64 Pull requests merged by 26 people

• [Pipeliner] Skip async_wait when there is no async_cp op

  #6681 merged May 3, 2025

• [NFC] Remove #include "triton/Analysis/Utility.h" from LoopUnroll.cpp as unused

  #6684 merged May 3, 2025

• [BACKEND] backwardRematerialization cost model

  #6667 merged May 2, 2025

• [Blackwell] Fix the TMEM message heuristic

  #6692 merged May 2, 2025

• Support hoist ConvertDot Operand when the leaves of the slice are DescriptorLoadOp

  #6690 merged May 2, 2025

• [BUILD] Abort installation for Python <= 3.8

  #6649 merged May 2, 2025

• [Blackwell] Fix TMEM message size selection heuristic

  #6687 merged May 2, 2025

• [Backend] Use NVVM::MapaOp over call intrinsic (NFC)

  #6682 merged May 2, 2025

• [Blackwell] Fix TMEM allocation when captured buffer is not a subview

  #6686 merged May 2, 2025

• [BACKEND] Fixes removeZeroBasesAlongDim outDims and surjectivity check

  #6669 merged May 2, 2025

• [BACKEND] Update LLVM version to https://github.com/llvm/llvm-project/commit/092b6e73e651469527662443b592f98f442ece72

  #6572 merged May 2, 2025

• Reorder tma lowering pass after fence insertion.

  #6679 merged May 2, 2025

• Disable test_nvidia_tool for AMD and simplify default ptxas check

  #6676 merged May 2, 2025

• [Testing] Update test_compile_stats to allow custom library complier paths

  #6672 merged May 2, 2025

• [NFC] Factor core matrix layout out of nvmma conversion for clarity

  #6677 merged May 2, 2025

• Avoid calling self.knobs during reset

  #6673 merged May 2, 2025

• [PROTON-DEV] Trace reader for high-level parsing

  #6662 merged May 2, 2025

• [lit][cfg] Declare mlir-translate in lit config

  #6671 merged May 2, 2025

• [NVIDIA] WGMMA prefetch pass: Support asyn_token in LocalLoad

  #6640 merged May 2, 2025

• [BACKEND] Fix wrong linear layout for 3D transposed nvmma

  #6668 merged May 1, 2025

• [BACKEND] Fix tma lowering crash due to wrong insert point

  #6666 merged May 1, 2025

• [NFC][Layouts] Cleanup leading offset references + exponential iteration

  #6665 merged May 1, 2025

• [Blackwell] Refactor/slightly generalize warp specialization

  #6597 merged May 1, 2025

• [BACKEND] Support memdesc_reshape to allow different HBM layout for mmav5

  #6482 merged May 1, 2025

• [LAYOUTS] Improve split inference and fix comments

  #6663 merged May 1, 2025

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

  #5957 merged May 1, 2025

• [Analysis] divisibility handling for dividing by a power-of-two constant

  #6657 merged May 1, 2025

• [Hopper][WS] Automatic task partition for anchor ops

  #6658 merged Apr 30, 2025

• [TEST] Enable test-alignment.mlir checks & remove failing checks

  #6661 merged Apr 30, 2025

• [TEST][AMD] Only check kpack=1 on the cdna4

  #6651 merged Apr 30, 2025

• [AMD][BACKEND] Filter barriers in Membar between LocalLoads and AsyncCopy when pipelining

  #6639 merged Apr 30, 2025

• Update README.md for vscode intellisense

  #6648 merged Apr 30, 2025

• Revert "[BE][PIPELINE] Enabling mmav5 pipelining for 2 dots in the loop by default (#6599)"

  #6653 merged Apr 30, 2025

• Revert "[LAYOUTS] Use divideLeft on layout inference (#6577)"

  #6652 merged Apr 30, 2025

• [RELAND][TritonNVIDIAGPU] Add missing memory effects to some ops (#6518)

