[Quantization] Add TRT-ModelOpt as a Backend

[ishan-modi]

What does this PR do?

WIP, aimed at adding new backend for quantization #11032. For now, this PR just works for on-the-fly quantization. Loading pre-quantized models errors out and it is to be fixed by NVIDIA team in next release early may

Depends on

• this to support latest diffusers
• this to enable INT8 quantization
• this to enable NF4 quantization

Code

# !pip install "git+https://github.com/ishan-modi/diffusers.git@add-trtquant-backend#egg=diffusers[nvidia_modelopt]"importtorchfromtqdmimporttqdmfromdiffusersimportSanaPipeline, SanaTransformer2DModel, StableDiffusion3Pipeline, , SD3Transformer2DModelfromdiffusers.quantizers.quantization_configimportNVIDIAModelOptConfigcheckpoint="Efficient-Large-Model/Sana_600M_1024px_diffusers"model_cls=SanaTransformer2DModelpipe_cls=SanaPipeline# checkpoint = "stabilityai/stable-diffusion-3-medium-diffusers"# model_cls = SD3Transformer2DModel# pipe_cls = StableDiffusion3Pipelineinput= {"prompt":"A capybara holding a sign that reads Hello World", "num_inference_steps":28, "guidance_scale":3.5}
quant_config_fp8= {"quant_type": "FP8", "quant_method": "modelopt"}
quant_config_int8= {"quant_type": "INT8", "quant_method": "modelopt"}
quant_config_int4= {"quant_type": "INT4", "quant_method": "modelopt", "block_quantize": 128, "channel_quantize": -1, "modules_to_not_convert": ["conv", "patch_embed"]}
quant_config_nf4= {"quant_type": "NF4", "quant_method": "modelopt", "block_quantize": 128, "channel_quantize": -1, "scale_block_quantize": 8, "scale_channel_quantize": -1, "modules_to_not_convert": ["conv"]}
quant_config_nvfp4= {"quant_type": "NVFP4", "quant_method": "modelopt", "block_quantize": 128, "channel_quantize": -1, "modules_to_not_convert": ["conv"]}
deftest_quantization(config, checkpoint, model_cls):
quant_config=NVIDIAModelOptConfig(**config)
print(quant_config.get_config_from_quant_type())
quant_model=model_cls.from_pretrained(checkpoint, subfolder="transformer", quantization_config=quant_config, torch_dtype=torch.bfloat16).to('cuda')
returnquant_modeldeftest_quant_inference(model, input, pipe_cls, iter=1):
inference_memory=0for_intqdm(range(iter)):
withtorch.no_grad():
output=pipe_cls.from_pretrained(checkpoint, transformer=model, torch_dtype=torch.bfloat16).to('cuda')(**input).images[0]
inference_memory+=torch.cuda.max_memory_allocated()
inference_memory/=iteroutput.save("test.png")
print("Inference Memory: ", inference_memory/1e6)
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
test_quant_inference(test_quantization(quant_config_fp8, checkpoint, model_cls), input, pipe_cls)
# test_quant_inference(test_quantization(quant_config_int8, checkpoint, model_cls), input, pipe_cls)# test_quant_inference(test_quantization(quant_config_int4, checkpoint, model_cls), input, pipe_cls)# test_quant_inference(test_quantization(quant_config_nf4, checkpoint, model_cls), input, pipe_cls)# test_quant_inference(test_quantization(quant_config_nvfp4, checkpoint, model_cls), input, pipe_cls)# test_quant_inference(model_cls.from_pretrained(checkpoint, subfolder="transformer", torch_dtype=torch.bfloat16).to('cuda'), input, pipe_cls)

Following is a discussion on speedups while using real_quant with NVIDIA team here

[ishan-modi]

@sayakpaul, would you mind giving a quick look and suggestions

[sayakpaul]

Thanks for getting started on this. I guess there is a problem here: NVIDIA/Model-Optimizer#165? Additionally, the API should have a TRTConfig in place of just a dict being the quantization config.

