Kimi-K3 training support

Author: yuzhongw-nvidiaCreated Aug 26, 2026Updated Sep 18, 2026
Labelsenhancement

This issue tracks Kimi-K3 training support in Megatron Core.

Kimi-K3 combines a hybrid Kimi Delta Attention (KDA) / gated Multi-Head Latent Attention (MLA) backbone, Attention Residuals (AttnRes), Stable LatentMoE, global-batch Quantile Balancing (QB), MoonViT-V2, per-head Muon, and native MXFP4 quantization-aware training.

Overall status: the baseline KDA/gated-MLA hybrid path, Stable LatentMoE, and non-CUDA-Graph global-batch QB are available in dev. The September 18 preview adds MTP, per-head Muon for KDA/MLA, and initial long-context training support, including TP/CP numerical alignment with packed THD inputs up to 256K tokens. Full-model and 9L-proxy MXFP8 recipes are published with measured performance. K3-specific upstream integration, full-model checkpoint resume, and broader correctness and convergence qualification remain in progress.

Status: ✅ available in the stated scope / merged / resolved · in progress / open / draft · planned / pending validation

Availability in the frozen preview and upstream merge status are stated separately. A merged supporting PR does not necessarily mean the entire capability has completed end-to-end qualification.

Primary scope follows the Kimi-K3 section of the 2026 Q3 Megatron Core MoE roadmap. Lower-priority MoonViT and per-head Muon work, plus native MXFP4 training, are retained here so the tracker covers the intended Kimi-K3 training stack without turning into a complete paper gap list.

Last updated: 2026-09-18

Status at a glance

Core functionality

Capability Status Priority Summary
Model architecture P0 Text backbone integrated in the preview; K3 attention and AttnRes upstream PRs remain open
MTP ✅ Preview P0 Hybrid KDA/MLA + AttnRes MTP is supported and enabled in both published recipes
Distributed and long-context training ✅ Preview P0 TP/CP + packed THD alignment up to 256K; variable-length AttnRes with non-interleaved PP
Quantile Balancing P0 Core QB and TE fixes are merged; preview adds padding-aware statistics; end-to-end CUDA Graph qualification remains open
Per-head Muon P0.5 Available in preview recipes; #7085 and #7395 remain unmerged
MoonViT and multimodal integration P1 MoonViT-V2 and the multimodal projector/token-integration path are planned
Training validation and recipe P0 Full-model and 9L-proxy 256K recipes and benchmarks published; resume, broader correctness and convergence qualification remain pending

Optimization work

Area Status Priority Current focus
Pipeline-parallel efficiency P0.5 Highest optimization priority: reduce the substantial PP communication bubbles and improve utilization across stages
Long-context CP layout P0.5 Adapt the Qwen3.5 layout/routing-metadata optimizations from #6387 to Kimi K3
Fine-grained memory controls P0.5 Selective recompute and chunked optimizer-state offload are used in the recipes; explore more activation-offload/recompute choices
Kernel and runtime optimizations P0.5 Updated compiled/FLA AttnRes, fused KDA pre-GDR, MLA gate JIT fusion and configurable validation frequency are in the preview; upstreaming continues
MLA latent context parallelism P1 Experimental implementation remains draft in #6829
CUDA Graphs P1 Qualify the merged TE bounds fix in the K3 stack and validate end-to-end KDA + MoE capture/replay
Native MXFP4 training P1 Add the K3 MXFP4-weight / MXFP8-activation QAT, optimizer, and checkpoint path

Kimi-K3 Preview Branch

The rolling Kimi-K3 preview development branch is kimi_k3_dev. It is the latest integration point for Kimi-K3-specific work ahead of upstream completion and should not be treated as an immutable release.

Preview history

Preview Commit Abstract
kimi_k3_preview_20260907 ba527e6ef2cb2b34eabd77c4aeb86fb1a7ae76a4 First frozen Kimi-K3 preview: integrates low-rank KDA, gated MLA with per-layer NoPE, native Block AttnRes, Stable LatentMoE with SiTU-GLU, and global-batch QB in one training stack.
kimi_k3_preview_20260918 4adbeff83d019ebe242faba5b5d02e15ffa7513e Extends functional support with MTP and per-head Muon, establishes TP/CP + packed THD training up to 256K, updates AttnRes/KDA/MLA optimizations, and publishes full-model and 9L-proxy recipes with measured throughput.

1. Core Functionality

1.1 Model Architecture

Status: ✅ The text-backbone composition is available in the frozen preview; K3 attention extensions and native AttnRes are still being upstreamed.

Component Status Current work
KDA ✅ Preview / upstream First-class hybrid KDA landed in #6556; K3 low-rank F-decay/output-gate projections from open #6877 are integrated in the preview
Gated MLA ✅ Preview / upstream Baseline gated MLA landed in #6556; the preview includes per-layer NoPE, latent-projection extensions and elementwise output-gate JIT fusion from open #6877
Attention Residuals ✅ Preview / upstream Native Block AttnRes from draft #6840, updated compiled/FLA paths, and packed-sequence support with non-interleaved PP are integrated in the preview
MTP ✅ Preview / upstream Hybrid KDA/MLA + AttnRes MTP from #6840 is integrated; both published recipes enable one MTP prediction depth
Stable LatentMoE ✅ Available in dev Latent RMSNorm and dense/routed/shared-expert SiTU-GLU support are merged; focused BF16 forward/backward parity passed

Stable LatentMoE work

Capability Status Tracking / PRs
Kimi-K3 latent-MoE RMSNorm ✅ Merged #6804
SiTU-GLU MCore integration ✅ Merged #6673
SiTU-GLU Transformer Engine support ✅ Merged TransformerEngine#3402
SiTU-GLU cuDNN Frontend backend ✅ Merged cudnn-frontend#645
SiTU-GLU backend API-contract fixes ✅ Merged cudnn-frontend#670

The optional post-combine output normalization in #6449 is not required by the current Kimi-K3 target configuration and is no longer on this tracker's K3 critical path.

