Search results

[Bug] Gemma4 MoE backend missing vision-aware attention mask for use_bidirectional_attention="vision" models

bugSomething isn't workingSomething isn't working

Status: Open.

#1891 In NVIDIA-NeMo/Automodel;

· jQizhang opened on Apr 17, 2026

NaN loss with SDPA attention when local_batch_size > 1 on Gemma 4 4B

Status: Open.

#1883 In NVIDIA-NeMo/Automodel;

· HuiyingLi opened on Apr 16, 2026

Gemma4 26B-A4B MoE Expert Weights float32 After FSDP2 Prefetch PR

bugSomething isn't workingSomething isn't working

Status: Open.

#1863 In NVIDIA-NeMo/Automodel;

· sharonyu-115 opened on Apr 15, 2026

bug: MoE expert LoRA checkpoints incompatible with HF PEFT and vLLM (Nemotron Super 120B)

community-request

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

Status: Open.

#1814 In NVIDIA-NeMo/Automodel;

· pst2154 opened on Apr 13, 2026

Optimize causal-conv1d/transformer engine installation in CI

Automation

enhancementNew feature or requestNew feature or request

good first issueGood for newcomersGood for newcomers

Status: Open.

#1796 In NVIDIA-NeMo/Automodel;

· akoumpa opened on Apr 13, 2026

Fix PEFT adapter outputs for v5 support

bugSomething isn't workingSomething isn't working

good first issueGood for newcomersGood for newcomers

MoE

Status: Open.

#1755 In NVIDIA-NeMo/Automodel;

· adil-a opened on Apr 9, 2026

·  

26.06

Guidance on adapting a fine-tuning recipe into a pretraining recipe

community-request

enhancementNew feature or requestNew feature or request

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

Status: Open.

#1746 In NVIDIA-NeMo/Automodel;

· wswaq opened on Apr 9, 2026

Support LLaVA-OneVision-1.5 Model Training

community-request

enhancementNew feature or requestNew feature or request

good first issueGood for newcomersGood for newcomers

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

Status: Open.

#1744 In NVIDIA-NeMo/Automodel;

· randydl opened on Apr 9, 2026

Add optimized DSA kernels for GLM 5/5.1, DSv3.2

enhancementNew feature or requestNew feature or request

Status: Open.

#1715 In NVIDIA-NeMo/Automodel;

· hemildesai opened on Apr 7, 2026

Validate gemma4 31B and 26B MoE recipes on DGXSpark

Status: Open.

#1713 In NVIDIA-NeMo/Automodel;

· athitten opened on Apr 7, 2026

Support Qwen-Image finetune (T2I)

enhancementNew feature or requestNew feature or request

good first issueGood for newcomersGood for newcomers

Status: Open.

Feature

#1700 In NVIDIA-NeMo/Automodel;

· akoumpa opened on Apr 6, 2026

Remove separate moe_mesh references after https://github.com/NVIDIA-NeMo/Automodel/pull/1684

enhancementNew feature or requestNew feature or request

Status: Open.

#1699 In NVIDIA-NeMo/Automodel;

· hemildesai opened on Apr 6, 2026