Repository navigation

chore: nightly sync main into dev (08_04_2026) - #4213

Closed

Phlip79 wants to merge 84 commits into

devfrom

main2dev/08_04_2026

Closed

chore: nightly sync main into dev (08_04_2026)#4213
Phlip79 wants to merge 84 commits into
devfrom
main2dev/08_04_2026

Commits on Mar 25, 2026
- Add ability to perform local gradient accumulation in FP32 for a subset of parameters in the model (#4028)
  
  deepakn94
  and
  claude
  authored
  c586f6d
Commits on Mar 26, 2026
- Miscellaneous MXFP8 inference fixes (#4017)
  santhnm2
  authored
  09cce75
- Use `torch.int64` for grad_num_zero accumulation (#4015)
  
  WanZzzzzz
  and
  deepakn94
  authored
  a01a6c5
- Make text generation server hostname configurable (#3935)
  santhnm2
  authored
  548028b
- Add --muon-coefficient-type argument for Muon optimizer (#3927)
  
  authored
  0842ca2
- Pass gracefully if token_id not found in message (#3862)
  
  i-riyad
  and
  santhnm2
  authored
  606afda
- Improve load balancing behavior for prefix cache-aware routing (#3930)
  santhnm2
  authored
  0528a40
Commits on Mar 27, 2026
- Refactor setup.py to use get_pybind_include (#3658)
  sakgoyal
  authored
  58e0b85
- build: Bump TE to 2.14 (#4025)
  ko3n1g
  authored
  3758b54
Commits on Mar 30, 2026
- chore(beep boop 🤖): Bump (main) (2026-03-30)
  github-actions[bot]
  committed
  d863b7b
- fix traceback when interrupting run (#3439)
  dimapihtar
  authored
  a61ce5f
- chore: update goldenvalues (#4059)
  ko3n1g
  authored
  4dcd7d6
- Fix TemporalAsyncCaller pin_memory lifetime in async checkpointing (#2288)
  
  authored
  fc61ce5
- chore: Move to Py3.12 (#3826)
  
  ko3n1g
  and
  claude
  authored
  4bde3a4
Commits on Mar 31, 2026
- Update copy-pr-bot.yaml [skip ci]
  github-actions[bot]
  committed
  8256553
- Adding NVRx as a dependency and keeping the current code base optionally (#3899)
  
  authored
  704c7ee
- chore: Bump versions
  ko3n1g
  committed
  e1db321
- build: Set `ENV NVTE_BUILD_NUM_PHILOX_ROUNDS=3` (#4074)
  ko3n1g
  authored
  8dd65cd
- fix checkpointing conversion (#4058)
  dimapihtar
  authored
  a3a7a0c
- chore: Bump versions
  ko3n1g
  committed
  f09f5c9
- fix(ci): replace actions/setup-python with apt-get to avoid 429 rate limits (#4072)
  
  ko3n1g
  and
  claude
  authored
  dc113cf
- ci: Fix package name for code-freeze workflow (#4077)
  ko3n1g
  authored
  8f3bee5
- chore: bump `_code_freeze` workflow to `v0.86.0` (#4078)
  
  ko3n1g
  and
  claude
  authored
  1533beb
- Fix checkpoint inspector (#4079)
  janEbert
  authored
  ef2c8a0
- Update docs to conform to NVIDIA style guides (#4068)
  megnvidia
  authored
  fd1888b
- Miscellaneous inference fixes (#4030)
  
  authored
  2b85d0a
- fix fine_grained_callables with fused rmsnorm residual (#4026)
  CarlosGomes98
  authored
  15f14fc
- [Main][feat] Support overlapping A2A Combine backprop with wgrad GEMM (#3795)
  Wohox
  authored
  97e36aa
Commits on Apr 1, 2026
- chore: rotate oncall schedule
  github-actions[bot]
  committed
  7086b61
- Modify mfsdp default data-parallel-sharding-strategy (#3691)
  
  authored
  3499efe
- Fix fsdp_dtensor conversion for pretrained-only checkpoints (#3912)
  
  DAISY-gh
  and
  cspades
  authored
  1284d25
- Guard NVshmem issues (#4093)
  
  authored
  f9a61e3
- m-fsdp: wire use_precision_aware_optimizer from ddp_config to ParamAn… (#4024)
  
  rapatel
  and
  cspades
  authored
  fe5291f
- Megatron-FSDP: Add MXFP8 transpose helper buffer for Hybrid FSDP (#3918)
  
  shjwudp
  and
  cspades
  authored
  606ac26
- feat(fsdp): use TE general_gemm for mixed-precision wgrad in FSDP path (#3822)
  
  authored
  8d7a3f8
- Megatron-FSDP: Make _pre_forward_param_unshard and _register_post_backward_hook formal (#4029)
  
