Add FP8 kernel acceleration for compressed-tensors quantized models

[jiqing-feng]

What does this PR do?

This PR adds native FP8 matmul kernel support for compressed-tensors FP8 quantized models in transformers. Previously, compressed-tensors FP8 models were loaded via the compressed-tensors library and dequantized back to FP16/BF16 for inference. With this change, FP8 weights are kept in FP8 format and inference uses hardware-accelerated FP8 matmul kernels (torch._scaled_mm on XPU, fbgemm.f8f8bf16_rowwise on CUDA).

Key changes:

New file: src/transformers/integrations/compressed_tensors_fp8.py

• CTFP8Linear: FP8 linear layer that stores weights in FP8 and uses row-wise FP8 matmul kernels. Activations are dynamically quantized per-row via quantize_fp8_per_row.
• Weight converters (CompressedTensorsScaleConvert, CompressedTensorsFp8Dequantize) to handle the checkpoint format conversion (e.g. weight_scale → weight_scale_inv).
• CTFP8PerRowQuantize: Online quantization support — quantize BF16 weights to FP8 per-row on-the-fly during model loading.

Modified: src/transformers/quantizers/quantizer_compressed_tensors.py

• CompressedTensorsHfQuantizer now detects FP8 quantization configs (float type, num_bits=8) and automatically routes to the FP8 kernel path when GPU/XPU is available. Falls back to the default compressed-tensors dequantize path on CPU.
• Added get_weight_conversions() and get_quantize_ops() to support both pre-quantized loading and online quantization.
• No changes to the non-FP8 code path — existing INT8/INT4 compressed-tensors models are unaffected.

Modified: src/transformers/quantizers/auto.py

• Minor formatting change (no functional change).

Supported models

• Per-channel dynamic: e.g. RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic
• Per-tensor static: e.g. RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8
• Online quantization: Any BF16 model can be quantized to FP8 on-the-fly by passing a CompressedTensorsConfig with FP8 quantization scheme.

Usage

Pre-quantized model (no config needed)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

Online quantization

from transformers import AutoModelForCausalLM, CompressedTensorsConfig
from compressed_tensors.quantization import QuantizationScheme, QuantizationArgs, QuantizationType, QuantizationStrategy

ct_config = CompressedTensorsConfig(
    config_groups={
        "group_0": QuantizationScheme(
            weights=QuantizationArgs(
                num_bits=8, type=QuantizationType.FLOAT, strategy=QuantizationStrategy.CHANNEL,
            ),
        ),
    },
    run_compressed=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    quantization_config=ct_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

Devices

• XPU (Intel Data Center Max / Arc): uses torch._scaled_mm
• CUDA (SM89+): uses fbgemm.f8f8bf16_rowwise
• CPU: falls back to default compressed-tensors dequantize path

@sywangyi

[Rocketknight1]

cc @SunMarc

[SunMarc]

Thanks, left a comment !

[SunMarc]

Indeed, the model is dequantized on the fly even if run_compressed=True (it worked before but compressed-tensors preferred to delegate this work to vllm as it was duplicate work). We could add support for the most used methods here in transformers but it would be nice to use kernels from vllm if possible. Also, I don't know if this is worth using fbgemm. torchao don't use this lib anymore and it was quite a struggle to get this installed. The best would be to use kernels hosted on kernels-community as we do for our current fp8 support.

[jiqing-feng]

I've removed the fbgemm op and only use kernels-community.

[jiqing-feng]

Hi @SunMarc . Please check it the integration is ok. I'll clean the tests and doc after you approved the integration.