Unified text-to-image and image-conditioned generation through one pipeline
class: a 7.1B single-stream DiT with block-causal attention and an exact
cross-step prefix KV cache, a Qwen3-VL-8B text encoder, and a 16x RGBA
autoencoder.

Verified end to end on one NVIDIA GB300 built from PR #7759. The guide carries
the facts an operator acts on; the evidence for each is below.

What was measured (1024x1024, 40 steps, CFG off, seed fixed, --vae-use-tiling):

- serves and returns a valid PNG, checked by decoding it and asserting magic
  bytes, byte length and IHDR dimensions rather than accepting HTTP 200
- 4.5 s per image, 34.0 GB peak in BF16
- DiT FP8 (img_mlp kept BF16): 32.0-32.7 GB, slightly slower
- text-encoder FP8: 27.5-28.1 GB -- about three times the DiT-only saving
- prefix KV cache: a repeated prompt costs 0.31x the first submission; changing
  only the prompt tail returns it to 2432 ms, which rules out warm-up
- prefix_kv_cache_dtype=fp8_v serves, and returns 0.50x instead of 0.31x
- Ulysses SP across 2 GPUs: 3851 ms at 33.1 GB, against 4492 ms and 34.0 GB on 1
- step execution: 4 concurrent requests in 4019 ms against ~5870 ms serial
- dimensions floored to a multiple of 32 (1000 -> 992, 1328 -> 1312), never
  rejected; 1024x1536 and n=2 both fine
- seed determinism: byte-identical output
- RGBA survives the round trip -- PNG colour-type 6
- up to 4 reference images; a 5th is refused by the server

Bugs found by running it, neither visible from reading the source:

- --revision is not honoured. omni_base.py:110 calls
  download_weights_from_hf_specific() with no revision=, so the engine resolves
  "main"; a cache staged by commit sha has no refs/main and fails offline with
  LocalEntryNotFoundError despite every shard being present.
- The PR #7759 images set WORKDIR /opt/vllm-omni and do not inherit
  vllm/vllm-openai's `vllm serve` entrypoint, so passing only `<model> --flags`
  execs the model id as a binary and exits 127.

Corrections to claims that source-reading alone got wrong:

- /v1/images/edits works, but takes multipart/form-data. A JSON body resets the
  connection, which reads as a broken endpoint. A data: URL in the `url` form
  field is additionally capped at 1024 KB per part, which a 512x512 PNG exceeds.
- true_cfg_scale defaults to 4.0 when omitted but only applies alongside a
  negative_prompt, so a plain request behaves as if 1.0 had been sent.
- The first-request cost is not CUDA-graph capture: it persists at 1.26x with
  --enforce-eager and no graphs at all.
- Per-component FP8 is serve-reachable via --diffusion-quantization-config;
  --ignored-layers exists only in text_to_image.py, not in `vllm serve`.
- A plain clone lands on main, which does not contain this model.
- Without model.docker_image the builder falls back to vllm/vllm-openai:latest.

Not measured, and not claimed: throughput (vllm bench serve drives only
/v1/completions and /v1/chat/completions, so the t2i task carries a benchmark
override saying so), pipeline parallelism, and the GB200 FP8 figures, which stay
attributed to the vLLM-Omni contributors.

meta.hardware records gb300 as verified and default_hardware points there, so
the page opens on the hardware the numbers came from. vram_minimum_gb is 40 from
the repo formula ceil(33.1 GB x 1.2), which the measured 34.0 GB sits under. No
hardware: entry for any GPU that was not run.

Validated with `node scripts/build-recipes-api.mjs`: 191 -> 192 models, no
warnings naming this file.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Tiezhen WANG <tiezhen@inferact.ai>