COMMITS
July 26, 2026
E
July 25, 2026
K
build: exclude opencv-python-headless from shipped container (#3339)
kajalj22 committed
July 24, 2026
T
test(functional): relax grpo_topp_topk gen_kl_error bound 0.03 -> 0.032 (#3341)
Terry Kong committed
Y
feat(sc): NeMo-Gym path (#3267)
Yuki Huang committed
Y
feat(sc): setup + entrypoint (#3266)
Yuki Huang committed
J
perf: speed up TransformerEngine source builds (#3333)
Jonas Yang CN committed
Y
chore: Bump Gym (#3326)
Yi-Fu Wu committed
July 23, 2026
A
feat: Adding fast token (#3181)
arnavk-nvidia committed
Y
feat(sc): train path — StalenessSampler + _train_pump rewrite (#3220)
Yuki Huang committed
S
feat(generation): add TensorRT-LLM rollout backend (#2420)
shuyixiong committed
J
test: pass required NeMo Gym rollout arguments (#3317)
Jonas Yang CN committed
W
fix(async-grpo): add startup barrier to prevent checkpoint-resume deadlock (#3066)
Wenwen Gao committed
K
fix: forward calibrated KV cache scales on colocated ZMQ IPC refit path (#3226)
Kurt Yang committed
July 22, 2026
Y
feat(sc): rollout path — TQReplayBuffer, rollout_manager, _rollout_pump (#3219)
Yuki Huang committed
July 21, 2026
A
feat: GDPO vs GRPO multi-reward example and collapse demo (#2681)
Anjali Shah committed
Y
feat(grpo): stream rollout batches by prompt group (#3000)
Yi-Fu Wu committed
H
perf: increase Qwen3-235B async rollout GPU memory utilization on H100 (#3278)
humairafirdowse18 committed
S
ci: AUT-895 gate GB200 tests on DISABLE_GB200_TESTS variable (#3269)
svcnemo-autobot committed
A
feat: configurable GDPO per-reward weights and multi-reward NeMo Gym bridge (#2680)
Anjali Shah committed
P
fix: fix optimizer step working on random initialized model weights (#2372)
pengdurice committed
ℍ
feat: add checkpoint-engine refit interface and integrate NIXL (#2608)
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 committed
July 20, 2026
ℍ
fix(vllm): record logprob for the sampled token (#3297)
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 committed
ℍ
feat(modelopt): support real NVFP4 QAT rollout for MoE and Mamba (#2983)
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 committed
July 19, 2026
ℍ
feat: add delta-compressed collective refit (#2444)
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 committed
July 18, 2026
K
chore(deps): bump cryptography and mooncake to address CVEs (#3230)
kajalj22 committed
K
chore: replace decord2 with torchvision/torchaudio for media loading (#3107)
kajalj22 committed
July 17, 2026
K
fix: preserve runtime tools after arm64 Apptainer build (#3253)
kajalj22 committed
Z
feat: size routed_experts carry dtype from the model's expert count (#3113)
Zeyu Zhou committed
S
ci: bump async-gym megatron test wall-clock budget to 120 min (#3258)
Shanmugam Ramasamy committed