[DONT MERGE] Debug branch for the Qwen3 + full model compile

[anijain2305]

Take this PR. This just turns of compile and full model compilation, and then sets up the config values that you finish the program quickly. Use the instructions at the beginning of this file - https://github.com/pytorch/torchtune/blob/main/recipes/configs/qwen3/8B_full_single_device.yaml

torch.compile takes 198 seconds to fine-tune (with warmup). Eager takes only 48 seconds. If we do regional compile (compile only the transformer layers), it takes 35 seconds. So, something really bad is going on.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2831

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures, 2 Cancelled Jobs

As of commit 39b9193 with merge base 2344509 ():

NEW FAILURES - The following jobs have failed:

• GPU tests / gpu_test (3.11, stable) (gh)
  tests/recipes/test_full_finetune_distributed.py::TestFullFinetuneDistributedRecipe::test_training_state_on_resume_from_distributed_checkpoint_multi_rank[llama3/8B_full-llama3_hf_138m-1-4-False]
• Lint / lint (3.10) (gh)
  torchtune/training/_compile.py:13:1: F401 'torchtune.modules.TransformerSelfAttentionLayer' imported but unused

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• GPU tests / gpu_test (3.10, stable) (gh)
  tests/recipes/test_full_finetune_distributed.py::TestFullFinetuneDistributedRecipe::test_training_state_on_resume_from_distributed_checkpoint_multi_rank[llama3/8B_full-llama3_hf_138m-1-4-False]
• GPU tests / gpu_test (3.9, stable) (gh)
  tests/recipes/test_full_finetune_distributed.py::TestFullFinetuneDistributedRecipe::test_training_state_on_resume_from_distributed_checkpoint_multi_rank[llama3/8B_full-llama3_hf_138m-1-4-False]

This comment was automatically generated by Dr. CI and updates every 15 minutes.