[RFC]: vLLM-Omni XPU 2026 Q2 Roadmap

Motivation.

Model coverage:

• Toward >80% of all models at https://docs.vllm.ai/projects/vllm-omni/en/latest/models/supported_models/
• exclude only because:
  • memory usage can't be held with 8 * Intel ARC Pro B70
  • 3rd modeling with CUDA hard-code

Features to enable on Intel GPU(XPU)

• Sequence Parallel
• HSDP / FSDP
• torch compile - https://github.com/vllm-project/vllm-omni/pull/3113
• XPU Graph
• sleep mode - pt2.11 - https://github.com/vllm-project/vllm-omni/issues/2545
• lora support
• Large Scale Deployment - follow https://github.com/vllm-project/vllm-omni/issues/2136 (3) Large Scale Deployment (Disaggregated Serving))
• EPLB (TBD)
• Prefix Cache (TBD)

Quantization

• https://github.com/vllm-project/vllm-omni/issues/1325
• fp8 KV cache support - https://github.com/vllm-project/vllm-omni/pull/1413

Kernel list

• Sparse ATTN - https://github.com/vllm-project/vllm-omni/issues/2233

Performance optimization models

• Wan 2.2 - https://github.com/vllm-project/vllm-omni/issues/1355
  • I2V
  • T2V
  • S2V - https://github.com/vllm-project/vllm-omni/pull/2751
• Flux-dev (T2I)

Memory Management

• expand layerwise offloading model coverage
• multi stage weight sharing (Hunyuan as example)

UX

• Streaming input/output
• All supported models with xpu validated stage_config / document

misc:

• vllm omni xpu - model triage Agent skill

Proposed Change.

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

Feedback Period.

No response

CC List.

@hsliuustc0106 @Gaohan123

cc: @pi314ever @Joshna-Medisetty @yiliu30 @yma11

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.