[adapter] refactor: tidy _resolve_checkpoint_path call sites

[Jayce-Ping]

Summary

Two small follow-up refactors to BaseAdapter._resolve_checkpoint_path and BaseAdapter.load_checkpoint, both prompted by re-reading the code after PR #160 was merged. No behavior change.

What changes

A. Hoist os.path.expanduser into _resolve_checkpoint_path (commit 1)

Currently load_checkpoint does:

path = os.path.expanduser(path)
path = self._resolve_checkpoint_path(path)

The two normalization steps belong together. _resolve_checkpoint_path is the one entry point that takes a raw user-provided spec (local path OR HF repo spec, with or without hf://) and returns a usable local directory; making it own expanduser too means callers no longer have to remember to normalize ~ separately.

Safe across all three input branches:

• Local path with ~: expanded correctly (unchanged behavior).
• Local path without ~: no-op.
• HF spec (owner/repo or hf://...): no-op since expanduser only acts on a leading ~.

ModelArguments.__post_init__ still calls expanduser(resume_path) defensively at config-parse time, so other code that displays/logs model_args.resume_path keeps seeing the normalized form. The double call is idempotent.

B. Drop dead os.path.exists guard + collapse spec local (commit 2)

After the un-gating in PR #160, _resolve_checkpoint_path has these exit paths:

1. Returns a local path it has just confirmed with os.path.exists.
2. Returns a downloaded path that download_hf_checkpoint has verified with os.path.isdir.
3. Raises FileNotFoundError with the actual root cause.

So the post-resolve check in load_checkpoint was unreachable AND its error message lied about the failure mode (it said "not found locally or on HF Hub", but _resolve_checkpoint_path would already have raised the real reason). Removed:

path = self._resolve_checkpoint_path(path)
if not os.path.exists(path):
    raise FileNotFoundError(...)   # unreachable; misleading message

Same commit also collapses the redundant spec local in _resolve_checkpoint_path. parse_hf_checkpoint_path already strips the hf:// prefix internally, and the local-path os.path.exists check is gated on not force_hf, so passing path directly (with the prefix still attached on the HF branch) is safe. One fewer variable for the reader to track.

Diff

Net: -9/+8 in src/flow_factory/models/abc.py. No other files touched.

Test plan

• Pre-existing parser unit cases (8 happy + 5 error + 6 path-traversal) still pass — these refactors don't touch the parser.
• No new imports, no public API change, no config change.
• Single-GPU smoke: load LoRA from resume_path: "owner/repo" (will exercise the cleaned-up flow).

Out of scope

• ModelArguments.__post_init__'s expanduser call — kept; defensive normalization at config-parse time, idempotent with the resolver, removing it could surprise other readers of model_args.resume_path.

Made with Cursor

[Copilot]

Pull request overview

Small follow-up refactor to BaseAdapter checkpoint path handling. Moves os.path.expanduser into _resolve_checkpoint_path, removes a now-unreachable os.path.exists guard with a misleading error message in load_checkpoint, and collapses a redundant spec local variable since parse_hf_checkpoint_path strips the hf:// prefix internally. No behavioral change.

Changes:

• Hoist expanduser into _resolve_checkpoint_path so callers don't need to remember to normalize ~ separately.
• Drop the unreachable post-resolve os.path.exists check in load_checkpoint whose error message lied about the failure mode.
• Pass path directly to parse_hf_checkpoint_path (which already handles the hf:// prefix) and remove the redundant spec variable.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.