Search results

[QUESTION] How to convert model from hf to support run with --expert-tensor-parallel-size= 2 with attention tp=1

questionFurther information is requestedFurther information is requested

Status: Open.

#1766 In NVIDIA/Megatron-LM;

· hxq2333 opened on Aug 22, 2025

[BUG] Context parallel <code>nan</code> loss

bugSomething isn't workingSomething isn't working

Status: Open.

#1764 In NVIDIA/Megatron-LM;

· leoleoasd opened on Aug 21, 2025

[BUG] --async-save mistakenly recorded the time when the model was saved in the log

bugSomething isn't workingSomething isn't working

Status: Open.

#1763 In NVIDIA/Megatron-LM;

· 861482002 opened on Aug 21, 2025

[QUESTION] Is there official support to convert a self-defined llama-style model from HuggingFace format to megatron core format?

questionFurther information is requestedFurther information is requested

Status: Open.

#1760 In NVIDIA/Megatron-LM;

· isaacveg opened on Aug 20, 2025

[ENHANCEMENT] Support native mxfp4 training of gpt-oss

enhancementNew feature or requestNew feature or request

Status: Open.

#1759 In NVIDIA/Megatron-LM;

· Opdoop opened on Aug 20, 2025

[BUG] RuntimeError: CUDA error: operation not supported

bugSomething isn't workingSomething isn't working

Status: Open.

#1758 In NVIDIA/Megatron-LM;

· EvilCalf opened on Aug 20, 2025

[QUESTION] Multiple nodes with different GPU type, How to configure?

questionFurther information is requestedFurther information is requested

Status: Open.

#1757 In NVIDIA/Megatron-LM;

· justalittlenoob opened on Aug 20, 2025

[BUG]Tensor parallelism produces divergent loss with non-divisible vocabulary sizes

bugSomething isn't workingSomething isn't working

Status: Open.

#1754 In NVIDIA/Megatron-LM;

· Tona8941 opened on Aug 19, 2025

[BUG]--cross_entropy_fusion_impl te results in grad_norm 0

bugSomething isn't workingSomething isn't working

module: transformer engine

Status: Open.

#1750 In NVIDIA/Megatron-LM;

· Jintao-Huang opened on Aug 17, 2025

[BUG]MoE FA supporting for DS V2 Lite(16 attention header) is incorrect

bugSomething isn't workingSomething isn't working

module: moe

Status: Open.

#1749 In NVIDIA/Megatron-LM;

· joyjie818 opened on Aug 16, 2025

[BUG]Load DCP OOM

bugSomething isn't workingSomething isn't working

Status: Open.

#1746 In NVIDIA/Megatron-LM;

· zjjott opened on Aug 14, 2025

[QUESTION] In MoE, how is the correctness of Sinkhorn guaranteed when enabling sequence parallelism?

questionFurther information is requestedFurther information is requested

Status: Open.

#1740 In NVIDIA/Megatron-LM;

· IamArabbit opened on Aug 12, 2025