migrate

name: migrate version: 1.4.0 description: >- This skill should be used when the user wants to migrate, transfer, copy, or mirror repos between HuggingFace and ModelScope using Modal. Triggers on "migrate model to ModelScope", "transfer dataset to HuggingFace", "copy from HF to MS", "mirror repo", "move to modelscope", "batch migrate multiple repos", "parallel migration", "detached migration", "fire and forget", "verify migration integrity", or "clean up org after migration".

HF-Modal-ModelScope Migration

Migrate repos between HuggingFace and ModelScope using Modal as an ephemeral cloud compute bridge. No files touch the local machine — everything transfers cloud-to-cloud.

Architecture

Local Machine              Modal Container              Platforms
┌──────────┐  modal run  ┌─────────────────┐  API     ┌──────────┐
│ Claude   │ ──────────> │ snapshot_download│ <──────> │ HF Hub   │
│ Code     │             │   OR git clone   │          │          │
│          │             │ upload_folder    │ <──────> │ MS Hub   │
└──────────┘             └─────────────────┘          └──────────┘

Each migration spins up a fresh Modal container, transfers files via platform SDKs, then destroys the container. No persistent storage between runs.

Supported Directions

• HuggingFace -> ModelScope: Download via huggingface_hub.snapshot_download (auto-falls back to git clone + git lfs pull if API returns 403). Upload via ModelScope HubApi.upload_folder()
• ModelScope -> HuggingFace: Download via modelscope.hub.snapshot_download, upload via HfApi.upload_folder()

Migration Modes

Standard (default)

Single container downloads the entire repo, then uploads it. Tested up to 58.5 GB; 24-hour timeout supports larger repos. Use --parallel for anything over ~50 GB.

Parallel (--parallel)

Splits the repo into chunks, each processed by an independent container. Up to 100 containers run concurrently. Each chunk worker: clones repo structure (no LFS data) -> selectively pulls its assigned LFS files -> uploads to destination. Best for repos over 50 GB.

                    ┌─ Container 0: clone + pull chunk 0 + upload ─┐
HuggingFace ────────┼─ Container 1: clone + pull chunk 1 + upload ─┼──── ModelScope
  (source)          ├─ Container 2: clone + pull chunk 2 + upload ─┤    (dest)
                    └─ ...up to 100 concurrent containers...       ┘

Guardrails:

• Chunk count auto-capped at 100 (chunk size increased automatically for very large repos)
• Repos with >500K files get a warning (git clone overhead per chunk)
• Post-upload verification compares file count and size against source manifest

Built-in Safety & Reliability

These features run automatically — no flags needed:

• Auto-fallback to git clone: If HuggingFace's API returns 403 (org storage limit lockout), the script automatically retries via git clone + git lfs pull.
• Destination validation (fail-fast): Before downloading, validates the destination namespace exists. Catches typos before wasting time on a download.
• Visibility preservation: Private repos stay private on the destination. Source visibility is auto-detected and mapped.
• SHA256 verification: Every LFS file is hash-checked after upload. Skipped files (no extractable hash) are reported separately.
• Download progress monitoring: For git-based downloads, a background thread monitors directory size and prints real-time progress.
• Size estimation: Before starting, estimates migration duration based on benchmark data and prints an ETA.
• 24-hour timeout: All migration functions have an 86400s (24h) timeout. Tested up to 58.5 GB single-container; use --parallel for larger repos.

Supported Repo Types

Prerequisites

Three sets of credentials must be available as environment variables:

Tokens can be set in the shell or placed in ${CLAUDE_PLUGIN_ROOT}/.env — the validation script and /migrate command auto-load this file. Install huggingface_hub and modelscope locally for token validation. The migration itself runs entirely on Modal.

Executing a Migration

Use the /migrate command for the guided interactive workflow. It handles token validation, parameter extraction, user confirmation, and execution.

For direct CLI usage, all commands below use this prefix (sources .env and prevents Unicode errors on Windows):

PREFIX="set -a && source \"${CLAUDE_PLUGIN_ROOT}/.env\" 2>/dev/null; set +a; PYTHONIOENCODING=utf-8"

Single repo:

eval $PREFIX modal run "${CLAUDE_PLUGIN_ROOT}/scripts/modal_migrate.py::main" --source "username/my-model" --to ms

Parallel chunked (large repos):

eval $PREFIX modal run "${CLAUDE_PLUGIN_ROOT}/scripts/modal_migrate.py::main" --source "org/big-dataset" --to ms --repo-type dataset --parallel --chunk-size 30

Batch (one container per repo):

eval $PREFIX modal run "${CLAUDE_PLUGIN_ROOT}/scripts/modal_migrate.py::batch" --source "user/model1,user/model2" --to ms --repo-type model

Detached (fire & forget):

eval $PREFIX modal run --detach "${CLAUDE_PLUGIN_ROOT}/scripts/modal_migrate.py::main" --source "username/my-model" --to ms

Key Differences: Single vs Batch

Batch mode uses the source repo ID as the destination. If the HF org name differs from the MS org name, use individual ::main runs with --dest.

Detached Mode (Fire & Forget)

Add --detach before the script path. The migration continues in Modal's cloud after the session ends. Monitor with modal app logs hf-ms-migrate or the Modal dashboard.

Edge Cases

• Repo already exists on destination: Single mode proceeds with a warning (files are updated/overwritten). Batch mode auto-skips existing repos.
• Private source repo: Works if the source token has read access. Visibility is preserved.
• 403 Forbidden (storage limit lockout): Auto-detected. Falls back to git clone + git lfs pull.
• Spaces to ModelScope: Skipped with a warning. To force migration as a model repo, use --repo-type model (only files transfer, not the app runtime).
• Large repos (>50GB): Use --parallel. Tested up to 58.5 GB single-container and 3.3 TB parallel (113 chunks).
• Very large repos (>1TB): Use --parallel. Chunk size auto-adjusts to keep within 100 chunks (e.g., 3.3 TB -> ~30 GB chunks -> 113 chunks, processed in waves of up to 100 concurrent containers).
• Batch with mismatched namespaces: Batch mode does not support --dest. If source org differs from dest org, use individual ::main runs with --dest.

Troubleshooting

Verification & Org Cleanup

For SHA256 cross-platform verification and org cleanup workflows (inventory, verify, migrate-then-delete), read: ${CLAUDE_PLUGIN_ROOT}/skills/migrate/references/verification-and-cleanup.md

Scripts Reference

SDK Reference

Before searching the web for HuggingFace or ModelScope SDK methods, read the bundled reference file. It contains complete Python SDK signatures for both platforms (list, info, create, download, upload, files, branches) and a key differences table.

Read first: ${CLAUDE_PLUGIN_ROOT}/skills/migrate/references/hub-api-reference.md