First Version Release

What's Changed

• hipblaslt auto tune by @wenxie-amd in #23
• fix(examples): fix Megatron path used in prepare_dataset by @Xiaoming-AMD in #24
• fix(torch_fsdp): get ddp_config failed when use torch_fsdp by @Xiaoming-AMD in #27
• preflight by @wenxie-amd in #25
• merge trace file by @wenxie-amd in #31
• add inter-node ring p2p test by @limou102 in #30
• feat(megatron): Align TFLOPs calculation for megatron by @Xiaoming-AMD in #28
• [Fix] fp8 option not work by @RuibinCheung in #33
• feat(HSA ENV): tune ROCm runtime with HSA_NO_SCRATCH_RECLAIM and HSA_ENABLE_SDMA by @Xiaoming-AMD in #32
• feat(config parse): replace yaml value(int/float ) from env('KEY') by @Xiaoming-AMD in #34
• feat(fsdp): patch Megatron torch_FSDP2 with Primus implementation by @Xiaoming-AMD in #35
• Add README by @wenxie-amd in #36
• fix typo error of preflight script by @wenxie-amd in #37
• [Feat] Add tensile tuning example by @RuibinCheung in #38
• refactor(examples): simplify usage and improve structure for clarity by @Xiaoming-AMD in #39
• test:Add model-specific Megatron trainer test cases with isolated test config by @Xiaoming-AMD in #40
• Primus benchmark by @xiaobochen-amd in #43
• docs(contributing): add initial contributing guidelines by @Xiaoming-AMD in #42
• Dev/yaoc/mixtral by @ChengYao-amd in #44
• fix fast async checkpoint on ROCm by @limou102 in #46
• docs & refactor: reorganize README, unify config usage, and improve Megatron pretrain scripts for Primus by @Xiaoming-AMD in #45
• optimize: reduce FP8 training memory usage via scoped TE layer config overrides by @Xiaoming-AMD in #47
• fix(megatron): add missing import for 'inspect' in TE kwargs patch by @Xiaoming-AMD in #49
• chore(submodule): update Megatron-LM from 20250324 to 20250522 by @Xiaoming-AMD in #50
• refactor: improve benchmark runner and report parser with multi-node support by @Xiaoming-AMD in #51
• feature(model): add mixtral pretrain config by @ChengYao-amd in #52
• update trace_moe_metric call to fit new megatron interface by @ChengYao-amd in #53
• fix(Megatron): fix interleaved virtual pipeline training error and add corresponding UT by @lhzhang333 in #54
• opt(UT): add num_workers=1 in UT yaml to save most of the time on exit by @lhzhang333 in #56
• Update mixtral pretrain configs by @yuankaichen-amd in #55
• refactor(docker): Update docker image to v25.5_py310 by @wenxie-amd in #57
• feat(config): Update LLaMA pretrain configs by @Xiaoming-AMD in #58
• feature(RDMA): Add filtering for gpu RDMA network adapters by @chaojhou in #59
• fix(trainer-test): improve training script success detection using stdout … by @Xiaoming-AMD in #63
• chore(license): add MIT LICENSE file for Primus by @Xiaoming-AMD in #61
• refactor: move Megatron run scripts to examples root and add --backend parameter for multi-backend support by @Xiaoming-AMD in #64
• feat(torchtitan): Add TorchTitan Backend Support (Initial Stub) by @Xiaoming-AMD in #65
• feat(torchtitan): add --local-ranks-filter support in torchrun launcher by @Xiaoming-AMD in #67
• fix(slurm): remove --reservation flag and quote variables in run_slurm_pretrain.sh by @Xiaoming-AMD in #68
• feat(megatron): enable manual pipeline split in (interleaved) 1F1B-PP by monkey patching by @lhzhang333 in #69
• rebase main to instella branch by @wenxie-amd in #71
• fix(ip-interface): socket interface env regression by @Xiaoming-AMD in #70
• feat: Add run_k8s_pretrain interface for Kubernetes workload submission by @Xiaoming-AMD in #72
• feat(run_k8s_pretrain): support --workspace and improve job spec defaults by @Xiaoming-AMD in #73
• feat(megatron): support mock_data mode to skip dataset preparation by @Xiaoming-AMD in #74
• feat(k8s_pretrain): support log in stdout and file by @chaojhou in #76
• feat(k8s): Support for Node Selection via --nodelist and Add nodes by @Xiaoming-AMD in #75
• docs: add TorchTitan backend support entry to README by @Xiaoming-AMD in #78
• add benchmark for checkpoint saving by @limou102 in #81
• feat(torchtitan):Add model configs for LLaMA3-405B and LLaMA3-70B (TorchTitan) by @Xiaoming-AMD in #82
