Generated banner for Kermt Embed
Generated logo for Kermt Embed

Kermt Embed

NVIDIAOfficialPublished by NVIDIA

  • Skill
  • Workflow
  • Science and research
  • Apache-2.0
  • Medium risk
  • Has scripts

From the author

Extract per-molecule embeddings from any encoder-bearing KERMT checkpoint. Use a local checkpoint or optionally download a pinned Hugging Face model bundle using HF_TOKEN if configured. Run containerized embedding extraction and write model bundles, per-readout .npy embeddings, canonical SMILES, and validity arrays to user-selected host directories.
— the description by NVIDIA

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The kermt-embed skill provides a structured, containerized workflow for generating per-molecule embeddings from KERMT model checkpoints. It manages the entire pipeline, including system validation, data preparation, checkpoint verification, and execution. The skill supports both custom checkpoints and the automatic, consent-based download of pretrained NVIDIA KERMT models. It outputs standardized .npy embedding files for four distinct readouts, along with canonical SMILES and validity metadata, ensuring consistent results for downstream chemical analysis.

What it does

  • Validates system requirements and container environment for embedding extraction.
  • Automates the download and verification of pretrained KERMT model bundles.
  • Cleans and prepares input SMILES data from CSV files for model processing.
  • Executes embedding extraction on a single GPU using specified or default checkpoints.
  • Generates standardized .npy embedding arrays and metadata for molecular analysis.
  • Provides replayable run manifests for reproducibility.

When to use it

  • When you need to generate numerical vector representations for a set of molecules.
  • When you want to use a pretrained KERMT model to featurize chemical data.
  • When you have a custom encoder-bearing KERMT checkpoint and need to extract embeddings from it.
  • When you require a reproducible, containerized pipeline for molecular embedding tasks.

Advantages

  • Ensures consistent environment configuration via containerization.
  • Automates complex multi-step validation of checkpoints and input data.
  • Provides clear, standardized output formats for downstream integration.
  • Includes built-in safety checks for model downloads and hardware compatibility.

Usage examples

“Extract embeddings for the molecules in my_molecules.csv using the default pretrained model.”
The agent will prompt for consent to download the NV-KERMT-70M-v2 model, validate the CSV, and generate embedding files in a new run directory.
“Run embedding extraction on molecules.csv using my local checkpoint at /path/to/model.pt.”
The agent validates the local checkpoint and input CSV, then executes the extraction workflow using the provided model file.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/nvidia/kermt-embed/download
  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip kermt-embed.zip -d .agents/skills

Documentation by NVIDIA

kermt-embed

Extract per-molecule embeddings from any encoder-bearing KERMT checkpoint. The skill is the workflow orchestrator: validate ckpt, validate CSV, clean SMILES, launch the runner blocking, return the per-readout .npy files.

Skill and runtime paths

Set SKILL_DIR to the absolute path of this installed skill directory. Export KERMT_REPO as the absolute path to the KERMT checkout used for model execution. The bundled container helper mounts that checkout at /workspace and this skill at /skill (read-only). Commands inside the container use /skill/scripts/; defaults are bundled in config/. See Released models for checkpoint bundle requirements.

Downloads and local outputs

The optional released-model branch reads config/released_model.json for the Hugging Face repository, pinned revision, and filenames. The bundled scripts/fetch_released_model.py downloads the model bundle over HTTPS into the host directory the user selects. Public models work without credentials; if HF_TOKEN is set, the container helper forwards it for Hugging Face authentication. Prepared data, logs, and workflow results go into the chosen run directory.

Hardware requirements

  • GPUs: 1 (single-GPU).

  • VRAM: ≥ 4 GB for the default batch_size 64.

  • Disk: depends on output size — roughly a few MB per 1k molecules at hidden 800 per readout, so ~10–20 MB per 1k molecules across the 4 readouts. Plus a small canonical_smiles.npy + validity.npy per run.

  • Driver / CUDA: any host supporting CUDA 12.6.

Inputs

Required:

  • --csv <path> — SMILES CSV. First column is smiles; other columns are ignored (no targets needed).

Checkpoint (optional — defaults to the released model if omitted):

  • --ckpt <path> — any encoder-bearing checkpoint. Grover_base, cmim, hybrid, and finetuned ckpts are all accepted. The validator only refuses ckpts with no encoder. If omitted, the skill offers to download the released pretrained hybrid model nvidia/NV-KERMT-70M-v2 and embed with it — see "Resolve & validate the checkpoint" (workflow step 3).

