Integrate with pt_job_queue

[xmfan]

Changed needed for: drisspg/pt_job_queue#9

There's 2 usage modes:

• control and run locally, on devgpu (directly call ptq)
• control locally and run remotely on ODC (prefix ptq commands with uv run)

Quick Start: Using ptq with TorchTitan (Remote on ODC)

Prerequisites

• pt_job_queue cloned locally

1. Set up workspace

uv run ptq setup <machine>

2. Register TorchTitan

uv run ptq repo add git@github.com:xmfan/torchtitan.git --machine <machine>
uv run ptq repo list

3. Run a job

# Investigate a GitHub issue
uv run ptq run --issue 2818 --repo torchtitan --machine <machine>

# Run an adhoc task
uv run ptq run --repo torchtitan --machine <machine> -m "fix the FSDP OOM bug"

4. Check results

uv run ptq list                  # see all jobs
uv run ptq peek <job_id>         # check progress
uv run ptq results <job_id>      # fetch results
uv run ptq web                   # web dashboard at http://127.0.0.1:8000

Testing plan in upstream PR: drisspg/pt_job_queue#9.

[fegin]

torchtitan/parallelisms/ and torchtitan/config_manager.py do not exist. Do you mean torchtitan/distributed/, torchtitan/config/job_config.py?

[tianyu-l]

discussed offline:

• since ptq is still new, as torchtitan maintainer I prefer this not to be exposed to OSS publicly yet
• on ptq side @drisspg seems happy with torchtitan specifics added to ptq repo, so technically this is unblocked
• more than happy to revisit later

[xmfan]

Landed as a hardcoded repo in drisspg/pt_job_queue#9