A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 00:45 UTC
[Bugfix] fix moe_wna16 get_quant_method (#12648)
Fix https://github.com/vllm-project/vllm/issues/12647 The `get_quant_method` of `moe_wna16` always return moe method, GPTQ-based linear method or AWQ-based linear method, even when the target module is attention layer. https://github.com/vllm-project/vllm/blob/baeded25699f9f4851843306f27f685c4d4ee7c5/vllm/attention/layer.py#L86-L92 Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
Jinzhen Lin committed
Feb 2, 2025 at 15:29 UTC
baaa2b24da86d63965dbffc34c97c7c4b50288db
1 file changed
+12
-15
vllm/model_executor/layers/quantization/moe_wna16.py
+12
-15
@@ -6,16 +6,13 @@ from vllm.distributed import get_tensor_model_parallel_rank, get_tp_group
6
from vllm.model_executor.layers.fused_moe.layer import (
7
FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported)
8
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
9
-from vllm.model_executor.layers.quantization.awq import (AWQConfig,
10
- AWQLinearMethod)
11
-from vllm.model_executor.layers.quantization.awq_marlin import (
12
- AWQMarlinConfig, AWQMarlinLinearMethod)
9
+from vllm.model_executor.layers.quantization.awq import AWQConfig
10
+from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
11
from vllm.model_executor.layers.quantization.base_config import (
12
QuantizationConfig, QuantizeMethodBase)
15
-from vllm.model_executor.layers.quantization.gptq import (GPTQConfig,
16
- GPTQLinearMethod)
13
+from vllm.model_executor.layers.quantization.gptq import GPTQConfig
14
from vllm.model_executor.layers.quantization.gptq_marlin import (
18
- GPTQMarlinConfig, GPTQMarlinLinearMethod)
15
+ GPTQMarlinConfig)
16
from vllm.model_executor.utils import set_weight_attrs
17
from vllm.platforms import current_platform
18
@@ -131,18 +128,18 @@ class MoeWNA16Config(QuantizationConfig):
128
else:
129
if self.linear_quant_method == "gptq":
130
if self.use_marlin:
134
- return GPTQMarlinLinearMethod(
135
- GPTQMarlinConfig.from_config(self.full_config))
131
+ return GPTQMarlinConfig.from_config(
132
+ self.full_config).get_quant_method(layer, prefix)
133
else:
137
- return GPTQLinearMethod(
138
- GPTQConfig.from_config(self.full_config))
134
+ return GPTQConfig.from_config(
135
+ self.full_config).get_quant_method(layer, prefix)
136
elif self.linear_quant_method == "awq":
137
if self.use_marlin:
141
- return AWQMarlinLinearMethod(
142
- AWQMarlinConfig.from_config(self.full_config))
138
+ return AWQMarlinConfig.from_config(
139
+ self.full_config).get_quant_method(layer, prefix)
140
else:
144
- return AWQLinearMethod(
145
- AWQConfig.from_config(self.full_config))
141
+ return AWQConfig.from_config(
142
+ self.full_config).get_quant_method(layer, prefix)
143
else:
144
raise ValueError("moe_wna16 only support gptq and awq.")
145