feat(smoke-test): mount qwen3-0.6b base model via weights API instead of HF

[William-Gao1]

This PR makes two changes to unblock and harden the training smoke test.

1. Mount qwen3-0.6b base model via the weights API (BDN)

The smoke test pulled Qwen/Qwen3-0.6B directly from HuggingFace inside the training container on every run (via axolotl resolving base_model). With the daily schedule plus the H100/H200 matrix, this was getting rate-limited by HF.

• config.py — add WeightsSource(source="hf://Qwen/Qwen3-0.6B", mount_location="/mnt/user/Qwen3-0.6B") to the TrainingJob, and pass the mount path to the container via the MODEL_MOUNT_PATH env var. Baseten mirrors the repo into BDN once (deduplicated at the WeightVersion level) and CSI-mounts it.

• train.py — seed_hf_cache_from_mount() wires the mounted weights into the HF hub cache under the canonical repo id (real commit from a cheap metadata call → snapshots/<commit> symlinked to the mount). from_pretrained("Qwen/Qwen3-0.6B") then loads from the mount (only lightweight HEADs, no 1.5 GB download) and records the HF id in the saved LoRA adapter, so the checkpoint is deployable — deploy resolves the base from the BDN mirror by namespace/model id.

  (Recording the id at train time is deliberate: a post-hoc rewrite of adapter_config.json loses a race — the job is COMPLETED and deployed before the re-upload + ensure_checkpoints re-read propagate the new base_model.)

2. Fix deploy-and-infer: pass run_id

truss's create_model_version_from_inference_template() gained a required run_id positional arg, so the deploy step failed with missing 1 required positional argument: 'run_id' even after training completed and checkpointed. The smoke test deploys training-job checkpoints, and _hydrate_deploy_config raises if run_id is set alongside training-job checkpoints, so the correct value is None.

Notes

• First run still seeds the BDN mirror from HF once; the win is on repeated runs. The dataset (winglian/pirate-ultrachat-10k) is still an HF pull.
• Separately, there's an intermittent dataset-prep failure (FileNotFoundError on an arrow shard under the project cache mount during axolotl's 128-process tokenization) that predates this PR and is not addressed here.

[github-actions]

Example Validation Results

✅ examples/qwen3-0.6b-axolotl/training/config.py

Status: TRAINING_JOB_COMPLETED | Job: q8o1yy3 | Project: pr-118-examples--qwen3-0.6b-axolotl

Cache Summary

Error fetching cache summary: 404 Client Error: Not Found for url: https://api.baseten.co/v1/training_projects/6wg6gg3/cache/summary
Client error: No cache summary data found for this project. Ensure you're running jobs with cache enabled.
ERROR HTTPError: 404 Client Error: Not Found for url: 
https://api.baseten.co/v1/training_projects/6wg6gg3/cache/summary

Checkpoints

2 checkpoint(s)

ID                   Type       Size         Created
-------------------- ---------- ------------ -------------------
.                    lora       996.93 MB    2026-06-30 16:32:18
checkpoint-10        lora       981.05 MB    2026-06-30 16:33:31