TheBloke

/

Llama-2-70B-Chat-GPTQ

like 260

Yes, that will work. Recommended to use ExLlama for maximum performance. You need to load less of the model on GPU1 - a recommended split is 17.2GB on GPU1, 24GB on GPU 2. This leaves room for context on GPU1.

@TheBloke how to spread workload to multiple GPU? Default example is:

model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,
        model_basename=model_basename,
        inject_fused_attention=False, # Required for Llama 2 70B model at this time.
        use_safetensors=True,
        trust_remote_code=False,
        device="cuda:0",
        use_triton=use_triton,
        quantize_config=None)

how to define this spilt?

Yes, that will work. Recommended to use ExLlama for maximum performance. You need to load less of the model on GPU1 - a recommended split is 17.2GB on GPU1, 24GB on GPU 2. This leaves room for context on GPU1.

This is probably a dumb question, but using ExLlama or ExLlama HF isn't enough to run this on a 4090, is it?
Maybe if I can split it with my 11900k, but I don't know how to do that.

@TheBloke can you please help with this?

@neo-benjamin
add max_memory parameter.
reference: https://huggingface.co./TheBloke/Llama-2-70B-GPTQ/discussions/9