Get the model

You're Claude Code, working in my repository.

Task
Write a script for this project that downloads a Muse Spark Open checkpoint from Hugging Face into a local folder, verifies that every file landed (the shard count the page lists, then each file's SHA-256 against the Hub's metadata), and loads the tokenizer from that folder to render one chat prompt. Make the checkpoint (BF16 or NVFP4) and the folder arguments.

Read this first
Meta's docs page "Get the model" as Markdown: https://dev.meta.ai/docs/muse-spark-open/get-the-model.md
It's the source of truth for request and response shapes. The same page for people: https://dev.meta.ai/docs/muse-spark-open/get-the-model

Setup
- Server: the model runs on hardware you control, behind an OpenAI-compatible server (SGLang, vLLM, llama.cpp or another runtime this page covers). Code takes the server's base URL from configuration (such as http://127.0.0.1:8000/v1) and sends the model name it serves (GET <base URL>/models lists it).
- Auth: a local server takes any placeholder key unless it was started with one; read a real key, and any token a download needs, from the environment. Never hard-code, print or commit a key or token.
- Python: the OpenAI SDK (pip install openai) with base_url set to the server's /v1 and a placeholder api_key, or plain requests.

Facts to get right
- Repos: meta-models/Muse-Spark-Open-816B-A42B (BF16, about 1.65 TB, 65 weight shards) and meta-models/Muse-Spark-Open-816B-A42B-NVFP4 (about 537 GB, 127 shards).
- CLI: pip install -U huggingface_hub, then hf download <repo> --local-dir <folder>; running it again resumes.
- Meta's download page serves the same files: keep the original names, all in one folder, and pass the folder where commands take the repo id.
- The tokenizer check needs transformers 5.19 or later. Each rendered message starts with <|start|> and its role, content after <|message|>, and ends with <|eot|>.

Starter code (Python)
import hashlib, pathlib
from huggingface_hub import HfApi, snapshot_download
from transformers import AutoTokenizer

repo, folder = "meta-models/Muse-Spark-Open-816B-A42B-NVFP4", pathlib.Path("./muse-spark-open-nvfp4")
snapshot_download(repo_id=repo, local_dir=folder)

print(len(list(folder.glob("*.safetensors"))), "shards")  # 127 for NVFP4, 65 for BF16

def sha256(path):
    digest = hashlib.sha256()
    with path.open("rb") as fh:
        for chunk in iter(lambda: fh.read(1 << 24), b""):  # 16 MB at a time: a BF16 shard is about 25 GB
            digest.update(chunk)
    return digest.hexdigest()

for f in HfApi().model_info(repo, files_metadata=True).siblings:
    path = folder / f.rfilename
    if not path.exists():
        print("missing:", f.rfilename)
    elif f.lfs and sha256(path) != f.lfs.sha256:
        print("mismatch:", f.rfilename)

tok = AutoTokenizer.from_pretrained(folder)
messages = [{"role": "user", "content": "In one sentence, what is Muse Spark good at?"}]
print(tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False))

Done when
- The script reports 65 (BF16) or 127 (NVFP4) shards and no missing files.
- Every LFS file's SHA-256 matches HfApi().model_info(repo, files_metadata=True).
- apply_chat_template prints a prompt containing <|start|>user<|message|>.
- Any key or token comes from the environment and never appears in code, logs or commits.