Repository navigation

feat(model): add Nemotron 3.5 Super VL support - #6100

Merged

cuichenx merged 65 commits into

mainfrom

chcui/sage/nmvl

Sep 21, 2026

Merged

feat(model): add Nemotron 3.5 Super VL support#6100
cuichenx merged 65 commits into
mainfrom
chcui/sage/nmvl

Commits on Aug 24, 2026
- feat(model): add multimodal bridge support
  cuichenx
  committed
  ef2c56f
Commits on Aug 25, 2026
- fix(model): remove video token pruning
  cuichenx
  committed
  f375ef5
- feat(recipes): add Nemotron 3.5 Super VL SFT
  cuichenx
  committed
  c7d458d
- docs(verification): add Nemotron 3.5 Super VL card
  cuichenx
  committed
  985dda4
- feat(perf): add Nemotron 3.5 Super VL H100 recipe
  cuichenx
  committed
  6de9975
- fix(recipes): relocate Super VL performance config
  cuichenx
  committed
  8a03bc0
- refactor(recipes): separate Super VL configuration
  cuichenx
  committed
  e10aedf
- fix(model): use dynamic temporal video token counts
  cuichenx
  committed
  58cb15a
- fix(data): default Omni video sizing to processor
  cuichenx
  committed
  33cd41c
- fix(nemotron-omni): align media anchors and video prompts
  cuichenx
  committed
  b48cce9
- fix(nemotron-omni): simplify video prompt policy
  cuichenx
  committed
  c0d7744
- fix(nemotron-omni): align temporal video inference
  cuichenx
  committed
  e9f10ff
Commits on Aug 26, 2026
- perf(nemotron-omni): release cached Super VL allocations
  cuichenx
  committed
  a499189
- perf(nemotron-omni): shard Super VL training with TP2
  cuichenx
  committed
  b809ee2
- docs(nemotron-omni): record Super VL H100 performance
  cuichenx
  committed
  ab3fb66
Commits on Aug 27, 2026
- perf(nemotron-omni): tune Super VL H100 training
  cuichenx
  committed
  26b2694
- docs(nemotron-omni): update Super VL H100 performance
  cuichenx
  committed
  d95975e
Commits on Aug 28, 2026
- feat(nemotron-omni): add Super VL pretraining recipes
  cuichenx
  committed
  1c459be
- docs(nemotron-omni): track real Super VL pretraining
  cuichenx
  committed
  68dd1a5
- fix(nemotron-omni): align GB200 HybridEP fusion
  cuichenx
  committed
  69ce2dc
- fix(nemotron-omni): disable unsafe VLM DP overlap
  cuichenx
  committed
  428eb5c
- docs(nemotron-vl): record real H100 pretraining
  cuichenx
  committed
  1500d7c
- fix(nemotron-omni): expose decoder for CUDA graphs
  cuichenx
  committed
  d30760a
- perf(nemotron-vl): align GB200 sequence length
  cuichenx
  committed
  a3fc1e2
- docs(nemotron-vl): record GB200 pretraining verification
  cuichenx
  committed
  1d01237
Commits on Aug 29, 2026
- feat(nemotron-vl): add SFT and PEFT verification recipes
  cuichenx
  committed
  5aab22f
- fix(nemotron-vl): disable shared expert overlap for PEFT
  cuichenx
  committed
  9c4c7b3
- fix(nemotron-vl): fit H100 PEFT with selective recompute
  cuichenx
  committed
  2f1c677
- fix(nemotron-vl): fit H100 PEFT without recompute
  cuichenx
  committed
  b1d2932
- docs(nemotron-vl): record post-training verification
  cuichenx
  committed
  0e7a2ca
Commits on Aug 30, 2026
- chore: merge upstream main
  cuichenx
  committed
  2f887a0
- feat(nemotron-vl): add GB200 long-context SFT recipe
  cuichenx
  committed
  ab021c1
- fix(nemotron-vl): preserve Super VL export entrypoint
  cuichenx
  committed
  24edc11
Commits on Aug 31, 2026
- feat(inference): support multi-GPU HF verification
  cuichenx
  committed
  a62b583
- fix(inference): load VLM exports with image-text auto class
  cuichenx
  committed
  289f7b6
- fix(data): unwrap MCore tokenizer for chat SFT
  cuichenx
  committed
  71a63b3
- fix(nemotron-vl): close RADIO summary buffer on export
  cuichenx
  committed
  7aba5a8
- fix(conversion): postprocess derived export buffers
  cuichenx
  committed
  388dc02
- fix(recipe): use HF tokenizer for CoderForge SFT
  cuichenx
  committed
  0445eb2
- feat(verification): support separate image processing
  cuichenx
  committed
  30b2b37
- fix(verification): omit processor-only image metadata
  cuichenx
  committed
  215e20e
- allow 2d patchified inputs in nemotron-omni (#5894)
  