  shjwudp
  and
  cspades
  authored
  748ac49
- Megatron-FSDP: Fix insufficient double buffers during gradient reduce (#4054)
  
  shjwudp
  and
  cspades
  authored
  3dc2251
- Fix M-FSDP MXFP8 related BUGs (#3991)
  
  shjwudp
  and
  cspades
  authored
  41f3b6f
- FIX: Use decoupled gradients for precision-aware M-FSDP grad norm (#3746)
  
  authored
  a52ceeb
Commits on Apr 2, 2026
- [Megatron-FSDP] Fix compatibility with frozen parameters and add unit tests (#3287)
  
  shjwudp
  and
  cspades
  authored
  150e37a
- Align chat completions endpoint with vLLM (#4063)
  santhnm2
  authored
  cb3bb41
- [M-FSDP] Refactor uneven dtensor to full tensor and add UT (#3190)
  
  shjwudp
  and
  cspades
  authored
  159e347
- Add agent instruction files (#4102)
  
  Phlip79
  and
  claude
  authored
  5a7f520
- Bump eopt version (#4100)
  skyw
  authored
  17a67b9
- Refactor emerging optimizer integration (#4113)
  skyw
  authored
  5b512b4
Commits on Apr 3, 2026
- Update copy-pr-bot.yaml [skip ci]
  github-actions[bot]
  committed
  8b8ceb5
- Fix over provisioning of Mamba state memory when max_requests is set (#4114)
  santhnm2
  authored
  dcc6d62
- base strategy simplification (#4001)
  dimapihtar
  authored
  2697b82
- add support for DCP and FSDP async save (#4027)
  dimapihtar
  authored
  69f3b34
- Add more emerging optimizers (#3907) (#4119)
  skyw
  authored
  c9797ad
- Fix FSDP checkpoint conversion and loading for Qwen3.5-VL (#3936)
  
  DAISY-gh
  and
  cspades
  authored
  76e4daa
- docs: update mcore optimizer docstrings to google style (#2799)
  Akshat8510
  authored
  a025a69
- Update oncall schedule (#4117)
  Phlip79
  authored
  07db9f7
- Set tensor-parallel attributes irrespective of perform_initialization (#4084)
  
  authored
  a72c027
- docs: add developer-guide skill with CI/CD and failure navigation guidance (#4035)
  
  ko3n1g
  and
  claude
  authored
  c0e3134
- chore: Move skills (#4136)
  ko3n1g
  authored
  8758d16
- ci: Let Claude react to comment (#4135)
  ko3n1g
  authored
  fd76254
- Nemotron3 Super GB200 release config (#4118)
  maanug-nv
  authored
  d865bba
- Enable CUDA graph for ADAM optimizer (#3429)
  
  authored
  3d87bfc
- Claude review should recommend testing (#4137)
  Phlip79
  authored
  499266a
Commits on Apr 4, 2026
- cleanup: remove unused `scatter_gather_tensors_in_pipeline` argument (#4140)
  
  Phlip79
  and
  claude
  authored
  10e7b74
Commits on Apr 5, 2026
- fix: Remove fail-fast (-x) and guard distributed teardown against deadlock (#4139)
  
  ko3n1g
  and
  claude
  authored
  0b8306b
Commits on Apr 6, 2026
- chore(beep boop 🤖): Bump (main) (2026-04-06)
  github-actions[bot]
  committed
  1d43284
- Claude: add respond-to-issue skill (#4141)
  
  Phlip79
  and
  claude
  authored
  7d536c0
- Fix muon getter backward compatability (#4157)
  skyw
  authored
  6652d57
- Audit of user guide (#4098)
  megnvidia
  authored
  97cd326
- Fix `RerunStateMachine` crash (`TypeError: 'NoneType' object is not subscriptable`) by not saving a checkpoint after a transient NaN / Inf (#3981)
  
  yezhengmao1
  and
  dimapihtar
  authored
  0b5e3ae
- Preserve type of decorated methods/classes (#4062)
  nschank
  authored
  fa5103c
Commits on Apr 7, 2026
Commits on Apr 8, 2026
Commits on Apr 11, 2026
- fix: remove files that dev intentionally deleted
  Phlip79
  committed
  40d3094
- Merge remote-tracking branch 'origin/dev' into main2dev/08_04_2026
  Phlip79
  committed
  4e8407e

Commits

Commits on Mar 25, 2026

Commits on Mar 26, 2026

Commits on Mar 27, 2026

Commits on Mar 30, 2026

Commits on Mar 31, 2026

Commits on Apr 1, 2026

Commits on Apr 2, 2026

Commits on Apr 3, 2026

Commits on Apr 4, 2026

Commits on Apr 5, 2026

Commits on Apr 6, 2026

Commits on Apr 7, 2026

Commits on Apr 8, 2026

Commits on Apr 11, 2026