,
desc_act=False,
dynamic={},
)
if isinstance(layer, RoutedExperts):
if use_marlin:
return GPTQMarlinMoEMethod(quant_args_marlin, layer.moe_config)
return AutoGPTQMoEMethod(quant_args_marlin, layer.moe_config)
else:
from vllm.model_executor.layers.quantization.moe_wna16 import (
MoeWNA16Config,
@@ -406,9 +393,13 @@ class INCConfig(QuantizationConfig):
if isinstance(layer, (LinearBase, ParallelLMHead)):
if use_marlin:
return GPTQMarlinLinearMethod(quant_args_marlin)
return AutoGPTQLinearMethod(quant_args_marlin)
else:
return GPTQLinearMethod(quant_args)
raise NotImplementedError(
f"INC quantization with bits={weight_bits}, sym={sym} "
"is not supported. Only 4-bit and 8-bit symmetric "
"quantization is supported with Marlin kernels."
)
return None