## NVIDIA/Megatron-LM — v0.12.0rc3
v0.15.3 _1290+ commits._ ### Features - **cp: `docs: Add changelog for 0.15.3 (3286)` into `core_r0.15.0` (#3288)** (309ffca) - **cp: `docs: Add changelog for 0.15 (2499)` into `core_r0.15.0` (#2502)** (4ef3ffa) - **ADLR/megatron-lm!4191 - Add sequence packing for RL** (ee8e930) - **ADLR/megatron-lm!4208 - Added attribute to support latest TE** (2c10442) - **ADLR/megatron-lm!4180 - chore: Add merge-queues** (8477817) - **ADLR/megatron-lm!4184 - chore: Add MoE changenotes** (737a23f) - **ADLR/megatron-lm!3978 - feat: Add Non-1D RoPE Frequency Indexing for Mixed Resolution VLM Batches** (6fe43ce) - **ADLR/megatron-lm!4103 - ci: Add install-test** (e35c195) - **ADLR/megatron-lm!3889 - Add num_tokens_total as alternative to num_tokens_to_generate** (ee0b4a2) - **ADLR/megatron-lm!4098 - ci: Add dependabot to GitHub CI** (844ecbd) - **ADLR/megatron-lm!3865 - Kan-MR1: Add chunked prefill** (e189664) - **ADLR/megatron-lm!4044 - YaRN support for gpt-oss** (2c1b77a) - **Revert "ADLR/megatron-lm!4067 - Add a throughput test for the dynamic engine case with decode-only graphs"** (56021bd) - **ADLR/megatron-lm!4067 - Add a throughput test for the dynamic engine case with decode-only graphs** (4cf968c) - **ADLR/megatron-lm!4053 - Add fp8_dpa option for CurrentScaling FP8 recipe** (61047e6) ### Fixes - **fix: Replace yaml.FullLoader with yaml.SafeLoader in quant_config.py (#3136)** (59511bf) - **Cherry-pick bug fixes into 0.15.X. (#2858)** (45b404c) - **Ko3n1g/fix/version numbers (#2550)** (5594439) - **ADLR/megatron-lm!4173 - Fix Sink Attention TP** (3b1b9b2) - **ADLR/megatron-lm!4196 - Fix yarn superclass init** (b9d237d) - **ADLR/megatron-lm!4159 - Fix ProcessGroupCollection missing initialization** (5f2becf) - **ADLR/megatron-lm!4182 - fix ckpt metadata** (d8c6aa4) - **fix** (44af130) - **fix** (ceb8c42) - **ADLR/megatron-lm!4213 - ci(fix): Iteration testing** (46a90b0) - **ADLR/megatron-lm!4210 - fix 1f1b overlap ut for mtp standalone** (44bc753) - **ADLR/megatron-lm!4174 - ci: Fix thresholds for iteration-time** (696f164) - **ADLR/megatron-lm!4171 - Fix typos: vritual → virtual** (fe2b34f) - **ADLR/megatron-lm!4102 - fix inference performance test** (10e9224) - **ADLR/megatron-lm!3819 - checkpointing: fix error propagation** (ca64390) - **ADLR/megatron-lm!4126 - ci(fix): Send dev-notifications to dev-channels** (1650adc) - **ADLR/megatron-lm!3937 - fix typos in megatron/core/transformer/multi_token_prediction.py** (efb2919) - **ADLR/megatron-lm!3939 - Fix duplicate init for self.module in DistributedDataParallel** (8301dfd) - **ADLR/megatron-lm!3958 - Fix: feat(MoE) Add the missing fp8 init part in MTP** (d6c6e54) - **ADLR/megatron-lm!4107 - ci(fix): Install check** (fd0e0df) - **ADLR/megatron-lm!4104 - chore: Fix script for generating files locally** (da2d32f) - **ADLR/megatron-lm!4064 - Fix gradient sync in example training script (GitHub #1685)** (feb68a3) - **ADLR/megatron-lm!4100 - ci(fix): build-test-wheel job** (c767b38) - **ADLR/megatron-lm!3813 - Fix add_BOS in dynamic inference** (18f8b2e) - **ADLR/megatron-lm!4083 - ci: Fix nemo-tests** (275854c) - **ADLR/megatron-lm!3774 - Fix router input jitter dtype** (20b3954) - **ADLR/megatron-lm!4086 - Fix UT: Compare shared experts overlap with A2A overlap** (697fcef) - **ADLR/megatron-lm!4062 - Fix mis-set of model_auto_sync and add basic gradient existence assertion to fully_shard tests.