Megatron-LM

* Features * Inference * Add async support for DynamicInferenceEngine ([MR \!3187](https://github.com/NVIDIA/Megatron-LM/commit/05079d55a5bfcc7a43f4619e36a40a9e8db3f882)) * Pad input tensors and enable FP8 weights for FP8 inference ([MR \!3341](https://github.com/NVIDIA/Megatron-LM/commit/6a6cd478839d90cf09a837adf8c79cbc844bc920)) * Force inference to always gather logits with tensor parallelism ([MR \!3442](https://github.com/NVIDIA/Megatron-LM/commit/7c9cdcb794089968278c7272e0261a68edf5d369)) * Multi batch size CUDA Graphs for Dynamic Inference ([MR \!3402](https://github.com/NVIDIA/Megatron-LM/commit/30aabe5e3133c6d70aa55aaabad4ea8cb04ce63c)) * Post-training * ModelOpt updates ([MR \!3268](https://github.com/NVIDIA/Megatron-LM/commit/550ed5243c3a18e39430c15e8918ee63e41d7eaf)) * Add speculative decoding AR validation feature * Add DeepSeek and Qwen model configs * Performance * ModelCommProcessGroup integration ([MR \!3391](https://github.com/NVIDIA/Megatron-LM/commit/26adc2dfde53fbc2b063e2fdd1d9ed26578811a6)) * Add HyperCommGrid: N-Dimensional Communication Grid for Model Parallelism ([MR \!3398](https://github.com/NVIDIA/Megatron-LM/commit/45400df7da7fa23e3aff86804e5ac254d9a8d3c0)) * Flexible creation and management of communication groups * Add support for Spike No More embedding initializations and weight decay skipping ([MR \!3500](https://github.com/NVIDIA/Megatron-LM/commit/ee74aa66a06b24e511270f285db475941ef63bfd)) * MoE * We're actively optimizing large-scale fine-grained MoE performance on Blackwell Platform. * Features: * Support Expert Parallel A2A Overlapping ([MR \!3470](https://github.com/NVIDIA/Megatron-LM/commit/0c6c1176fb3e3e00534b3591f1ad023d4ecad6fb); [MR \!3074](https://github.com/NVIDIA/Megatron-LM/commit/4b30ec54aba97e16a083eca33d2df1dd48e1b48f)) * Support CP and recompute for MTP ([MR \!3330](https://github.com/NVIDIA/Megatron-LM/commit/650ab87d04105869f197f2ddc441e3b18ca93724)) * Add support for global aux loss ([MR \!3318](https://github.com/NVIDIA/Megatron-LM/commit/e58d9080ea212e005ccba0b6607bfcc86451285d)) * Memory Optimization * Support recomputation for FP8 layernorm/moe\_act/shared\_experts ([MR \!3465](https://github.com/NVIDIA/Megatron-LM/commit/6850cc6a739d168f8c84db6cdacf4fe2931c0c49)) * Support optimizer offloading for DSV3 FP8 training ([MR \!3659](https://github.com/NVIDIA/Megatron-LM/commit/abbde02f54b62a5194ebe951218e98feceba6d42)) * Performance Optimization * Add MoE router fusion ([MR \!3809](https://github.com/NVIDIA/Megatron-LM/commit/d93743a9f11d5d17824b8b49868cc90f2904896f)) * Updates for MoE cudagraph ([MR \!3631](https://github.com/NVIDIA/Megatron-LM/commit/95452706d7aa16dc174813e12639a8c8356fbe87)) * Bug fixes: * Fix router input jitter dtype ([MR \!3774](https://github.com/NVIDIA/Megatron-LM/commit/20b395424d2e2bbfaab57b2f954294eb57c90c82)) * Model support * Add MiMo video VLM train example ([MR \!3543](https://github.com/NVIDIA/Megatron-LM/commit/786f5629d3462aff2f8855f51db70e882c475116)) * Add AVLM