monster/kimodo-mrp.ccc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Publish Kimodo GGUF bundle workflow

Browse Source
This commit is contained in:
Richard Palethorpe 2026-08-23 08:46:52 +01:00
parent 2558baec65
commit e3eb09fec6
5 changed files with 319 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
README.md
View File

@ -51,7 +51,7 @@ translations and local XYZW rotations.
## Demo
After building the debug preset and converting the text bundle:
After building the debug preset and downloading the native GGUF bundle:
```sh
go run ./demo -addr 0.0.0.0:8094
@ -61,11 +61,28 @@ Open `http://localhost:8094`. The left sidebar contains the prompt and a
persistent history; choosing a previous animation restores its prompt for a
new generation.
## Licensed weights
## Weights
The SMPL-X checkpoint and Llama base model are gated. After accepting their
Hugging Face licences and authenticating, download the exact revisions and
hash manifests with:
The ready-to-run native GGUF bundle is published under the Hugging Face
`LocalAI-io` organisation (not GitHub's `localai-org`). Download it directly;
this avoids recreating the conversion locally:
```sh
nix develop path:. --command scripts/download_gguf_weights.sh --output "$PWD"
```
The installer verifies the published manifest and SHA-256 hashes. Use
`--motion-only` when supplying a precomputed 4096-float LLM2Vec embedding.
The GGUF bundle includes converted Meta Llama 3 material and Kimodo is
non-commercial research-only. Review the published model card and upstream
licences before downloading or redistributing.
### Regenerating the bundle
This is only needed to reproduce a conversion. The SMPL-X checkpoint and Llama
base model are gated. After accepting their Hugging Face licences and
authenticating, download the exact revisions and hash manifests with:
```sh
nix develop path:. --command hf auth login
@ -79,3 +96,12 @@ Convert the local LLM2Vec model to the native component bundle with:
nix develop path:. --command scripts/convert_llm2vec_bundle.sh \
"$PWD/models/llama3-8b-instruct-base" "$PWD/generated/llm2vec-text-bundle"
```
Validate a prospective release without network access, then explicitly upload
it from an account allowed to publish to `LocalAI-io`:
```sh
nix develop path:. --command python scripts/publish_gguf.py
nix develop path:. --command python scripts/publish_gguf.py \
--upload --confirm-upstream-licences
```

68
scripts/download_gguf_weights.sh Executable file
View File

@ -0,0 +1,68 @@
#!/usr/bin/env bash
# Download the published native GGUF bundle into Kimodo's standard paths.
set -euo pipefail
export HF_HUB_DISABLE_PROGRESS_BARS=1
ORG="${GGUF_ORG:-LocalAI-io}"
REPO_DEFAULT="$ORG/Llama-3-Kimodo-SMPLX-RP-v1-GGUF"
usage() {
printf '%s\n' "usage: $0 --output DIR [--repo HF_REPO] [--revision REVISION] [--motion-only]" >&2
exit 2
}
output='' repo="$REPO_DEFAULT" revision='main' motion_only=0
while [ "$#" -gt 0 ]; do
case "$1" in
--output) [ "$#" -ge 2 ] || usage; output=$2; shift 2 ;;
--repo) [ "$#" -ge 2 ] || usage; repo=$2; shift 2 ;;
--revision) [ "$#" -ge 2 ] || usage; revision=$2; shift 2 ;;
--motion-only) motion_only=1; shift ;;
*) usage ;;
esac
done
[ -n "$output" ] || usage
command -v hf >/dev/null || { echo "hf not found; enter the Nix shell first" >&2; exit 1; }
mkdir -p "$output"
patterns=("MANIFEST.json" "SHA256SUMS" "models/kimodo-smplx-rp-v1-f32.gguf")
if [ "$motion_only" -eq 0 ]; then
patterns+=("generated/llm2vec-text-bundle/*")
fi
echo "Downloading $repo at $revision into $output"
args=(download "$repo" --revision "$revision" --local-dir "$output")
for pattern in "${patterns[@]}"; do args+=(--include "$pattern"); done
hf "${args[@]}" >/dev/null
manifest="$output/MANIFEST.json"
[ -f "$manifest" ] || { echo "missing MANIFEST.json from $repo" >&2; exit 1; }
python - "$manifest" "$output" "$motion_only" <<'PY'
import hashlib
import json
import sys
from pathlib import Path
manifest_path = Path(sys.argv[1])
output = Path(sys.argv[2])
motion_only = sys.argv[3] == "1"
root = output.resolve()
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
if manifest.get("format") != "kimodo-gguf-manifest-v1":
raise SystemExit("unsupported or malformed GGUF manifest")
for entry in manifest.get("files", []):
relative = Path(entry.get("path", ""))
if relative.is_absolute() or ".." in relative.parts or relative.suffix != ".gguf":
raise SystemExit(f"unsafe manifest path: {relative}")
if motion_only and str(relative).startswith("generated/"):
continue
path = root / relative
if not path.is_file() or path.stat().st_size != entry.get("bytes"):
raise SystemExit(f"missing or wrong-sized file: {relative}")
h = hashlib.sha256()
with path.open("rb") as f:
for chunk in iter(lambda: f.read(1 << 20), b""):
h.update(chunk)
if h.hexdigest() != entry.get("sha256"):
raise SystemExit(f"checksum mismatch: {relative}")
print("verified native Kimodo GGUF bundle")
PY

7
scripts/hf/Llama-3-Kimodo-SMPLX-RP-v1-GGUF/NOTICE Normal file
View File

@ -0,0 +1,7 @@
Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
Built with Meta Llama 3.
