v0.9.2rc2

Full Changelog: v0.9.2rc1...v0.9.2rc2

v0.9.2

Full Changelog: v0.6.0...v0.9.2

v0.9.2rc1

Full Changelog: v0.6.0...v0.9.2rc1

v0.9.1+rocm

What's Changed

• Upstream merge 2025 06 03 by @gshtras in #566
• Fix attention fp8 output fusion for split attention path in v1 by @gshtras in #569
• Upstream merge 2025 06 05 by @gshtras in #570
• Rocm 6.4.1 as base by @gshtras in #571
• Upstream merge 2025 06 09 by @gshtras in #575
• Updated README.md for June 10 release by @Mcirino1 in #574
• Upstream merge 2025 06 12 by @gshtras in #577

Full Changelog: v0.9.0.1+rocm...v0.9.1+rocm

v0.9.0.1+rocm

What's Changed

• Aiter mla cherrypick by @gshtras in #543
• Cherry pick skinny gemms by @gshtras in #544
• Upstream merge 2025 05 15 by @gshtras in #547
• Remove gradlib mention from pyproject by @gshtras in #549
• Fix input layer norm mismatch for Eagle Speculative Decoding compatib… by @mmkamani7 in #548
• Updated README.md by @Mcirino1 in #546
• Upstream merge 2025 05 19 by @gshtras in #550
• Reduce diff from upstream by @gshtras in #551
• Transformers 4.52 fix by @gshtras in #552
• Remove the option to compile cython during docker build by @gshtras in #553
• Fixing pre-commit in github. by @gshtras in #554
• Restrict FP8 attn fusion to triton FA by @gshtras in #545
• Upstream merge 2025 05 27 by @gshtras in #557
• Removing redundant parameters from the MIs side and fixing Navi build by @gshtras in #559
• Removing RPD in favor of torch profiler for V1 by @gshtras in #558
• Added benchmark results and commit hash by @Mcirino1 in #556
• Upstream merge 2025 06 02 by @gshtras in #565

New Contributors

• @mmkamani7 made their first contribution in #548

Full Changelog: v0.8.5+rocm...v0.9.0.1+rocm

v0.8.5+rocm

What's Changed

• Bump hiblaslt by @gshtras in #528
• Restrict setuptools version by @gshtras in #529
• Revert aiter commit by @gshtras in #530
• Update README.md by @t-parry in #531
• Restrict ray version due to a breaking release by @gshtras in #533
• Upstream merge 2025 04 29 by @gshtras in #536
• Upstream merge 2025 05 06 by @gshtras in #538

Full Changelog: v0.8.4+rocm...v0.8.5+rocm

v0.8.4+rocm

What's Changed

• Remove duplicate code in config.py by @sstamenk in #494
• In light of the breaking cmake v4 release by @gshtras in #495
• Docs_update_20250327 by @arakowsk-amd in #493
• Upstream merge 2025 03 31 by @gshtras in #497
• Triton MLA parameter tweak for AMD GPU by @qli88 in #498
• Upstream merge 2025 04 02 by @gshtras in #499
• Bump aiter version by @gshtras in #500
• Adding 2stage MoE support separately until it is added upstream by @gshtras in #501
• Fused FP8 conversion in attention for v1 by @gshtras in #502
• Upstream merge 2025 04 07 by @gshtras in #503
• Fix fused moe by @gshtras in #506
• Update moe_tune_script.sh by @divakar-amd in #507
• Doubled size to wa issue and preserve CAR perf by @maleksan85 in #510
• Re-enable custom paged attention for V0 by @charlifu in #511
• Updated README.md with April 10 results by @Mcirino1 in #512
• Update README.md by @faisalgulfam32 in #514
• Updating base image by @charlifu in #515
• Update test-template.j2 to enable building by @Alexei-V-Ivanov-AMD in #517
• Update test-template.j2 to fix new location of run-amd-test.sh by @Alexei-V-Ivanov-AMD in #518
• Rocm 6.4 docker by @gshtras in #519
• Update README.md by @t-parry in #521
• Update README.md by @t-parry in #523
• Upstream merge 2025 04 21 by @gshtras in #522
• Upstream merge 2025 04 25 by @gshtras in #524

New Contributors

• @sstamenk made their first contribution in #494
• @faisalgulfam32 made their first contribution in #514

