Repository navigation

build: update development PyTorch image to 26.09 - #7725

Merged

balasaajay merged 59 commits into

NVIDIA:mainfrom

balasaajay:bump-dev-image-26.09-pr7687-20260930

Oct 10, 2026

Commits on Sep 30, 2026
- chore: update development base image to 26.09-py3
  
  svcnemo-autobot
  authored and
  balasaajay
  committed
  84f81ea
- fix: update DeepGEMM for PyTorch C++20 requirements
  
  svcnemo-autobot
  authored and
  balasaajay
  committed
  1ff951b
- fix: install elfutils headers for DeepGEMM builds
  
  svcnemo-autobot
  authored and
  balasaajay
  committed
  2560c5a
- build: backport Mamba C++20 flags for PyTorch 26.09
  balasaajay
  committed
  7640398
- build: fix DeepGEMM FP8 headers on CUDA 12.9
  balasaajay
  committed
  dac0cff
- fix(ci): avoid failing permission prepass on CoreWeave runners
  balasaajay
  committed
  d58a15d
- fix(tests): restore MoE benchmark RNG and failure propagation
  balasaajay
  committed
  46d3cca
- fix(inference): honor FA4 split-KV and attention-sink requirements
  balasaajay
  committed
  67ee88c
- build: align CuTe and QuACK dependencies with FlashAttention 4
  balasaajay
  committed
  6ebe61d
- build: satisfy FA4 TVM FFI requirements with compatible TileLang
  balasaajay
  committed
  e9f01a0
- fix: backport FA4 packed subtraction compatibility
  balasaajay
  committed
  2f16cb1
- test: preserve historical MoE benchmark routing
  balasaajay
  committed
  71f91fc
- test: refresh NanoV3 GB200 batch128 performance baseline
  balasaajay
  committed
  64dd800
- test: refresh H100 FSDP context-parallel loss reference
  balasaajay
  committed
  0e2c866
- fix: handle missing FSDP checkpoint version metadata
  balasaajay
  committed
  e1e0c85
- test: refresh stable H100 DeepSeek loss references
  balasaajay
  committed
  87a35cd
- test: refresh stable H100 GPT performance improvement
  balasaajay
  committed
  a883731
- test: refresh H100 DeepSeek overlap loss reference
  balasaajay
  committed
  85f8fc5
- test: refresh H100 hybrid FSDP loss reference
  balasaajay
  committed
  b028d4e
- test: refresh H100 FSDP v2 overlap loss reference
  balasaajay
  committed
  28f8046
- test: refresh H100 HSDP loss reference
  balasaajay
  committed
  e071ec3
- test: refresh H100 MoE GRPO timing reference
  balasaajay
  committed
  74ad5bb
- test: refresh H100 pipelined GPT timing reference
  balasaajay
  committed
  070712a
- test: refresh stable T5 references at full precision
  balasaajay
  committed
  20838da
- build: backport NCCL EP window offsets for PyTorch 26.09
  balasaajay
  committed
  4906863
- test: clear inference mode after batch-invariance tests
  balasaajay
  committed
  9b8300d
- test: refresh stable A100 T5 reference at full precision
  balasaajay
  committed
  c9819a9
- fix: register CUDA graph generators on PyTorch 26.09
  balasaajay
  committed
  bd4fe43
- fix: remove unused Torch version import
  balasaajay
  committed
  ee07a6b
- fix: select compatible paged attention for small Hopper heads
  balasaajay
  committed
  513efe4
- fix: match native clamp boundary gradients in fused SwiGLU
  balasaajay
  committed
  45679b9
- fix: preserve token-only padding during SSM decode
  balasaajay
  committed
  39bb2bc
- test: dump thread stacks during stalled pytest teardown
  balasaajay
  committed
  718410b
- test: retain every rank stderr in CI console logs
  balasaajay
  committed
  0b07373
- test: restore existing tests for the image update
  balasaajay
  committed
  6fef589
Commits on Oct 1, 2026
- Merge main and retain Transformer Engine 2.20
  balasaajay
  committed
  16c1b9a
- fix: guard Hopper attention dispatch by CUDA device
  balasaajay
  committed
  12b1c0a
- test: disable dev cases that leak inference mode
  balasaajay
  committed
  3befeee
- fix: defer clamp boundary probe until backward
  balasaajay
  committed
  8f99759
Commits on Oct 2, 2026
- Merge branch 'main' into bump-dev-image-26.09-pr7687-20260930
  balasaajay
  authored
  4f4d214
Commits on Oct 3, 2026
- test: quarantine A100 LTS MoE reference mismatch
  balasaajay
  committed
  3e3e659
- test: cover PyTorch 26.09 kernel compatibility fixes
  balasaajay
  committed
  88879c2
Commits on Oct 6, 2026
Commits on Oct 7, 2026
- ci: validate PyTorch 26.09 with FlashAttention 2
  balasaajay
  committed
  d842b7a
- Merge branch 'main' into bump-dev-image-26.09-pr7687-20260930
  balasaajay
  authored
  090200d
- fix: detect CUDA graph generator registration at runtime
  
  ksivaman
  authored and
  balasaajay
  committed
  fab15a7
- test: re-enable six GPT CP2 GitHub cases
  balasaajay
  committed
  a7a6534
Commits on Oct 8, 2026

Commits

Commits on Sep 30, 2026

Commits on Oct 1, 2026

Commits on Oct 2, 2026

Commits on Oct 3, 2026

Commits on Oct 6, 2026

Commits on Oct 7, 2026

Commits on Oct 8, 2026