Issues

• Explore/benchmark FlashMask

  enhancement
  NVIDIA-NeMo/Automodel#4397 · Filter by author HuiyingLi opened on Oct 17, 2026
  

• Follow-up: add example config + discoverability for use_gigatoken (gigatoken backend)

  community-request
  NVIDIA-NeMo/Automodel#4367 · Filter by author aminehd opened on Oct 15, 2026
  

• NemotronH (force_hf: true): routed-expert weights silently not loaded — nemotron_h missing from MODELS_REQUIRING_TENSOR_MERGING

  community-requestwaiting-on-customer
  NVIDIA-NeMo/Automodel#4357 · Filter by author thejoecrabtree opened on Oct 14, 2026
  1

• Qwen3-30B-A3B QuACK+DeepEP benchmark exceeds 1h CI budget on 8xB200: first iteration ~48 min spent in QuACK gemm_tuned autotuning

  NVIDIA-NeMo/Automodel#4353 · Filter by author janbernloehr opened on Oct 14, 2026
  

• [bug] DeepSeek-V4-Pro: a few ranks on one node stall inside the HybridEP dispatch in step 0; every other rank times out in the next FSDP all-gather (2 occurrences, flight-recorder traces)

  community-request
  NVIDIA-NeMo/Automodel#4352 · Filter by author linmuchuiyang opened on Oct 14, 2026
  2

  

• Gemma4 audio eager mask: track HF #49404 until dependency upgrade

  NVIDIA-NeMo/Automodel#4342 · Filter by author HuiyingLi opened on Oct 13, 2026

• DSpark: non-colocated target serving for DeepSeek V4 / V4.1 training

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4325 · Filter by author khazic opened on Oct 13, 2026

• Expose chat_template_kwargs (e.g. thinking_mode) as a recipe option

  community-request
  NVIDIA-NeMo/Automodel#4324 · Filter by author khazic opened on Oct 13, 2026
  

• [Tracking] VR200 container readiness

  NVIDIA-NeMo/Automodel#4300 · Filter by author akoumpa opened on Oct 12, 2026
  1

  

• [RFC] Model-agnostic n-gram embedding (Engram-style) for training existing checkpoints

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4292 · Filter by author khazic opened on Oct 12, 2026

• [Feature Request/Discusion] Investigate and add LoRA fine-tuning support for Qwen-Image-Edit-2511

  community-requestenhancementwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4265 · Filter by author Alizestl opened on Oct 10, 2026
  2

• [recipe/docs] DeepSeek-V4-Pro recipes on H100: raise dist_env.timeout_minutes, enable the #3893 prewarm / shared HybridEP JIT cache, and document offline Hub resolution

  community-request
  NVIDIA-NeMo/Automodel#4229 · Filter by author linmuchuiyang opened on Oct 9, 2026
  

• [bug] distributed.defer_fsdp_grad_sync: false has no effect for MoE models under pipeline parallelism (r0.6.0 / NGC 26.08)

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4228 · Filter by author linmuchuiyang opened on Oct 9, 2026
  

• [bug] deepseek_v41_flash_tulu3_packed_cp8_32k.yaml: training completes on its target 16 × 4 GB200, but the update-99 checkpoint save OOMs in TE FusedAdam.state_dict (BF16 → FP32 moment expansion); the same save completes at 32 nodes (8.2 TB, 1:41)

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4209 · Filter by author linmuchuiyang opened on Oct 8, 2026
  3

  

• Audit packed-sequence metadata contracts across models, CP/PP, and MTP

  NVIDIA-NeMo/Automodel#4173 · Filter by author HuiyingLi opened on Oct 6, 2026
  

• ci: a PR that is green against a stale base can turn main red and fail other contributors' PRs

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4170 · Filter by author kabirvashisht4-glitch opened on Oct 6, 2026

• [Perf] Qwen3.8-Flash-Next: flex QSA mask memory at long sequences

  community-requestwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4147 · Filter by author rouchenzi opened on Oct 5, 2026
  6

  

• Qwen3.8-Flash-Next requires extra template handling for non-reasoning SFT data

  community-requestwaiting-on-customer
  NVIDIA-NeMo/Automodel#4145 · Filter by author rouchenzi opened on Oct 4, 2026
  2

  

• Explore supporting KohakuFA

  enhancement
  NVIDIA-NeMo/Automodel#4134 · Filter by author akoumpa opened on Oct 4, 2026
  

• [RFC] Explore deprecation of DDP in favor of FSDP2

  enhancement
  NVIDIA-NeMo/Automodel#4122 · Filter by author akoumpa opened on Oct 2, 2026
  1

• ChatDataset never supervises the end-of-turn token for GLM chat templates

  community-requestwaiting-on-customer
  NVIDIA-NeMo/Automodel#4108 · Filter by author waynel96 opened on Oct 1, 2026
  2

  

• fake_balanced_gate skips router computation, confounding load-balance benchmarks

  NVIDIA-NeMo/Automodel#4106 · Filter by author akoumpa opened on Oct 1, 2026
  

• [Tracking] followup checkpoint I/O architecture, correctness, performance, and memory

  ckptenhancementPerformancePoR
  NVIDIA-NeMo/Automodel#4096 · Filter by author akoumpa opened on Oct 1, 2026
  

• Kimi K3 MXFP4 support

  community-requestenhancementwaiting-on-maintainers
  NVIDIA-NeMo/Automodel#4078 · Filter by author sophiayyya opened on Oct 1, 2026
  1

  

• Support checkpointing and export of adapted MoE router biases

  NVIDIA-NeMo/Automodel#4075 · Filter by author yuhezhang-ai opened on Sep 30, 2026