  #6644 merged Apr 29, 2025

• [PROTON-DEV] Fill in some AMD op lowerings

  #6604 merged Apr 29, 2025

• [Dialect] Mark Join and Split as pure

  #6645 merged Apr 29, 2025

• Revert "[LAYOUTS] Fix mixed precision swizzling (#6565)"

  #6643 merged Apr 29, 2025

• RFC [python] Rename config.py > knobs.py

  #6641 merged Apr 29, 2025

• [NVIDIA] Enable Programmatic Dependent Launch in Triton

  #6394 merged Apr 29, 2025

• [Backend][Hopper] Add a skeleton third-party warp specialization pass

  #6624 merged Apr 29, 2025

• Use symlinks for external plugins to fix TRITON_PLUGIN_DIRS

  #6627 merged Apr 29, 2025

• [NFC] Split Membar tests into common and ttng specific files

  #6637 merged Apr 29, 2025

• [python][typehint] Simplify find_paths_if and add typehints to _utils

  #6497 merged Apr 29, 2025

• [python][typing] Specify mypy_path in pyproject.toml

  #6596 merged Apr 29, 2025

• [PROTON-DEV] Trace binary decoder and event parser

  #6603 merged Apr 29, 2025

• [BENCH] Ignore egg-info files

  #6631 merged Apr 29, 2025

• [Remarks][MMA] Add remark for not applying MMAv5 and improve wording on remark of MMAv3

  #6528 merged Apr 29, 2025

• [BACKEND] Fix dot-operand pattern

  #6632 merged Apr 29, 2025

• [TritonNVIDIAGPU] Add dependency tokens to TMEM ops

  #6520 merged Apr 29, 2025

• [AMD][BACKEND] Schedule AsyncWait in front of AsyncCopy and LocalLoad

  #6621 merged Apr 29, 2025

• [AMD][BACKEND] Add alias information to other stores from AsyncCopy/BufferLoadToLocal

  #6619 merged Apr 29, 2025

• [BENCH] Remove use of the unpack operator to make scripts compatible with Python earlier than 3.11.

  #6630 merged Apr 29, 2025

• [NVIDIA] WGMMA subtiling: distribute NumImpreciseAcc across subtiles when NumImpreciseAcc > 0

  #6469 merged Apr 28, 2025

• [Dialect] Mark Join, Split, and Cat as non speculatable

  #6629 merged Apr 28, 2025

• Fix TensorMemoryAllocation to correctly implement getAlloc

  #6614 merged Apr 28, 2025

• [python][config] Add cache.home_dir config

  #6626 merged Apr 28, 2025

• [AMD] Disable warp specialization tests for AMD target

  #6620 merged Apr 28, 2025

• [BACKEND] Fix fence insertion analysis

  #6622 merged Apr 28, 2025

• [python] Introduce config module for all env vars/hooks

  #6467 merged Apr 28, 2025

• Add deduction guide to SmallVector

  #6618 merged Apr 28, 2025

• Expose max_threads_per_block to avoid launch failures and improve autotune robustness

