(
model : MetaModel,
quant_conf : BitsAndBytesConfig,
)
| 93 | return output |
| 94 | |
| 95 | def quantize( |
| 96 | model : MetaModel, |
| 97 | quant_conf : BitsAndBytesConfig, |
| 98 | ): |
| 99 | module_list = [_ for _ in model.named_modules() if isinstance(_[1], |
| 100 | (LoraColumnParallelLinear, LoraRowParallelLinear, |
| 101 | ColumnParallelLinear, RowParallelLinear, torch.nn.Linear))] |
| 102 | quant_blocklist = model.get_quant_blocklist() |
| 103 | |
| 104 | for name, module in tqdm(module_list, desc="Qunatization Process"): |
| 105 | if "lora" in name or name in quant_blocklist: |
| 106 | continue |
| 107 | if isinstance(module, ( |
| 108 | LoraColumnParallelLinear, |
| 109 | LoraRowParallelLinear, |
| 110 | ColumnParallelLinear, |
| 111 | RowParallelLinear, |
| 112 | torch.nn.Linear |
| 113 | )): |
| 114 | # 1. Initialize quantization operator |
| 115 | if quant_conf.load_in_4bit: |
| 116 | quanted_layer = bnb.nn.Linear4bit( |
| 117 | module.in_features, |
| 118 | module.out_features, |
| 119 | bias=None, |
| 120 | compute_dtype=quant_conf.bnb_4bit_compute_dtype, |
| 121 | compress_statistics=True, |
| 122 | device=None) |
| 123 | if quant_conf.bnb_4bit_compute_dtype != None: |
| 124 | quanted_layer.compute_type_is_set = True |
| 125 | |
| 126 | quanted_layer.weight = bnb.nn.Params4bit( |
| 127 | module.weight.data.clone(), |
| 128 | requires_grad=False, |
| 129 | quant_type=quant_conf.bnb_4bit_quant_type, |
| 130 | ) |
| 131 | |
| 132 | elif quant_conf.load_in_8bit: |
| 133 | quanted_layer= bnb.nn.Linear8bitLt( |
| 134 | module.in_features, |
| 135 | module.out_features, |
| 136 | bias=None, |
| 137 | has_fp16_weights=quant_conf.llm_int8_has_fp16_weight, |
| 138 | threshold=quant_conf.llm_int8_threshold, |
| 139 | ) |
| 140 | quanted_layer.weight = bnb.nn.Int8Params( |
| 141 | module.weight.data.clone(), |
| 142 | requires_grad=False, |
| 143 | #has_fp16_weights=quant_conf.llm_int8_has_fp16_weight, |
| 144 | ) |
| 145 | else: |
| 146 | raise NotImplementedError(f'Please determine the proper quantization type.') |
| 147 | |
| 148 | # 2. Convert FP layer to quantized layer |
| 149 | module.quanted_layer = quanted_layer |
| 150 | |
| 151 | if isinstance(module, LoraColumnParallelLinear): |
| 152 | forward_func = forward_LoraColumnParallelLinear |
no test coverage detected