5 Pull requests opened by 5 people

• Fix llama_mistral loader by using args.true_vocab_size

  #1491 opened Mar 20, 2025

• [Bug Fix] fix p2p communication order error and stuck problems when pp 2 and vpp 2 with remove pad

  #1495 opened Mar 22, 2025

• fix: MultiLatentAttention cp_comm_type

  #1499 opened Mar 24, 2025

• fix for group_limited_topk: K_r is moe_router_topk instead of moe_router_num_groups

  #1502 opened Mar 25, 2025

• Fix typo on distrib_optimizer.py

  #1505 opened Mar 26, 2025

4 Issues closed by 4 people

• [BUG] kv_channels get wrong for MLA

  #1501 closed Mar 25, 2025

• [BUG] Failed to install megatron_core: Could not find a version that satisfies the requirement nvidia-resiliency-ext; platform_machine == "x86_64"

  #1493 closed Mar 25, 2025

• [BUG] Incorrect `seq_aux_loss` impl for DeepSeek-V3

  #1438 closed Mar 24, 2025

• [BUG] "ValueError: optimizer got an empty parameter list" under pipeline parallel

  #1166 closed Mar 20, 2025

7 Issues opened by 7 people

• [QUESTION]How to set the routed scaling factor

  #1504 opened Mar 26, 2025

• [BUG] Wrong attention gradient in Transformer Engine

  #1503 opened Mar 26, 2025

• [BUG] T5 Model does not work when TP size is 1

  #1500 opened Mar 24, 2025

• [ENHANCEMENT] Global Batch Load Balancing for MoE Models

  #1498 opened Mar 23, 2025

• [BUG] Can not load _extra_state with TorchDistLoadShardedStrategy

  #1497 opened Mar 23, 2025

• [QUESTION] fp8 can not set pp>1

  #1496 opened Mar 22, 2025

• [BUG] Load Balancing loss discrepancy with/without CUDA Graphs

  #1494 opened Mar 21, 2025

15 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [ENHANCEMENT] Multi-token Prediction(MTP) support

  #1404 commented on Mar 21, 2025 • 0 new comments

• [QUESTION] plan to implement zero bubble pipeline or dual pipeline and MoE comm-comp overlapping

  #1399 commented on Mar 21, 2025 • 0 new comments

• [BUG] Checkpoint state dict remapping is not applied for MLA layers

  #1417 commented on Mar 21, 2025 • 0 new comments

• [BUG] Token routing probability all-gather precision in token_dispatcher causes differing results between EP ranks

  #1421 commented on Mar 21, 2025 • 0 new comments

• [QUESTION]convert LLaMA2-7B to the Megatron format failed: the converted model only repeats meaningless numbers

  #1365 commented on Mar 23, 2025 • 0 new comments

• [BUG] When trying to convert llama2-7b model from HF format to megatron format

  #1348 commented on Mar 23, 2025 • 0 new comments

• [QUESTION] checkpointing/loading memory overhead

  #1380 commented on Mar 24, 2025 • 0 new comments

• [BUG] can't load saved fp8 checkpoint when resume training

  #1350 commented on Mar 24, 2025 • 0 new comments

• [BUG] Using fp16 uses more memory than using fp32

  #1349 commented on Mar 24, 2025 • 0 new comments

• [QUESTION] Performance Impact of Using item() in `total_num_tokens += num_tokens.item()` in megatron/core/pipeline_parallel/schedules.py

  #1403 commented on Mar 25, 2025 • 0 new comments

• [QUESTION] Does MLA in Megatron-Core support PackedSeqParams?

  #1398 commented on Mar 25, 2025 • 0 new comments

• [ENHANCEMENT] Replace the hardcoded top-2 sum group selection strategy with configurable top-k

  #1441 commented on Mar 25, 2025 • 0 new comments

• [BUG]loss error when using MLA

  #1445 commented on Mar 26, 2025 • 0 new comments

• Enabling LR scaling for a specific layer (ex. down-projection...) during pretraining

  #1262 commented on Mar 25, 2025 • 0 new comments

• Add Mamba TRTLLM support

  #1320 commented on Mar 25, 2025 • 0 new comments