  • --pretrained-release — explicit opt-in to use the released model without the interactive prompt (for non-interactive / agent runs). Mutually exclusive with --ckpt.

  • --model-dir <dir> — where to save the downloaded bundle (default $KERMT_REPO/models/NV-KERMT-70M-v2/). An already-complete bundle there is reused, not re-downloaded.

Optional:

  • --batch-size N — override the configured default (64).

  • --gpus 0 — single GPU id (default 0).

  • --from-prepare <dir> — skip the prepare step and reuse an existing prepare_data.json in <dir>.

Workflow

Let $KERMT_REPO be the path to your kermt repo checkout.

  1. Pre-flight: container + system probe.

    "$SKILL_DIR/scripts/kermt_container.sh" check_system
  2. Compute run directory.

    RUN_DIR=$KERMT_REPO/runs/embed_$(date -u +%Y-%m-%dT%H-%M-%SZ)
  3. Resolve & validate the checkpoint.

    Resolve — only if --ckpt was omitted. Default to the released pretrained hybrid model nvidia/NV-KERMT-70M-v2:

    • Consent gate. Unless --pretrained-release was passed, ask the user: "No checkpoint given — download the released model nvidia/NV-KERMT-70M-v2 (NVIDIA Open Model License, https://huggingface.co/nvidia/NV-KERMT-70M-v2) and embed with it? [y/N]". Never download without an explicit yes (or --pretrained-release). If both --ckpt and --pretrained-release are given, abort — they conflict.

    • Save location. Default $KERMT_REPO/models/NV-KERMT-70M-v2/; honor --model-dir <dir> if given. An already-complete bundle is reused.

    • Download (foreground; ~282 MB on first fetch):

      "$SKILL_DIR/scripts/kermt_container.sh" run --model-dir <save-dir> -- \
          "python /skill/scripts/fetch_released_model.py --out /model"

      Parse the JSON; abort on ok: false (surface errors). On success set <user-ckpt> = <save-dir>/kermt_contrastive_v2.0.pt.

    Validate the resolved (or user-provided) ckpt:

    "$SKILL_DIR/scripts/kermt_container.sh" run --ckpt <user-ckpt> -- \
        "python /skill/scripts/check_checkpoint.py --mode embed --ckpt /ckpt"

    Parse JSON. Abort on ok: false. The validator only refuses encoder-less ckpts (rare).

  4. Validate the data.

    "$SKILL_DIR/scripts/kermt_container.sh" run --data <user-csv> -- \
        "python /skill/scripts/check_data.py --mode embed --csv /data/<basename>"
  5. Prepare the data (clean-only — no features step).

    "$SKILL_DIR/scripts/kermt_container.sh" run --data <user-csv> --run-dir $RUN_DIR -- \
        "python /skill/scripts/prepare_data.py --mode embed \\
             --csv /data/<basename> --out /runs/data"

    Outputs land at $RUN_DIR/data/prepare_data.json with a single clean_csv path. task/extract_embeddings.py featurizes from SMILES on the fly.

  6. Launch the runner (blocking).

    "$SKILL_DIR/scripts/kermt_container.sh" run \\
        --ckpt <user-ckpt> --run-dir $RUN_DIR -- \\
        "python /skill/scripts/run_extract_embeddings.py \\
             --ckpt /ckpt \\
             --prepare-manifest /runs/data/prepare_data.json \\
             --out /runs \\
             [--gpus 0 --batch-size N]"
  7. Report to the user.

    • Embeddings directory: $RUN_DIR/out/

      • atom_from_atom.npy, bond_from_atom.npy, atom_from_bond.npy, bond_from_bond.npy (the 4 standard readouts; each shape (N_rows, hidden_size))

      • metadata.pkl — pickle of a dict containing canonical_smiles (RDKit-canonicalized SMILES per row), valid (boolean per-row: did RDKit parse it), plus other run metadata.

    • Manifest: $RUN_DIR/run.json

    • Log: $RUN_DIR/logs/embed.log

Hard rules

  • Never download the released model without consent. When --ckpt is omitted, download nvidia/NV-KERMT-70M-v2 only after an explicit user "yes" or an explicit --pretrained-release flag. --ckpt and --pretrained-release are mutually exclusive.

