mosaicml

/

mpt-7b-storywriter

like 822

Follow

Mosaic ML, Inc. 351

In the model cards we explain how to take advantage of ALiBi so that you can increase maximum sequence length during inference:

Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:

config = transformers.AutoConfig.from_pretrained(
  'mosaicml/mpt-7b-instruct',
  trust_remote_code=True
)
config.update({"max_seq_len": 4096})
model = transformers.AutoModelForCausalLM.from_pretrained(
  'mosaicml/mpt-7b-instruct',
  config=config,
  trust_remote_code=True
)

Many thanks @jacobfulano .

Is it possible to directly deploy with that configuration using Sagemaker, or would I have to take a route like building a docker image including the config update? Thanks

HI @RonanMcGovern , I'm not sure what the limitations of Sagemaker are, but as long as you can pass in a custom HF config, you can set config.max_seq_len=4096 dynamically at start time. It's just a config arg, no need for custom Docker images or anything.

Hi @abhi-mosaic I've managed to get MPT-7B running well on Google collab with:

    # use cache directory while loading model and tokenizer
    self.model = AutoModelForCausalLM.from_pretrained(
        model_name,
        cache_dir=cache_dir,
        torch_dtype=torch_dtype,
        trust_remote_code=trust_remote_code,
        use_auth_token=use_auth_token,
    )

I'm then trying to add in the config for max_seq_len but not getting any joy:

        # Load the configuration
        config = transformers.AutoConfig.from_pretrained(model_name,
                                                        trust_remote_code=trust_remote_code)
        # Explicitly set the max_seq_len
        config.max_seq_len = 4096

        # Load the model with the updated configuration
        self.model = transformers.AutoModelForCausalLM.from_pretrained(
            model_name,
            config=config,
            cache_dir=cache_dir,
            torch_dtype=torch_dtype,
            trust_remote_code=trust_remote_code,
            use_auth_token=use_auth_token,
        )