Open lzcchl opened 2 weeks ago
My use case and GPU: model: Qwen2-72B-Instruct max_token_len (input+output): 20000 gpus: 4xA100
when I use code from https://github.com/casper-hansen/AutoAWQ/blob/main/docs/examples.md and change parameters in function model.quantize as below:
model = AutoAWQForCausalLM.from_pretrained( model_path, **{"low_cpu_mem_usage": True, "use_cache": False} ) model.quantize( tokenizer, quant_config=quant_config, calib_data=load_my_data(), n_parallel_calib_samples=1, max_calib_samples=128, max_calib_seq_len=20000 )
but It run OOM, and use only one gpu, I set device_map='auto', but It OOM again, how can I change parameters for run well ?
You can't use such a long sequence length because it does not fit in memory.
My use case and GPU:
model: Qwen2-72B-Instruct max_token_len (input+output): 20000 gpus: 4xA100
when I use code from https://github.com/casper-hansen/AutoAWQ/blob/main/docs/examples.md and change parameters in function model.quantize as below:
but It run OOM, and use only one gpu, I set device_map='auto', but It OOM again, how can I change parameters for run well ?