azure-jobs 0.1.39

A fast CLI for submitting and managing Azure ML jobs via pure REST APIs

Azure Jobs

A fast, lightweight CLI for submitting Azure ML jobs through pure REST APIs — no azure-ai-ml SDK and no amlt runtime required.

aj run adds a template inheritance layer on top of three submission backends:

• native — direct Azure ML REST (default for AML / Singularity).
• amlt — delegates to the amlt CLI for compatibility.
• volcano — submits to a Kubernetes Volcano cluster via kubectl.

Install

pipx install azure_jobs

Requires az login. The volcano backend additionally needs kubectl configured against your cluster.

Quickstart

mkdir my-project && cd my-project
aj init                          # scaffold .azure_jobs/, register workspace
aj pull <user>/<repo>            # (optional) clone shared templates
aj run -t gpu train.py           # submit using the "gpu" template

.py scripts run via uv run, .sh via bash. Drop a .codeignore (or .amltignore) at the project root to exclude paths from the upload.

Templates

Templates live under .azure_jobs/template/ as YAML files.

1. aj init — scaffolds .azure_jobs/ and (optionally) pulls a starter template repo.
2. aj pull <user>/<repo> — clone a shared template repo into .azure_jobs/.
3. Hand-author — drop a YAML file into .azure_jobs/template/.

Minimal leaf template (.azure_jobs/template/gpu.yaml):

base: [account.default, storage.default, environment.aml]
config:
  target:
    name: my-cluster
  jobs:
    - name: train
      sku: "{nodes}xA100-80GB"

base chains other YAML files in .azure_jobs/ (dotted name → .azure_jobs/<dir>/<name>.yaml); {nodes} / {processes} are substituted from CLI flags. Inheritance, merge rules, and SKU formats are documented in docs/configuration.md.

aj template list                # see what's available
aj template show <name>         # resolved config (after inheritance)
aj template validate            # sanity check
aj template push -m "msg"       # commit + push back upstream

aj run

aj run -t gpu train.py           # submit via REST
aj run train.py                  # reuse last template
aj run -t gpu -n 4 -p 8 train.py # 4 nodes × 8 GPUs/node
aj run -d train.py               # dry run — print config, don't submit
aj run --amlt -t gpu train.py    # submit via amlt instead

Flag	Purpose
-t	Template name
-n	Number of nodes
-p	GPUs per node (drives SKU + AJ_PROCESSES)
--ppn	Launcher processes per node (e.g. torchrun --nproc-per-node)
-d	Dry run
--amlt	Submit via amlt

Positional args after the script are forwarded verbatim to your command.

How it works

1. Resolve the template, walk the base chain, merge configs.
2. Apply CLI overrides (-n / -p / --ppn).
3. Build a normalized SubmitRequest.
4. Dispatch by backend:
   • native — register environment (SHA-deduped) → upload code (content-addressed) → PUT /jobs/{name}.
   • volcano — render Volcano Job YAML → upload code to a PVC via kubectl exec + tar → kubectl create.
   • amlt — write a submission YAML and shell out to amlt run.
5. Append a SubmitRecord to record.jsonl and print the portal URL.

Code uploads are content-addressed: identical (template + command + code) → identical hash → re-runs reuse the prior asset.

AJ_* environment variables

Exported into every job. Read them in your training script.

Variable	Meaning
AJ_NAME	Job display name
AJ_ID	Submission ID (matches record.jsonl)
AJ_TEMPLATE	Template name used
AJ_NODES	Number of nodes
AJ_GPUS_PER_NODE	-p value
AJ_PROCESSES	AJ_NODES × AJ_GPUS_PER_NODE
AJ_PROCESSES_PER_NODE	--ppn value
AJ_SUBMIT_TIMESTAMP_UTC	Submission timestamp

Example — torchrun with whatever the user requested:

torchrun \
  --nnodes=$AJ_NODES \
  --nproc_per_node=$AJ_GPUS_PER_NODE \
  --node_rank=$RANK \
  --master_addr=$MASTER_ADDR \
  train.py

aj dash

Interactive TUI dashboard for browsing and managing cloud jobs.

aj dash

Key	Action
↑ ↓	Move selection
← →	Prev / next page
enter / i	Job detail panel
l	Open logs (auto-streams if the job is running)
o	Pick a different log file
c	Cancel the selected job
r	Refresh
f / e / w	Filter by status / experiment / workspace
/	Search
F	Clear all filters
esc	Help overlay
q	Quit

Use as a Python SDK

The same engine the CLI uses is exposed at the package root, so you can build and submit jobs from your own scripts:

from azure_jobs import (
    Template,
    build_submit_request,
    submit_via_native,   # also: submit_via_volcano, submit_via_amlt
    get_workspace_config,
)

template = Template.from_conf_path(".azure_jobs/template/gpu.yaml")
request = build_submit_request(
    template,
    name="my-job", sid="abc123", sku="2xA100-80GB",
    user_command="train.py", user_args=(),
    workspace=get_workspace_config(),
    template_name="gpu", nodes=2, processes=8,
    code_dir="/path/to/project",  # defaults to os.getcwd()
)
result = submit_via_native(request)
print(result.status, result.portal_url)

See docs/sdk.md for the full surface and a submit_and_record example.

Documentation

Document	Contents
Commands	aj job, aj template, aj quota, aj sku, aj dash, ...
SDK	Programmatic submission API
Architecture	Module layout, submission flow, backends
Configuration	Templates, inheritance, merge rules, SKU formats
REST API	REST client design, endpoints, job body shape
Comparison	aj vs amlt feature matrix
Roadmap	Planned features