torch._dynamo.exc.Unsupported: call_method UserDefinedObjectVariable(Logger)

[jsoto-gladia]

System Info

• transformers version: 4.43.0
• Platform: Linux-5.15.0-118-generic-x86_64-with-glibc2.35
• Python version: 3.10.14
• Huggingface_hub version: 0.23.3
• Safetensors version: 0.4.3
• Accelerate version: 0.24.1
• Accelerate config: not found
• PyTorch version (GPU?): 2.1.1+cu121 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?:
• Using GPU in script?:
• GPU type: Tesla V100S-PCIE-32GB

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
audio_sample = ds[0]["audio"]

processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
model = model.to('cuda')

model.generation_config.cache_implementation = "static"
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)

input_features = processor(
audio_sample["array"], sampling_rate=audio_sample["sampling_rate"], return_tensors="pt"
).input_features

input_features = input_features.to("cuda")
_ = model.generate(input_features)

Expected behavior

it does not raise an error
as of now it raises
File "/home/jsoto/.pyenv/versions/3.10.14/envs/triton-server/lib/python3.10/site-packages/torch/_dynamo/exc.py", line 172, in unimplemented
raise Unsupported(msg)
torch._dynamo.exc.Unsupported: call_method UserDefinedObjectVariable(Logger) warning_once [ConstantVariable(str)] {}

[jsoto-gladia]

related to #31987

[jsoto-gladia]

if I put fullgraph=False, it works fine

[jsoto-gladia]

the same problem arises with torch 2.2.1

[jsoto-gladia]

the same problem arises with torch 2.3.1

[jsoto-gladia]

a similar problem arises with torch 2.4.0
File "/home/jsoto/.pyenv/versions/3.10.14/envs/triton-server/lib/python3.10/site-packages/torch/_dynamo/exc.py", line 221, in unimplemented
raise Unsupported(msg)
torch._dynamo.exc.Unsupported: Logger not supported for non-export cases

[jsoto-gladia]

it seems the problem is related to multilingual model...
If I use tiny, it gives the error above
I I use tiny.en, everything works fine

[jsoto-gladia]

@sanchit-gandhi

[jsoto-gladia]

when putting a mulitlingual model I get the following error
File "/home/jsoto/.pyenv/versions/3.10.14/envs/triton-server/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 464, in inner
raise exc.UserError(
torch._dynamo.exc.UserError: Dynamic control flow is not supported at the moment. Please use functorch.experimental.control_flow.cond to explicitly capture the control flow. For more information about this error, see: https://pytorch.org/docs/main/generated/exportdb/index.html#cond-operands

[jsoto-gladia]

I found the problem. It lies in the fact that if you do not specify the language
torch compile will try to run detect language (that contains an if condition here) and thus fails
the solution is to provide the language

[jsoto-gladia]

for the record the versions are:
torch==2.3.0
torchaudio==2.3.0
transformers==4.44.1

[ArthurZucker]

Well, we can keep this open as language detection would be nice to have in compile as well no?