Adding torchao apis to gpt-fast

[HDCharles]

Summary:

adding torchao apis to gpt-fast and some minor tweaks

Test Plan:

(in progress)
export MODEL_REPO=meta-llama/Meta-Llama-3-8B

python quantize.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --mode torchao-int8 python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int8.pth --compile python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int8.pth python eval.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int8.pth --tasks wikitext --compile

For model checkpoints/meta-llama/Meta-Llama-3-8B/model_torchao-int8.pth
wikitext: {'word_perplexity,none': 7.900496793735154, 'word_perplexity_stderr,none': 'N/A', 'byte_perplexity,none': 1.4718578218273202, 'byte_perplexity_stderr,none': 'N/A', 'bits_per_byte,none': 0.5576383170121927, 'bits_per_byte_stderr,none': 'N/A', 'alias': 'wikitext'}

python quantize.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --mode torchao-int4-hqq python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4-hqq.pth --compile python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4-hqq.pth python eval.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4-hqq.pth --tasks wikitext --compile

For model checkpoints/meta-llama/Meta-Llama-3-8B/model_torchao-int4-hqq.pth
wikitext: {'word_perplexity,none': 8.44187872159186, 'word_perplexity_stderr,none': 'N/A', 'byte_perplexity,none': 1.4902143610748824, 'byte_perplexity_stderr,none': 'N/A', 'bits_per_byte,none': 0.575519871235033, 'bits_per_byte_stderr,none': 'N/A', 'alias': 'wikitext'}

python quantize.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --mode torchao-int4 python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4.pth --compile python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4.pth python eval.py --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4.pth --tasks wikitext --compile

For model checkpoints/meta-llama/Meta-Llama-3-8B/model_torchao-int4.pth
wikitext: {'word_perplexity,none': 8.59031159441983, 'word_perplexity_stderr,none': 'N/A', 'byte_perplexity,none': 1.4950796712267396, 'byte_perplexity_stderr,none': 'N/A', 'bits_per_byte,none': 0.5802223661766339, 'bits_per_byte_stderr,none': 'N/A', 'alias': 'wikitext'}

Reviewers:

Subscribers:

Tasks:

Tags:

[HDCharles]

@Chillee should i add info to the README or implement this differently?

[gchlebus]

Is the plan for this PR also to add fp8 support which is available in torchao?

[jerryzh168]

it seems that this one does not work with tp yet:

ENABLE_INTRA_NODE_COMM=1 torchrun --standalone --nproc_per_node=2 generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/model_torchao-int4.pth

[rank1]: NotImplementedError: AffineQuantizedTensor dispatch: attempting to run unimplemented operator/function: func=<OpOverload(op='aten.split_with_sizes', overload='default')>, types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>,), arg_types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>, <class 'list'>), kwarg_types={}

will need to implement this op in AQT to support this or change the tp implementation to DTensor I guess.

[meta-cla]

Hi @HDCharles!

Thank you for your pull request.

We require contributors to sign our Contributor License Agreement, and yours needs attention.

You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!