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, 05:45 UTC
[bugfix] fix the default value of llm_int8_threshold in BitsAndBytesConfig (#10657)
yansh97 committed
Nov 27, 2024 at 13:55 UTC
cfb3bf25fb981494fa6c575fb0714388c9df99b0
1 file changed
+2
-2
vllm/model_executor/layers/quantization/bitsandbytes.py
+2
-2
@@ -26,7 +26,7 @@ class BitsAndBytesConfig(QuantizationConfig):
26
llm_int8_enable_fp32_cpu_offload: bool = False,
27
llm_int8_has_fp16_weight: bool = False,
28
llm_int8_skip_modules: Optional[List[str]] = None,
29
- llm_int8_threshold: float = 0.0,
29
+ llm_int8_threshold: float = 6.0,
30
) -> None:
31
32
self.load_in_8bit = load_in_8bit
@@ -103,7 +103,7 @@ class BitsAndBytesConfig(QuantizationConfig):
103
["llm_int8_skip_modules"],
104
default_value=[])
105
llm_int8_threshold = get_safe_value(config, ["llm_int8_threshold"],
106
- default_value=0.0)
106
+ default_value=6.0)
107
108
return cls(
109
load_in_8bit=load_in_8bit,