/v1/audio/transcriptions rejects the model name the server was started with (404 unknown_model) since 0.3.0

未关闭
  • 哪里不对？
#94 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
3/5
预计耗时
1-2 天
新手友好度
55/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp
领域
api

调研方向

Start with the API behavior described in docs/api.md, then reproduce the issue using the supplied nemo-speech serve and curl commands. Trace how /v1/models builds the ASR model ID and how /v1/audio/transcriptions validates the model field. Done means clients can use a stable, discoverable model name for transcription without receiving unknown_model.

由索引模型根据 Issue 内容生成。

描述

Summary

Since 0.3.0, POST /v1/audio/transcriptions validates the model field against the IDs in GET /v1/models and returns 404 unknown_model for anything else (as described in docs/api.md).

However, the ID reported for the ASR model is not a name a client would know. When started with --asr-model nemotron-3.5, the server lists the model ID as .nemotron-3.5-asr-streaming-0.6b.q8_0.gguf, prepending a leading dot that the file on disk does not have. As a result, a client that sends the name used to start the server (the indexed name nemotron-3.5 or the GGUF file name) gets a 404. On 0.2.0, the same request succeeded because the model field was ignored.

/v1/audio/speech behaves differently: it accepts magpie, magpietts, and even nonsense.

Repro

nemo-speech pull nemotron-3.5
nemo-speech serve --port 18999 --no-ui --asr-model nemotron-3.5 &
curl -s http://127.0.0.1:18999/v1/models | jq '.data[].id'
# ".nemotron-3.5-asr-streaming-0.6b.q8_0.gguf"

for m in nemotron-3.5 nvidia/nemotron-3.5-asr-streaming-0.6b \
         nemotron-3.5-asr-streaming-0.6b.q8_0.gguf .nemotron-3.5-asr-streaming-0.6b.q8_0.gguf ""; do
  printf '%-45s ' "model='$m'"
  curl -s -o /dev/null -w '%{http_code}\n' -X POST http://127.0.0.1:18999/v1/audio/transcriptions \
    -F [email protected] -F "model=$m"
done

On 0.3.0:

model='nemotron-3.5'                          404
model='nvidia/nemotron-3.5-asr-streaming-0.6b' 404
model='nemotron-3.5-asr-streaming-0.6b.q8_0.gguf' 404
model='.nemotron-3.5-asr-streaming-0.6b.q8_0.gguf' 200
model=''                                      200

The 404 response body is {"error":{"code":"unknown_model","message":"requested model is not available","param":"model","type":"invalid_request_error"}}. On 0.2.0, all five requests return 200.

Starting the server with the full GGUF path instead of the indexed name produces the same ID, including the leading dot. The server log also prints [asr] model=.nemotron-3.5-asr-streaming-0.6b.q8_0.gguf.

Expected

Either of these would work:

  • Accept the name the model was loaded with (the indexed name such as nemotron-3.5, or the GGUF file name) in addition to the listed ID.
  • List a stable ID without the leading dot (for example, the indexed name) so clients can discover and send it.

The leading dot appears unintended in any case.

Impact

Upgrading from 0.2.0 to 0.3.0 breaks OpenAI-compatible clients that send model (which the OpenAI SDK requires) unless they first look up the leading-dot ID. We worked around this by omitting model, which the docs state selects the loaded model.

Environment

  • nemo-speech 0.3.0, macOS 26 on Apple M5, Metal build (also reproduced the ID listing on 0.2.0)
  • ASR: nemotron-3.5 (nvidia/nemotron-3.5-asr-streaming-0.6b, q8_0)
主要语言
C++
星标
207
派生
47
平均合并
13 小时 10 分钟
30 天内合并 PR
12

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

NVIDIA/NeMo-Speech.cpp 的其他 Issue

查看 NVIDIA/NeMo-Speech.cpp 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。