[RFC]: FLUX.2-dev Model Acceleration Support

Motivation.

FLUX.2-dev is a high-quality text-to-image model by Black Forest Labs. Following the successful integration in PR #1629, this RFC proposes acceleration feature support to improve inference efficiency and reduce deployment costs.

Current Status

Model Info:

• Name: FLUX.2-dev
• Source: https://huggingface.co/black-forest-labs/FLUX.2-dev
• Integration PR: #1629 (merged)
• Total Model Size: ~112.6 GB (Transformer ~64.3 GB + T5-XXL text encoder ~48.0 GB)

Currently Supported Features (from PR #1629):

Feature	Status	Notes
Text-to-Image	✅ Supported	1024x1024 generation
Image-to-Image	✅ Supported	Image editing capabilities
Tensor Parallel	✅ Supported	TP=1 & cpu-offload, TP=2 & cpu-offload, TP=4 working

Performance (from PR #1629):

Config	GPU Memory	Peak Memory	Time per Image	Status
Diffusers (baseline)	-	104.9s	✅
TP=1 + CPU Offload	66696MiB	67352MiB	89.8s	✅
TP=2	81112MiB	81182MiB MiB	39.1s	✅
TP=4	68160MiB	81116MiB MiB	29.1s	✅
TP=1 (no offload)	OOM	-	-	❌

---

Proposed Acceleration Features

1. Model x Feature Matrix

Following the format from issue #1217, here's the acceleration feature roadmap for FLUX.2-dev:

Feature	Priority	Status	Notes
⚡TeaCache	P0	✅ https://github.com/vllm-project/vllm-omni/pull/1871	Significant speedup potential
⚡Cache-DiT	P0	✅ https://github.com/vllm-project/vllm-omni/pull/1814	Alternative to TeaCache
🔀SP (Ulysses & Ring)	P2	⏳ https://github.com/vllm-project/vllm-omni/pull/3244	Sequence parallelism
🔀CFG-Parallel	P0	✅ https://github.com/vllm-project/vllm-omni/pull/2010	Verify compatibility
🔀Tensor-Parallel	Done	✅ https://github.com/vllm-project/vllm-omni/pull/1629	TP=1 & cpu-offload, TP=2 & cpu-offload, TP=4 working
💾VAE-Patch-Parallel	P1	🙋 Not supported	Reduce VAE memory pressure
💾GGUF-Quantization	P1	⏳ https://github.com/vllm-project/vllm-omni/pull/3027	Low-bit quantization

2. Feature Priority Analysis

High Priority (P0):

1. TeaCache / Cache-DiT
2. CFG-Parallel

Medium Priority (P1):

3. GGUF-Quantization
4. VAE-Patch-Parallel

Low Priority (P2):

5. Ulysses-SP
6. Ring-SP

Proposed Change.

Please provide the detailed design document of the RFC using the template.

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.