Fix push-hf command in CLI

[cmpatino]

This PR fixes two issues when using the push-hf command in the CLI:

1. Fixes how we dealt with the checkpoint_complete file because the previous version conflicted with the Click CLI.
2. The previous version required having the revision already available in the Hugging Face Hub. We now create the revision if it's not already available.

[kashif]

Suggested change

	checkpoint_complete_file = extract_dir / "checkpoint_complete"
	checkpoint_complete_file.unlink(missing_ok=True)
	checkpoint_complete.unlink(missing_ok=True)

[cmpatino]

Improved push-hf Command

Changes

1. Main Branch Initialization with Real Content

• Before: New repos had main branch with just a minimal README
• Now: Main branch gets the actual checkpoint content

2. Added --overwrite Flag (default: False)

• Prevents accidental overwrites of existing branches
• Requires explicit --overwrite flag to replace existing content

Key Scenarios on Revision Logic

New repo, no revision:

tinker checkpoint push-hf tinker://run/weights/0001 --repo user/model

→ Uploads to main

New repo with revision:

tinker checkpoint push-hf tinker://run/weights/0001 --repo user/model --revision step-0001

→ Uploads to main, creates step-0001 branch from main

Existing branch without --overwrite:

tinker checkpoint push-hf tinker://run/weights/0002 --repo user/model

→ Error: "Branch 'main' already exists. Use --overwrite to replace..."

Existing branch with --overwrite:

tinker checkpoint push-hf tinker://run/weights/0002 --repo user/model --overwrite

→ Overwrites main branch

New revision in existing repo:

tinker checkpoint push-hf tinker://run/weights/0002 --repo user/model --revision step-0002

→ Creates and uploads to step-0002 (main unchanged)

Benefits

• Main branch always has usable model content
• Prevents accidental overwrites
• More efficient branching (no duplicate uploads)

[kashif]

lgtm