Nemotron-H support

未关闭
  • 哪里不对？
#198 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

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
5/100
Issue 类型
功能
描述清晰度
需要澄清
活跃度
停滞
技术栈
python
领域
machine-learning

调研方向

This is blocked on two things: NVIDIA has not released Nemotron-H architecture details or weights, and Mamba-2 support (#68) is a prerequisite. Before starting, check whether the architecture is public and whether #68 has landed. The issue names no files or tests, so there is no clear entry point. Done would mean an 8B Nemotron-H model loads and trains in BF16 with forward and backward passes tested.

由索引模型根据 Issue 内容生成。

描述

enhancement need update

🎯 Goal (What & Why)

Add support for training Nemotron-H models.

Nemotron-H is a family of hybrid SSM-Transformer models (8B, 47B, 56B) trained by NVIDIA in FP8 on 20T tokens. They combine strong accuracy with up to 3x faster inference, making them attractive candidates for instruction tuning, continual pretraining, and up-cycling (e.g., via added layers). NVIDIA plans to release weights and reference implementations (HuggingFace, NeMo, Megatron-LM), but nothing is public yet.

We want to be ready to support Nemotron-H in Fast-LLM once details are available. These models require both Transformer and Mamba-2 layers. The forward and backward pass must be implemented. FP8 support is likely not required for initial work (we'll use BF16), but is tracked separately in #63. Mamba-2 support is tracked in #68 and is a prerequisite for this ticket.

Partially enabled by #242.

🚀 Execution Plan

Step 1: What is the smallest working version?
  • Wait for NVIDIA to release architectural details.
  • Implement Nemotron-H block structure (interleaved Mamba-2, MLP, and Transformer layers).
  • Use BF16 as default precision for now.
  • Validate by running short training in Fast-LLM with Nemotron-H-8B or 47B weights (if available).
Step 2: What additional optimizations are possible (but optional)?
  • Add support for FP8 training (#63) once the NVIDIA recipe is available and stable.
  • Extend support for VLM variants if we decide to do multimodal training.
  • Optimize long-context support (128k tokens) once basic functionality is in place.

📌 Acceptance Criteria

  • Nemotron-H models (at least 8B) can be loaded and trained in Fast-LLM.
  • Both forward and backward passes are implemented and tested.
  • Training is functional in BF16.
  • Code is documented clearly and includes a usage example.
  • Performance/impact summary is provided in the PR.
  • FP8 support is not part of this ticket but should not block anything here.

🛠️ Project Management

  • Assign the project to the Fast-LLM project.
  • Set the Estimate field (in days) in the GitHub project.
  • Use the Size field to categorize the PR size (Small/Medium/Large).
  • Assign an owner when opening the issue.
主要语言
Python
星标
337
派生
52
30 天内没有已合并 PR

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。

ServiceNow/Fast-LLM 的其他 Issue

查看 ServiceNow/Fast-LLM 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。