Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

Feature: Support fishaudio/drama-3 (multi-speaker TTS, 35B-A3B MoE)

オープン
  • 何か問題がありますか？
#1,700 コメント 4 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
6/100
issue の種類
機能追加
明瞭さ
説明が足りない
活発さ
活発
領域
machine-learning

調査の方向性

This is scoping work, not a code change. Start by checking whether the existing TTS task strategy covers reference-audio conditioning, speaker tags and streaming, and look at how the transform-model flow onboards a family. Identify the codec for the 10 codebooks and confirm whether TensorRT can consume the FP8 experts. Done means an agreed design and validation plan, not a patch.

索引モデルが issue の本文から書いたものです。

説明

Request type

New model or checkpoint support

Problem and use case

Request support for fishaudio/drama-3 (https://huggingface.co/fishaudio/drama-3), a multi-speaker, long-form text-to-speech model. TensorRT-Model-Connect has no family for it today, and no existing issue or PR covers it (searched for drama, fishaudio, fish audio).

Per the model card, the model:

  • has about 36B total parameters (safetensors, BF16 and F8_E4M3 tensors);
  • uses a 35B mixture-of-experts text backbone with 3B active parameters per token and hybrid full/linear attention;
  • uses a codebook-MTP audio decoder that decodes 10 codebooks causally over the whole sequence, with FP8 MoE experts and the remaining weights in BF16;
  • supports up to 16 speakers through <|speaker:0|> ... <|speaker:15|> tags, voice cloning from a reference clip plus transcript, point tags for emotion control, audio editing as continuation, and automatic chunking and streaming for long text;
  • is released under the Fish Audio Community License, and its reference deployment path is SGLang-Omni.
Proposed outcome

A drama_3 model family (or whatever name the maintainers prefer) that onboards the checkpoint through the normal transform-model flow and produces a bundle that passes family validation against the reference implementation.

Model and target details
  • Model: fishaudio/drama-3 (revision to be pinned at onboarding)
  • Task: text-to-speech with multi-speaker and reference-audio conditioning
  • Precision: BF16 with FP8 (E4M3) MoE experts as shipped; other precisions need separate evaluation
  • Target GPU: not decided
Open questions (unverified, need scoping before implementation)
  • Task fit: The current audio contract is mono-only. It is unclear whether the existing TTS task strategy covers reference-audio conditioning, speaker tags, and streaming output, or whether a new task or an extension is needed first.
  • Codec: The model card does not say which codec turns the 10 codebooks back into a waveform. It has to be identified and covered too.
  • Architecture overlap: The text backbone looks like a hybrid-attention MoE. Family ownership rules mean any code is duplicated into the new family rather than shared with qwen_moe or similar families.
  • FP8 experts: Need to confirm that TensorRT can consume the shipped FP8 expert weights, or whether a dequantized BF16 path is the baseline.
  • Size: 36B parameters is far above the current families. Single-GPU feasibility and the validation plan have to be settled.
  • License: The Fish Audio Community License terms need review before redistributing any bundle or fixture.
Alternatives considered

The model card's only documented runtime is SGLang-Omni.

主要言語
Python
スター
276
フォーク
73
平均マージ
1日 9時間
マージ済み PR（30日）
156

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

NVIDIA/TensorRT-Model-Connect のほかの issue

NVIDIA/TensorRT-Model-Connect の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。