aidsc 0.1.0

Package for Inferencing AI at DSC

aidsc

Small Python client for the Zeus control plane (vllm-orchestrator): typed language inference payloads, health checks, optional PAMD SSH when Zeus is not on localhost, and POST /api/requests.

Install (this monorepo)

From the repository root:

uv sync

aidsc is linked as an editable path dependency; then:

from aidsc import aidsc, LanguageArgs, LanguageInferenceRequest

Quick usage

from aidsc import aidsc, LanguageArgs, LanguageInferenceRequest

client = aidsc()  # reads .env via python-dotenv
client.ensure_ready()

req = LanguageInferenceRequest(
    args=LanguageArgs(
        model="gpt-oss-20b",
        questions=["Answer in one sentence: what is 2+2?"],
    ),
    max_context_length=8192,
    max_server_uptime=604800,
    max_new_tokens=512,
    total_concurrent_models=1,
    max_num_seqs=8,
)
out = client.infer(req, server=117)  # pin class117.sc.fsu.edu
print(out)

How this relates to Zeus

1. Your code builds a LanguageInferenceRequest and calls aidsc.infer, which sends JSON to POST /api/requests on Zeus (see app.py).
2. Zeus validates a ZeusRequest, picks a lab machine (or honors server), and drives the receiver on that host over SSH (see utils/ssh/orchestrator.py).
3. GET /api/zeus-health on Zeus returns {"healthy": true} for local liveness only.

Environment variables

Variable	Purpose
aidsc_ZEUS_LOCAL_URL	Default http://127.0.0.1:32553 — Zeus when you are on the Zeus host.
aidsc_ZEUS_REMOTE_URL	Default http://144.174.11.196:32553 — Zeus URL as reachable from PAMD (used when localhost is down).
aidsc_SSH_PAMD	Full SSH prefix, e.g. ssh [e-mail], used when localhost health fails.

Interactive bootstrap (writes .env and ensures .gitignore contains .env when in a git repo):

aidsc().setup_interactive()

Payload shape

Matches data/sample_payload.json: type is always "language" for now; args holds model and questions. These map to Zeus POST /api/requests fields:

Field	Zeus JSON key	Default
max_context_length	max_context_length	8192
max_model_len	(alias for max_context_length )	—
max_server_uptime	max_server_uptime	604800 (7 days)
max_new_tokens	max_new_tokens	512
total_concurrent_models	total_concurrent_models	1
max_num_seqs	max_num_seqs	8

aidsc.infer(..., max_new_tokens=...) still overrides the request’s max_new_tokens when passed.

Layout

Path	Role
src/aidsc/client.py	aidsc, mapping to Zeus JSON
src/aidsc/models.py	Pydantic payload models
src/aidsc/zeus_client.py	HTTP health + /api/requests
src/aidsc/connectivity.py	Localhost vs PAMD+remote health
src/aidsc/env_bootstrap.py	.env load, .gitignore for .env