@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 7576cd38df
Fork 0

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] Check bnb_4bit_quant_storage for bitsandbytes (#10642)

Michael Goin committed Nov 26, 2024 at 15:29 UTC 7576cd38dfdf1672d04f4fe659f8260a9d319e8b
1 file changed +11
vllm/model_executor/layers/quantization/bitsandbytes.py
+11
@@ -20,6 +20,7 @@ class BitsAndBytesConfig(QuantizationConfig):
20 load_in_8bit: bool = False,
21 load_in_4bit: bool = True,
22 bnb_4bit_compute_dtype: str = "float32",
23 + bnb_4bit_quant_storage: str = "uint8",
24 bnb_4bit_quant_type: str = "fp4",
25 bnb_4bit_use_double_quant: bool = False,
26 llm_int8_enable_fp32_cpu_offload: bool = False,
@@ -31,6 +32,7 @@ class BitsAndBytesConfig(QuantizationConfig):
32 self.load_in_8bit = load_in_8bit
33 self.load_in_4bit = load_in_4bit
34 self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype
35 + self.bnb_4bit_quant_storage = bnb_4bit_quant_storage
36 self.bnb_4bit_quant_type = bnb_4bit_quant_type
37 self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant
38 self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload
@@ -38,10 +40,15 @@ class BitsAndBytesConfig(QuantizationConfig):
40 self.llm_int8_skip_modules = llm_int8_skip_modules or []
41 self.llm_int8_threshold = llm_int8_threshold
42
43 + if self.bnb_4bit_quant_storage not in ["uint8"]:
44 + raise ValueError("Unsupported bnb_4bit_quant_storage: "
45 + f"{self.bnb_4bit_quant_storage}")
46 +
47 def __repr__(self) -> str:
48 return (f"BitsAndBytesConfig(load_in_8bit={self.load_in_8bit}, "
49 f"load_in_4bit={self.load_in_4bit}, "
50 f"bnb_4bit_compute_dtype={self.bnb_4bit_compute_dtype}, "
51 + f"bnb_4bit_quant_storage={self.bnb_4bit_quant_storage}, "
52 f"bnb_4bit_quant_type={self.bnb_4bit_quant_type}, "
53 f"llm_int8_skip_modules={self.llm_int8_skip_modules})")
54
@@ -80,6 +87,9 @@ class BitsAndBytesConfig(QuantizationConfig):
87 bnb_4bit_compute_dtype = get_safe_value(config,
88 ["bnb_4bit_compute_dtype"],
89 default_value="float32")
90 + bnb_4bit_quant_storage = get_safe_value(config,
91 + ["bnb_4bit_quant_storage"],
92 + default_value="uint8")
93 bnb_4bit_quant_type = get_safe_value(config, ["bnb_4bit_quant_type"],
94 default_value="fp4")
95 bnb_4bit_use_double_quant = get_safe_value(
@@ -99,6 +109,7 @@ class BitsAndBytesConfig(QuantizationConfig):
109 load_in_8bit=load_in_8bit,
110 load_in_4bit=load_in_4bit,
111 bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
112 + bnb_4bit_quant_storage=bnb_4bit_quant_storage,
113 bnb_4bit_quant_type=bnb_4bit_quant_type,
114 bnb_4bit_use_double_quant=bnb_4bit_use_double_quant,
115 llm_int8_enable_fp32_cpu_offload=llm_int8_enable_fp32_cpu_offload,