[ishan-modi]

I think the problem has been fixed the newest release, I just need to bump it up in diffusers requirements, also we can do the following for passing Config class

from diffusers.quantizers.quantization_config import ModelOptConfig
quant_config = ModelOptConfig(quant_type="FP8_WO", modules_to_not_convert=["conv"])
model = SanaTransformer2DModel.from_pretrained(checkpoint, subfolder="transformer", quantization_config=quant_config...

by TRTConfig did you mean including the config classes from ModelOptimizer here ?

[sayakpaul]

We use namings like BitsAndBytesConfig depending on the backend. See here:
https://github.com/huggingface/diffusers/blob/fb54499614f9603bfaa4c026202c5783841b3a80/src/diffusers/quantizers/quantization_config.py#L177C7-L177C25

So, in this case, we should be using TRTConfig or something similar.

[sayakpaul]

I think the problem has been fixed the newest release, I just need to bump it up in diffusers requirements

Alright, let's try with the latest fixes then.

[ishan-modi]

The newer version wasn't backward compatible hence the issues, I have fixed it.

Related to naming, package name is nvidia_modelopt, hence ModelOpt, but I can make it TRTModelOpt if you'd like ?

[sayakpaul]

Doesn't it have any reliance on tensorrt?

[ishan-modi]

No it doesn't, we can use TRT to compile the quantized model

[sayakpaul]

No it doesn't, we can use TRT to compile the quantized model

Could you elaborate what you mean by this?

[sayakpaul]

This looks nice. Could you demonstrate some memory savings and any speedups when using modelopt, please? We can then add tests, docs, etc.

[ishan-modi]

Could you elaborate what you mean by this?

Yeah, so for quantizing the model we dont use tensorRT, but once the model is quantized we can compile the model using tensorrt.

[ishan-modi]

💾 Model & Inference Memory (in MB)

Quantization Type	SanaTransformer2DModel		SD3Transformer2DModel		FluxTransformer2DModel
	Model Size (MB)	Inference (MB)	Model Size (MB)	Inference (MB)	Model Size (MB)	Inference (MB)
FP8	592.48	2060.80	2142.78	4405.61	13697.76	14438.87
INT4	306.20	2072.30	1160.89	3443.98	8803.12	9559.62
NVFP4	721.99	Err	1148.20	Err	8724.27	Err
Original (BF16)	1183.50	2642.37	4169.90	6457.41	23802.82	-

Following is the code

importtorchfromtqdmimporttqdmfromdiffusersimportSanaTransformer2DModel, SD3Transformer2DModel, FluxTransformer2DModelfromdiffusers.quantizers.quantization_configimportNVIDIAModelOptConfigcheckpoint="Efficient-Large-Model/Sana_600M_1024px_diffusers"model_cls=SanaTransformer2DModel# checkpoint = "stabilityai/stable-diffusion-3-medium-diffusers"# model_cls = SD3Transformer2DModel# checkpoint = "black-forest-labs/FLUX.1-dev"# model_cls = FluxTransformer2DModelinput=lambda: (torch.randn((2, 32, 32, 32), dtype=torch.bfloat16).to('cuda'), torch.randn((2,10,300,2304), dtype=torch.bfloat16).to('cuda'), torch.Tensor([0,0]).to('cuda'))
# input = lambda: (torch.randn((1,16,96,96), dtype=torch.bfloat16).to('cuda'), torch.randn((1,300,4096), dtype=torch.bfloat16).to('cuda'), torch.randn((1, 2048), dtype=torch.bfloat16).to('cuda'), torch.Tensor([0]).to('cuda'))# input = lambda: (torch.randn((1,1024, 64), dtype=torch.bfloat16).to('cuda'), torch.randn((1,300,4096), dtype=torch.bfloat16).to('cuda'), torch.randn((1, 768), dtype=torch.bfloat16).to('cuda'), torch.Tensor([0]).to('cuda'), torch.randn((300, 3)).to('cuda'), torch.randn((1024, 3)).to('cuda'), torch.Tensor([0]).to('cuda'))quant_config_fp8= {"quant_type": "FP8", "quant_method": "modelopt"}
quant_config_int4= {"quant_type": "INT4", "quant_method": "modelopt", "block_quantize": 128, "channel_quantize": -1}
quant_config_nvfp4= {"quant_type": "NVFP4", "quant_method": "modelopt", "block_quantize": 128, "channel_quantize": -1, 'modules_to_not_convert' : ['conv']}
deftest_quantization(config, checkpoint, model_cls):
quant_config=NVIDIAModelOptConfig(**config)
print(quant_config.get_config_from_quant_type())
quant_model=model_cls.from_pretrained(checkpoint, subfolder="transformer", quantization_config=quant_config, torch_dtype=torch.bfloat16, device_map="balanced").to('cuda')
print(f"Quant {config['quant_type']} Model Memory Footprint: ", quant_model.get_memory_footprint() /1e6)
returnquant_modeldeftest_quant_inference(model, input, iter=10):
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
inference_memory=0for_intqdm(range(iter)):
withtorch.no_grad():
output=model(*input())
inference_memory+=torch.cuda.max_memory_allocated()
inference_memory/=iterprint("Inference Memory: ", inference_memory/1e6)
test_quant_inference(test_quantization(quant_config_fp8, checkpoint, model_cls), input)
# test_quant_inference(test_quantization(quant_config_int4, checkpoint, model_cls), input)# test_quant_inference(test_quantization(quant_config_nvfp4, checkpoint, model_cls), input)# test_quant_inference(model_cls.from_pretrained(checkpoint, subfolder="transformer", torch_dtype=torch.bfloat16).to('cuda'), input)