Kimodo-SMPLX-RP-v1 source model: NVIDIA. This converted distribution remains
subject to the NVIDIA Internal Scientific Research and Development Model License
and is for non-commercial research use only.

66
scripts/hf/Llama-3-Kimodo-SMPLX-RP-v1-GGUF/README.md Normal file
View File

@ -0,0 +1,66 @@
---
license: other
library_name: ggml
tags:
- gguf
- ggml
- text-to-motion
- smplx
- llama-3
---
# Llama-3-Kimodo-SMPLX-RP-v1-GGUF
Native F32 GGUF conversion of NVIDIA's Kimodo-SMPLX-RP-v1 motion model and
its LLM2Vec text encoder. It is for the `kimodo.cpp` GGML runtime; it is not a
llama.cpp language-model conversion.
`Kimodo-SMPLX-RP-v1` is restricted to non-commercial research use. Before use
or redistribution, review and comply with the upstream NVIDIA model terms and
the Meta Llama 3 Community License below. This repository does not grant rights
beyond those upstream licences.
## Install
From the kimodo.cpp checkout:
```sh
nix develop path:. --command scripts/download_gguf_weights.sh --output "$PWD"
```
This places the motion model at `models/kimodo-smplx-rp-v1-f32.gguf` and text
components under `generated/llm2vec-text-bundle/`, which are the default paths
used by the library and demo. Add `--motion-only` for embedding-only inference.
`SHA256SUMS` and `MANIFEST.json` record every published artifact and its source
revision; the installer verifies the selected files after download.
## Contents
- `models/kimodo-smplx-rp-v1-f32.gguf` — Kimodo motion diffusion model (F32)
- `generated/llm2vec-text-bundle/` — tokenizer, embeddings, final norm, and 32
F32 transformer layers for native LLM2Vec inference
The text components are split deliberately: kimodo.cpp loads a bounded number
of layers at once for GPU memory control.
## Provenance
The conversion is generated by `kimodo.cpp` from these exact upstream commits:
- NVIDIA Kimodo-SMPLX-RP-v1: `1419ba56b734c48bbafb41fefa84088ca94583b5`
- Meta Llama-3-8B-Instruct: `8afb486c1db24fe5011ec46dfbe5b5dccdb575c2`
- McGill LLM2Vec MNTP adapter: `31474e395ada192e8ed1586db6be79fb3b70c9c0`
- McGill LLM2Vec supervised adapter: `baa8ebf04a1c2500e61288e7dad65e8ae42601a7`
The two McGill adapters are MIT-licensed. The merged text encoder includes
Meta Llama 3 material. **Built with Meta Llama 3.**
## Licences and notices
- Kimodo: [NVIDIA Internal Scientific Research and Development Model License](https://huggingface.co/nvidia/Kimodo-SMPLX-RP-v1)
(non-commercial research only).
- Text base: [Meta Llama 3 Community License](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct).
- LLM2Vec adapters: [MIT](https://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp).
`LICENSE-META-LLAMA-3.txt` and `NOTICE` accompany every published copy with
the required Meta licence and attribution.

147
scripts/publish_gguf.py Executable file
View File

@ -0,0 +1,147 @@
#!/usr/bin/env python3
"""Publish the reproducible Kimodo GGUF distribution to Hugging Face.
The default is deliberately a dry run: it validates the exact converter
outputs, prints every path, size and SHA-256, and performs no network I/O.
Use --upload only after reviewing the upstream licence obligations.
The text bundle contains merged Meta Llama 3 weights. It is therefore kept in
the same repository as the Kimodo motion GGUF and is published under the model
name required by the Meta Llama 3 Community License:
LocalAI-io/Llama-3-Kimodo-SMPLX-RP-v1-GGUF
"""
from __future__ import annotations
import argparse
import hashlib
import io
import json
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
HF_ORG = "LocalAI-io" # Hugging Face organisation; GitHub is localai-org.