Full Changelog: v0.8.2+rocm...v0.8.4+rocm

v0.8.2+rocm

What's Changed

• Stable aiter build by @gshtras in #450
• Remove batch padding on ROCm by @gshtras in #451
• Aiter whl fix branch by @gshtras in #452
• tuning adjustment for quantized skinny gemm. by @amd-hhashemi in #444
• Upstream merge 25 03 03 by @gshtras in #458
• Cython compilation fix by @gshtras in #460
• Building the base images for MI and Navi; Using aiter hotfix by @gshtras in #461
• Hipblaslt bench by @gshtras in #462
• Fixing the shape to use in padding calculation by @gshtras in #464
• Upstream merge 25 03 10 by @gshtras in #471
• V1 rocm support by @maleksan85 in #469
• nightly_fixed_aiter_integration_final_20250305 README update by @Mcirino1 in #470
• Updated README.md with config info and header font size by @Mcirino1 in #473
• Bump aiter version by @gshtras in #476
• upstream_merge_25_03_12 by @gshtras in #478
• Rocm vllm ci fix (new design) by @Alexei-V-Ivanov-AMD in #475
• Add @hongxiayang by @gshtras in #481
• use unified fp8 notation in FA by @divakar-amd in #483
• Removing the padding again after it had been overwritten by upstream merge by @gshtras in #485
• Upstream merge 2025 03 24 by @gshtras in #489
• Using temp dir for the intermediate P3L file by @gshtras in #491
• Using torch commit that supports running scaled_mm on Radeon by @gshtras in #492

New Contributors

• @Mcirino1 made their first contribution in #470

Full Changelog: v0.7.3+rocm...v0.8.2+rocm

v0.7.3+rocm

What's Changed

• Initial attempt to adjust codeowners to the ROCm fork by @gshtras in #420
• Applying weight padding to deepseek by @gshtras in #421
• [Model] DeepSeek Tunings by @rasmith in #423
• Removing bad config by @gshtras in #425
• Refine codeowners by @gshtras in #427
• Upstream merge 25 02 17 by @gshtras in #430
• Updating PR template to point people to the upstream repo. by @gshtras in #431
• Enabling the ROCm-vLLM CI on MI250 machines by @Alexei-V-Ivanov-AMD in #432
• Optimization for quantized gemm skinny sizes by @amd-hhashemi in #411
• Restricting FP8 wvSplitk to MI300x by @gshtras in #439
• Remove mi300a by @gshtras in #440
• resolve configs diff for mixtral8x7B by @divakar-amd in #437
• Torch version bump to fix tunable ops by @gshtras in #442
• Using AITER branch with fixed whl by @gshtras in #443
• Bump hipblaslt version. Minor fixes to printing the versions by @gshtras in #447
• Bumping the version in the right place by @gshtras in #448
• Upstream merge 25 02 24 by @gshtras in #449

Full Changelog: v0.7.2+rocm...v0.7.3+rocm

v0.7.2+rocm

What's Changed

• 20250127 docs update by @arakowsk-amd in #392
• Faster Custom Paged Attention kernels by @sanyalington in #372
• Improved memory profiling by @gshtras in #394
• Aiter readme by @gshtras in #400
• fix None dict for quark by @hliuca in #402
• Upstream merge 25 02 03 by @gshtras in #403
• Mbatch p3l by @Alexei-V-Ivanov-AMD in #401
• Fix quark fp8 format loading. by @fxmarty-amd in #395
• WARP_SIZE in sgl moe kernel by @gshtras in #406
• Update README.md 20250205_aiter by @arakowsk-amd in #407
• fix rocm get_device name by @divakar-amd in #359
• Fixing the output formatting in P3L by @gshtras in #414
• Add tuned moe config for qwen1.5_moe_A2.7B by @sky0530 in #398
• Update Benchmark Profiling Scripts by @AdrianAbeyta in #417
• updating 20250207 image manifiest by @arakowsk-amd in #416
• Upstream merge 25 02 10 by @gshtras in #418
• Aiter base by @gshtras in #419

New Contributors

• @arakowsk-amd made their first contribution in #392
• @fxmarty-amd made their first contribution in #395
• @sky0530 made their first contribution in #398

Full Changelog: v0.7.0+rocm...v0.7.2+rocm