v26.8.0

What's Changed

• test: cover Compressor overlap and non-overlap windowing (PRPUNDIT-25) by @jiagaoxiang in #1098
• test: cover fused clamped SwiGLU Triton autograd (PRPUNDIT-23) by @jiagaoxiang in #1106
• test: pin AttentionResidualMixer against an independent mix (PRPUNDIT-22) by @jiagaoxiang in #1105
• test: pin Llama3 TurboAttention GQA layout (PRPUNDIT-20) by @jiagaoxiang in #1103
• test: pin Wan CausalConv3d streaming cache (PRPUNDIT-24) by @jiagaoxiang in #1099
• test: pin DeepSeek-V3 TurboAttention BSHD layout (PRPUNDIT-18) by @jiagaoxiang in #1101
• test: cover GPT-OSS turbo sink attention forward (PRPUNDIT-19) by @jiagaoxiang in #1102
• test: pin AdaLNContinuous scale/shift order and gradients (PRPUNDIT-17) by @jiagaoxiang in #1100
• docs(networking): add AINIC bundle version guide by @yeandy in #1094
• fix(flux): clone the validation loss out of Megatron's rescale path by @olehtika in #1079
• fix(config): parse scientific-notation CLI overrides as floats (e.g. --lr 1e-5) by @SumonAMD in #1112
• test: pin Qwen3 TurboAttention layout and qk-norm (PRPUNDIT-21) by @jiagaoxiang in #1104
• testintel: add missing test for AvgDown3D.forward (created by linxwang) by @jiagaoxiang in #1096
• docs: bump current release tags from v26.5 to v26.6 by @amd-fuyuajin in #1114
• examples(megatron): add Llama-3.1-8B BF16 LoRA-SFT recipe for MI355X by @SumonAMD in #1113
• fix(backend): raise FileNotFoundError instead of assert False in setup_backend_path (survives python -O) by @SumonAMD in #1111
• testintel: add missing test for Decoder3d.forward (created by linxwang) by @jiagaoxiang in #1097
• fix(te): requantize_data raises UnboundLocalError on the captured inp by @Anai-Guo in #1108
• feat(gpt-oss): MXFP4 deosc BF16 snapshots and simplified 20B recipe by @HuangWei-95 in #1070
• [JAX] fix XLA_FLAGS precedence inversion, add XLA_FLAGS_APPEND, document flag layering by @yeandy in #1054
• corrections to Docs/networking/ainic bundle guide by @yeandy in #1118
• examples(megatron): add Qwen3-30B-A3B and Qwen2.5-72B MXFP4 pretrain recipes for MI355X by @SumonAMD in #1119
• update baremetal install v26.6 and update docs to v26.6 by @amd-fuyuajin in #1120
• Dev/jax maxtext v26.6 primus cli by @amd-fuyuajin in #1095
• skills/spur-cluster-status: report QoS node caps vs live usage by @wenxie-amd in #1124
• skills/spur-cluster-status: report QoS node caps vs live usage by @wenxie-amd in #1125
• docs: organize technical blogs and remove the deprecated docs tree by @Xiaoming-AMD in #1121
• docs: drop the Instella-MoE blog from the Primus article lists by @Xiaoming-AMD in #1126
• refactor(runner): retire the examples launchers in favour of primus-cli by @WangLingxun in #999
• fix: work around Triton buffer-store miscompile and ROCm compiled-training NaNs by @WangLingxun in #962
• feat(gpt-oss): run the dense QKVO GEMMs on Turbo FlyDSL by @HuangWei-95 in #1127
• Update docs for ainic predownloaded bundle by @yeandy in #1133
• [Fix] add support for amd-smi by @GeneDer in #1134
• Fix system validation link by @SwRaw in #1137
• Add SpecForge backend for offline DFlash capture and train by @yudigege86 in #1139
• SDMA support for Megatron with distributed optimizer by @Z-Y00 in #1128
• fix(lint): reformat the SDMA distributed-optimizer files by @Xiaoming-AMD in #1157
