sglang/docs/docs/sglang-diffusion/compatibility_matrix.mdx

---
title: "Supported Models"
description: "Browse model families and public checkpoints supported by SGLang Diffusion."
---

import { DiffusionModelCatalog } from '/src/snippets/diffusion/model-catalog.jsx';

## Qwen-Image 2.1 integration status

Qwen-Image 2.1 has a separate native pipeline for text-to-image and optional
reference-image conditioning. Full-checkpoint validation on H200 includes
1024×1024, 40-step generation and editing. Reference comparisons use
Transformers 4.57.3, BF16 and matching Torch SDPA settings with cuDNN SDPA
disabled. The native encoder explicitly preserves that version's pre-final-norm
hidden state and vision position interpolation semantics.

Full-checkpoint functional checks on up to two H200 GPUs cover DiT TP, Ulysses,
Ring + FA, CFG parallel, DiT layerwise offload, FA/Sage, parallel
tiled VAE decode, Cache-DiT, breakable CUDA graphs, and multiple reference
images and outputs. One H200 with resident weights also passed HTTP generation
and editing, including multiple outputs and reference images. These are
functional checks, not a broad quality evaluation of lossy optimizations.

Additional full-checkpoint checks on B200 cover native encoder TP, combined
TP2 × Ulysses2 and TP2 × Ring2 + FA on four GPUs, and
`--layerwise-offload-components all`. Two-GPU VAE `spatial_shard` decode passed
with TP2, CFG parallelism, and all-component offload. These use 512px/4-step
editing, plus a 1024px/4-step comparison of spatial and full VAE decode.
Single-GPU 1024px/40-step generation, editing, and transparent PNG output also
passed at source revision `d629d234f1`.

Two-B200 component regressions assert actual encoder weight sharding and compare
text and image conditioning with a single-rank reference. VAE regressions cover
both decoder structures and even/odd latent heights in FP32 and FP64. FP64
spatial decode matches the full decoder; BF16 full-checkpoint output is not
bit-exact. In the tested image, spatial versus full decode gave RGBA PSNR
60.69 dB; this is one sample, not a general quality guarantee.

Request-local prefix KV reuse and new-request graph replay have CUDA regression
tests, including batched targets and independent variable-length prefixes.
Compatible text-to-image requests support opt-in dynamic batching; image-edit
requests remain separate, while multiple outputs within one request are supported.
See the cookbook's [batching guidance](/cookbook/diffusion/Qwen-Image/Qwen-Image-2.1#batching)
for deployment settings and floating-point reproducibility limits.
The deployment picker marks only its exact tested HTTP combinations as
verified, including H200, B200, RTX PRO 6000 96GB, RTX 5090, and RTX 4090.
CLI-only combinations remain Unverified in the picker.

On `1eab5de5990`, one RTX PRO 6000 Blackwell Server Edition passed repeated
1024px/40-step HTTP generation, editing, transparent generation, and transparent
input editing with resident weights or DiT layerwise offload and Torch SDPA.
Repeated requests and corresponding outputs across placements matched RGBA
pixels; transparent outputs retained alpha from 0 to 255. Its SM120
RoPE, normalization, fusion-gate, and model regressions also passed. This check
does not establish multi-GPU or quantized-checkpoint coverage on that platform.

Disaggregated encoder, denoiser, and decoder roles passed full-checkpoint HTTP
generation, variable-length prompts, image editing, and CFG on three B200 GPUs,
using Mooncake's same-host TCP transport. At 512px/4 steps, their RGBA pixels
matched the monolithic server exactly. Multi-host RDMA and multi-rank roles
are not covered by that check.