• feat(tp-overlap): add te backend and support tp overlap for megatron. by @zhenhuang12 in #79
• feat(benchmark): update kernel benchmark and add llama405B config by @xiaobochen-amd in #77
• llama3.1_405B model config by @wenxie-amd in #84
• print training envs by @wenxie-amd in #85
• add checkpoint loading metrics by @limou102 in #86
• feat: add new ckpt args of megatron by @wenxie-amd in #88
• doc: Add Mistral Models and Fix Formatting in examples/README.md by @Xiaoming-AMD in #87
• refactor(cli): Enhance Primus CLI with --override Support & Simplify Platform Defaults by @Xiaoming-AMD in #89
• chore(license): add AMD license headers by @Xiaoming-AMD in #90
• feat(k8s launch):Support forwarding unrecognized --args to ENTRY_POINT by @Xiaoming-AMD in #91
• fix(megatron): sync initialize_megatron of primus with that of megatron by @lhzhang333 in #93
• enable deepseek qk_layernorm by @wenxie-amd in #94
• checkout Primus-Turbo by github secret by @wenxie-amd in #96
• feat(tp-overlap): support torchtitan by patch fused_all_gather_matmul of torch op by @zhenhuang12 in #92
• add deprecated_20251209 moe layer by @wenxie-amd in #98
• feat(megatron): add attn warmup to save iter1's time when pp is used by @lhzhang333 in #97
• Primus Config/Patch Document by @wenxie-amd in #100
• feat(megatron): enable dumping pp schedule data and add pp visualization tool by @lhzhang333 in #99
• add patch readme for attn_warmup and decoder_pipeline_manual_split_list by @lhzhang333 in #101
• feat(megatron): add model and pretrain config for LLaMA3.1-405B by @Xiaoming-AMD in #102
• refactor: Refactor Torchtitan Config & Launch: YAML Unification, Backend Auto-Selection by @Xiaoming-AMD in #106
• refactor(torchtitan): switch llama3 configs from TOML to YAML by @Xiaoming-AMD in #108
• doc(examples): Rename Torchtitan LLaMA3 Configs to LLaMA3.1 and Update README Links by @Xiaoming-AMD in #110
• Add tas k8s runner's ci file by @haishuok0525 in #109
• test(megatron): add Mixtral-8x22B/Mixtral-8x7B test and TRAIN_LOG override support by @Xiaoming-AMD in #114
• Speedup primus-turbo build in k8s-ci runner by @wenxie-amd in #113
• fix(trainer): auto-enable tensorboard when profiling is enabled by @Xiaoming-AMD in #116
• Code isolation from shared path by @haishuok0525 in #119
• feat(megatron): add moe_use_fused_router_with_aux_score by @ChengYao-amd in #111
• [UT] Add deterministic extra check and unit test by @RuibinCheung in #115
• feat(turbo): Primus-Torchtitan support Primus-Turbo backend. by @xiaobochen-amd in #118
• permute fusion and padded mla attention by @wenxie-amd in #120
• Add deepseek v2 config by @wenxie-amd in #121
• fix(megatron): Correct offset calculation when vpp degree is larger than 2 by @lhzhang333 in #122
• update config for llama2-7b and llama3.1-8b to align test by @llying-001 in #123
• fix(tests): fix async_tp tests for warning print. by @zhenhuang12 in #125
• feat(megatron): mixtral & llama parallel tuning by @Xiaoming-AMD in #126
• fix(megatron): sync pp related code in primus trainer with megatron by @lhzhang333 in #129
• feat(tp-overlap): support te and torchtitan by patch fused_matmul_reduce_scatter of torch op by @llying-001 in #112
• feat(turbo): Primus-Megatron support Primus-Turbo backend by @kyle-256 in #124
• ci(install-turbo): update commit to remove triton_dist. by @zhenhuang12 in #130
• Reduce trace file size in torch.profile by @wenxie-amd in #131
• feat(offline_tune): support generate reports by @RuibinCheung in #133
• feat(turbo): fit PrimusTurboAttention with cp by @ChengYao-amd in #134
• fix(config): update gradient_accumulation_fusion and moe_use_legacy_grouped_gemm comment by @RuibinCheung in #135
• feat(async_tp): support fp8 for all_gather + matmul. by @zhenhuang12 in #132
• add NCCL_IB_HCA, CLEAN_DOCKER_CONTAINER by @wenxie-amd in #138
• add 515B/1T/2T/4T proxy model; update launcher by @wenxie-amd in #150
• PR for primus/megatron v25.7 release by @vidushi8 in #145

Full Changelog: https://github.com/AMD-AIG-AIMA/Primus/commits/v0.1.0-alpha