NVIDIA-NeMo / NVIDIA-NeMo/Automodel

[Tracking] Kimi K3 roadmap

Open
#3,278 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
963
Forks
318
Avg merge
3d 20h
Merged PRs (30d)
143

Description

Kimi K3 training currently runs only on GB200. This roadmap tracks GB200 support and optimization.

Last refreshed: 2026-08-19.

Status: ✅ validated · 🚧 waiting on validation · 📋 planned · ❌ known broken

1. Core Training

Feature GB200 Notes
Basic training (EP + PP + FSDP) Implemented in #3259; GB200 recipe and CI topology corrected in #3420
Vision training 🚧 Implementation landed in #3259; full-model recipe validation pending

2. Fusion & Performance

Feature GB200 Notes
Transformer Engine Gated MLA Implemented in #3259; correctness validated without CP, full-model performance benchmark planned
FlashKDA 📋 P2; implementation PR pending
MoonEP 📋 P0; tracked in #3251, implementation PR pending
Attention Residuals fusion 📋 P2; implementation PR pending
Grouped GEMM + SiTU-GLU 📋 P2; base grouped-expert support landed in #3259, fused path PR pending

3. Parallelism

Feature GB200 Notes
PP + EP + FSDP Validated in #3259; GB200 recipe follow-up in #3420
Context Parallel 🚧 CP1/CP2 parity validated in #3259; CP recipe validation pending
HybridEP Validated in #3259

4. Related Kimi K3 Work

Feature Status PR
DSpark draft training #3263
EAGLE-3 draft training #3286
DFlash draft training 🚧 #3287
Expert LoRA HF export #3435

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing the status table and the linked PRs and issue, especially #3251, #3259, #3420, and #3435. This is a broad roadmap rather than a self-contained task; completion would require implementing or validating the remaining GB200 items and updating their statuses.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.