Fp8 error in accelerate test

benetoriol · March 10, 2025, 10:51pm

I am trying to run fp8 training with accelerate.

I have set up accelerate config with fp8 and all default.
then I run accelerate test
and get :

line 87, in setup_fp8_env
    values = value.strip("()").split(",")
             ^^^^^^^^^^^
AttributeError: 'bool' object has no attribute 'strip'

this does not happen when using fp16.
I am using a 4090, any help is appraciated

NVIDIA-SMI 555.58.02 Driver Version: 555.58.02 CUDA Version: 12.5
nvcc -V: 12.8

Pip freeze

accelerate==1.4.0
annotated-types==0.7.0
bitsandbytes==0.45.3
certifi==2025.1.31
charset-normalizer==3.4.1
diffusers==0.32.2
filelock==3.16.1
fsspec==2024.10.0
huggingface-hub==0.29.2
idna==3.10
importlib_metadata==8.6.1
Jinja2==3.1.4
MarkupSafe==2.1.5
mpmath==1.3.0
networkx==3.4.2
numpy==2.1.2
nvidia-cublas-cu12==12.8.3.14
nvidia-cuda-cupti-cu12==12.8.57
nvidia-cuda-nvrtc-cu12==12.8.61
nvidia-cuda-runtime-cu12==12.8.57
nvidia-cudnn-cu12==9.7.1.26
nvidia-cufft-cu12==11.3.3.41
nvidia-cufile-cu12==1.13.0.11
nvidia-curand-cu12==10.3.9.55
nvidia-cusolver-cu12==11.7.2.55
nvidia-cusparse-cu12==12.5.7.53
nvidia-cusparselt-cu12==0.6.3
nvidia-nccl-cu12==2.25.1
nvidia-nvjitlink-cu12==12.8.61
nvidia-nvtx-cu12==12.8.55
packaging==24.2
pillow==11.0.0
psutil==7.0.0
pydantic==2.10.6
pydantic_core==2.27.2
pytorch-triton==3.2.0+git4b3bb1f8
PyYAML==6.0.2
regex==2024.11.6
requests==2.32.3
safetensors==0.5.3
sympy==1.13.3
tokenizers==0.21.0
torch==2.7.0.dev20250310+cu128
torchaudio==2.6.0.dev20250310+cu128
torchvision==0.22.0.dev20250310+cu128
tqdm==4.67.1
transformer_engine==1.13.0
transformer_engine_cu12==1.13.0
transformer_engine_torch==1.13.0
transformers==4.49.0
typing_extensions==4.12.2
urllib3==2.3.0
zipp==3.21.0

EDIT: ““solved””

modified accelerate/utils/launch.py to

86                 if arg == "fp8_override_linear_precision":
87                     if not isinstance(value, list):
88                         value = [value, value, value]
89                     values = value
90                     current_env[prefix + "FP8_OVERRIDE_FPROP"] = str(values[0])
91                     current_env[prefix + "FP8_OVERRIDE_DGRAD"] = str(values[1])
92                     current_env[prefix + "FP8_OVERRIDE_WGRAD"] = str(values[2])

No idea if this is is how it’s supposed to be done

John6666 · March 11, 2025, 3:22am

github.com/huggingface/accelerate

src/accelerate/utils/launch.py

main


      
                  # Intel MPI and MVAPICH both use the same arg names
                  return mpi_app, "-f", "-n", "-ppn", ""
          
          
          def setup_fp8_env(args: argparse.Namespace, current_env: Dict[str, str]):
              """
              Setup the FP8 environment variables.
              """
              prefix = "ACCELERATE_"
              for arg in vars(args):
                  if arg.startswith("fp8_"):
                      value = getattr(args, arg)
                      if value is not None:
                          if arg == "fp8_override_linear_precision":
                              values = value.strip("()").split(",")
                              current_env[prefix + "FP8_OVERRIDE_FPROP"] = values[0].strip()
                              current_env[prefix + "FP8_OVERRIDE_DGRAD"] = values[1].strip()
                              current_env[prefix + "FP8_OVERRIDE_WGRAD"] = values[2].strip()
                          else:
                              current_env[f"{prefix}{arg.upper()}"] = str(getattr(args, arg))
              return current_env

github.com/huggingface/accelerate

src/accelerate/utils/launch.py

main


      
              current_env["MASTER_ADDR"] = args.main_process_ip if args.main_process_ip is not None else "127.0.0.1"
              current_env["MASTER_PORT"] = str(args.main_process_port) if args.main_process_port is not None else "29500"
          
          try:
              mixed_precision = PrecisionType(args.mixed_precision.lower())
          except ValueError:
              raise ValueError(
                  f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
              )
          
          current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
          if args.mixed_precision.lower() == "fp8":
              if not is_fp8_available():
                  raise RuntimeError(
                      "FP8 is not available on this machine. Please ensure that either Transformer Engine or MSAMP is installed."
                  )
              current_env = setup_fp8_env(args, current_env)
          
          try:
              dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
          except ValueError:

There is a possibility that there is a bug in the way CLI handles fp8-related arguments…
Unless you have a GPU that supports fp8 natively, such as 4090, you probably won’t reach the code, so you won’t have a problem.

The error itself will disappear if you lower the version of accelerate, as it probably occurred at the time of the following commit.

Topic		Replies	Views
Error while fine tuning with peft, lora, accelerate, SFTConfig and SFTTrainer Models	3	885	November 7, 2024
Why my Accelerate just doesn't work? 🤗Accelerate	2	6016	March 7, 2022
"Attempting to unscale FP16 gradients" error when using optimizer in mixed precision training with Accelerate 🤗Accelerate	1	2437	April 15, 2024
FSDP Auto Wrap does not work using `accelerate` in Multi-GPU Setup 🤗Accelerate	0	216	September 6, 2024
Issues when using `accelerate` with `fp16` Intermediate	4	11532	January 22, 2024

Fp8 error in accelerate test

Related topics