Feature: run the minimax_music3 DiT and depth decoder in bf16 under precision=bf16 (DiT 16x to 19x faster per call)

未关闭
  • 哪里不对？
#2,053 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
5/5
预计耗时
一周以上
新手友好度
42/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
python, pytorch
领域
machine-learning

调研方向

Start with the minimax_music3 family implementation and its precision handling, then trace how the DiT, depth decoder, and language-model KV cache are built and used. Reproduce the reported measurements on the specified model and GPUs, and check existing end-to-end quality thresholds before changing precision or attention lengths. Done means bf16 precision speeds up the proposed components while meeting measured quality thresholds, and decode attention processes only filled cache rows.

由索引模型根据 Issue 内容生成。

描述

enhancement

Request type: Improvement to existing behavior

Problem and use case

families/minimax_music3 requires precision=bf16, but only the language model is built in bf16. The DiT, the RVQ depth decoder, the condition encoder and the vocoder are built in float32 with TF32 off, so their GEMMs run on CUDA cores. For a 20 s song (the minimax-music3-l0 prompt, native runtime, 500 frames) the engine time is dominated by these float32 parts:

GPU Engine time per song Language model Depth decoder DiT Vocoder
B200 49.4 s 13.44 s 12.19 s 23.57 s 0.23 s
B300 43.3 s 12.34 s 12.22 s 18.56 s 0.21 s

The same graphs built with bf16 blocks (a test-harness build, not offered by the family) are much faster per call:

Engine GPU Default, float32 (ms) Same graph, bf16 (ms) Speedup
DiT call B200 / B300 98.36 / 77.25 5.28 / 4.67 18.6x / 16.5x
Depth decoder call B200 / B300 1.628 / 1.641 0.405 / 0.410 4.0x / 4.0x

A second, separate cost: every language-model decode step attends over the whole static 9216-row KV cache. In a B200 kernel profile of the bf16 step (11.0 ms), the decode attention kernel takes 5.4 ms (36 calls over 9216 rows), the cache concatenation 0.46 ms and the GEMMs about 4.3 ms, so the step is bound by attention over unused cache rows, not by weight reads.

Proposed outcome

  1. Under precision=bf16, build the DiT blocks and the depth decoder in bf16 (keeping float32 only where measurement shows it is needed, for example the timestep embedding, stem convolutions or norms), with E2E thresholds set from measured bf16-vs-reference numbers. Optionally the same for the condition encoder and vocoder if quality allows.
  2. Limit the language-model decode attention to the filled cache length instead of all 9216 rows.

Model and target details

MiniMaxAI/MiniMax-Music3 @ fbdf52fbaaca799592917417eb05f1899f1255ec (the minimax-music3-l0 manifest); TensorRT 11.4; B200 / B300. The DiT has 36 blocks (hidden 2048, 60 calls per 8 s window); the depth decoder has 4 layers (hidden 4096, 15 calls per frame, 7515 calls per 20 s song).

Alternatives considered

FP8 for these parts was measured (not merged): it gets most of its gain over the default from running the DiT blocks in bf16 (FP8 DiT call 4.25 ms against 5.28 ms for the bf16 graph, 1.24x), so a plain bf16 build captures most of the benefit with no quantization.

Additional context

Found during an FP8 study of this family (not merged).

主要语言
Python
星标
288
派生
73
平均合并
1 天 3 小时
30 天内合并 PR
273

环境准备

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/TensorRT-Model-Connect 的其他 Issue

查看 NVIDIA/TensorRT-Model-Connect 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。