molmim-nim

Files (13)

• BENCHMARK.md
• SKILL.md
• skillspector-baseline.yml · config/
• config.yml · evals/
• evals.json · evals/
• trigger_evals.json · evals/
• api.md · references/
• examples.md · references/
• parameters.md · references/
• science.md · references/
• validation.md · references/
• skill-card.md
• skill.oms.sig

SKILL.md (first lines)

# MolMIM NIM

Generate, sample, embed, and decode small molecules with MolMIM. Use this
guide for first-pass hosted/local usage; load supplemental files only when
needed:

- `references/api.md`: endpoints, schema, Docker flags, response fields.
- `references/science.md`: use cases, strengths, limits, and handoffs.
- `references/parameters.md`: generation, sampling, and optimization effects.
- `references/validation.md`: SMILES/property/artifact checks.
- `references/examples.md`: compact hosted/local request patterns.

## Choose Mode

Ask only when context is unclear:

> Hosted NVIDIA API or local Docker NIM?

See [`references/api.md`](references/api.md) under **Endpoints** for the full
hosted/local endpoint list.

Mode difference: the hosted API reference exposes `/generate`; the local
container exposes the broader latent-space workflow (`/embedding`, `/hidden`,
`/decode`, `/sampling`, `/generate`). Do not invent hosted latent endpoints.

Hosted requests use `Authorization: Bearer $NGC_API_KEY`. Local inference uses
no auth header after readiness.

## Local Docker

Use shell env first; source repo-root `.env` only if present. Do not print keys.
MolMIM docs use `NGC_CLI_API_KEY` for the local container; this repo accepts
`NGC_API_KEY` or `NVIDIA_API_KEY` and maps to `NGC_CLI_API_KEY` for startup.
Mount `LOCAL_NIM_CACHE` at `/home/nvs/.cache/nim`.

For the exact startup preflight (the `NGC_API_KEY`/`NVIDIA_API_KEY` →
`NGC_CLI_API_KEY` mapping, `docker login`, and the `docker run` for
`nvcr.io/nim/nvidia/molmim:1.0.0`), copy the command block in
[`references/api.md`](references/api.md) under **Local Docker** verbatim.

Readiness check:

```bash
until curl -sf http://localhost:8000/v1/health/ready; do sleep 5; done
```

Local embedding smoke test after readiness. Local inference uses no
`Authorization` header:

```python
import requests

seed = "CN1C=NC2=C1C(=O)N(C(=O)N2C)C"
response = requests.post(
    "http://localhost:8000/embedding",
    headers={"Content-Type": "application/json"},
    json={"sequences": [seed]},
    timeout=60,
)
response.raise_for_status()
embedding_data = response.json()
embeddings = embedding_data["embeddings"]
print(f"received {len(embeddings)} embedding vector(s)")
```

## Hosted Generation Pattern

Use hosted `/generate` for seed-SMILES generation or optimization. Use
`algorithm: "CMA-ES"` for guided property optimization and `algorithm: "none"`
for unguided sampling around the seed.

```python
import os
import requests

hosted = True
url = (
    "https://health.api.nvidia.com/v1/biology/nvidia/molmim/generate"
    if hosted else "http://localhost:8000/generate"
)
headers = {"Content-Type": "application/json"}
if hosted:
    headers["Authorization"] = f"Bearer {os.getenv('NGC_API_KEY')}"

payload = {
    "smi": "CN1C=NC2=C1C(=O)N(C(=O)N2C)C",
    "algorithm": "CMA-ES",
    "num_molecules": 10,
    "property_name": "QED",
    "minimize": False,
    "min_similarity": 0.4,
    "particles": 8,
    "iterations": 3,
}

response = requests.post(url, headers=headers, json=payload, timeout=180)
response.raise_for_status()
result = response.json()
```

Generation gotchas:

- Field name is `smi`, not `smiles`.
- `algorithm` is `"CMA-ES"` or `"none"`.
- `property_name` is `"QED"` or `"plogP"`.
- `num_molecules` is 1-100. `iterations` is 1-1000. `particles` is 2-1000.
- `min_similarity` is 0-1 in the hosted API reference; local docs emphasize
  common values up to 0.7 for constrained optimization.
- `scaled_radius` is 0-2 and is mainly used with `algorithm: "none"` or local
  `/sampling`.

## Local Latent Workflow

Use local-only endpoints for embedding, hidden-state manipulation, and decode.
This is also the surface used by the guided optimization example package.
For local latent workflows, state explicitly that the hosted API reference
exposes `/generate`; `/embedding`, `/hidden`, `/decode`, and `/sampling` are
local-only in the current docs.

From NVIDIA/skills, Apache-2.0 AND CC-BY-4.0.

More from NVIDIA/skills