  • Never modify the user's ckpt. The runner reads-only via task/extract_embeddings.py's --checkpoint <path> flag.

  • Arch comes from the ckpt. No --hidden-size flag etc. on this runner; task/extract_embeddings.py reads arch from the ckpt's saved_args.

Common errors

  • prepare_data manifest is missing required output 'clean_csv' → prepare ran with --skip-clean but no source CSV given. Re-run prepare without it.

  • --gpus '0,1' is single-GPU only → pass a single id.

Replayability

$(jq -r .cmd_replay $RUN_DIR/run.json)

If ok_to_replay: false (dirty kermt repo worktree at launch time), pin the commit via repo.commit and git checkout it first.

Files

scripts/fetch_released_model.py

#!/usr/bin/env python3
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

"""Download a released KERMT model bundle from Hugging Face.

Runs INSIDE the kermt container (huggingface_hub is part of the image env).
Writes the released-model directory bundle — the checkpoint plus its vocab
files — into the directory mounted at `--out` (the skills mount the user's
chosen save location there via `kermt_container.sh --model-dir <host>`), then
emits a single JSON object to stdout that the calling skill parses.

The downloaded directory is exactly the repo's "released model bundle" layout
(see skills/README.md "Released models"): `<ckpt>.pt` + the three
`pretrain_*_vocab.*` files in one flat directory. The downstream skill then
feeds it through the existing `--ckpt <out>/<ckpt_name>` flow; for
continue-pretrain the bundled vocab files are auto-detected in the ckpt's
parent directory. No runner changes are needed.

Defaults (repo id, pinned revision, ckpt + vocab filenames) come from
`config/released_model.json` so the pin lives in one place; every value
is overridable on the CLI.

Idempotent: if the bundle is already complete in `--out` (ckpt + all vocab
files present), nothing is downloaded and `reused: true` is reported — so a
re-invocation never re-fetches the 282 MB checkpoint.

Authentication: the repo is public (no token needed). If `HF_TOKEN` is set in
the environment (forwarded into the container by `kermt_container.sh`),
huggingface_hub picks it up automatically — useful against shared-IP rate
limits or if the repo is ever gated.

Output (stdout)
---------------
{
  "ok": true | false,
  "repo_id": str,
  "revision": str,
  "out": str,                 // container path of the bundle dir (e.g. /model)
  "ckpt": str | null,         // container path of the checkpoint file
  "vocab_dir": str | null,    // == out (where the vocab files live)
  "ckpt_name": str,
  "vocab_files": [str, ...],
  "files_present": [str, ...],
  "ckpt_bytes": int | null,
  "reused": bool,             // true if the bundle already existed (no download)
  "license": str | null,
  "license_url": str | null,
  "errors": [str, ...]
}

Exit code: 0 on `ok: true`, 1 on `ok: false`.

CLI
---
    fetch_released_model.py [--out /model]
                            [--repo-id <id>] [--revision <sha|tag|branch>]
                            [--ckpt-name <name>] [--config <path>]
"""

from __future__ import annotations

import argparse
import json
import sys
import traceback
from pathlib import Path
from typing import Any

# Default config lives at config/released_model.json (one dir up from
# scripts/). Resolved relative to this file so the script is
# location-independent.
DEFAULT_CONFIG = (
    Path(__file__).resolve().parent.parent / "config" / "released_model.json"
)


def _load_config(config_path: Path) -> dict[str, Any]:
    if not config_path.is_file():
        raise FileNotFoundError(f"released-model config not found at {config_path}")
    return json.loads(config_path.read_text())


def fetch(
    *,
    out: Path,
    repo_id: str,
    revision: str,
    ckpt_name: str,
    vocab_files: list[str],
    license_name: str | None = None,
    license_url: str | None = None,
) -> dict[str, Any]:
    """Resolve-or-download the released bundle into `out`. Returns the manifest
    dict (never raises for the expected failure modes — they land in
    `errors[]` with `ok: false`)."""
    result: dict[str, Any] = {
        "ok": False,
        "repo_id": repo_id,
        "revision": revision,
        "out": str(out),
        "ckpt": None,
        "vocab_dir": None,
        "ckpt_name": ckpt_name,
        "vocab_files": list(vocab_files),
        "files_present": [],
        "ckpt_bytes": None,
        "reused": False,
        "license": license_name,
        "license_url": license_url,
        "errors": [],
    }

    required = [ckpt_name, *vocab_files]

    def _present() -> list[str]:
        return [name for name in required if (out / name).is_file()]

