LTX Inference Async Azure

Telegram-driven, cost-optimized LTX-2.3 inference on Azure Spot with a durable async pipeline.

What it does

• Always-on control plane (Azure Functions) accepts generation requests
• Durable state in Azure Blob + Table + Service Bus
• Spot GPU worker boots on demand, runs ComfyUI + LTX, then auto-deallocates
• Telegram notifications deliver results (or Blob links for large outputs)

Architecture (high level)

1. Telegram webhook -> Azure Functions
2. Function writes input + manifest to Blob, state to Table, enqueues Service Bus
3. Spot worker pulls job, runs ComfyUI workflow, uploads MP4, notifies user
4. Worker deallocates after idle timeout

Key docs

• docs/env-setup.md - required env values and setup steps
• docs/muse-studio-integration-plan.md - integration plan for Muse Studio

Repo layout

• src/ltx_async_azure/ - control plane, worker, shared models
• infra/bicep/ - Azure infrastructure
• ops/ - image build, bootstrap, RBAC, deploy scripts
• workflows/ - ComfyUI workflow templates

Local setup

uv venv .venv --python 3.11
uv sync --group dev
uv run pytest

Status

GPU quota is required for A100 Spot in most subscriptions. If your subscription has zero quota, request it first, then proceed with deployment.