Perf Config for 1 node GB200 DSV3 - #3545

Closed

gautham-kollu wants to merge 61 commits into

mainfrom

gk/1node_gb200_dsv3

Commits on Apr 27, 2026

first draft
gautham-kollu
committed
b55883a
add debug code
gautham-kollu
committed
8ea7c14
cleanup
gautham-kollu
committed
ea40b22
gb200 x4
gautham-kollu
committed
ac0319d
review comments
gautham-kollu
committed
77cfa8c

Commits on Apr 28, 2026

Commits on May 5, 2026

Merge branch 'main' into gk/1node_gb200_dsv3
gautham-kollu
authored
e37c1f5

Commits on May 6, 2026

add scripts changes
gautham-kollu
committed
6c4e18e

Commits on May 12, 2026

[perf] update: switch GPT-OSS B200 V2 dispatcher default to alltoall (#3614)

authored and
gautham-kollu
committed
1903ad7
chore(beep boop 🤖): Bump (main, mcore-main) (2026-05-05)

github-actions[bot]
authored and
gautham-kollu
committed
0291878
[model, ckpt, docs] fix: support HF→Megatron conversion under decentralized PGs (#3673)

authored and
gautham-kollu
committed
098dbd5
[training] feat: forward MoE/MTP metrics to MLFlow and Comet (#3647)

authored and
gautham-kollu
committed
dfc6fe7
[recipe, test] feat: add unit tests for qwen2_audio and kimi_k25_vl recipes (#3646)

lonexreb
authored and
gautham-kollu
committed
892086a
[PEFT, ckpts] feat: modelopt for LoRA & deepseek arch (#3612)

HollowMan6
authored and
gautham-kollu
committed
162da60
[docs, ci] feat: add bump-dependency skill and fix L2 trigger docs (#3686)

ko3n1g
authored and
gautham-kollu
committed
5b589e4
ci: update maintainers (#3694)

authored and
gautham-kollu
committed
e38cda4
Add Gemma 4 (26B-A4B) LLM and VLM bridge (#3148)

authored and
gautham-kollu
committed
deed27c
refactor(docker): move hybridep install from fw_base to Dockerfile.ci (#3702)

ko3n1g
authored and
gautham-kollu
committed
f44df4c
Add Gemma4VL News and docs (#3708)

suiyoubi
authored and
gautham-kollu
committed
9cced02
[ci] refactor: consolidate PR + release workflows; use validate-only (#3658)

authored and
gautham-kollu
committed
7507731
fix: vllm 0.20.1 build (#3700)

chtruong814
authored and
gautham-kollu
committed
102fa6d
fix(perf): set expandable_segments:True for nemotron_3_nano h100 fp8_cs (cuda-oom on 26.04 base) (#3706)

ko3n1g
authored and
gautham-kollu
committed
dbb9e80
Fix Gemma3 example folder (#3724)

suiyoubi
authored and
gautham-kollu
committed
337dc54
[ci, build] chore: Revert PR #3702 (hybridep docker move) to investigate VLM perf regression (#3729)

yaoyu-33
authored and
gautham-kollu
committed
b78d47d
track mem-max-reserved (#3705)

authored and
gautham-kollu
committed
8d4c380
chore(beep boop 🤖): Bump (main, mcore-main) (2026-05-07)

github-actions[bot]
authored and
gautham-kollu
committed
ba7c213
[ci, build] chore: revert revert of PR #3702 (restore hybridep docker move) (#3734)

ko3n1g
authored and
gautham-kollu
committed
2371390
[ci] fix: use GPU_COUNT instead of unread CI_RUNNER for deepseek perf launcher

ko3n1g
authored and
gautham-kollu
committed
70fbb99
GPTOSS openmathinstruct and eval (#2990)

weijiac0619
authored and
gautham-kollu
committed
50f741c
fix(build): restore rdma-core v60 + bump NVSHMEM 3.6.5 for multi-node B200 DeepEP (#3704)

ko3n1g
authored and
gautham-kollu
committed
0008aaf
chore(beep boop 🤖): Bump (main, mcore-main) (2026-05-08)

github-actions[bot]
authored and
gautham-kollu
committed
cf99da5
Reorganize ModelOpt docs (#3715)

kevalmorabia97
authored and
gautham-kollu
committed
5c81c44
ci: Remove unnecessary taint node job (#3745)

chtruong814
authored and
gautham-kollu
committed
ed420cb
build: pin flashinfer-python + flashinfer-cubin in pyproject.toml (#3750)

ko3n1g
authored and
gautham-kollu
committed
a2b0a85
[training] fix: aggregate TB memory metrics across PP group (#3645)

lonexreb
authored and
gautham-kollu
committed
75799e7
[training, test] test: add MLA, MTP, and provider-override coverage for FLOPs calculator (#3695)

authored and
gautham-kollu
committed
5f9447c
[training, test] feat: add unit tests for pg_utils helpers (#3650)

authored and
gautham-kollu
committed
5c12732
[recipe, test] feat: add unit tests for CometPlugin (#3652)

authored and
gautham-kollu
committed
0573ee0
[model, ckpt] fix: align GPT-OSS BF16 down_proj orientation on import (#3743)

cuichenx
authored and
gautham-kollu
committed
71ebb51
[MXFP8]Update param buffer before AG in eval (#3727)

WanZzzzzz
and
gautham-kollu
committed
6aebe35
chore(beep boop 🤖): Bump (main, mcore-main) (2026-05-09)

github-actions[bot]
authored and
gautham-kollu
committed
bd8f9f4
build: upgrade mamba-ssm to 2.3.2.post1, causal-conv1d to 1.6.2.post1 (#3764)

ko3n1g
authored and
gautham-kollu
committed
35ab5a1
[docs] feat: add cherrypick policy for release branches (#3739)

ko3n1g
authored and
gautham-kollu
committed
5975d6f
chore(beep boop 🤖): Bump (main, mcore-main) (2026-05-10)

github-actions[bot]
authored and
gautham-kollu
committed
51453a1
[performance] feat: add 405B B200/B300 V2 aliases + 405B GB200 NVFP4 256x gpu scale expendable segments addition due to CUDA OOM issue (#3759)

rsalagame-nvidia
authored and
gautham-kollu
committed
9230619
ci: Major refactor of release-workflows (#3725)

authored and
gautham-kollu
committed
986c646
nccl-algo reset
gautham-kollu
committed
2f44bd3
1
gautham-kollu
committed
b706332
run as module
gautham-kollu
committed
9659e4a
llama31 405b gb200 nvfp4 no pg overlap (#3713)

malay-nagda
authored and
gautham-kollu
committed
7831e35
[ci] fix: check-imports action lookup on ARC container runner (#3774)

ko3n1g
authored and
gautham-kollu
committed
23cfc35
[build] chore: bump transformer-engine to release_v2.15 (#3672)

ko3n1g
authored and
gautham-kollu
committed
b358ee1
[ci, test] chore: Split L0 converter launch script into conversion and generation (#3758)

authored and
gautham-kollu
committed
1dc260b
[model]Add Qwen3‑Omni training support (#3317)

authored and
gautham-kollu
committed
31e432e
[docs, ci] chore: rename state labels and retire remove-needs-author workflow (#3779)

yaoyu-33
authored and
gautham-kollu
committed
8109cf3
Malay/move expandable flag (#3781)

malay-nagda
authored and
gautham-kollu
committed
b565504
Update L0_Launch_recipes_deepseek_perf.sh
gautham-kollu
committed
b27acc5
1
gautham-kollu
committed
d9f7c49