for MIMO ([MR \!3624](https://github.com/NVIDIA/Megatron-LM/commit/db41707430bff743f986b5779712c74242b99caa)) * Ease of use * Add uv support for source installs ([MR \!3615](https://github.com/NVIDIA/Megatron-LM/commit/164204cd7216e642bdef7299c569d95f02f9a79e)) * Automated weekly prereleases ([MR \!3574](https://github.com/NVIDIA/Megatron-LM/commit/7e59266c70ef34a246438640af690b55c7ecac28)) * Bug fixes * Use mscale\_all\_dim for softmax\_factor ([MR \!2800](https://github.com/NVIDIA/Megatron-LM/commit/e96a358f60c82b8ac8d965d91c3cc4ad0230a4e0)) * Fix FP8 param blockwise scaling unit test ([MR \!3480](https://github.com/NVIDIA/Megatron-LM/commit/57082f946a04c3390fcfc43634dc546ec3ded033)) * Fix unit test blockwise scaling ([MR \!3491](https://github.com/NVIDIA/Megatron-LM/commit/6d95fe63658f967e56a3fda88a9c30a424fcb520)) * Optimize prefill for token-less requests ([MR \!3499](https://github.com/NVIDIA/Megatron-LM/commit/daaa650a9ac4291d4027ca2fdeb4298ce024efd2)) * Add default values for Fp8Padding and Fp8Unpadding ([MR \!3501](https://github.com/NVIDIA/Megatron-LM/commit/42b2b1d10a9cb699b7e5aa40f6bfba9c2a1348aa)) * Fix CUDA graph logic for flexible pp layout ([MR \!3505](https://github.com/NVIDIA/Megatron-LM/commit/020d85e50ddf0f0282802002acb3662129a519c5)) * Load FP8 models with strict=False ([MR \!3508](https://github.com/NVIDIA/Megatron-LM/commit/1ab876ddc4c1893c76f26d775226a8d1dcdfb3d2)) * Skip rope check for torch \< 1.4.0 ([MR \!3528](https://github.com/NVIDIA/Megatron-LM/commit/d8180ef8ed0bb6f305dcdedf1b27d91304f361a3)) * Disable Apex tests for stability ([MR \!3539](https://github.com/NVIDIA/Megatron-LM/commit/d1256277fe378add0a2cfd7251f5a350b6d126ec)) * Fix typo in parallel\_state expert parallelism ([MR \!3548](https://github.com/NVIDIA/Megatron-LM/commit/5783ff32af759b8102cf0cb0bb82b30c48b9da26)) * Guard modelopt on macOS ([MR \!3549](https://github.com/NVIDIA/Megatron-LM/commit/76144fe1106e4fb0e69aa75b7a6ab66e71e8f37f)) * Retry on CUDA function failure ([MR \!3554](https://github.com/NVIDIA/Megatron-LM/commit/809aab68307a64c1386d68cc78ef70f8f4e12a80)) * Fix NCCL mem pool creation error ([MR \!3557](https://github.com/NVIDIA/Megatron-LM/commit/b61e21153146a563309b5d44cb5d7f7425806072)) * Fix get\_rotary\_seq\_len return type ([MR \!3559](https://github.com/NVIDIA/Megatron-LM/commit/1fa6bc83c7aeae95abc8e86ff0aac596985a01c3)) * Retry on CUDA function failure ([MR \!3560](https://github.com/NVIDIA/Megatron-LM/commit/7da88d74865c3f1a59894173246f26e7b3bf91b9)) * Fix NCCL allocator attribute error ([MR \!3565](https://github.com/NVIDIA/Megatron-LM/commit/6b656114795d74c3353cb007c59af49b1752f447)) * Ensure multi-prompt inference works ([MR \!3568](https://github.com/NVIDIA/Megatron-LM/commit/0fae48931000c9c7af06f7dcf037b5b7d96e0cd6)) * Fix MD5 on FIPS systems ([MR \!3577](https://github.com/NVIDIA/Megatron-LM/commit/83ee8c2848a3b1d42b40086a64da11e19f4b191f)) * Fixes dynamic context and inference bugs ([MR \!3582](https://github.com/NVIDIA/Megatron-LM/commit/e9c1da60a1ccc85376666d58568ed1d3e5a4f9db)) * Fix