feat: add vLLM batch inference and text-to-image batch job bundles

[anika-suman-amazon]

What was the problem/requirement? (What/Why)

AWS Deadline Cloud had no sample job bundles for GPU-based AI batch workloads — specifically batch LLM inference and batch image generation. Users building pipelines around large language models or diffusion models had no reference for how to structure an OpenJD job that loads a model once per worker and distributes prompt tasks across a fleet efficiently.

What was the solution? (How)

Two new job bundles:

job_bundles/vllm_batch — Batch LLM inference using vLLM. A step environment starts a vLLM OpenAI-compatible server on each worker (loading the model once), and each task sends a single HTTP request to that server. An Aggregate step collects per-task results into a combined output.jsonl and results.html viewer. Default model: Qwen/Qwen2.5-7B-Instruct.

job_bundles/text_to_image_batch — Batch image generation using diffusers. A step environment loads a diffusion pipeline once per worker (via a tiny stdlib HTTP server), and each task POSTs a prompt and receives PNG bytes back. Optionally composites caption text over the image via PIL with auto-vibe font selection. An Aggregate step produces gallery.html. Default model: black-forest-labs/FLUX.2-klein-4B (Apache 2.0, no auth required, 4-step distilled, ~13 GB VRAM).

The two bundles are designed to chain: vllm_batch generates slogans or captions → text_to_image_batch renders them as images with the generated text overlaid as crisp typography. Both bundles include sample_prompts.jsonl, scripts/run_task.py, scripts/aggregate.py, and tools/prompt_builder.html (a zero-dependency browser tool for building JSONL input files).

What is the impact of this change?

Users get two ready-to-submit samples for GPU AI workloads:

• vllm_batch demonstrates the "load model once, serve many tasks" pattern for LLM inference at scale on Deadline Cloud.
• text_to_image_batch demonstrates the same pattern for diffusion-based image generation, and doubles as a standalone batch image generator or a downstream stage in a vLLM → image pipeline.

How was this change tested?

Both bundles were submitted to a live Deadline Cloud farm with an SMF fleet with NVIDIA GPUs and a Conda queue environment configured.

vllm_batch — submitted with sample_prompts.jsonl and Prompts=1-10, model Qwen/Qwen2.5-7B-Instruct. Verified that:

• The vLLM server started and was polled ready before tasks began
• Each task completed and wrote a per-task result_N.jsonl
• The Aggregate step combined results into output.jsonl and results.html

text_to_image_batch — submitted with sample_prompts.jsonl (bakery slogans from vllm_batch output) and Prompts=1-8, model black-forest-labs/FLUX.2-klein-4B. Verified that:

• The diffusers server loaded the pipeline and came up healthy before tasks began
• Each task generated a PNG with caption overlay composited via PIL
• The Aggregate step produced gallery.html with thumbnails, metadata, and search
• Also works independently as a job bundle to produce images with optional captions in batch

The chain was also tested end-to-end: vllm_batch output.jsonl fed directly into text_to_image_batch as InputFile, with generated_text auto-detected as the overlay caption.

Was this change documented?

Yes. Both bundles include a README.md covering:

• How it works (ASCII architecture diagram)
• Farm setup prerequisites
• Quick start (GUI and CLI submitter)
• Prompt Range syntax
• Input/output format
• All parameters with defaults and descriptions
• Chaining instructions (vllm_batch → text_to_image_batch)
• Troubleshooting section

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.