34 Pull requests merged by 19 people

• [BACKEND] Revert smem layout heuristic added in PR#5924

  #5983 merged Feb 21, 2025

• [LLVM] Bump to llvm/llvm-project@c78cb3028363

  #5981 merged Feb 21, 2025

• [Blackwell] Fallback to MMAv2 for numWarps other than 4 or 8

  #5978 merged Feb 21, 2025

• [AMD] Rework MFMA intrinsic mapping queries

  #5937 merged Feb 20, 2025

• Propagate DotOp thru Join & improve shmem load into LinearEnc

  #5924 merged Feb 20, 2025

• [Backend] Try to fix infinite loop in membar

  #5973 merged Feb 20, 2025

• [Analysis] Use verify-diagnostics for print-based tests (NFC)

  #5970 merged Feb 20, 2025

• [BACKEND] Enable generic reduction on all layouts

  #5962 merged Feb 20, 2025

• [AMD]Support Scale is None in DotScaledOp in gfx950

  #5931 merged Feb 19, 2025

• [AMD] Update smem size for cdna4

  #5964 merged Feb 19, 2025

• [BACKEND] enable lld

  #5907 merged Feb 19, 2025

• Remove git commit hash in wheel name when building from release branch

  #5953 merged Feb 19, 2025

• [Frontend] Support returning tensor descriptor from functions

  #5958 merged Feb 19, 2025

• [AMD]Make AMDGPUAccelerateMatmul depend on TritonAMDGPUDialect

  #5959 merged Feb 18, 2025

• [NFC][BACKEND] Simplify reduce helpers

  #5954 merged Feb 18, 2025

• [AMD]Fix an error in the cache modifier bit setting

  #5948 merged Feb 18, 2025

• [AMD] Skip scalar and 1D tensor load for sinkSecondLoad

  #5955 merged Feb 18, 2025

• [AMD] Fix loop trip count for scf.while in ConvertToBufferOps

  #5952 merged Feb 18, 2025

• [RELAND] [BC breaking] [FRONTEND] Throw an error when we would downcast an integral constant to a dtype it does not fit in (#5866)

  #5926 merged Feb 18, 2025

• [TritonGPU] Fix crash in Accelerate matmul

  #5949 merged Feb 18, 2025

• [BACKEND] Fix dereference nullptr

  #5944 merged Feb 17, 2025

• [AMD] Revert using llvm.intr.masked.{load|store}

  #5913 merged Feb 17, 2025

• [AMD-Pipeline] Add multi-stage global/local prefetch

  #5353 merged Feb 17, 2025

• [Blackwell] Fix test_pipeliner.py breakage

  #5940 merged Feb 17, 2025

• [FRONTEND] Cache and annotate the TRITON_F32_DEFAULT env variable

  #5942 merged Feb 17, 2025

• [AMD] Fix failing tests due to mid-air collision

  #5943 merged Feb 17, 2025

• [AMD] Improve ConvertToBufferOps with range analysis

  #5563 merged Feb 16, 2025

• [NFC] Remove duplicate test parameters for test_dot

  #5938 merged Feb 16, 2025

• Fix incorrect kernel compilation in batched matmul (#5620)

  #5936 merged Feb 16, 2025

• [BACKEND] Add arith::CeilFloorDivExpandOpsPatterns

  #5934 merged Feb 16, 2025

• [AMD] Add MLIR Remark Messages when the Ping Pong Scheduler Succeeds

  #5914 merged Feb 16, 2025

• [LAYOUTS] Make operator* associative and dimension-order-preserving

  #5928 merged Feb 15, 2025

• [AMD] Fix buffer cache modifier test index out of range

  #5904 merged Feb 15, 2025

• [FRONTEND] Fix default values of tl.range

  #5932 merged Feb 14, 2025

16 Pull requests opened by 11 people

• [Blackwell] Support narrower TMEM messages and shapes

  #5945 opened Feb 17, 2025

• [Blackwell] Propagate TMA attributes from MMA operand

  #5947 opened Feb 18, 2025

• typeConverter to llvm support addressSpace attribute

  #5951 opened Feb 18, 2025

• [PROTON-DEV] proton dialect to protongpu dialect lowering

  #5956 opened Feb 18, 2025

• [python][compiler] Implement CompilationListener to report compile times

  #5957 opened Feb 18, 2025

• [AMD] Turn buffer ops support on by default

  #5960 opened Feb 18, 2025

• [FRONTEND] [BC Breaking] Require global variables to be insantiated as constexpr ob…

  #5961 opened Feb 19, 2025

• [Backend] Plumb `ttg.warp_specialize` through LLVM lowering

  #5963 opened Feb 19, 2025

• [AMD] replace `rocm_lld` with lld API call

  #5966 opened Feb 19, 2025

• [WIP][DNR] Codegen for `ttg.warp_specialize`

  #5968 opened Feb 20, 2025

• [backend] Update LLVM version to https://github.com/llvm/llvm-project/commit/386af4a5c64ab75eaee2448dc38f2e34a40bfed0