1.2 Quantile Balancing

Status: ✅ The eager/non-graph K3 global-batch path and fused histogram backend are available.

Capability Status PR
Baseline Quantile Balancing ✅ Merged #5349
Kimi-K3 global-batch Quantile Balancing ✅ Merged #6637
Dense routing maps for Flex dispatch ✅ Merged #6614
Fused QB router histogram path ✅ Merged TransformerEngine#3395
Mutable QB bounds under CUDA Graphs ✅ Merged; K3 graph qualification pending TransformerEngine#3426

The preview also excludes padded tokens from QB histograms and aligns padded SBHD validation with packed THD semantics. End-to-end CUDA Graph readiness is tracked separately below.

1.3 Per-head Muon

Status: ✅ Available in the frozen preview and both released recipes; upstream integration and checkpoint continuity qualification remain in progress.

Open #7085 provides per-head Q/gate/K/V orthogonalization for supported MHA/GQA/MLA layouts and TP handling for fragmented query-group blocks. The preview adds KDA projection splitting and the replicated Transformer Engine weight metadata fix from draft #7395. Both published 256K recipes enable per-head Muon with duplicated TP mode. Upstreaming the K3-specific integration and validating checkpoint/resume continuity remain open.

1.4 MoonViT and Multimodal Integration

Status: P1 work planned.

  • Add the MoonViT-V2 vision encoder training path.
  • Add the multimodal projector and packed image/video-token integration with the Kimi-K3 language model.

No public MCore implementation PR is currently tracked.

1.5 Training Validation and Recipe

Status: Distributed correctness coverage and published recipes have progressed; complete full-model training qualification remains pending.

  • BF16 module-level forward/backward parity for KDA, MLA NoPE + output gate, Stable LatentMoE, AttnRes, SiTU-GLU, and QB.
  • Align supported TP/CP + packed THD configurations with the padded SBHD baseline, establishing training support up to 256K context length.
  • Exercise MTP and per-head Muon in full-model and 9L-proxy 256K MXFP8 runs.
  • Publish full-model and 9L-proxy recipes, shared container build/launch instructions, and recorded throughput and memory results.
  • Qualify full-model checkpoint save/resume, including optimizer state continuity.
  • Extend full-model correctness coverage and establish short-convergence/training-quality evidence with real data and normal router balancing.

Released recipes and benchmarks

Both recipes use MXFP8, packed THD sequences of 262144 tokens, a global batch size of 64, one MTP prediction depth, and per-head Muon. The 9L proxy keeps the full model's widths and expert count while reducing the backbone to nine decoder blocks.

Recipe Measured hardware TP / PP / EP / CP Median TFLOP/s/GPU
Full text backbone, 93 decoder blocks 256 NVIDIA GB300 GPUs 4 / 4 / 64 / 16 498
9L proxy 64 NVIDIA GB300 GPUs 4 / 1 / 64 / 16 661

The proxy configuration targets GB200; the reported measurement was collected on GB300. Results are medians over steps 6–18 of 20-step runs using mock data and forced balanced routing. The benchmark guide records iteration times, peak memory and the measurement environment, including the HybridEP revision difference from the published Dockerfile. These runs establish executable recipes and measured performance; convergence qualification remains open.

The published recipes use Megatron-LM's pretrain_hybrid.py. A separate Megatron Bridge model provider or multimodal delivery path remains outside these text-backbone recipes.

2. Optimization Work

The next priorities are PP communication bubbles, Kimi-K3 adaptation of the Qwen3.5 CP layout work, and further activation-offload/recompute tradeoffs.

Area Status Plan / dependency
Pipeline-parallel efficiency Planned Reduce the substantial communication-related bubbles in the full-model PP4 run and improve utilization across pipeline stages
Long-context CP layout Planned Adapt the merged Qwen3.5 work in #6387 to Kimi K3, reducing layout conversions and reusing packed-sequence routing metadata
Fine-grained recompute/offload In progress Selective KDA/MoE/MLP/MLA-up-projection recompute and chunked optimizer-state offload are used in the recipes. Explore finer choices with AttnRes to balance memory and throughput; the preview also includes the FP8 gather/offload fix from open #7392
KDA fusion and runtime validation ✅ Preview / upstream Both recipes enable fused pre-GDR preprocessing from draft fork PR #7. Configurable validation frequency from draft #7202 and its KDA integration reduce repeated synchronization
AttnRes and MLA output gate ✅ Preview / upstream Updated compiled/FLA AttnRes and the single-source fast path are integrated; FLA is the default, while the recipes use compiled AttnRes. Elementwise MLA gating uses JIT fusion. #6840 and #6877 remain unmerged
MLA latent context parallelism Draft Complete #6829 and validate correctness, memory, communication and scaling; this is additional optimization work beyond the standard CP path used by the 256K recipes
CUDA Graphs In progress TransformerEngine#3426 is merged. Qualify a K3 stack containing the fix and validate end-to-end KDA + MCore MoE capture/replay, building on cudnn-frontend#556
Native MXFP4 training Planned Support K3's MXFP4 weights / MXFP8 activations from QAT through optimizer update, distributed execution and checkpointing; the current published recipes use MXFP8

References