Speed Ups

There is no significant speedup between the different quantizations because internally modelopt still uses high precision arithmetic (float32).

Sorry for being a bit late on this, @sayakpaul let me know next steps !

[sayakpaul]

@ishan-modi let us know if this is ready to be reviewed.

[ishan-modi]

@sayakpaul, I think it is ready for preliminary review, on-the-fly quantization works fine. But loading pre-quantized models errors out and will be fixed in next release here (early may) by NVIDIA team.

@jingyu-ml, just so that you are in the loop

[sayakpaul]

Looking good so far!

Could you also demonstrate some memory and timing numbers with the modelopt toolkit and some visual results?

No need, just saw #11173 (comment). But it doesn't measure the inference memory which is usually done via torch.cuda.max_memory_allocated(). Could we also see those numbers? Would it be possible to make it clear in the PR description that

on-the-fly quantization works fine. But loading pre-quantized models errors out and will be fixed in next release NVIDIA/Model-Optimizer#185 (early may) by NVIDIA team.

@jingyu-ml is it expected to not see any speedups in latency?

[SunMarc]

Thanks for this Just some nits, it could be nice to add this quantization scheme to transformers after this gets merged !

[sayakpaul]

@ishan-modi just a quick question. Do we know if the nunchaku SVDQuant method is supported through modelopt? From https://github.com/NVIDIA/TensorRT-Model-Optimizer?tab=readme-ov-file#quantization-examples-docs, it seems like it is supported. But could you confirm?

[ishan-modi]

@sayakpaul, yes modelopt does support SVDQuant, but in this integration we support only min-max based calibration see here. I think we should iteratively add advanced quantizations like svd_quant and awq once we have the base going, let me know if you think otherwise.

[sayakpaul]

That's fine. I wanted to because I think if we can support svd_quant through our modelopt backend, I am happy to drop #12207. Hence wanted to check.

[sayakpaul]

Will merge after @DN6 has had a chance to review. @ishan-modi can we also include a note in the docs that just performing the conversion step with modelopt won't lead to speed improvements (as pointed out here)?

@realAsma@jingyu-ml after this PR is merged, we could plan writing a post/guide on how to take a modelopt converted diffusers pipeline and use in deployment settings for realizing the actual speed gains.

[DN6]

Excellent work @ishan-modi 👍🏽 Thank you 🙏🏽

[sayakpaul]

@ishan-modi can we fix the remaining CI problems and then we should be good to go.

[ishan-modi]

@sayakpaul, should be fixed now.

[sayakpaul]

Congratulations on shipping this thing, @ishan-modi! Thank you!

Let's maybe now focus on the following things to maximize the potential impact:

• SVDQuant Support
• Guide to actually benefit from speedups

Happy to help.

[sayakpaul]

@ishan-modi I hope things are well on your end.

I think we never got around to showing how to deploy / use the TRT-serialized models through an engine to realize actual speedups? @kevalmorabia97 / @jingyu-ml do you think we could collaborate on that?