Search results

[QUESTION] Missing CLI argument for <code>outer_dp_sharding_strategy</code> in Megatron-LM training scripts

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#3038 In NVIDIA/Megatron-LM;

· cailun01 opened on Jan 22, 2026

Fake communication support for MoE models with all_to_all_single

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#3025 In NVIDIA/Megatron-LM;

· skeyyk opened on Jan 21, 2026

Feature Request: Dynamic Inference Context Support for T5

enhancementNew feature or requestNew feature or request

Status: Open.

#3016 In NVIDIA/Megatron-LM;

· sbhavani opened on Jan 20, 2026

[QUESTION]How can I confirm that I have completed one epoch in Megtron

community-request

questionFurther information is requestedFurther information is requested

Status: Open.

#3012 In NVIDIA/Megatron-LM;

· Xiaotianhe123 opened on Jan 20, 2026

[QUESTION] how to correctly reproduce deepseek-r1 pre-training performance on GB200

bugSomething isn't workingSomething isn't working

community-request

module: moe

Status: Open.

#3000 In NVIDIA/Megatron-LM;

· yuqie opened on Jan 19, 2026

[QUESTION] Question for DSA

community-request

module: moe

questionFurther information is requestedFurther information is requested

Status: Open.

#2998 In NVIDIA/Megatron-LM;

· xhjhggybz opened on Jan 19, 2026

[QUESTION] Unexpected updates on unused token embeddings when using --untie-embeddings-and-output-weights

community-request

needs-follow-upIssue needs follow-upIssue needs follow-up

questionFurther information is requestedFurther information is requested

Status: Open.

#2994 In NVIDIA/Megatron-LM;

· bugm opened on Jan 19, 2026

[QUESTION] Support for Xiaomi MiMo-V2-Flash Architecture (Hybrid Attention + Fine-Grained MoE)

community-request

enhancementNew feature or requestNew feature or request

Status: Open.

#2976 In NVIDIA/Megatron-LM;

· Eisenhower opened on Jan 16, 2026

Qwen3-Next-80B saving checkpoint ShardedTensor.flattened_range is not supported.

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#2956 In NVIDIA/Megatron-LM;

· yangyuxiang1996 opened on Jan 15, 2026

Muon optimizer enhancement

enhancementNew feature or requestNew feature or request

module: optimizer

Status: Open.

Enhancement

#2936 In NVIDIA/Megatron-LM;

· BoxiangW opened on Jan 13, 2026

·  

Core 0.16

Design Proposal: mHC (Manifold-Constrained Hyper-Connections)

dev branchDev branch related issues and developmentDev branch related issues and development

enhancementNew feature or requestNew feature or request

module: moe

Status: Open.

#2919 In NVIDIA/Megatron-LM;

· Victarry opened on Jan 13, 2026

[QUESTION] Why is Multi-Token Prediction not supported with the mRoPE

community-request

questionFurther information is requestedFurther information is requested

Status: Open.

#2918 In NVIDIA/Megatron-LM;

· liudl85 opened on Jan 13, 2026