** (03045f2) - **ADLR/megatron-lm!4047 - fix(FSDP): avoid redundant meta device materialization for Megatron FSDP** (cabf5d1) - **ADLR/megatron-lm!4066 - ci(fix): Run inference tests** (16e19d0) ### Backend - **cp: Explicitly set minimum torch version to >= 2.6.0 (#3085) to `core_r0.15.0` (#3156)** (e8d782a) - **beep boop đŸ€–: Bumping versions** (27838c5) - **beep boop đŸ€–: Bumping versions** (1221b91) - **Revert "beep boop đŸ€–: Bumping versions" (#2562)** (0d7e02b) - **beep boop đŸ€–: Bumping versions** (d742a7d) - **Revert "beep boop đŸ€–: Bumping versions" (#2557)** (56345b7) - **beep boop đŸ€–: Bumping versions** (8a22ddf) - **Revert "beep boop đŸ€–: Bumping versions" (#2554)** (4cf56ac) - **beep boop đŸ€–: Bumping versions** (c56b7a7) - **beep boop đŸ€–: Bumping versions** (d4702de) - **cp: `ci: Run on push to release branch (1960)` into `core_r0.15.0` (#1961)** (fbbe62e) - **Ko3n1g/ci/core r0.15.0 (#1948)** (2d94aa8) - **Replay !4110 - Renaming Chunk to Block During Memory Management** (950aa43) - **ADLR/megatron-lm!4194 - Inference: throughput tests** (47d5049) - **ADLR/megatron-lm!4279 - ci: Build for sm80** (c9746e4) - **ADLR/megatron-lm!4193 - Automate ModelOpt restore** (2263b5d) - **ADLR/megatron-lm!3816 - Moe padding inference** (56818f9) - **Revert "ADLR/megatron-lm!4110 - Renaming Chunk to Block During Memory Management"** (bcdd591) - **ADLR/megatron-lm!4160 - docs: Update installation** (9a79a45) - **ADLR/megatron-lm!4110 - Renaming Chunk to Block During Memory Management** (f759111) - **ADLR/megatron-lm!4255 - ci: Cherrypick main into dev** (cf2087c) - **ADLR/megatron-lm!4226 - Use up-to-date model state dictionary when pre-processing the state dict for DCP.** (0ba8470) - **ADLR/megatron-lm!4228 - chore: Upgrade dependencies (2025-10-13)** (e2d0c1b) - **ADLR/megatron-lm!4218 - Refactor MambaMixer into separate training, prefill, and decode paths** (4e6e8b9) - **ADLR/megatron-lm!4240 - chore: Tool for upgrading deps** (fdd0a8a) - **tests: Update golden values** (15afa9f) - **ADLR/megatron-lm!4108 - Bump pytorch base container to 25.09** (44ab057) - **ADLR/megatron-lm!3823 - Unify enable/external cudagraph with cuda-graph-impl** (5769804) - **ADLR/megatron-lm!4188 - Disable singleton_local_shards by default** (43a09cb) - **ADLR/megatron-lm!4059 - Update README.md - Latest News Section** (c637e2e) - **ADLR/megatron-lm!4217 - Pop ModelOpt state after pruning + example readme update** (1d9896f) - **ADLR/megatron-lm!4224 - ci: Cluster-specific golden values** (dabfbc4) - **ADLR/megatron-lm!4068 - Compute shared experts before router** (e8024d7) - **cp: CI changes back to main** (c12c92f) - **ADLR/megatron-lm!4195 - build: All groups** (227395a) - **ADLR/megatron-lm!4204 - Make all dynamic inference unit tests internal** (272753b) - **ADLR/megatron-lm!3818 - Deprecate legacy inference** (4d51353) - **ADLR/megatron-lm!4164 - clone loss mask instead of assigning it to cached_loss_mask** (c94d58f) - **ADLR/megatron-lm!4190 - Update RL CODEOWNERS** (636179d) - **ADLR/megatron-lm!4183 - tests: Disable broken LTS nightly tests** (93aa335) - **Replay `ADLR/megatron-lm!3689 - Inference-only full model CUDA graphs`** (3d0a16f) - **ADLR/megatron-lm!3989 - Ensure dynamic engine returns requests in consistent order** (9e669f6) - **ADLR/megatron-lm!4116 - Changes to update timeout after init to shorten the timeout period closer to iteration time** (48b1bbf) - **ADLR/megatron-lm!4172 - Typo fixes from community** (5db6704) - **remove internal note** (0b4a568) - **update** (c8a36f3) - **ADLR/megatron-lm!4165 - set HF_HUB_CACHE env var** (3c29362) - **ADLR/megatron-lm!4082 - Lookup tables for Inference CudaGraphs** (ab43252) - **ADLR/megatron-lm!4163 - chore: Upgrade dependencies (2025-10-06)** (c7bc889) - **ADLR/megatron-lm!3850 - Bridge Communicator: Enable joint training of independently parallel modules [2/2].