  rohitrango
  and
  claude
  authored
  3961f39
Commits on Sep 15, 2026
- chore: merge main into model support branch
  cuichenx
  committed
  baa319a
- perf(nemotron-vl): tune GB200 training topology
  cuichenx
  committed
  2b0c803
- docs(nemotron-vl): refresh verification guidance
  cuichenx
  committed
  9b28582
- docs(verification): allowlist provenance hashes
  cuichenx
  committed
  c22b31a
- chore: merge published model branch
  cuichenx
  committed
  87bac3d
- fix(nemotron-vl): preserve shared MTP block semantics
  cuichenx
  committed
  1eb8899
- docs(nemotron-vl): clarify MTP verification refresh
  cuichenx
  committed
  f6b7f35
Commits on Sep 16, 2026
- fix(nemotron-vl): preserve optimizer precision and weight decay
  cuichenx
  committed
  d80db46
- fix(nemotron-vl): tighten model support scope and adapter placement
  cuichenx
  committed
  a4b4757
Commits on Sep 17, 2026
- fix(nemotron-vl): align recipe environments and verification docs
  cuichenx
  committed
  11773d7
- fix(nemotron-omni): retain last-stage positions for MTP
  cuichenx
  committed
  4c29e74
Commits on Sep 18, 2026
- fix(nemotron-omni): preserve provider identity for training callbacks
  cuichenx
  committed
  5fb6151
- docs(nemotron-vl): mark training verification pending callback refresh
  cuichenx
  committed
  154d16b
- Revert "fix(nemotron-omni): preserve provider identity for training callbacks"
  cuichenx
  committed
  7757f62
- fix(training): bind runtime callbacks to the model configuration
  
  authored and
  cuichenx
  committed
  d1a1b72
- Merge main into chcui/sage/nmvl
  cuichenx
  committed
  eb6f9a0
Commits on Sep 19, 2026
Commits on Sep 20, 2026
- fix(launchers): preserve direct-Enroot task environment
  cuichenx
  committed
  5b5ef52
Commits on Sep 21, 2026
- fix(launchers): share allocation-level Slurm parameters
  cuichenx
  committed
  90072e5
- fix(nemotron-omni): resolve review findings and merge main
  cuichenx
  committed
  b049c95
- fix(tests): add missing NVIDIA copyright headers
  cuichenx
  committed
  46269e7

Commits

Commits on Aug 24, 2026

Commits on Aug 25, 2026

Commits on Aug 26, 2026

Commits on Aug 27, 2026

Commits on Aug 28, 2026

Commits on Aug 29, 2026

Commits on Aug 30, 2026

Commits on Aug 31, 2026

Commits on Sep 15, 2026

Commits on Sep 16, 2026

Commits on Sep 17, 2026

Commits on Sep 18, 2026

Commits on Sep 19, 2026

Commits on Sep 20, 2026

Commits on Sep 21, 2026