  #5974 opened Feb 20, 2025

• [AMD] [DEBUG] Added LLVM Debug messages for when the pingpong scheduler fails

  #5975 opened Feb 20, 2025

• [AMD][NFC] refactor RangeAnalysis

  #5977 opened Feb 20, 2025

• [AMD] Remove non-linear-layout-based local load pattern

  #5979 opened Feb 21, 2025

• [LAYOUTS] Allow DistributedEncoding attributes to override get[Total]ElemsPerThread()

  #5980 opened Feb 21, 2025

• [Interface] Add dot interface methods to get A/B tensor

  #5984 opened Feb 21, 2025

3 Issues closed by 3 people

• Setting `TRITON_F32_DEFAULT` does not trigger recompilation

  #5941 closed Feb 17, 2025

• tl.dot with batched (3D) input is working only in emulation mode when os.environ['TRITON_INTERPRET'] = '1' is set.

  #5620 closed Feb 16, 2025

• I couldn't find this package in PIP... why...

  #5935 closed Feb 15, 2025

9 Issues opened by 9 people

• Addition Incorrect

  #5972 opened Feb 20, 2025

• Bug in tutorials/06-fused-attention.py: test_op assertion fails for specific input.

  #5971 opened Feb 20, 2025

• Nightly install

  #5967 opened Feb 19, 2025

• errors introduced by scalars in Interpreter mode

  #5965 opened Feb 19, 2025

• Dose Triton supports new features of Blackwell for RTX5090 and 5080?

  #5950 opened Feb 18, 2025

• Cache Modifier '.cs' Not Supported for LOAD

  #5946 opened Feb 18, 2025

• Triton kernel not compiling with multiple threads and GPUs

  #5933 opened Feb 15, 2025

• Upstream LLVM SLP vectorizer change requires the correct triple

  #5930 opened Feb 14, 2025

• fatal : Unsupported .version 8.6; current version is '8.5'

  #5929 opened Feb 14, 2025

22 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [AMD] [FrontEnd] Optimize is_within_2gb and only enable with buffer ops supported

  #5898 commented on Feb 18, 2025 • 8 new comments

• [AMD][Atomic] Fix fp16 atomic operation

  #5839 commented on Feb 20, 2025 • 2 new comments

• Try Blackwell CI

  #5922 commented on Feb 20, 2025 • 0 new comments

• [PIPELINE] Refactor loop lowering.

  #5918 commented on Feb 14, 2025 • 0 new comments

• cache: add the triton version to the json metadata

  #5912 commented on Feb 20, 2025 • 0 new comments

• Protect autotuner with synchronization

  #5893 commented on Feb 15, 2025 • 0 new comments

• Add triton 3.13t builds - DO NOT Merge

  #5455 commented on Feb 18, 2025 • 0 new comments

• [NVIDIA][Backend] fix the wrong comment

  #5305 commented on Feb 15, 2025 • 0 new comments

• Is Triton unable to install in python 3.10 versions?

  #1057 commented on Feb 21, 2025 • 0 new comments

• IndexError: map::at with RTX 2080Ti

  #4813 commented on Feb 20, 2025 • 0 new comments

• FP8 GEMM implemented with triton is slower on Ada (SM89)

  #5583 commented on Feb 20, 2025 • 0 new comments

• Gather does not work if index is much longer than value

  #5836 commented on Feb 20, 2025 • 0 new comments

• SystemError: PY_SSIZE_T_CLEAN macro must be defined for '#' formats

  #5919 commented on Feb 19, 2025 • 0 new comments

• `tl.cumsum(i1)` computes `tl.cumsum_xor`

  #5897 commented on Feb 19, 2025 • 0 new comments

• No wheels for arm

  #5561 commented on Feb 19, 2025 • 0 new comments

• Wrong source read by triton.jit in some python versions

  #1589 commented on Feb 19, 2025 • 0 new comments

• Triton import is broken in Python 3.7 in triton 2.0.0.post1

  #1727 commented on Feb 18, 2025 • 0 new comments

• Potential Bug in **_attn_fwd_tma** Function

  #5816 commented on Feb 17, 2025 • 0 new comments

• Accessing slices of a tensor

  #656 commented on Feb 17, 2025 • 0 new comments

• Questions about the tutorial fused-attention

  #3700 commented on Feb 15, 2025 • 0 new comments

• Back-to-back BMMs failed with Triton nightly

  #5424 commented on Feb 15, 2025 • 0 new comments

• Is there a plan to support Windows?

  #1640 commented on Feb 14, 2025 • 0 new comments