    # 1. Idempotent reuse — bundle already complete in `out`.
    if out.is_dir() and set(_present()) == set(required):
        result["reused"] = True
    else:
        # 2. Download. Import here so a stale image (missing huggingface_hub)
        #    surfaces a clean, actionable JSON error rather than a traceback.
        try:
            from huggingface_hub import snapshot_download
        except ImportError:
            result["errors"].append(
                "huggingface_hub is not available in the container image. The "
                "released-model download needs it; rebuild the image with "
                "`kermt-setup` (it now ships huggingface_hub) and retry."
            )
            return result

        out.mkdir(parents=True, exist_ok=True)
        try:
            # local_dir gives a flat copy (the bundle layout) rather than the
            # opaque blob/snapshot cache. HF_TOKEN, if set, is read by the lib.
            snapshot_download(repo_id=repo_id, revision=revision, local_dir=str(out))
        except Exception as exc:  # noqa: BLE001
            result["errors"].append(
                f"download failed for {repo_id}@{revision}: {type(exc).__name__}: {exc}"
            )
            return result

    # 3. Verify the bundle is complete regardless of download/reuse path.
    present = _present()
    result["files_present"] = present
    missing = [name for name in required if name not in present]
    if missing:
        result["errors"].append(
            f"bundle at {out} is missing expected file(s): {missing}. "
            f"Present: {present}."
        )
        return result

    ckpt_path = out / ckpt_name
    result["ckpt"] = str(ckpt_path)
    result["vocab_dir"] = str(out)
    try:
        result["ckpt_bytes"] = ckpt_path.stat().st_size
    except OSError:
        result["ckpt_bytes"] = None

    result["ok"] = True
    return result


def main(argv: list[str] | None = None) -> int:
    parser = argparse.ArgumentParser(
        description="Download a released KERMT model bundle from Hugging Face (runs in-container)."
    )
    parser.add_argument(
        "--out",
        default="/model",
        help="Directory to write the bundle into (default: /model, the --model-dir mount).",
    )
    parser.add_argument(
        "--config",
        default=str(DEFAULT_CONFIG),
        help="Path to released_model.json (default: config/released_model.json).",
    )
    parser.add_argument(
        "--repo-id", default=None, help="Override the HF repo id from the config."
    )
    parser.add_argument(
        "--revision",
        default=None,
        help="Override the pinned revision (sha/tag/branch).",
    )
    parser.add_argument(
        "--ckpt-name",
        default=None,
        help="Override the checkpoint filename from the config.",
    )
    args = parser.parse_args(argv)

    try:
        cfg = _load_config(Path(args.config))
        repo_id = args.repo_id or cfg["repo_id"]
        revision = args.revision or cfg["revision"]
        ckpt_name = args.ckpt_name or cfg["ckpt_name"]
        vocab_files = list(cfg.get("vocab_files", []))
        result = fetch(
            out=Path(args.out),
            repo_id=repo_id,
            revision=revision,
            ckpt_name=ckpt_name,
            vocab_files=vocab_files,
            license_name=cfg.get("license"),
            license_url=cfg.get("license_url"),
        )
    except Exception as exc:  # noqa: BLE001
        print(traceback.format_exc(), file=sys.stderr)
        print(
            json.dumps(
                {
                    "ok": False,
                    "out": args.out,
                    "errors": [
                        f"unhandled exception in fetch_released_model: {type(exc).__name__}: {exc}"
                    ],
                },
                indent=2,
            )
        )
        return 1

    print(json.dumps(result, indent=2))
    return 0 if result["ok"] else 1


if __name__ == "__main__":
    raise SystemExit(main())

Requirements and compatibility

Written for: Claude Code, OpenAI Codex

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python

The author says: Requires docker, nvidia-container-toolkit, and a CUDA-capable NVIDIA GPU. Designed for Claude Code, Codex, and Nemotron.

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by
NVIDIA
Publisher
NVIDIA (Official)
Source
NVIDIA/skills · skills/bionemo-kermt-embed · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by NVIDIA or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download kermt-embed.zip