Feature: nemotron_omni decode is bound by 2944 conditional expert subgraphs per token (about 50 ms/token for a 3B-active model)

未关闭
  • 哪里不对？
#2,060 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
35/100
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 families/nemotron_omni/moe.py and inspect how each routed expert is built as an IIfConditional subgraph. Compare this with another family’s gathered MoE implementation, then investigate whether the selected experts’ weights can be run as a grouped or batched GEMM. Done means reducing per-token decode time and decoder plan and build size; the issue gives B200/B300 decode measurements for comparison.

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/nemotron_omni builds each MoE layer of the Nemotron-H language model as 128 IIfConditional expert subgraphs (moe.py, one per routed expert; top-6 run). With 23 MoE layers that is 2944 conditional expert subgraphs per decode step, and the decoder plan has about 25k to 64k layers.

On the official NVFP4 checkpoint (the manifests' nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4), decode takes about 48 to 53 ms per token on B200 / B300 (about 20 tokens/s), for a model with about 3B active parameters (reading 3B parameters at about 8 TB/s would take well under 1 ms). Prefill is token by token through the same step, so it has the same cost per token.

The step time does not depend on the weight precision, which shows the step is bound by the conditional structure and kernel count, not by weight reads (decoder only, TensorRT 11.4, batch 1, ms per token at cache positions 100 / 2000):

GPU fp16 (from the BF16 checkpoint, built with the family decoder code) FP8 (official FP8 checkpoint) NVFP4 (as on main)
B200 47.7 / 47.8 48.3 / 48.4 49.7 / 49.8
B300 48.9 / 48.8 48.9 / 48.7 49.7 / 49.6

Decoder build time is also large: 24 to 52 minutes on B200 / B300 and about 101 minutes on H100 for NVFP4.

Proposed outcome

Build the routed experts of each MoE layer without one conditional subgraph per expert, for example a gathered/batched expert formulation (gather the selected experts' weights and run them as one grouped or batched GEMM, as other families do with a gathered MoE), so the decode step time scales with the active weights and the plan and build time shrink.

Model and target details

nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 @ 16993199e436da4ba75ddc410855f87e0d996ee6 (the nemotron-omni-nvfp4-native manifest). Language model: 52 layers (23 Mamba2, 23 MoE with 128 routed experts top-6 plus 1 shared, 6 attention), hidden 2688. TensorRT 11.4; B200 / B300 (also seen on H100).

Alternatives considered

FP8 weights (official FP8 checkpoint) and fp16 weights give the same step time as NVFP4 (table above), so weight precision does not help while the MoE stays conditional.

Additional context

Found during an FP8 study of this family (not merged). Separate note for anyone building a Hugging Face reference for Nemotron-H: without mamba-ssm, the checkpoint's own PyTorch fallback in modeling_nemotron_h.py diverges after the first 128-token chunk (KL about 1e-8 on positions 0 to 127, 0.07 to 0.39 after); a plain sequential scan matches the fp16 engine (KL 0.0032).

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。