• update mamba-370m config file for MI325X by @amd-fuyuajin in #1140
• examples(megatron): add gfx1250 native LoRA SFT recipes by @kailashg26 in #1155
• fix(ci): skip SpecForge unit tests until the overlay image is in CI by @Xiaoming-AMD in #1163
• [OOB Release] add JAX v26.7 dockerfile by @hlira-amd in #1143
• [OOB Release] add primus v26.7 dockerfile by @hlira-amd in #1142
• fix(deepseek-v4): stop multiplying V4 TFLOPs by hc_mult by @wenxie-amd in #1158
• Expose per-head max attention logit on the FLA fused backend for qk_clip by @vanshbhatia-amd in #1162
• feat: add fused grouped gemm support by @RuibinCheung in #1165
• Jax docker release - align the JAX backends with v26.7 by @amd-fuyuajin in #1172
• docs(release): update documentation for v26.7 by @amd-fuyuajin in #1171
• feat(release-docs): add the release docs agent by @amd-fuyuajin in #1170
• feat(tools): add perf batch runner and result extractor by @amd-fuyuajin in #1173
• fix: add MI350X GPU env file for gfx950 WarpSpeed by @coketaste in #1178
• feat(gpt-oss): enable RCCL-SDMA parameter gather by default by @suachong in #1168
• ci: gate GPU unit tests on the ci:gpu PR label by @WangLingxun in #1182
• Enable fused MXFP4 weight-gradient accumulation by @suachong in #1174
• examples(megatron): bake MXFP4 TE env into YAML env: blocks by @coketaste in #1179
• refactor(hybrid): rename Hylo stack back to Zebra-Llama, and fix the hybrid eval/convert tooling by @vanshbhatia-amd in #1167
• fix(torchtitan): initialize GPT-OSS router bias by @WangLingxun in #1183
• Add qk_clip (MuonClip) patch for hybrid models and the distributed optimizer by @vanshbhatia-amd in #1164
• feat(turbo): Llama-3.1-8B FP4 dense MLP, residual RMSNorm, and skip-y by @kylasa in #1156
• fix: apply FLA Triton autotune patch on the primus-cli pretrain path by @coketaste in #1181
• perf(megatron): match MXFP4 de-osc to the forward grid and cut its step cost by @HuangWei-95 in #1184
• ci: bump Primus-Turbo to pick up MegaMoE FlyDSL ownership (#525) by @Xiaoming-AMD in #1195
• fix(deepseek-v4): honour turbo_mega_moe_precision for MegaMoE experts by @Xiaoming-AMD in #1196
• Add SpecForge backend for online training (Mooncake + SGLang) by @yudigege86 in #1191
• perf(megatron): skip overwritten grad clears and isolate gradient reduce-scatter by @suachong in #1186
• Add MiniMax-M3 (text tower) model support by @wenxie-amd in #1197
• [Fix] Tune batch size in Zebra Llama-Mamba 3B BF16 pretraining config to avoid MI300 OOM by @kailashg26 in #1209
• fix(megatron): disable Triton recompiles for zebra Llama KDA 1B by @kailashg26 in #1215
• feat(projection): Adding a MaxText/JAX backend to Primus Projection by @araina-amd in #1203
• Revert "feat(projection): Adding a MaxText/JAX backend to Primus Projection" by @wenxie-amd in #1220
• feat(projection): Adding a TorchTitan backend to Primus Projection by @araina-amd in #1202
• [v26.8] perf(megatron): tune MI325X env settings and extend perf configs by @alfuyao-amd in #1227
• docs(release): backport v26.8 documentation (#1235) by @alfuyao-amd in #1239

New Contributors

• @Anai-Guo made their first contribution in #1108
• @SwRaw made their first contribution in #1137
• @coketaste made their first contribution in #1178

Full Changelog: v26.7.0...v26.8.0