- Added Mixture-of-Kittens MoE backend (#8210)
- TransformerEngine upgraded from 2.20 to 2.21
- Various bug fixes
Changelog Details
- beep boop 🤖: Bumping Megatron Core to v0.20.1 by @nvidia-megatron-lm-release-bot[bot] :: PR: #8714
- cp:
build: use PyPI nvidia-cudnn-frontend 1.29.1 in CI dev image (8303)intocore_r0.20.0by @ko3n1g :: PR: #8425 - fix: backport MFSDP v2 EP checkpoint support to core_r0.20.0 by @cuichenx :: PR: #8465
- cp: Refactor Schedule Node activation ownership (#8302) into core_r0.20.0 by @vthumbe1503 :: PR: #8911
- build: update Transformer Engine to release_v2.21 tip on core_r0.20.0 by @balasaajay :: PR: #8792
- Backport MoK, THD attention graphs, and DSv4 FLOPs to core_r0.20.0 by @cuichenx :: PR: #8929
- cp: fix(pipeline): preserve input storage on core_r0.20.0 by @svcnemo-autobot :: PR: #9089
- cp: Backport fused MLA KV RoPE MXFP8 quantization (#8144) to core_r0.20.0 by @malay-nagda :: PR: #9416
- cp: fix(rl): backport FastAPI environment contract to core_r0.20.0 by @svcnemo-autobot :: PR: #9404
- cp: Load the MLA query RoPE dims contiguously in the forward kernel (#8141) to core_r0.20.0 by @dingqingy-nv :: PR: #9487
- cp: Accumulate the embedding weight gradient into main_grad (#8142) to core_r0.20.0 by @dingqingy-nv :: PR: #9486
- build: fix NLTK CVEs on core_r0.20.0 by @balasaajay :: PR: #9492
- build: upgrade Black to 26.10.1 on core_r0.20.0 by @balasaajay :: PR: #9496
- docs: align core_r0.20.0 docs and latest pointer with 0.20.1 by @balasaajay :: PR: #9605
- beep boop 🤖: Bumping Megatron Core to v0.20.1 [skip ci] by @github-actions[bot]
- cp:
build: use PyPI nvidia-cudnn-frontend 1.29.1 in CI dev image (8303)intocore_r0.20.0(#8425) by @ko3n1g - fix: backport MFSDP v2 EP checkpoint support to core_r0.20.0 (#8465) by @cuichenx
- Refactor Schedule Node activation ownership (#8302) by @vthumbe1503
- cp: Refactor Schedule Node activation ownership (#8302) into core_r0.20.0 (#8911) by @ko3n1g
- build: update Transformer Engine to release_v2.21 tip on core_r0.20.0 (#8792) by @balasaajay
- Backport MoK, THD attention graphs, and DSv4 FLOPs to core_r0.20.0 (#8929) by @cuichenx
- cp: fix(pipeline): preserve input storage on core_r0.20.0 (#9089) by @svcnemo-autobot
- cp: Backport fused MLA KV RoPE MXFP8 quantization (#8144) to core_r0.20.0 (#9416) by @malay-nagda
- cp: fix(rl): backport FastAPI environment contract to core_r0.20.0 (#9404) by @svcnemo-autobot
- cp: Load the MLA query RoPE dims contiguously in the forward kernel (#8141) to core_r0.20.0 (#9487) by @dingqingy-nv
- cp: Accumulate the embedding weight gradient into main_grad (#8142) to core_r0.20.0 (#9486) by @dingqingy-nv
- build: fix NLTK CVEs on core_r0.20.0 (#9492) by @balasaajay
- build: upgrade Black to 26.10.1 on core_r0.20.0 (#9496) by @balasaajay
- docs: align core_r0.20.0 docs and latest pointer with 0.20.1 (#9605) by @balasaajay