Online FP8 for the DiT, encoder, and both together passed 1024px/40-step HTTP
generation and editing on one B200; both together also passed transparent PNG
generation. Serialized E4M3FN FP8 component folders with scalar weight scales
and dynamic activations passed generation, editing, and transparent output for
each component and both together. The combined export also passed separate
checks with TP2 plus encoder folding and with single-GPU all-component
layerwise offload. The 476 loaded quantized matrices and
scales matched their checkpoint values. Quantization and TP change numerical
results. Revision `f1f3366c7c` fixes vision rotary initialization rounding during
CPU/GPU transfer. In 1024px/40-step B200 checks, all-component layerwise offload
matched resident generation, editing, and transparent RGBA pixels exactly
for both native precision and the combined serialized export. Revision
`81c8c550fa` additionally preserves FP8 weights and FP32 buffers during
whole-encoder CPU offload; separate checks of that mode also matched resident
pixels. Resident output was unchanged by these fixes. See the cookbook for
sample quality metrics.

Native-name Q4_0 GGUF files passed 1024px/40-step B200 HTTP generation,
editing, and transparent output for DiT-only, encoder-only, and both together.
The 476 packed matrices matched their exported bytes. The combined export also
passed TP2 with encoder folding; separate all-component layerwise and
whole-encoder CPU offload checks matched resident RGBA pixels exactly.
Quantization and TP change output pixels, including alpha. These checks cover
the documented private exports, not arbitrary community checkpoints. Other
GGUF types and hardware remain unverified for this model.

Calibrated ModelOpt-format NVFP4 component directories also passed the same
six-configuration B200 HTTP matrix: each component, both together, two offload
modes, and TP2 with encoder folding. All 476 packed matrices and their scales
matched the exports after runtime layout transforms. Both offload modes
matched resident RGBA pixels exactly. NVFP4 and TP2 change image and alpha
values; this private export uses max calibration on only six requests, without
SVDQuant or AWQ. See the cookbook for measured errors. Native FP4 requires
Blackwell, and this model's NVFP4 exports remain unverified on RTX PRO 6000 and
RTX 5090.

Synthetic Diffusers-format LoRA passed dynamic loading, merging,
and removal on one B200 and TP2 with encoder folding; both removal paths
restored the base image exactly. Trained-adapter quality remains unverified.
Platform latency measurements
and their exact workloads are in the cookbook. Breakable graphs
replay only matching warmup signatures, including condition-prefix length;
other shapes run eagerly. VAE tiling is opt-in and can change numerical results.
Do not infer compatibility from the older Qwen-Image row.

Use the public `Qwen/Qwen-Image-2.1` checkpoint; no Hugging Face token is required.
See the [Qwen-Image 2.1 cookbook](/cookbook/diffusion/Qwen-Image/Qwen-Image-2.1)
for checkpoint layout and usage. The standard two-GPU E2E suite includes `qwen_image21_t2i_tp2`
with TP2, 1024px/40-step generation, two requests, and image consistency checks.

The additional opt-in HTTP cases check repeated generation, editing, and real
RGBA alpha output from a local checkpoint:

```bash
export SGLANG_QWEN_IMAGE21_TEST_MODEL=/models/Qwen-Image-2.1
export SGLANG_QWEN_IMAGE21_TEST_IMAGE=/path/to/reference.png
python -m pytest -q python/sglang/multimodal_gen/test/server/test_server_qwen_image21.py
```

The module skips without the model variable; only editing skips when the image
variable is absent. Credentials stay in the runner environment.

## Supported model inventory

Use a listed checkpoint as `--model-path` with `sglang generate` or
`sglang serve`. This registry-backed list contains known public entry points;
family detection may also support compatible local directories. Open the linked
Cookbook recipe for launch commands, optimizations, adapters, and model-specific
notes.

<Tabs>
  <Tab title="Image and 3D">
    <DiffusionModelCatalog category="image" />
  </Tab>
  <Tab title="Video and audio">
    <DiffusionModelCatalog category="video" />
  </Tab>
  <Tab title="World and action">
    <DiffusionModelCatalog category="world" />
  </Tab>
</Tabs>