Triton AssertionError: duplicate template name during inference with Torch 2.8.0, 2.90 cu129

[myzsqn]

Running a RTX 3080 Ti

Traceback (most recent call last):
  File "C:\Users\user\intellijprojects\chatterbox\example_tts.py", line 27, in <module>
    audio = model.generate("fast generation using cudagraphs-manual, warmup")
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\user\intellijprojects\chatterbox\src\chatterbox\tts.py", line 253, in generate
    speech_tokens = self.t3.inference(
                    ^^^^^^^^^^^^^^^^^^
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\utils\_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\user\intellijprojects\chatterbox\src\chatterbox\models\t3\t3.py", line 519, in inference
    torch.compiler.cudagraph_mark_step_begin()
    from torch._inductor import cudagraph_trees
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\cudagraph_trees.py", line 60, in <module>
    from torch._inductor.compile_fx import (
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\compile_fx.py", line 111, in <module>
    from .fx_passes.joint_graph import joint_graph_passes
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\fx_passes\joint_graph.py", line 27, in <module>
    from ..pattern_matcher import (
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\pattern_matcher.py", line 80, in <module>
    from .lowering import fallback_node_due_to_unsupported_type
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\lowering.py", line 7232, in <module>
    import_submodule(kernel)
    importlib.import_module(f"{mod.__name__}.{filename[:-3]}")
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\kernel\flex_attention.py", line 692, in <module>
    flex_attention_template = TritonTemplate(
                              ^^^^^^^^^^^^^^^
  File "C:\Users\user\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\_inductor\select_algorithm.py", line 1432, in __init__
    assert name not in self.all_templates, "duplicate template name"
AssertionError: duplicate template name

Happens specifically in the t3.py when calling
torch.compiler.cudagraph_mark_step_begin()

Solved the issue by downgrading to Torch 2.7.0 using cu128

pip uninstall torch torchaudio

pip install torch==2.7.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128 --no-cache-dir

Leaving here in case anyone else has similar issues