TE version for interleaved fused RoPE ([MR \!3586](https://github.com/NVIDIA/Megatron-LM/commit/b72b6cc161f5273b545bca09677382917cf20492)) * Fix MTP with MoE and TP logging ([MR \!3594](https://github.com/NVIDIA/Megatron-LM/commit/9af96623b66693e058f6bfce8d0094dc976792d8)) * Guard TE import fix ([MR \!3596](https://github.com/NVIDIA/Megatron-LM/commit/1bf946b1ec3f11e71459c7c0d06a97edbed96a1a)) * Add assertion for NCCL UB case ([MR \!3599](https://github.com/NVIDIA/Megatron-LM/commit/e11d28592f19c122859be764b7afe7c208d9acc1)) * Remove Encoder PP related Functions ([MR \!3604](https://github.com/NVIDIA/Megatron-LM/commit/9e49aa4446a58cc21c4dc0c5d0806551ad075ca7)) * Fix segfaults in tests ([MR \!3605](https://github.com/NVIDIA/Megatron-LM/commit/f6492fe8164fd5b9ad55007d435ccfc66cb98cc7)) * Fix TE error in distributed optimizer ([MR \!3625](https://github.com/NVIDIA/Megatron-LM/commit/e6c510ff3c1159f8955589b26f7c395bdf0607d9)) * Remove redundant barrier in checkpoint flow ([MR \!3626](https://github.com/NVIDIA/Megatron-LM/commit/26869feb6a3ac7f5616cb7253c37a4244d107d70)) * Support VPP MTP, fix logging ([MR \!3630](https://github.com/NVIDIA/Megatron-LM/commit/c351a473c7eedac2c43eab0815afb9759f4f8187)) * Retry mechanism for free(): invalid pointer errors ([MR \!3632](https://github.com/NVIDIA/Megatron-LM/commit/ec35b41b2df145a7ccb84afc48d94e0786e094da)) * Fix test\_replication.py issues ([MR \!3633](https://github.com/NVIDIA/Megatron-LM/commit/f7b50b271b2e0e396069e02551b21aa6fb374b43)) * Fix typo in parallel\_state ([MR \!3634](https://github.com/NVIDIA/Megatron-LM/commit/3c79a2c330290df58804c33e28e7c197fcc1f0b9)) * Fix CUDA graph logic determination ([MR \!3635](https://github.com/NVIDIA/Megatron-LM/commit/90efa3ef8a3c4f9e0f1db9f67ab9348bfa501387)) * Fix TE installation error ([MR \!3636](https://github.com/NVIDIA/Megatron-LM/commit/7e7322c01c9cb8ec254ecd9042700b22b70fe5c8)) * Ensure correct sharding type in local tests ([MR \!3643](https://github.com/NVIDIA/Megatron-LM/commit/946357f8dd7fdc12424b3a66bc999e6c0a02696c)) * Fix cudagraphed backward buffer reuse for last layer ([MR \!3645](https://github.com/NVIDIA/Megatron-LM/commit/ee61cf450d24760952e8995aab045ab6d55b986e)) * Set default for packed\_seq\_params in get\_rotary\_seq\_len ([MR \!3651](https://github.com/NVIDIA/Megatron-LM/commit/510d58c46664f44c556005ac928c5c531e12f761)) * Fix dynamic example script errors ([MR \!3653](https://github.com/NVIDIA/Megatron-LM/commit/72e290bf1f4bbf0c8047bb10a51da6ea6372e163)) * Guard TE import fix ([MR \!3666](https://github.com/NVIDIA/Megatron-LM/commit/ac198fc0d60a8c748597e01ca4c6887d3a7bcf3d)) * Breaking changes: * `megatron.core.distributed.custom_fsdp` refactored as breaking change to `megatron.core.distributed.fsdp.src.megatron_fsdp` * Known issues

Megatron-LM

NVIDIA Megatron Core 0.14.0

More Python Projects

AutoGPT

yt-dlp

markitdown

NVIDIA Megatron Core 0.14.0

More Python Projects

AutoGPT

yt-dlp

markitdown

HelloGitHub