DEFAULT_REPO = f"{HF_ORG}/Llama-3-Kimodo-SMPLX-RP-v1-GGUF"
MOTION_NAME = "kimodo-smplx-rp-v1-f32.gguf"
TEXT_NAMES = (
"tokenizer.gguf", "embedding.gguf", "final-norm.gguf",
*(f"layer-{index:02d}.gguf" for index in range(32)),
)
SOURCE_REVISIONS = {
"nvidia/Kimodo-SMPLX-RP-v1": "1419ba56b734c48bbafb41fefa84088ca94583b5",
"meta-llama/Meta-Llama-3-8B-Instruct": "8afb486c1db24fe5011ec46dfbe5b5dccdb575c2",
"McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp": "31474e395ada192e8ed1586db6be79fb3b70c9c0",
"McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised": "baa8ebf04a1c2500e61288e7dad65e8ae42601a7",
}
CARD = ROOT / "scripts/hf/Llama-3-Kimodo-SMPLX-RP-v1-GGUF/README.md"
NOTICE = ROOT / "scripts/hf/Llama-3-Kimodo-SMPLX-RP-v1-GGUF/NOTICE"
LLAMA_LICENSE = ROOT / "models/llama3-8b-instruct-base/LICENSE"
def digest(path: Path) -> str:
value = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
value.update(chunk)
return value.hexdigest()
def require_revision(repo: str) -> None:
revision = ROOT / "models" / {
"nvidia/Kimodo-SMPLX-RP-v1": "Kimodo-SMPLX-RP-v1",
"meta-llama/Meta-Llama-3-8B-Instruct": "llama3-8b-instruct-base",
"McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp": "llm2vec-mntp-adapter",
"McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised": "llm2vec-adapter",
}[repo] / "REVISION"
if not revision.is_file():
raise ValueError(f"missing provenance file: {revision}")
actual = revision.read_text(encoding="utf-8").split()[0]
if actual != SOURCE_REVISIONS[repo]:
raise ValueError(f"unexpected {repo} revision: {actual} (expected {SOURCE_REVISIONS[repo]})")
def artifacts(motion: Path, bundle: Path) -> list[tuple[Path, str]]:
result = [(motion, f"models/{MOTION_NAME}")]
result.extend((bundle / name, f"generated/llm2vec-text-bundle/{name}") for name in TEXT_NAMES)
for source, destination in result:
if not source.is_file() or source.stat().st_size == 0:
raise ValueError(f"missing or empty GGUF: {source}")
if source.suffix != ".gguf":
raise ValueError(f"not a GGUF: {source}")
with source.open("rb") as handle:
if handle.read(4) != b"GGUF":
raise ValueError(f"invalid GGUF magic: {source}")
if ".." in Path(destination).parts:
raise ValueError(f"unsafe destination: {destination}")
return result
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--motion", type=Path, default=ROOT / "models" / MOTION_NAME)
parser.add_argument("--text-bundle", type=Path,
default=ROOT / "generated/llm2vec-text-bundle")
parser.add_argument("--repo", default=DEFAULT_REPO)
parser.add_argument("--upload", action="store_true",
help="actually create/update the Hugging Face model repo")
parser.add_argument("--confirm-upstream-licences", action="store_true",
help="required with --upload; confirms authority to redistribute all inputs")
args = parser.parse_args()
try:
if not CARD.is_file() or not NOTICE.is_file() or not LLAMA_LICENSE.is_file():
raise ValueError("model card, NOTICE, or Meta Llama 3 licence is missing")
for repo in SOURCE_REVISIONS:
require_revision(repo)
files = artifacts(args.motion, args.text_bundle)
except ValueError as error:
print(f"error: {error}", file=sys.stderr)
return 1
entries = []
for source, destination in files:
entries.append({"path": destination, "bytes": source.stat().st_size, "sha256": digest(source)})
manifest = {
"format": "kimodo-gguf-manifest-v1",
"repository": args.repo,
"source_revisions": SOURCE_REVISIONS,
"files": entries,
}
sums = "".join(f"{entry['sha256']} {entry['path']}\n" for entry in entries)
total = sum(entry["bytes"] for entry in entries)
print(f"repo: https://huggingface.co/{args.repo}")
print(f"files: {len(entries)} GGUFs, {total / 1e9:.2f} GB")
for entry in entries:
print(f" {entry['sha256']} {entry['bytes']:>12} {entry['path']}")
if not args.upload:
print("\n[dry-run] nothing uploaded. Re-run with --upload --confirm-upstream-licences to publish.")
return 0
if not args.confirm_upstream_licences:
print("error: --upload requires --confirm-upstream-licences", file=sys.stderr)
return 2
from huggingface_hub import HfApi
api = HfApi()
api.create_repo(args.repo, repo_type="model", exist_ok=True)
uploads = [(CARD, "README.md"), (NOTICE, "NOTICE"),
(LLAMA_LICENSE, "LICENSE-META-LLAMA-3.txt")]
for source, destination in uploads + files:
print(f"uploading {destination} ...", flush=True)
api.upload_file(path_or_fileobj=str(source), path_in_repo=destination,
repo_id=args.repo, repo_type="model",
commit_message=f"Add {destination}")
for payload, destination in ((json.dumps(manifest, indent=2, sort_keys=True).encode() + b"\n", "MANIFEST.json"),
(sums.encode(), "SHA256SUMS")):
print(f"uploading {destination} ...", flush=True)
api.upload_file(path_or_fileobj=io.BytesIO(payload), path_in_repo=destination,
repo_id=args.repo, repo_type="model",
commit_message=f"Add {destination}")
print(f"done -> https://huggingface.co/{args.repo}")
return 0
if __name__ == "__main__":
sys.exit(main())