[ITerydh]

Title: [diffusion] Support LongCat-Image-Edit and LongCat-Image-Edit-Turbo (I2I)

Motivation

SGLang diffusion currently supports LongCat-Image for text-to-image only. This PR adds image-editing (I2I) support for meituan-longcat/LongCat-Image-Edit and the distilled meituan-longcat/LongCat-Image-Edit-Turbo, mirroring the diffusers LongCatImageEditPipeline reference implementation.

Modifications

• New LongCatImageEditPipeline (runtime/pipelines/longcat_image.py): composed I2I pipeline — input validation → joint VL prompt encoding → image VAE encoding → denoising → decoding.
• New LongCatImageEditPipelineConfig (configs/pipeline_configs/longcat_image.py), matching the diffusers reference:
  • Output resolution derived from the condition image (~1MP area, ceil to /16), which intentionally differs from the shared /32 helper.
  • Reference image is VAE-encoded (argmax sampling), packed, and concatenated after the noisy latents along the sequence dim; predictions over reference tokens are sliced off.
  • RoPE position ids use modality 1 for noisy tokens and modality 2 for reference tokens, both offset by the full text sequence length (VL image tokens + 512 body).
• New LongCatImageEditTextEncodingStage (stages/model_specific_stages/longcat_image_edit.py): joint text+image prompt encoding via Qwen2.5-VL, loading the vision tower and masking padded prompt
  tokens; the negative prompt is encoded against the same image when CFG is enabled. Ulysses SP is supported.
• Registry / sampling defaults (registry.py, configs/sample/longcat_image.py): Edit (50 steps, CFG 4.5) and Edit-Turbo (8 steps, CFG off); Turbo detector registered first so it wins.
• Docs: compatibility matrix updated with both model ids.
• Unit tests: test_longcat_image_edit_config.py (CPU-only config-hook tests) and updated test_qwen2_5vl_generation.py.

Usage:

CUDA_VISIBLE_DEVICES=0 \
sglang generate \
--model-path meituan-longcat/LongCat-Image-Edit \
--image-path demo.jpg \
--prompt "将图片中人物的衣服改变为蓝色,头发改为红色,保证图像整体细节不变" \
--num-gpus 1 \
--num-inference-steps 50 --guidance-scale 4.5 --negative-prompt "" --seed 42 \
--output-file-path sp1.png --save-output

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
sglang generate \
--model-path meituan-longcat/LongCat-Image-Edit \
--image-path demo.jpg \
--prompt "将图片中人物的衣服改变为蓝色,头发改为红色,保证图像整体细节不变" \
--num-gpus 8 --sp-degree 8 --ulysses-degree 8 \
--num-inference-steps 50 --guidance-scale 4.5 --negative-prompt "" --seed 42 \
--output-file-path sp8.png --save-output

Accuracy Tests

base:

result:

Speed Tests and Profiling

No changes to existing model paths; the new pipeline follows the standard denoising loop. LongCat-Image (T2I) behavior is unchanged. Currently, we are introducing a new model, so there is no need for performance testing. Our main focus is on completing correctness testing.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the
  speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ✅ Run #32805157172
Latest PR Test (Extra): ✅ Run #32805157023
Latest PR Test (AMD ROCm 7.2): ⏳ Run #32805157198

[BBuf]

Does this handle num_outputs_per_prompt > 1? prompt_embeds stays at batch 1 here, while the image and noise latents use the effective batch size. The DiT will hit a batch mismatch. Could we repeat both positive and negative embeds and add an n=2 test?

[ITerydh]

Thanks, i was confirmed and fixed. LongCatImagePipelineConfig was missing an expand_conditioning_to_sample_batch override, so on the multi-output path the
noise/reference latents were built at batch N while prompt_embeds stayed at batch 1, breaking the joint-attention concat. Fixed by mirroring QwenImagePipelineConfig (repeat pos/neg embeds, masks, seq-lens; no-op at n=1), covering both T2I and Edit. Added an n=2 test, and --num-outputs-per-prompt 2
now works end-to-end.

[BBuf]

This also changes the cache-free path for existing LongCat T2I: longcat_postprocess_text keeps all 512 body tokens, including padding. Could we add a T2I main/PR or Diffusers parity check and update the PR description? The current claim that T2I behavior is unchanged does not hold.

[ITerydh]

You're right — the "T2I unchanged" claim doesn't hold, and I've corrected the PR description. Ran the Diffusers parity check you asked for:

LongCat T2I does feed the padded 512-token body to the DiT (longcat_postprocess_text keeps padding), so masking it changes T2I output. But this is a fix toward parity, not a regression — Diffusers' LongCatImagePipeline._encode_prompt passes attention_mask to the text encoder (pipeline_longcat_image.py:328), so masking the padding is the reference behavior.

Parity check (LongCat-Image, 1024², seed 42, 50 steps, prompt-rewrite off):

• sglang with mask vs Diffusers ref: mean|Δ| = 13 (chaotic-sampler noise level)
• sglang without mask vs Diffusers ref: mean|Δ| = 44 — visibly hallucinated background content

So the mask brings T2I closer to Diffusers. No separate T2I mechanism is needed since the fix lives in the shared Qwen2.5-VL encoder. Qwen-Image(-Edit) stays unaffected (dynamic padding=True, no pads for a single prompt).

[BBuf]

Could we scope the cache-free mask change to LongCat? Passing a non-None mask switches LocalAttention to the PyTorch SDPA path and changes existing Qwen-Image outputs across multiple backends. I don't think we should update the Qwen-Image GTs in this PR without separate Diffusers parity validation.

[ITerydh]

Done — scoped the cache-free mask to LongCat in ec3693e06. Added a honor_cache_free_padding_mask flag that TextEncoderLoader sets only for LongCat (T2I + Edit); it gates both the attn_mask passing and the fallback mask build. Every other pipeline (Qwen-Image included) now keeps the original attn_mask = attention_mask if use_cache else None, so LocalAttention stays on its fast path and no Qwen-Image GTs change in this PR. Verified LongCat sp1/sp8 outputs are byte-identical to before scoping; the unit test asserts both flag states.

[BBuf]

LGTM

[BBuf]

Please fix lint error

[ITerydh]

Please fix lint error

OK