Unable to compile model for Apple GPU, failed to resolve 'MOGGKernelAPI'

[walkerbrown]

This tutorial was a great intro to GPT-2's architecture, thanks! I'm able to run inference (slowly) on my CPU.

Bug report

I wasn't able to get it running on my Apple GPU as the model fails to compile with current nightly build of Max.

File "modular/max-llm-book/.pixi/envs/default/lib/python3.14/site-packages/max/engine/api.py", line 424, in load
    _model = self._impl.compile_from_object(
        model._module._CAPIPtr,  # type: ignore
        custom_extensions_final,
        model.name,
    )
ValueError: Failed to resolve module path for `MOGGKernelAPI`

The above exception was the direct cause of the following exception:

...

RuntimeError: Failed to compile the model. Please file an issue, all models should be correct by construction and this error should have been caught during construction.
For more detailed failure information run with the environment variable `MODULAR_MAX_DEBUG=True`.

I have confirmed identical error output when running pixi run main, please see attached below.

Environment

MAX version	26.2.0.dev2026021505
Hardware	Apple M2 Max, 38-core GPU
OS	macOS 26.2 (25C56)
Metal Version	Metal 4
Python Version	3.14.3

[walkerbrown]

run-main-with-debug-env-set.txt

[BradLarson]

When other people have encountered this, the source of the error seemed to be something weird with the local Python environment. Are you happening to run the Pixi process inside of an existing activated Python virtual environment? Someone also encountered this in a NixOS setup, but you're probably not running into that on macOS.

[walkerbrown]

Hi @BradLarson, thanks for pointing out that forum post! I missed it and it does look like the same issue or similar. On my side, I've just gone back to confirm there's no virtual env weirdness needed to reproduce.

The procedure was to clone the repo fresh at 5301afe to avoid any .pixi/ or ~/.pixi/ interference, then

• pixi install -vv with verbose flag
• pixi list | grep modular to get the as-installed versions, and finally
• pixi run main

dylan@mac max-llm-clean-test % git diff --stat
dylan@mac max-llm-clean-test % git log --oneline -n 1
5301afe (HEAD -> main, origin/main, origin/HEAD) [MAX] [Examples] Extend LLM book GPT-2 model for bfloat16 To let the GPT-2 example model run on GPUs, we need to align the model weights to the datatype used by the default device. Without overriding anything, that's bfloat16, so this selects the bfloat16 version of the GPT-2 model weights. The use of float64 for the probabilities is also incompatible with certain GPUs, so this calculation is now performed in float32.

