Handle module names from Dynamo compiler in FP8 Quantizer

[sandeep-maddipatla]

Type of Change

• Bug Fix, with No API change.

Description

• The Measure component generates stats with original module names from the model, as opposed to the altered module names used by torch.compile to reference the original model layers in a compiled model.
• We need the quantizer to reference the original module names to be able to use the measured stat dumps correctly. This PR captures the Quantizer change to do this

Expected Behavior & Potential Risk

Current behavior is that is the quantizer searches for the stats with altered module names from compiled model and fails with an exception such as below

Exception: Error - Layer '_orig_mod.x_embedder' was called but was not quantized because no measures were supplied.

With this PR, this error is no longer generated.

How has this PR been tested?

The test script below reproduces the problem. Re-run this with the PR in place to verify there is no error.

# Fetch sources, install dependencies
pip install optimum-habana sentencepiece
git clone https://github.com/huggingface/optimum-habana
cd /path/to/working-dir
cp -r /path/to/optimum-habana/examples/stable-diffusion/quantization .
huggingface-cli login --token YourHFTokenGoesHere

Have below test script in working directory.

import os
import torch
from optimum.habana.diffusers import  GaudiFlowMatchEulerDiscreteScheduler, GaudiFluxPipeline

mode = os.environ.get('MODE', 'quant')

# load model
model_name = "black-forest-labs/FLUX.1-dev"
scheduler = GaudiFlowMatchEulerDiscreteScheduler.from_pretrained(
    model_name,
    subfolder="scheduler"
)
pipe = GaudiFluxPipeline.from_pretrained(
    model_name,
    scheduler=scheduler,
    use_habana=True,
    use_hpu_graphs=False,
    gaudi_config="Habana/stable-diffusion",
    bf16_full_eval=True,
    torch_dtype=torch.bfloat16
)

if mode == 'measure':
    # dump measure stats through INC
    os.environ["QUANT_CONFIG"] = "quantization/flux/measure_config.json"
    pipe(
        prompt="A picture of sks dog in a bucket",
        quant_mode="measure",
    )
    print('Measurement step done')
elif mode == 'quant':
    # quantize with INC (from measured stats)
    os.environ["QUANT_CONFIG"] = "quantization/flux/quantize_config.json"
    pipe.transformer = torch.compile(pipe.transformer, backend="hpu_backend")
    image = pipe(
        prompt="A picture of sks dog in a bucket",
        quant_mode="quantize"
    ).images[0]
    image.save(f"output_image.png")
    print('Quant Step done')
else:
    print(f'Unrecognized setting for MODE={mode}')

Run the two-step quantization with below commands - run measure step first, and then quant.

MODE=measure PT_HPU_LAZY_MODE=0 python reproducer.py
MODE=quant PT_HPU_LAZY_MODE=0 python reproducer.py

Dependency Change?

No library / dependency changes

[skaulintel]

looks good to me.

[xin3he]

I'd like to get comments from Habana team, @ulivne and @linoybu , please take a look~

[sandeep-maddipatla]

Hi @yiliu30 .. Thank you for your response! Is this limitation of INC (expecting an eager model) documented? If so, can you pls point me to it?

I also see that the INC measure counterpart for this script handles the altered layer names from torch.compile - so I assumed this flow was supported.

[yiliu30]

Hi @sandeep-maddipatla, We haven't documented it for now. But the INC FP8 quantization logic is to replace the Torch module with a patched one. For example, we replace torch.Linear with PatchedLinear. When creating the patched module, we copy some attributes from the original one. However, the compile process may modify some of them.

Have you encountered any issues with the recommended flow?

[yiliu30]

Closing this due to no response for a while. Please feel free to reopen it. @sandeep-maddipatla