** (53008b8) - **ADLR/megatron-lm!4120 - Importing unified_memory does not fail if no cuda** (ae44e49) - **ADLR/megatron-lm!4154 - cp MR !4128 from dev: `ci: Auto-cherrypick MR into main`** (d5d3a8a) - **ADLR/megatron-lm!4105 - chore: Remove automation team from testonboarding** (c561ece) - **ADLR/megatron-lm!4125 - chore: Update classifiers in pyproject.toml** (caa6bd2) - **ADLR/megatron-lm!3921 - Avoid split and concat with RoPE** (71a09cc) - **Revert "ADLR/megatron-lm!3689 - Inference-only full model CUDA graphs"** (366633a) - **ADLR/megatron-lm!3689 - Inference-only full model CUDA graphs** (e0efff9) - **ADLR/megatron-lm!4117 - Port 1574 from Github** (e002b5c) - **ADLR/megatron-lm!4113 - Make override opt_param arguments consistent** (d3c06bd) - **ADLR/megatron-lm!3886 - Update Megatron-LM to use UB COMM+GEMM overlap for first-last BF16 layers** (d47b838) - **ADLR/megatron-lm!4020 - change experimental logging to log only once** (c3f1d76) - **ADLR/megatron-lm!3985 - Dynamic inference context | Unified memory.** (ef4ae45) - **ADLR/megatron-lm!4095 - Fixing HF import default dtype and checkpoint loading issue with singleton_local_shards=True.** (435e7e0) - **ADLR/megatron-lm!4077 - Take actual minimum value for max_requests_override when --inference-dynamic-batching-max-requests-override < 4** (d915c90) - **ADLR/megatron-lm!4099 - ci: Retry on HF connect failures** (3aac2c3) - **ADLR/megatron-lm!3888 - Change inference request ID to integer from string-integer** (12d0b41) - **ADLR/megatron-lm!4073 - Deprecate max_position_embeddings and its check** (cb6ab12) - **ADLR/megatron-lm!4097 - Adjust train-samples in functional test model config to guarantee 100B tokens training** (d9f5ec4) - **ADLR/megatron-lm!3913 - Megatron-LM: Make flash_decode=True + inference_context=None work** (42a56ec) - **ADLR/megatron-lm!3912 - Optimize attention preproc** (f0d9fa9) - **ADLR/megatron-lm!4087 - chore: Improvements to GitHub CI** (74d9bcf) - **ADLR/megatron-lm!4084 - ci: Send dev alerts to separate channel** (a00300f) - **ADLR/megatron-lm!4071 - tests: Onboard DSv3 test** (1d2a44a) - **ADLR/megatron-lm!4081 - chore: Upgrade dependencies (2025-09-29)** (bf1d5b2) - **ADLR/megatron-lm!4079 - build: Test TE2.7 wheel** (ce8185c) - **ADLR/megatron-lm!3972 - Handle pre-forward and post-forward hooks in TE fused MLP** (8503180) - **Replay (!3904) Cudagraph code refactor** (98b6f0e) - **Revert "ADLR/megatron-lm!3904 - Cudagraph code refactor"** (793b89a) - **ADLR/megatron-lm!3904 - Cudagraph code refactor** (c173615) ### Docs - **docs: Add changelog** (0a296d1) ### Chore - **build: Pin nvshmem (#2863)** (3b3a3a1) - **chore: Version bump for 0.15.1 release (#2844)** (512da5d) - **chore: Bump to 0.15 (#2504)** (d4e9e14) - **chore: Update settings for release testing (#2096)** (301a773) - **build: Bump modelopt (#2046) (#2050)** (b5d7c62) - **ci: Update copyright checker (#1975)** (de61ba0) - **ci: Increase timeout for MRs** (d013dcf) - **ci: Add golden values for inference** (a76f08e) - **ci: Fix nemo build** (fc06460) - **chore: Add error message** (572f019) - **ci: Retry on unknown_failure** (807cd66) - **ci: Restart jet jobs on failure** (6adc9e2) - **ci: Restart jet jobs on failure** (8156f26) - **ci: Allow failure on merge-to-dev** (90e3abd) - **ci: Check deterministically** (8a35969) - **ci: No batch short anymore** (3914532) - **ci(fix): Cherrypicking from forks** (9ae2982) - **ci: Restart on `corrupted unsorted chunks`** (66b170c) - **ci: Update A100 values** (08a0a19) - **ci: Fix notification channel** (da5fd9d) - **ci: Retry on DeepEP launch failure** (5e03143) - **ci: Restart on msgpack error** (b8581a3) - **chore: Version bump** (f345294) - **ci(hotfix): Nightly runs** (f497bac) - **chore: Version bump** (e656693) _Recap by [Repo Wrapped](https://repowrapped.com/gh/NVIDIA/Megatron-LM?utm_source=github-action)._