Choose a runtime from product constraints

This is an OpenMOSS third-party model, not a FunASR model; no external VAD requirement does not imply an absence of internal segmentation. The vLLM path pins 0.27.1 and the SGLang Omni path pins 3f819f9c; revalidate real long multi-speaker audio before upgrades. SGLang Omni verbose_json keeps the speaker identifier in the [Sxx] prefix of segments[].text, which the FunASR adapter parses only after validating that the segment is backed by raw_text. LocalAI / moss-transcribe.cpp is another independent third-party reimplementation.

Current commands use official FunAudioLLM/Fun-ASR-Nano-2512-vllm@a4362c943d48951f98ca2a62181cc028970270c5, tested 2026-09-07 in an existing environment: vLLM 0.27.1+cu129 / Torch 2.13.0+cu129 / Transformers 5.15.0 / CUDA 12.9. This is not a clean installation validation; preparation only checks versions and downloads a local snapshot. Eight HTTP requests establish functional compatibility only; first Chinese was not warmed, and the 0.911 s two-request probe is not capacity evidence. The 2026-08-13 benchmark below remains a historical community-converted checkpoint result, neither relabeled nor compared with the new run. #54944 updates main docs/test registry only; v0.28.0 still referenced the community artifact at audit time, not a claim of released official references.

A TensorRT plan is not portable across arbitrary GPU architectures or TensorRT versions; rebuild and revalidate accuracy, memory, and capacity on the target stack.

v0.2.6 provides a dedicated sm_120 Blackwell package with official cuBLAS and a static MSVC runtime; ten-platform build, archive, and dependency audits do not prove operation on physical Blackwell hardware. A positive RX 9070 XT v0.2.5 retest also does not resolve every AMD Windows 0xC0000005 report; Android/Mali remains outside the prebuilt and validated targets.

Official verification currently covers Linux source builds; the server does not include production authentication, TLS, tenant quotas, or durable queues, so provide them at the gateway and infrastructure layers.

SenseVoice is merged into audio.cpp main@979e070f but is not yet in a tagged release. Pin that main commit until a release is published. Neither path currently provides word-level timestamps.

The example server does not include production authentication, tenant quotas, or full observability; provide them at the gateway and infrastructure layers.

Capacity depends on realtime ratio, chunks, VAD, connection count, and client backpressure; file RTFx does not replace concurrency tests.

The CPU example is not a CUDA production image; GPU drivers, image, scheduling, and capacity still require platform validation.

Model conversion, operator support, and thread settings vary by platform; measure capacity on the target CPU with real audio.

A checklist does not replace workload, accuracy, failure, and security exercises for the target product.