scGPT — Single-Cell Foundation Model

Prerequisites

How to run

Loading the vocabulary and checkpoint

scGPT checkpoints are raw directories (args.json, best_model.pt, vocab.json) — not Hugging Face hub repos. Point at the directory, not an HF repo id.

from scgpt.tokenizer.gene_tokenizer import GeneVocab
gv = GeneVocab.from_file("/path/to/scgpt-human/vocab.json")
print(len(gv))   # 60697 for the released human checkpoint

Embedding an AnnData

import anndata as ad
from scgpt.tasks import embed_data

adata = ad.read_h5ad("dataset.h5ad")        # var must contain a gene-name column
emb = embed_data(
    adata,
    model_dir="/path/to/scgpt-human",
    gene_col="feature_name",
    use_fast_transformer=False,             # see Gotchas
)
# emb is an AnnData with .obsm["X_scGPT"]

Output format

embed_data returns an AnnData whose .obsm["X_scGPT"] is the per-cell embedding (n_cells × emb_dim, 512 by default). Downstream: feed to scanpy.pp.neighbors / scanpy.tl.umap.

Remote compute

Needs ≥24 GB VRAM and the released human checkpoint (~200 MB: args.json, best_model.pt, vocab.json). Read compute_details({provider, mode:'read'}) for an environment with scgpt and a pre-cached checkpoint directory, then:

c = host.compute.create(provider)
job = c.submit_job(
    intent="scGPT embed 50k cells — 1×GPU, ~5 min",
    inputs=[
        {"src": "dataset.h5ad", "dst_filename": "dataset.h5ad"},
        {"src": "embed.py", "dst_filename": "embed.py"},
    ],
    command="python3 embed.py",
    environment=...,   # env name from compute_details
    outputs=["embedded.h5ad"],
    timeout_seconds=1800,
)
print(job.job_id)   # cell ends here — kernel never blocks on compute

Then call the wait_for_notification brain-tool. When the compute_done notification arrives, act on its payload:

save_artifacts(payload["featured_files"])   # paths under hpc/<job_id>/

For the full result dict (output_files, remote_workdir, …), re-enter the kernel and bind the compute handle separately — .close() lives on the handle, not on the job object:

h = host.compute.create(provider)
res = h.attach_job(job_id).result()
h.close()

See the remote-compute-ssh / remote-compute-modal skill for the orchestration details.

In embed.py, pass model_dir= the checkpoint path from compute_details. If flash-attn is unavailable in that environment, set use_fast_transformer=False.

Gotchas

• use_fast_transformer default is True but resolves to a FlashAttention path that may not import in every env. Pass use_fast_transformer=False unless you've confirmed flash_attn loads cleanly.
• The package historically depended on torchtext.vocab.Vocab; in environments without torchtext a pure-Python shim provides Vocab — functionally identical for GeneVocab, but if you hit AttributeError: 'Vocab' object has no attribute …, you're on a stale shim.
• Gene names must match the vocab; unmatched genes are dropped. Set gene_col to the column in adata.var that holds symbols.

Troubleshooting

Next: cluster/annotate the embedding with the scanpy library (sc.pp.neighbors → sc.tl.leiden / sc.tl.umap), or compare to an scvi-tools latent space on the same data.