  #6522 merged Apr 28, 2025

• Support None in getitem slices

  #6616 merged Apr 28, 2025

• [BACKEND] Do not use C++20 designated initializers

  #6615 merged Apr 28, 2025

15 Pull requests opened by 14 people

• Add `TRITON_IGNORE_LIBTRITON_HASH` for hash stability between Python versions

  #6617 opened Apr 28, 2025

• [AMD] Remove f8 dtypes in type checking for gfx12

  #6628 opened Apr 28, 2025

• [Python:compiler] Replace readbytes with MMAP

  #6650 opened Apr 30, 2025

• [AMD] Remove is within_2gb check for specialization

  #6655 opened Apr 30, 2025

• [AMD][LLVM] unpack `fmul`/`fadd` near MFMA

  #6656 opened Apr 30, 2025

• [WIP][DNR] Optimize partitioning and scheduling for attention

  #6660 opened Apr 30, 2025

• [knobs] Fix environment propagation & scope() API

  #6664 opened May 1, 2025

• [frontend] Fix autotune cache lookup when interpreter enabled

  #6678 opened May 2, 2025

• [Cuda] Small test fixes for Consumer Blackwell Cards and Hopper variants

  #6680 opened May 2, 2025

• Add proton profiling to tutorials 06 and 08

  #6685 opened May 2, 2025

• [AMD] Optimize to use 128-bit stores in epilogue for CDNA4

  #6688 opened May 2, 2025

• Aref Automatic Warp Specialization [AutoWS] Implementation

  #6689 opened May 2, 2025

• [PROTON-DEV] Time shift trick

  #6693 opened May 3, 2025

• [Backend] Improve how dynamic register reallocation is implemented

  #6694 opened May 3, 2025

• Add support for masked histograms

  #6695 opened May 3, 2025

5 Issues closed by 5 people

• Hopper: For TMA load, the layout LHS does not hoist for RS (register-smem) WGMMA

  #6646 closed May 2, 2025

• Triton matmul_ogs Kernel missing multiplying expert weight

  #6527 closed May 1, 2025

• ROCm: ImportError: cannot import name 'intel' from 'triton._C.libtriton

  #6634 closed Apr 29, 2025

• Adding third party TRITON_PLUGIN_DIRS is broken

  #6612 closed Apr 29, 2025

• errors introduced by scalars in Interpreter mode

  #5965 closed Apr 27, 2025

5 Issues opened by 5 people

• Ordering of lines matters with num_stages > 1.

  #6691 opened May 2, 2025

• Reduction is duplicated in TTIR -> TTGIR with num_stages>1 causing strange inconsistencies

  #6647 opened Apr 29, 2025

• TMA Store Non-deterministic results

  #6638 opened Apr 29, 2025

• [Feature Request] Print autotune keys with TRITON_PRINT_AUTOTUNING=1

  #6636 opened Apr 29, 2025

• Triton tl.argmax Accuracy Discrepancy in Top-K Selection, when input contains infinity number

  #6635 opened Apr 29, 2025

14 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] Add Concat op to AMDGPU dialect

  #6590 commented on May 2, 2025 • 18 new comments

• [LAYOUTS] Generic stmatrix lowering

  #6609 commented on Apr 29, 2025 • 13 new comments

• [BACKEND] BF16 atomic_add support

  #6519 commented on May 3, 2025 • 8 new comments

• Add default pre_run hooks as a JITFunction class variable

  #6434 commented on Apr 28, 2025 • 2 new comments

• [AMD] Relax conditions in ExtractSlice verifier

  #6417 commented on May 1, 2025 • 1 new comment

• xFormers for CUDA 12.8 : AttributeError: Cannot set attribute 'src' directly. Use '_unsafe_update_src()' and manually clear .hash of all callersinstead.

  #6123 commented on Apr 26, 2025 • 0 new comments

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

  #5919 commented on Apr 30, 2025 • 0 new comments

• tl.dot on transposed matrix tries to rearrange matrix in shared memory

  #6569 commented on Apr 30, 2025 • 0 new comments

• fail instal on arm platform

  #6606 commented on May 1, 2025 • 0 new comments

• Add Structured Tracing Mechanism for Triton Kernel Compilation

  #6364 commented on Apr 30, 2025 • 0 new comments

• [Draft] Enable `atomic_add` for `bf16`

  #6418 commented on Apr 28, 2025 • 0 new comments

• [PROTON] Simplify proton viewer APIs for bench_mlp analysis

  #6452 commented on Apr 30, 2025 • 0 new comments

• [WIP][AMD]Aggregated load of scales in DotScaledOp

  #6529 commented on Apr 29, 2025 • 0 new comments

• [PIPELINE] Enable MMAv5 pipelining for multi-dot loops by default

  #6613 commented on May 2, 2025 • 0 new comments