live: full-duplex personaplex-7b over gpt-live websocket sessions

[xiaguan]

Full-duplex serving, end to end: NVIDIA PersonaPlex-7B over OpenAI's GPT-Live WebSocket protocol, with a browser demo.

What

Crate / file	Change
omni-engine::live	The live contract. Sessions go over a channel. Frame n of a session is the nth frame the engine ran, so the agent's audio is one contiguous stream. Each session has a jitter buffer that starts playout after a prebuffer, refills it after an underrun, and discards the caller's audio of overrun ticks. A drift-free 80 ms clock skips overrun ticks instead of bursting. drive is the one loop both live engines run (clock, admission, Pulse counters) over a Ticker; Line is a running session. Property-tested.
omni-frontend	GET /v1/live/sessions (GPT-Live: strict session.start, errors naming the param, unsupported events refused by name, server_busy keeps the socket for a retry, mute forwards silence). Demo page at /. GET /live/config and GET /live/stats. Tick counters in /metrics. The socket's paths (abort, engine stop, close grace, close before start, full queue, mute) are tested against an engine the test plays by hand.
omni-sim::live	CPU live engine: a pure per-session echo core (the caller late, a word per half second heard) under drive.
omni-bench duplex	Concurrent sessions at real time. Reports frame lateness and playback underrun per level; --ramp finds capacity.
omni-kern	Manifest builder, weight loading, common kernels and host helpers, moved out of omni-qwen3-tts so both GPU engines share them.
omni-personaplex	Mimi encoder → Helium → depformer → Mimi decoder as one kern manifest with its own kernels, one CUDA graph per tick over all sessions. Each session has a fixed ring of the model's 3000-frame context plus streaming codec state. The engine leases a session on arrival, prefills it in the next tick batched with other newcomers, and streams audio and the model's text. The role-prompt limit is derived so any allowed prompt fits the context. tests/host.rs covers batching, the tokenizer, the voice reader and the prompt without a checkpoint, so CI runs it.
pega-omni personaplex	Server subcommand behind the personaplex feature.
tools/live_check.py	GPT-Live protocol oracle. CI runs it plus a duplex ramp against sim-live.
docs/	duplex.md: how full duplex works, the protocol, options, design choices. personaplex.md: the engine. PersonaPlex section in bench.md.

Correctness

crates/omni-personaplex/tests/golden.rs runs the serving path against the reference implementation's recorded run. It uses three sessions: two in lockstep (bit-identical to each other) and one joining mid-run.

Check	Result
Input rows and emitted frames	exact
Text logits	cosine ≥ 0.998
Depformer logits	cosine ≥ 0.9996
Depformer argmax	98.8% agree
Voiced caller semantic codes	100% agree
Session prefilled in other batches vs its lockstep twin	> 40 dB
Decoded audio	SNR 31.8 dB

The qwen3-tts golden test still passes after the omni-kern move.

Performance (one GB300)

128 concurrent full-duplex sessions, each streaming real speech at real time, run clean:

Metric	Result
Playback underrun	none
Late ticks	none once warm (the first call at each batch size captures its graph)
Frame lateness p99	12.6 ms
Tick at 128 sessions	15 ms of its 80 ms budget

The ceiling is memory (1.5 GiB of KV per session), not compute. Details are in docs/bench.md.

Try it

cargo build --release -p omni-server --features personaplex
target/release/pega-omni personaplex --model-path personaplex-7b-v1
# open http://127.0.0.1:8000/ (for a remote GPU: ssh -L 8000:127.0.0.1:8000)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.