Skip to content

Record MossFormer2 Colab feasibility evidence - #66

Merged
Tinnci merged 1 commit into
mainfrom
agent/mossformer2-feasibility-spike
Jul 12, 2026
Merged

Record MossFormer2 Colab feasibility evidence#66
Tinnci merged 1 commit into
mainfrom
agent/mossformer2-feasibility-spike

Conversation

@Tinnci

@Tinnci Tinnci commented Jul 12, 2026

Copy link
Copy Markdown
Owner

Summary

  • add a Colab-only bounded MossFormer2_SR_48K feasibility workflow
  • prove the minimal checkpoint set and exact safetensors round trips
  • record offline CPU mono/stereo/native-48k behavior and alignment constraints
  • complete the P3 spike without implementing a backend

Remote evidence

  • source: modelscope/ClearerVoice-Studio@6b3774dc79c46ae8bed2a4fa5f706f0ac8c75c61
  • model: alibabasglab/MossFormer2_SR_48K@39eb1f25ea84f5e0315ade9ac0070fff216fc690
  • required files: pointer + m + g; 1.74 GB training file not downloaded
  • checkpoint states: 929 tensors / 54,497,666 parameters and 311 tensors / 55,150,402 parameters
  • both state dicts load with weights_only=True and exactly round-trip through safetensors
  • CPU inference remained offline after explicit download
  • 16 kHz stereo preserves two channels
  • 16 kHz and native 48 kHz fixtures both expose a -224 sample output alignment delta
  • repeated same-process inference is bit-exact; cross-process parity requires tolerances

All checkpoint, model, audio, and HPE work ran via Colab CLI.

Local verification

  • git diff --check
  • pixi run lint
  • pixi run format-check
  • pixi run typecheck
  • pixi run test (163 passed, 6 skipped)

@Tinnci
Tinnci marked this pull request as ready for review July 12, 2026 09:49
Copilot AI review requested due to automatic review settings July 12, 2026 09:49
@Tinnci
Tinnci merged commit 6846423 into main Jul 12, 2026
3 checks passed
@Tinnci
Tinnci deleted the agent/mossformer2-feasibility-spike branch July 12, 2026 09:49

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR records the completed MossFormer2_SR_48K feasibility spike (as Colab-only evidence) without implementing a new backend, and updates roadmap/candidate documentation to reflect the results and constraints (minimal checkpoints, safe conversion, offline CPU inference, and alignment/parity considerations).

Changes:

  • Adds a Colab-only spike script to reproduce checkpoint download/inspection, safetensors round-trip validation, and offline CPU inference evidence.
  • Updates roadmap and backend-candidate documentation to mark the spike as complete and capture key findings (minimal files, alignment delta, tolerance-based parity).
  • Documents how to run/download the spike evidence via the existing Colab CLI workflow and notes it in the changelog.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file
File Description
ROADMAP.md Marks the candidate spike milestone as completed and summarizes the spike outcomes/constraints.
examples/colab_mossformer2_spike.py New Colab-only script that reproduces and archives the MossFormer2 feasibility evidence.
docs/BACKEND_CANDIDATES.md Updates the MossFormer2 candidate entry with completed feasibility evidence and implementation constraints.
docs/ACCELERATORS.md Adds Colab CLI instructions for running/downloading the spike evidence archive.
CHANGELOG.md Notes the addition of the MossFormer2 feasibility workflow under Unreleased.

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

Comment thread ROADMAP.md
The evaluation policy lives in [docs/EVALUATION.md](docs/EVALUATION.md).

## Priority 3: One Candidate Spike, Not Three Implementations
## Completed: One Candidate Spike, Not Three Implementations
Comment on lines +61 to +65
def download_checkpoints() -> list[dict[str, object]]:
from huggingface_hub import hf_hub_download

CHECKPOINT_DIR.mkdir(parents=True, exist_ok=True)
records = []
Comment on lines +91 to +92
conversion_dir = EVIDENCE_DIR / "converted"
conversion_dir.mkdir()
Comment on lines +265 to +276
run(
"install-dependencies",
[
"uv",
"pip",
"install",
"--system",
"-e",
str(SOURCE_DIR / "clearvoice"),
"safetensors",
],
)
Comment on lines +223 to +226
def archive() -> None:
with tarfile.open(ARCHIVE_PATH, "w:gz") as bundle:
bundle.add(EVIDENCE_DIR, arcname="evidence")
print(f"[mossformer2-spike] evidence archive: {ARCHIVE_PATH}", flush=True)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants