feat: vLLM-Omni backend (native omni: text/audio/vision) + chat UI

[ramkrishna2910]

Draft — backend fully validated end-to-end on gfx1151; UI typechecks but hasn't been visually run (see Validation). Depends on the vllm-omni* release from lemonade-sdk/vllm-rocm#23 (merged).

What

Adds vLLM-Omni as a first-class backend so omni / any-to-any multimodal models (Qwen2.5-Omni today; Cosmos3 later) run with ROCm acceleration on AMD gfx1151 (Strix Halo) — a single model doing text + audio + vision in, text + native voice out.

vLLM-Omni is a pure-Python layer on the same base vLLM+PyTorch+Triton, shipped as a separate vllm-omni* release artifact, so it gets its own recipe + pin (see vllm-rocm#23).

Backend (54788cc9) — uses the new self-describing descriptor model

Per docs/dev/adding-a-backend.md — one folder + a few appends, no router/CLI/doc edits:

• src/cpp/{include,server}/backends/vllm_omni/ — descriptor + VLLMOmniServer (mirrors vllm). Launches vllm-omni-server serve <model> --omni --deploy-config <yaml> --served-model-name --max-model-len, resolves the per-model deploy config via the model's deploy_config extra key + get_resource_path, forwards chat. Native voice/vision ride through the chat body.
• CMakeLists.txt LEMON_BACKENDS += "vllm-omni|vllm_omni"
• backend_versions.json pin vllm-omni0.23.0rc1-rocm7.14.0
• server_models.json Qwen2.5-Omni-3B-vLLM-Omni
• resources/omni_deploy/qwen2_5_omni_1gpu.yaml — single-GPU stage colocation (upstream defaults are multi-GPU)

Chat UI (a423eab3)

• ChatWindow derives isAudioOutput from a chat-speech label (mirrors chat-transcription).
• LLMChatPanel: voice toggle + picker (Chelsie/Ethan, off by default — speech gen is slow); handleAudioChat sends a non-streaming modalities:[text,audio]+voice request and captures message.audio (handles vLLM-Omni's two-choice shape and OpenAI's single-choice) → rendered via the existing MessageAudio player. No new rendering code.
• Model labels vision/chat-transcription/chat-speech/omni light up vision input, audio input, and native voice output via existing capability flags.

Validation

• Backend: end-to-end on real gfx1151. lemond builds, registers the recipe, lists the model, loads it (launches the subprocess), and all four modalities pass: text, native voice out (24 kHz WAV), audio in, vision in — with the audio choice passed through lemonade's handler intact.
• UI: tsc --noEmit clean. Reuses proven audio-player/artifact path. Not yet run in the app (needs a webview) — reviewers should eyeball rendering/UX.

Notes for reviewers

• UI uses emoji icons (🔊/🔈) + one inline style on the voice-control wrapper — swap for house icon components / CSS.
• Omni over-generates audio length today (talker max_tokens tuning in the deploy YAML) — cosmetic.
• gfx1151-only (the qualified/hardware-validated omni target).

🤖 Generated with Claude Code