Add BlackForest Flux Support

[JingyaHuang]

What does this PR do?

Fixes #763, #676

Compilation

• Export of Flux pipeline (TP=8 for Flux transformer 2D) with neuronx_distributed.trace.model_builder.ModelBuilder

Export via CLI

• Regular

optimum-cli export neuron --model black-forest-labs/FLUX.1-dev --tensor_parallel_size 8 --batch_size 1 --height 768 --width 1360 --num_images_per_prompt 1 --torch_dtype bfloat16 flux_neuron/

• Tiny test

optimum-cli export neuron --model hf-internal-testing/tiny-flux-pipe-gated-silu --tensor_parallel_size 2 --batch_size 1 --height 8 --width 8 --num_images_per_prompt 1 --sequence_length 256 --torch_dtype bfloat16 tiny_flux_neuron/

Export with NeuronFluxPipeline API

from optimum.neuron import NeuronFluxPipeline

if __name__ == "__main__":
    compiler_args = {"auto_cast": "none"}
    input_shapes = {"batch_size": 1, "height": 1024, "width": 1024}

    pipe = NeuronFluxPipeline.from_pretrained(
        "black-forest-labs/FLUX.1-dev",
        torch_dtype=torch.bfloat16,
        export=True,
        tensor_parallel_size=8,
        # disable_neuron_cache=True,
        **compiler_args,
        **input_shapes
    )

    # Save locally
    pipe.save_pretrained("flux_dev_neuron_1024_tp8/")

    # Upload to the HuggingFace Hub
    pipe.push_to_hub(
        "flux_dev_neuron_1024_tp8/", repository_id="Jingya/FLUX.1-dev-neuronx-1024x1024-tp8"  # Replace with your HF Hub repo id
    )

You can find an example of compiled artifacts here (Jingya/flux.1-dev_neuronx_tp8 )

Inference

• Flux Inference

For generating an image with NeuronFluxPipeline:

from optimum.neuron import NeuronFluxPipeline

pipe = NeuronFluxPipeline.from_pretrained("flux_neuron")
prompt = "A cat holding a sign that says hello world"
out = pipe(
    prompt,
    guidance_scale=3.5,
    num_inference_steps=50,
    generator=torch.Generator("cpu").manual_seed(0)
).images[0]
out.save("flux_optimum.png")

Other

• Tests
• Doc

cc. @yahavb

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

This PR is stale because it has been open 15 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[github-actions]

This PR is stale because it has been open 15 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[dacorvo]

Massive pull-request: impressive !!!! I only have a few minor optional comments, and I think one test should be renamed.
Eventually, we should think about how some of the ModelBuilder related code can be reused for non-decoder models: it is unclear to me yet if there are some specifics that would prevent that though.

[dacorvo]

Eventually this could be omitted: the weights can be sharded at loading time only, which makes export a lot faster. You only need to remember where the weights are (local dir or hub).

[JingyaHuang]

Do you mean we shard the weights during the loading time instead of after the tracing? Indeed, it increases the whole export time, but somehow, I would rather spend more time exporting one shot and have faster loading + warmup during the deployment.

[dacorvo]

Maybe in your case it is not that bad, because sharding does not mean loading the weights on the device, does it ?
It is just that you don't cache the sharded weights, so it is only useful when you want to push the exported model to the hub.
This is what I do for decoders:

• when using the optimum-cli: export (cache or fetch NEFFs)
• when using from_pretrained: export (cache or fetch NEFFs) + load_weights
  For large models like llama 70B this makes a huge difference as loading weights takes several minutes.

[JingyaHuang]

na na, shard_checkpoint just shard the weights, then we either serialize things to disk like I do here or load to Neuron.

[JingyaHuang]

Yeah, I should definitly look into the neff cache part of ModelBuilder (not yet the case) and the work you have done that I could reuse. will do it next!

[tengomucho]

Few nits, otherwise LGTM!

[tengomucho]

why 24x if we only do TP8?

[JingyaHuang]

Because inf2 jump from either you want to have just 1 neuron device, or you want to have 6

[tengomucho]

oh, right

[dacorvo]

LGTM, thanks !