dylan@mac max-llm-clean-test % pixi install -vv
DEBUG pixi_config: Loading config from /etc/pixi/config.toml
DEBUG pixi_config: Loading config from /Users/dylan/Library/Application Support/pixi/config.toml
DEBUG pixi_config: Loading config from /Users/dylan/.pixi/config.toml
DEBUG pixi_config: Loading config from /Users/dylan/Developer/modular/max-llm-clean-test/.pixi/config.toml
DEBUG pixi_config: Failed to load local config: /Users/dylan/Developer/modular/max-llm-clean-test/.pixi/config.toml (error: no file was found at /Users/dylan/Developer/modular/max-llm-clean-test/.pixi/config.toml)
DEBUG pixi_core::environment: verifying prefix location is unchanged, with prefix file: /Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/conda-meta/pixi_env_prefix
 INFO pixi_core::lock_file::update: the lock-file is up-to-date
DEBUG pixi_core::lock_file::virtual_packages: Required virtual packages of environment 'default': __osx >=11.0, __unix, __osx >=11.0, __unix, __osx >=11.0, __unix, __osx >=11.0, __unix, __osx >=11.0, __unix, __osx >=11.0, __unix, __osx >=11.0
DEBUG pixi_core::lock_file::virtual_packages: Generic system virtual packages for env: 'default' : [__osx: __osx=26.3=0, __archspec: __archspec=1=m2, __unix: __unix=0=0]
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::lock_file::virtual_packages: Skipping virtual package: __unix as it's not in the accepted packages
DEBUG pixi_core::lock_file::virtual_packages: Required virtual package: __osx >=11.0 matches the system
DEBUG pixi_core::environment::conda_prefix: updating prefix for 'default'
DEBUG pixi_utils::rlimit: Increased RLIMIT_NOFILE to 2048
DEBUG pixi_core::environment::conda_metadata: Prefix file updated with: '/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/conda-meta'.
DEBUG pixi_core::environment::conda_metadata: Verify history file exists: /Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/conda-meta/history
DEBUG pixi_uv_context: uv keyring provider is disabled
 INFO pixi_core::lock_file::update: Installed environment 'default' in 3.9654515s
DEBUG pixi_core::environment: Wrote environment file to: "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/conda-meta/pixi"
✔ The default environment has been installed.
dylan@mac max-llm-clean-test % cat pixi.lock | grep modular
    - url: https://conda.modular.com/max-nightly/
      - conda: https://conda.modular.com/max-nightly/linux-64/max-26.2.0.dev2026012905-3.12release.conda
      - conda: https://conda.modular.com/max-nightly/linux-64/max-core-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/max-pipelines-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/mblack-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/modular-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/linux-64/mojo-0.26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/linux-64/mojo-compiler-0.26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/mojo-python-0.26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/osx-arm64/max-26.2.0.dev2026012905-3.12release.conda
      - conda: https://conda.modular.com/max-nightly/osx-arm64/max-core-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/max-pipelines-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/mblack-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/modular-26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/osx-arm64/mojo-0.26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/osx-arm64/mojo-compiler-0.26.2.0.dev2026012905-release.conda
      - conda: https://conda.modular.com/max-nightly/noarch/mojo-python-0.26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/linux-64/max-26.2.0.dev2026012905-3.12release.conda
- conda: https://conda.modular.com/max-nightly/osx-arm64/max-26.2.0.dev2026012905-3.12release.conda
- conda: https://conda.modular.com/max-nightly/linux-64/max-core-26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/osx-arm64/max-core-26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/noarch/max-pipelines-26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/noarch/mblack-26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/noarch/modular-26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/linux-64/mojo-0.26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/osx-arm64/mojo-0.26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/linux-64/mojo-compiler-0.26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/osx-arm64/mojo-compiler-0.26.2.0.dev2026012905-release.conda
- conda: https://conda.modular.com/max-nightly/noarch/mojo-python-0.26.2.0.dev2026012905-release.conda
dylan@mac max-llm-clean-test % pixi run main
✨ Pixi task (main): python main.py
Using device: Device(type=gpu,id=0), dtype: DType.bfloat16
Loaded HuggingFace model:
GPT2LMHeadModel(
  (transformer): GPT2Model(
    (wte): Embedding(50257, 768)
    (wpe): Embedding(1024, 768)
    (drop): Dropout(p=0.1, inplace=False)
    (h): ModuleList(
      (0-11): 12 x GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D(nf=2304, nx=768)
          (c_proj): Conv1D(nf=768, nx=768)
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D(nf=3072, nx=768)
          (c_proj): Conv1D(nf=768, nx=3072)
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
    )
    (ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
  )
  (lm_head): Linear(in_features=768, out_features=50257, bias=False)
)
Model has 12 layers, 12 heads, 768 embedding dim
Transposing transformer.h.0.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.0.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.0.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.0.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.1.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.1.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.1.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.1.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.2.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.2.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.2.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.2.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.3.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.3.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.3.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.3.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.4.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.4.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.4.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.4.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.5.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.5.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.5.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.5.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.6.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.6.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.6.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.6.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.7.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.7.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.7.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.7.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.8.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.8.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.8.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.8.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.9.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.9.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.9.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.9.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.10.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.10.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.10.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.10.mlp.c_proj.weight: torch.Size([3072, 768])
Transposing transformer.h.11.attn.c_attn.weight: torch.Size([768, 2304])
Transposing transformer.h.11.attn.c_proj.weight: torch.Size([768, 768])
Transposing transformer.h.11.mlp.c_fc.weight: torch.Size([768, 3072])
Transposing transformer.h.11.mlp.c_proj.weight: torch.Size([3072, 768])
Traceback (most recent call last):
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/engine/api.py", line 424, in load
    _model = self._impl.compile_from_object(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: Failed to resolve module path for `MOGGKernelAPI`

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 439, in <module>
    main()
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 384, in main
    max_model = MaxGPT2LMHeadModel(config)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 253, in __init__
    self.transformer = MaxGPT2Model(config)
                       ^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 226, in __init__
    self.h = Sequential(*(GPT2Block(config) for _ in range(config.n_layer)))
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 226, in <genexpr>
    self.h = Sequential(*(GPT2Block(config) for _ in range(config.n_layer)))
                          ^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 198, in __init__
    self.ln_2 = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/main.py", line 172, in __init__
    self.bias = Tensor.zeros([dim])
                ^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/tensor.py", line 696, in zeros
    return cls.full(shape, value=0, dtype=dtype, device=device)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/tensor.py", line 617, in full
    cls.constant(value, dtype=dtype, device=device), shape
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/tensor.py", line 575, in constant
    return F.constant(value, dtype, device)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/functional.py", line 203, in wrapped
    with contextlib.ExitStack() as stack:
         ^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/contextlib.py", line 610, in __exit__
    raise exc_details[1]
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/contextlib.py", line 595, in __exit__
    if cb(*exc_details):
       ^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/_realization_context.py", line 346, in __exit__
    F._run(self.realize_all())
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/functional.py", line 97, in _run
    return asyncio.run(coro)
           ^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/asyncio/runners.py", line 195, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/asyncio/base_events.py", line 691, in run_until_complete
    return future.result()
           ^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/_realization_context.py", line 255, in realize_all
    model = _session().load(graph)
            ^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/dylan/Developer/modular/max-llm-clean-test/.pixi/envs/default/lib/python3.12/site-packages/max/engine/api.py", line 430, in load
    raise RuntimeError(
RuntimeError: Failed to compile the model. Please file an issue, all models should be correct by construction and this error should have been caught during construction.
For more detailed failure information run with the environment variable `MODULAR_MAX_DEBUG=True`.

---

I think it's notable this fresh clone (dropping the .pixi/ cache) did change by rolling back a bit both the Max build and Python version from my initial report, but the resulting error is the same.

MAX version	26.2.0.dev2026012905
Hardware	Apple M2 Max, 38-core GPU
OS	macOS 26.2 (25C56)
Metal Version	Metal 4
Python Version	3.12.12