perf: Add DeepEP support to Megatron Policy#1645

Open

parthmannan wants to merge 55 commits intoNVIDIA-NeMo:mainfrom

parthmannan:pmannan/deepep_mcore_training

Commits on Jan 15, 2026

Add DeepEP support to Megatron Policy
parthmannan
committed
Add example recipe and add to config dict
parthmannan
committed
docs: Revise news section for nemotron v3 and DAPO algorithm support (#1640)

snowmanwwg
authored and
parthmannan
committed
chore: fix grpo functional test metric (#1643)

RayenTian
authored and
parthmannan
committed
feat: add support from building images using vllm from private repos (#1605)

authored and
parthmannan
committed
feat: Necessary changes for Gym GRPO tutorial (#1630)

authored and
parthmannan
committed
perf: Add qwen3 30b-a3b async-8-off recipe (#1642)

youngeunkwon0405
authored and
parthmannan
committed
feat: Add GPT-OSS support via mcore (#1452)

ashors1
authored and
parthmannan
committed
chore: Bump vllm to 0.11.2, torch to 2.9, transformers to 4.57.1 (#1563)

authored and
parthmannan
committed
fix: Support datasets saved with save_to_disk in ResponseDataset (#1610)

authored and
parthmannan
committed
Recipe update
parthmannan
committed
fix: Handle disabled validation in SFT training (#1611)

sahgerlad
authored and
parthmannan
committed
fix: Fix crash when using cp in dtensor path (#1663)

yfw
authored and
parthmannan
committed
fix: Fix Fp8 sequence padding for PP>1 case (#1579)

authored and
parthmannan
committed
test: Perf recipe for v0.5 (#1667)

authored and
parthmannan
committed
fix: Fix fp8 after vllm v0.11.2 bump (#1660)

guyueh1
authored and
parthmannan
committed
fix: Fix crash when using activation_checkpointing (#1676)

yfw
authored and
parthmannan
committed
feat: add dapo recipe and test (#1617)

ZhiyuLi-Nvidia
authored and
parthmannan
committed
feat: DTensorPolicyV2 GPT-OSS SFT support (#1470)

authored and
parthmannan
committed
fix: grad norm calculation for dtensor v2 (#1693)

hemildesai
authored and
parthmannan
committed
feat: Add Nemotron‑3 Nano 30B A3B BF16 SFT nightly tests (FSDP2, +LoRA) (#1648)

RayenTian
authored and
parthmannan
committed
feat: Support prefetching of specific envs (#1692)

hemildesai
authored and
parthmannan
committed
Upgrade DeepEP version to match
parthmannan
committed
Lint fix
parthmannan
committed
fix: Fix DTensor slice crash after PyTorch 2.9 bump (#1689)

zpqiu
authored and
parthmannan
committed
fix: grad norm check for automodel gpt oss nightly (#1708)

hemildesai
authored and
parthmannan
committed
fix: relax nanov3 nightly test metrics strict (#1712)

RayenTian
authored and
parthmannan
committed
fix: on GB200 use single-thread checkpoint save to avoid Cpu OOM (#1703)

guyueh1
authored and
parthmannan
committed
perf: [Perf recipe] Change TP 16->32 for deepseek GB200 sync benchmark (#1715)

guyueh1
authored and
parthmannan
committed
docs: Add doc for nano-v3 (#1694)

yfw
authored and
parthmannan
committed
fix: Disable cudnn sdpa backend when using activation checkpointing (#1717)

yfw
authored and
parthmannan
committed
fix: log metrics that can be coerced to scalars (#1723)

terrykong
authored and
parthmannan
committed
fix: use median instead of mean for logprob error for stability in nightlies (#1722)

terrykong
authored and
parthmannan
committed
fix: gemma3 27b must now have skip_tokenizer_init=False in vllm (#1721)

terrykong
authored and
parthmannan
committed
fix: fix several nightly tests that were flaky (#1724)

terrykong
authored and
parthmannan
committed
fix: apply offloading change from v2 to v1 (#1726)

authored and
parthmannan
committed
fix: mcore generation config restored in nightly test (#1720)

terrykong
authored and
parthmannan
committed
feat: Megatron SFT LoRA (#1629)

authored and
parthmannan
committed
build: Update aiohttp and urlib3 (#1746)

chtruong814
authored and
parthmannan
committed
fix: patch pytorch aten.alias.default shard strategy (#1728)

RayenTian
authored and
parthmannan
committed
feat: RL support for custom moe models in dtensor v2 (#1695)

hemildesai
authored and
parthmannan
committed
fix: split dtensorv1 vllm dependency (#1638)

yuki-97
authored and
parthmannan
committed
build: Resolve CVEs for gnupg and aiohttp (#1755)

chtruong814
authored and
parthmannan
committed
build: Bump mamba to d68d16e and causal-conv1d to 67e0a9d (#1759)

chtruong814
authored and
parthmannan
committed
Update uv
parthmannan
committed
ci: Clean up disk space for lint check (#1768)

chtruong814
authored and
parthmannan
committed
docs: Adding dtensor TP debugging summary (#1767)

authored and
parthmannan
committed
Fix lock conflict resolution after signoff
parthmannan
committed
Merge branch 'main' of https://github.com/NVIDIA-NeMo/RL into pmannan/deepep_mcore_training
parthmannan
committed
Merge branch 'main' of https://github.com/NVIDIA-NeMo/RL into pmannan/deepep_mcore_training
parthmannan
committed

Commits on Jan 16, 2026

Commits on Jan 19, 2026