[corticalstack]

Summary

Step 5 of #31: the fine-tune lab (15) on current torch and peft, with the GPU jobs pinned so they stop changing under the lab.

• Local fine-tune group: uv.lock moves torch 2.12.0 → 2.14.1 and peft 0.19.1 → 0.21.2. transformers stays at 4.53.3 because Phi-4-mini's remote code needs a release below 5.
• The GPU jobs were broken on main. They ran on acpt-pytorch-2.2-cuda12.1:latest and installed accelerate, datasets, peft and olive-ai with no versions.
  • Microsoft rebuilt that image on 2026-09-29. Despite the repository name it now ships torch 2.13.0, CUDA 12.6 and Python 3.10.
  • On the rebuilt image, pip can't replace the image's own packages. The fine-tune job failed in pip install before training started: Permission denied: 'InputStream.py' while removing antlr4-python3-runtime 4.9.3 to make way for azure-cli.
• What changed in the jobs:
  • The image is pinned to tag 58 instead of :latest.
  • accelerate, datasets and peft are pinned to the lock's versions and olive-ai to 0.13.0, so 15-04 loads the adapter with the versions that trained it.
  • The packages install into a venv created with --system-site-packages, so they use the image's torch without pip having to remove anything from the image.
• 15-02 and 15-03 now raise when their GPU job fails. Before, they printed the failure or fell back to 0% accuracy, and passed.

Type of change

• Bug fix (notebook, Bicep, or doc)
• New lab or sub-lab
• Extension to an existing lab
• Documentation / README / CONTRIBUTING update
• CI / tooling

Checklist

• On a feature branch (not main)
• Notebook(s) run top-to-bottom on a clean kernel (15-01 to 15-04; see below for 15-02 and 15-03)
• uv sync succeeds with no new heavyweight deps without justification
• New env vars added to .env.example with placeholders (none)
• Lab added (or updated) in the index table of README.md (no change)
• No secrets in committed files (.env, keys, connection strings)
• Screenshots placed under docs/screenshots/ (none)

Verification

Notebook	Where	Result
15-01	Local + APIM teacher	PASS, 124 s, on torch 2.14.1
15-02	A100 serverless GPU	PASS, 1210 s. The job installs the pins, Olive trains 300 LoRA steps on gpu-cuda, and the adapter (369 MB) is uploaded
15-03	A100 serverless GPU	PASS, 660 s. Accuracy is 65.7% (23/35) (the lab doc's last recorded result was 51.4%), with a SEVERITY: line in every response
15-04	Local CPU	PASS, 134 s. peft 0.21.2 loads the new adapter on torch 2.14.1 and classifies 2013-05-09 (the 2B PVTCS ammonia leak) as expected

• Region: swedencentral had no capacity (ManagedEnvironmentCapacityHeavyUsageError), so 15-02 and 15-03 ran against an A100 environment in italynorth. I used throwaway copies that changed only LOCATION and the environment name. The lab's code still targets swedencentral.
• Offline checks before the GPU run:
  • The pinned set resolves against the image's torch 2.13.0 and numpy 1.26.4.
  • Every olive finetune option the job passes exists in olive 0.13.0.
  • check_imports.py passes.

The 15-02 comment "ACA GPU NC24-A100 only available in Sweden Central" is out of date. A100 serverless GPU is now also in westus3, westus, eastus, canadacentral, italynorth, uksouth and australiaeast. I left it for a separate change.

Related issues

Refs #31