feat: Enable simulated user for multi-turn GRPO#606

Closed

jialei777 wants to merge 61 commits intoNVIDIA-NeMo:mainfrom

jialei777:jialei/simulated-user

Commits on Jul 23, 2025

ini

xxman-google
authored and
jialei777
committed
fix: load HF model only on rank 0 (#544)

parthchadha
authored and
jialei777
committed
feat: supports evaluation of multiple-choice benchmarks (#559)

authored and
jialei777
committed
fix: enable expandable segments for hopper+ (#594)

parthchadha
authored and
jialei777
committed
feat: Enable vLLM cudagraphs (#498)

authored and
jialei777
committed
docs: Update guide to include minimum compute requirement (#505)

authored and
jialei777
committed
fix: skip HelpSteer3 unit test if downloading failed (#612)

yuki-97
authored and
jialei777
committed
feat: optimize get logprobs when cp enabled. (#528)

joyang-nv
authored and
jialei777
committed
enable mcore rope fusion (#608)

jiemingz
authored and
jialei777
committed
fix: fix non-colocated with vllm tp>1 (#601)

yuki-97
authored and
jialei777
committed
feat: Refit: reduce the number of IPC calls by packing weights (#589)

authored and
jialei777
committed
feat: add flash-attn==2.7.4.post1 to backend dependencies (#622)

authored and
jialei777
committed
fix: Fix crash for logprob error plot (#623)

yfw
authored and
jialei777
committed
refactor: remove fsdp1 path (#614)

yuki-97
authored and
jialei777
committed
fix: fix a answer parsing bug in MMLU-Pro. (#598)

xxman-google
authored and
jialei777
committed
feat: add MMMLU eval benchmark. (#596)

xxman-google
authored and
jialei777
committed
fix: pytest_sessionfinish hook in case there is no _unit_test_data. (#628)

ffrujeri
authored and
jialei777
committed
fix: Don't call broadcast on dtensor (#627)

parthchadha
authored and
jialei777
committed
fix: Fix eval when using async engine (#626)

parthchadha
authored and
jialei777
committed
feat: Megatron MoE Support (#590)

authored and
jialei777
committed
chore: exclude ray.remote from coverage (#624)

terrykong
authored and
jialei777
committed
feat: guide to configure custom vllm version (#529)

terrykong
authored and
jialei777
committed
feat: Deepseek Support (#591)

authored and
jialei777
committed
feat: decouple checkpointing from validation (#575)

ashors1
authored and
jialei777
committed
feat: dynamically detect --gres=gpu:8 arg to work on clusters that don't need it (#642)

terrykong
authored and
jialei777
committed
fix: fix nccl P2P initialization error for non-colocated (#636)

authored and
jialei777
committed
fix: Mcore: Added functional grpo test and typing fixes (#527)

SahilJain314
authored and
jialei777
committed
feat: plumb environment variables to RayWorkerGroup (#631)

authored and
jialei777
committed
feat: Qwen3 support (#592)

authored and
jialei777
committed
fix: Fix megatron llama3.1-8b config (#652)

authored and
jialei777
committed
fix: update qwen32b config (#658)

authored and
jialei777
committed
fix: Make trust_remote_code default true in checkpoint (#663)

parthchadha
authored and
jialei777
committed
feat: add script to redact hparam paths from tensorboard logs (#347)

terrykong
authored and
jialei777
committed
test: add a unit test that verifies that the correct keys are present in configs (#587)

ashors1
authored and
jialei777
committed
add adk

xxman-google
authored and
jialei777
committed
rollout stable and faster

xxman-google
authored and
jialei777
committed
cleanup

xxman-google
authored and
jialei777
committed
docs: Add GitHub icon and link to top bar (#669)

aschilling-nv
authored and
jialei777
committed
fix: Tie weights after set_model_state_dict if required (#666)

parthchadha
authored and
jialei777
committed
feat: optimize refit by reducing set of IPC handles sent to each device (#634)

authored and
jialei777
committed
fix: adjust temperature scaling logic based on engine version (#660)

jubick1337
authored and
jialei777
committed
feat: introduce megatron checkpoint dir precedence (#665)

authored and
jialei777
committed
feat: optimize refit by preparing refit info ahead of time (#638)

yuki-97
authored and
jialei777
committed
docs: update converter path in README. (#672)

xxman-google
authored and
jialei777
committed
fix: make mcore lr scheduler configuration consistent with dtensor (#681)

authored and
jialei777
committed
fix: fix mcore LR increment (#685)

ashors1
authored and
jialei777
committed
fix: upgrade datasets to fix squad download (#692)

ashors1
authored and
jialei777
committed
fix: Megatron config updates to avoid OOM (#687)

ashors1
authored and
jialei777
committed
fix: fix lr scheduler for config that was missed in #681 (#693)

ashors1
authored and
jialei777
committed
fix: Fix gemma models broken by HF update (#676)

yfw
authored and
jialei777
committed
chore: add CP+SP (sequence parallel) assertion in DTensor worker (#689)

authored and
jialei777
committed
feat: MLFlow Integration for experiment tracking (#697)

authored and
jialei777
committed
fix: Fix activation checkpointing for mcore path (#703)

yfw
authored and
jialei777
committed
feat: Enable Context Parallelism and Sequence Packing for MCore and Dtensor (#704)

authored and
jialei777
committed
fix: SyntaxWarning: invalid escape sequence '\s' (#705)

RayenTian
authored and
jialei777
committed
cleanup and add docstring
jialei777
committed
remove image
jialei777
committed
ci: add a job that checks if submodules are fast forwarded (#695)

terrykong
authored and
jialei777
committed
add uv.lock file
jialei777
committed
update uv lock
jialei777
committed
Merge branch 'main' into jialei/simulated-user
jialei777
authored