Search results

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

community-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

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

questionFurther information is requestedFurther information is requested

Status: Open.

#2918 In NVIDIA/Megatron-LM;

· liudl85 opened on Jan 13, 2026

Feature Request: mHC (Manifold-Constrained Hyper-Connections)

enhancementNew feature or requestNew feature or request

module: moe

Status: Open.

#2890 In NVIDIA/Megatron-LM;

· sbhavani opened on Jan 9, 2026

Is there a hard blocker preventing Muon to be used with gated deltanet and gated attention?

community-request

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

module: optimizer

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

questionFurther information is requestedFurther information is requested

Status: Open.

#2885 In NVIDIA/Megatron-LM;

· MisterKloudy opened on Jan 9, 2026

VPP causes redundant dataset initialization leading to mmap virtual memory exhaustion

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#2860 In NVIDIA/Megatron-LM;

· excitingx opened on Jan 8, 2026

[QUESTION]How to convert a model from torchdist to hugginface？

bugSomething isn't workingSomething isn't working

community-request

Status: Open.

#2845 In NVIDIA/Megatron-LM;

· Listen-WLS opened on Jan 7, 2026

Feature Request: μP (Maximal Update Parameterization)

enhancementNew feature or requestNew feature or request

Status: Open.

#2824 In NVIDIA/Megatron-LM;

· sbhavani opened on Jan 6, 2026

How to convert checkpoint from fsdp_dtensor to torch_dist or huggingface in megatron-fsdp mode?

community-request

enhancementNew feature or requestNew feature or request

module: megatron-fsdp

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

Status: Open.

#2805 In NVIDIA/Megatron-LM;

· zhujian19891203 opened on Jan 5, 2026

Feature Request: Dual Chunk Attention for Long Context

enhancementNew feature or requestNew feature or request

Status: Open.

#2797 In NVIDIA/Megatron-LM;

· sbhavani opened on Jan 2, 2026