RuntimeError: HIP error: invalid device function

[faljse]

Checks

• This template is only for bug reports, usage problems go with 'Help Wanted'.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

arch
Radeon 8050S
torch==2.5.1+rocm6.2 torchaudio==2.5.1+rocm6.2

Steps to Reproduce

.) Followed instructions for amd gpu
pip install torch==2.5.1+rocm6.2 torchaudio==2.5.1+rocm6.2 --extra-index-url https://download.pytorch.org/whl/rocm6.2
pip install -e .
...
Successfully installed f5-tts-1.1.10

(f5-tts) [user@fw F5-TTS]$ f5-tts_infer-gradio 
Download Vocos from huggingface charactr/vocos-mel-24khz

vocab :  /home/user/F5-TTS/src/f5_tts/infer/examples/vocab.txt
token :  custom
model :  /home/user/.cache/huggingface/hub/models--SWivid--F5-TTS/snapshots/84e5a410d9cead4de2f847e7c9369a6440bdfaca/F5TTS_v1_Base/model_1250000.safetensors 

Traceback (most recent call last):
  File "/home/user/.conda/envs/f5-tts/bin/f5-tts_infer-gradio", line 3, in <module>
    from f5_tts.infer.infer_gradio import main
  File "/home/user/F5-TTS/src/f5_tts/infer/infer_gradio.py", line 90, in <module>
    F5TTS_ema_model = load_f5tts()
                      ^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/infer_gradio.py", line 68, in load_f5tts
    return load_model(DiT, F5TTS_model_cfg, ckpt_path)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/utils_infer.py", line 272, in load_model
    model = load_checkpoint(model, ckpt_path, device, dtype=dtype, use_ema=use_ema)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/utils_infer.py", line 197, in load_checkpoint
    model = model.to(dtype)
            ^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1340, in to
    return self._apply(convert)
           ^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 900, in _apply
    module._apply(fn)
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 900, in _apply
    module._apply(fn)
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 900, in _apply
    module._apply(fn)
  [Previous line repeated 1 more time]
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 927, in _apply
    param_applied = fn(param)
                    ^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1326, in convert
    return t.to(
           ^^^^^
RuntimeError: HIP error: invalid device function
HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3
Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.

✔️ Expected Behavior

No response

❌ Actual Behavior

No response

[SWivid]

F5-TTS/src/f5_tts/infer/utils_infer.py

Lines 189 to 197 in 9ae46c8

	if dtype is None:
	dtype = (
	torch.float16
	if "cuda" in device
	and torch.cuda.get_device_properties(device).major >= 7
	and not torch.cuda.get_device_name().endswith("[ZLUDA]")
	else torch.float32
	)
	model = model.to(dtype)

Not sure why, but maybe try print out to see if

F5-TTS/src/f5_tts/infer/utils_infer.py

Line 194 in 9ae46c8

and not torch.cuda.get_device_name().endswith("[ZLUDA]")

[faljse]

device="AMD Radeon 8060S"
dtype=None

i forced it to torch.float32 (was using float16)

    if dtype is None:
        dtype = (
            torch.float32
            #if "cuda" in device
            #and torch.cuda.get_device_properties(device).major >= 7
            # and not torch.cuda.get_device_name().endswith("[ZLUDA]")
            #else torch.float32
        )
    model = model.to(dtype)

now it launches, but synthesizing yields another/similar? situation:

Starting app...
* Running on local URL:  http://127.0.0.1:7860
* To create a public link, set `share=True` in `launch()`.
`torch_dtype` is deprecated! Use `dtype` instead!
Device set to use cuda
Using `chunk_length_s` is very experimental with seq2seq models. The results will not necessarily be entirely accurate and will have caveats. More information: https://github.com/huggingface/transformers/pull/20104. Ignore this warning with pipeline(..., ignore_warning=True). To use Whisper for long-form transcription, use rather the model's `generate` method directly as the model relies on it's own chunking mechanism (cf. Whisper original paper, section 3.8. Long-form Transcription).
Traceback (most recent call last):
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/gradio/queueing.py", line 759, in process_events
    response = await route_utils.call_process_api(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/gradio/route_utils.py", line 354, in call_process_api
    output = await app.get_blocks().process_api(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/gradio/blocks.py", line 2191, in process_api
    result = await self.call_function(
             ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/gradio/blocks.py", line 1698, in call_function
    prediction = await anyio.to_thread.run_sync(  # type: ignore
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/anyio/to_thread.py", line 61, in run_sync
    return await get_async_backend().run_sync_in_worker_thread(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 2525, in run_sync_in_worker_thread
    return await future
           ^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 986, in run
    result = context.run(func, *args)
             ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/gradio/utils.py", line 915, in wrapper
    response = f(*args, **kwargs)
               ^^^^^^^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/infer_gradio.py", line 290, in basic_tts
    audio_out, spectrogram_path, ref_text_out, used_seed = infer(
                                                           ^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/infer_gradio.py", line 160, in infer
    ref_audio, ref_text = preprocess_ref_audio_text(ref_audio_orig, ref_text, show_info=show_info)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/utils_infer.py", line 366, in preprocess_ref_audio_text
    ref_text = transcribe(ref_audio)
               ^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/F5-TTS/src/f5_tts/infer/utils_infer.py", line 179, in transcribe
    return asr_pipe(
           ^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/automatic_speech_recognition.py", line 275, in __call__
    return super().__call__(inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/base.py", line 1459, in __call__
    return next(
           ^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/pt_utils.py", line 126, in __next__
    item = next(self.iterator)
           ^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/pt_utils.py", line 271, in __next__
    processed = self.infer(next(self.iterator), **self.params)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/base.py", line 1374, in forward
    model_outputs = self._forward(model_inputs, **forward_params)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/pipelines/automatic_speech_recognition.py", line 542, in _forward
    tokens = self.model.generate(**generate_kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/models/whisper/generation_whisper.py", line 704, in generate
    init_tokens = self._retrieve_init_tokens(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/models/whisper/generation_whisper.py", line 1572, in _retrieve_init_tokens
    lang_ids = self.detect_language(
               ^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/.conda/envs/f5-tts/lib/python3.11/site-packages/transformers/models/whisper/generation_whisper.py", line 1671, in detect_language
    torch.ones((batch_size, 1), device=self.device, dtype=torch.long)
torch.AcceleratorError: HIP error: invalid device function
Search for `hipErrorInvalidDeviceFunction' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__HIPRT__TYPES.html for more information.
HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3
Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.

[SWivid]

torch.AcceleratorError: HIP error: invalid device function

Seems compatibility issue with transformers/models/whisper/generation_whisper.py. Likely, need to modify code of transformers package for your device.
Try also open an issue there (transformers repo) see if anyone got a workaround for AMD Radeon 8060S.