Releases · ROCm/FlyDSL

What's Changed

[kernels] Structure-only cleanup: unify naming, dedup helpers, split monolith module by @Phil-amd in #814
[Fix]: mla_decode test aborts whole collection on aiter API drift by @Phil-amd in #840
add mxfp8 support by @solinzby1 in #838
[CI] run_benchmark: treat runtime pytest.skip as skipped, not failed by @jhinpan in #842
[Enh] Add rocdl.buffer_rsrc utilities: make_buffer_ptr & get_buffer_rsrc by @sjfeng1999 in #841
[kernels]: add common/mem_ops for shared memory/atomic helpers by @Phil-amd in #839
[Docs] Remove duplicate Documentation table from README by @Peter9606 in #846
[Misc] Propagate signatures through expr decorators by @simondanielsson in #835
[CI] skip build/test for docs-only changes by @Phil-amd in #852
[kernels] Relocate shared MMA infrastructure under common by @Phil-amd in #849
[CI] Make benchmark baselines fork-safe and bounded by @jhinpan in #857
add xcd remap by @solinzby1 in #863
[Test] Expand RMSNorm backward configs, cache reuse, and Torch benchmarks by @jhinpan in #854
[Doc] Add FlyDSL Project Status and Compatibility Notice by @jamestangg in #865
fix(kernels): correct stale kernels.mma import in mxfp4_preshuffle by @coderfeli in #870
[kernels] Migrate pa_decode_swa reduce kernel to SharedAllocator smem by @xudoyuan in #812
[FMHA] Refactor kernels into shared helper framework by @yanguahe in #845
[Docs] Arithmetic types: authoritative spec + conformance suite by @sjfeng1999 in #868
[2/5] autotune: add opt-in search and RMSNorm waves-per-EU tuning (#770) by @jhinpan in #785
[Perf] Optimize RMSNorm backward with staged reduction and vec8 I/O by @jhinpan in #855
[layernorm] Add backward pass for training (PR 3/3, #769) by @jhinpan in #801
[Feat] Export shuffle xor/up/down/idx ops by @sjfeng1999 in #871
Atom 1250 refactor by @coderfeli in #866
[Perf]optimize flydsl flash attention kernel by @binding7012 in #850
Fix Vector list()-iteration hang by @adityas-amd in #873
Add FlyDSL kernel code cleanup skill by @coderfeli in #881
Skill: align kernel-authoring with code-cleanup (drop buffer_ops/ArithValue) by @coderfeli in #882
refactor(expr): move buffer_ops and split gfx1250 mcast policy to kernels.common by @Phil-amd in #880
Release: bump version to 0.3.0 by @coderfeli in #893
autotune: keep CI search opt-in (#770) by @jhinpan in #788
[gfx1250] Refactor MoE/GEMM kernels with layout/atom API by @aoli26 in #885
perf(pa tile): readable tile-programming PA decode kernel, matches/beats production by @fsx950223 in #825
[CI] Share MLIR cache across GPU jobs by @Phil-amd in #890
[Kernel] conv3d: parameterize tile size + add autotuner by @jiacao-amd in #820
[fix] Temporarily quarantine rmsnorm smoothquant on gfx1201 by @cschenjunlin in #895
[rmsnorm] Support FP32 weights with FP16/BF16 activations by @jhinpan in #884
[CI] Make MLIR cache fallback reliable by @Phil-amd in #897
Refactor mxmoe by @coderfeli in #892
[3/5] autotune: add offline config artifacts (#770) by @jhinpan in #786
[Refactor] Align RMSNorm with current FlyDSL kernel style by @jhinpan in #902
Add forward LSE output to FlyDSL flash attention kernels by @amd-wsung102 in #844
[Prune] Move arch-specific FP8 type selection to kernels_common by @sjfeng1999 in #908
[Doc] document fast-math flags in the arithmetic_types by @sjfeng1999 in #911
[perf]optimize flash attention fp8 for gfx950 by @binding7012 in #905
[ROCDL] Add arch-aware universal s_waitcnt by @sjfeng1999 in #915
[Enh] Add rounding-mode control for float-to-float casts by @sjfeng1999 in #916
[fix] Autotuner: propagate the tuned function's return value by @aryaman-gupta in #899
Port GEMM/MoE/conv kernels to the layout API by @coderfeli in #913
[Feat] Add to_llvm_ptr op to convert fly.ptr to llvm.ptr by @sjfeng1999 in #921
[compiler] carry Python containers through dynamic if/for/while/ifexp by @xudoyuan in #874
MegaMoE on Gfx950 by @GwilliamHu in #876
[Doc] Define API stability contract by @sjfeng1999 in #925
perf(attention): keep a head's Q-blocks on one XCD in the dualwave mapping by @JohnQinAMD in #919
[Dialect][Kernel][Perf] Add gfx120x.wmma atom and port the RDNA4 GEMM to the layout API by @vlluvia in #943
[Bugfix] Resolve reg offset on dynamic add_offset chains by @Phil-amd in #926

No results found