Search results

[MoE] Revive FP16 training for MoE

enhancementNew feature or requestNew feature or request

module: moe

Status: Open.

Enhancement

#2102 In NVIDIA/Megatron-LM;

· yanring opened on Nov 3, 2025

[Docs-only] Clarify seq_length behavior when variable_seq_lengths=True under pipeline parallelism (PP>1)

enhancementNew feature or requestNew feature or request

module: documentation

Status: Open.

#2064 In NVIDIA/Megatron-LM;

· adol001 opened on Oct 31, 2025

[QUESTION] Enabling <code>--fp8-param-gather</code> during model training with <code>mxfp8</code> consumes more GPU memory compared to leaving it disabled.

module: transformer engine

questionFurther information is requestedFurther information is requested

Status: Open.

#2063 In NVIDIA/Megatron-LM;

· cailun01 opened on Oct 31, 2025

Add assertion to check for rerun_state_machine.check_for_nan_in_loss and ddp.check_for_nan_in_grad to False with full-iteration cuda graph

bugSomething isn't workingSomething isn't working

Status: Open.

#2056 In NVIDIA/Megatron-LM;

· gautham-kollu opened on Oct 30, 2025

Multi stream sync problem in model init

bugSomething isn't workingSomething isn't working

Status: Open.

#2044 In NVIDIA/Megatron-LM;

· liddk opened on Oct 30, 2025

[QUESTION] Is post-training currently supported on Megatron-LM?

questionFurther information is requestedFurther information is requested

Status: Open.

#2040 In NVIDIA/Megatron-LM;

· kailashg26 opened on Oct 30, 2025

Attribute Error on Saving FSDP_DTensor with EP8

bugSomething isn't workingSomething isn't working

module: moe

Status: Open.

#2035 In NVIDIA/Megatron-LM;

· Skylion007 opened on Oct 29, 2025

RFC: Async overlap of CPU and GPU compute during dynamic inference step

enhancementNew feature or requestNew feature or request

Status: Open.

#2019 In NVIDIA/Megatron-LM;

· tdene opened on Oct 29, 2025

Loading a legacy pre014 optimizer state corrupts any future optimizer states

bugSomething isn't workingSomething isn't working

Status: Open.

#2016 In NVIDIA/Megatron-LM;

· Skylion007 opened on Oct 29, 2025

Do not auto select NVTE FUSED if CUDNN_FRONTEND version is less than 1.15.0

bugSomething isn't workingSomething isn't working

Status: Open.

#2011 In NVIDIA/Megatron-LM;

· Skylion007 opened on Oct 28, 2025

Megatron FSDP does not work with CPU_OFFLOADING

bugSomething isn't workingSomething isn't working

Status: Open.

#1986 In NVIDIA/Megatron-LM;

· Skylion007 opened on Oct 27, 2025

moe routing loss incorrect for post-training

bugSomething isn't workingSomething isn't working

Status: Open.

Bug

#1984 In NVIDIA/Megatron-LM;

· gshennvm opened on Oct 27, 2025

· 

Core 0.15