NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #473 most downloaded on PyPI
A language and compiler for custom Deep Learning operations
Latest stable
3.9.0
Last release 26 days ago
27 Oct 2026
Ships on a steady schedule
a new release about every 2 months
Some releases are documented
notes for 8 of 23 stable releases
Nothing withdrawn
no release was ever pulled
5 years old
23 releases · first in 2021
Drag across the ruler to compare, or tap a release to open it.
Atomic loads and stores: Added tl.atomic_load and tl.atomic_store , also available in Gluon and the interpreter. Loads default to acquire and stores d
tl.atomic_load and tl.atomic_store, also available in Gluon and the interpreter. Loads default to acquire and stores default to release (#11591, #11650, #11676, #11821)tl.atomic_poll: Added tl.atomic_poll, which waits until the value at a pointer equals an expected value and returns False on timeout (#10723, #10755, #11540)tl.fdiv(..., approx=True) for fast FP32 division (#11900)0 <= x <= 8, keyword-only parameters, tuple unpacking in list comprehensions, and operator methods on aggregate types (#12002, #11418, #11076, #10990)tl.clamp: tl.clamp accepts integer tensors (#12005)tl.topk k values, tl.cat shape mismatches, and non-32-bit tl.histogram inputs now fail at compile time (#11053, #11074, #11353, #10891)tl.softmax with dim other than 0, integer stores to boolean pointers, list-comprehension if filters, negative-zero constants, and stale compilations when a captured constexpr shares a global's name (#11409, #11853, #10888, #11426, #11224)0x00 now decodes to 2^-127 and 0xFF to NaN (previously zero and infinity), which changes results for kernels that use those values (#11624)TRITON_INTERPRET=1 now matches compiled kernels for integer overflow, tl.fma, tl.umulhi, and tl.histogram (#10923, #11194, #11210, #11390, #11424, #11531, #11641, #11403, #11218, #11751)tl.histogram, and ordered atomics in warp-specialized kernels (#10656, #10735, #10843, #10975, #10559, #11775, #10914)acquire, release, or acq_rel semantics now insert CTA-wide barriers so the ordering covers the whole block. Read-modify-write and CAS atomics default to acq_rel; pass sem="relaxed" to avoid the barriers (#10816, #11022)tl.histogram, gathers, axis analysis, and optimizations that moved or duplicated loads (#11521, #11688, #11617, #11621, #10744, #11186, #11269, #11682, #11842, #11637, #11981, #11990, #11943, #12011, #11767, #10838, #11546, #11548, #11337, #11314, #11906, #11975, #11240)tl.clamp, pure inline assembly, and Gluon embedding ops. Narrowing casts now fold discarded bits by default; set TRITON_FPSAN_HOMOMORPHIC_CASTS=1 to keep truncation (#10707, #10627, #10908, #10834, #11005, #11083, #11101, #11159, #11170, #11489, #11627, #11665, #11716, #11974)extend_with and extend_dialects_with, so a package can register itself on import (#10775)MmaEncodingTrait and SharedEncodingTrait (#11025, #11764)gfx1250-strict target (#11945, #12134)tdm.async_load_fused and pipelining of in-loop descriptor stores and scatters. Fixed rank-reducing loads, multi-CTA gather/scatter, and size-1 block dimensions. Non-zero padding (padding_option="nan") now raises an error (#10628, #10564, #11077, #11833, #11963, #12331, #12441, #10870)scaled_downcast to MXFP4/MXFP8 on CDNA4 and CDNA5, and scaled_upcast accepts compact scales (#11377, #11442, #10979)gfx117x) targets and RDNA3 buffer atomics. bf16 multiplication now rounds to nearest even, and dot_scaled rounds its bf16 downcast toward zero (#10878, #11474, #11535, #11227, #11629)TRITON_AMD_CODEGEN_PATH to a prebuilt libtriton_amd_codegen library (#11485, #11717, #11745, #11912, #11925, #11916, #12003, #12016)rocm_sdk) unless TRITON_* paths are set. It raises an error if a different HIP runtime is already loaded (#11028, #11398, #11492)tl.dot, CDNA barriers emitted on RDNA, buffer-atomic fence ordering, and miscompiles in range analysis and canonicalization (#10747, #11783, #10906, #11081, #11313, #11246, #11239)tl.load(..., volatile=True) keeps its volatility, which fixes hanging polling loops. Also fixed LDS races, wrong shared-memory addresses on RDNA3/3.5, out-of-bounds XF32 MFMA access, and a deadlock when loading HIP modules in parallel (#11223, #11169, #11765, #11719, #11651, #11566, #11727, #11265)warp_group_dot_wait with num_warps > 4 (#11151, #11245, #10988, #11691, #11056)clc=True launch option turns a one-tile-per-program kernel into a persistent Cluster Launch Control loop, including with warp_specialize=True (#11124, #11116)float8e5 and float8e4nv (#11340, #11647, #11699, #11924)redux.sync for 64-bit integers (#11299, #11302, #11537, #11788, #11813, #11816)cluster_barrier phase race, misplaced barriers around mbarrier initialization and TMEM deallocation, TMEM allocations overwritten while in use, and a dropped MMA operand initializer (#11334, #11498, #11525, #11636, #11887)gl.inline_asm, which runs inline assembly per thread and can take shared-memory and tensor-memory descriptors (#11623)cache_policy argument. On NVIDIA it controls L1 and L2 policies, fractional eviction, and L2 prefetch size (#11482, #12000)triton.compile accepts Gluon IR (.glir), so kernels can be compiled for a target without a GPU (#11628, #11806, #11198)atomic_load, atomic_store, atomic_poll, cumsum, and cumprod (#11591, #10723, #12007)scaled_upcast layout helpers, load_shared_fp4_repacked, ttgl.amd.hint.sched_barrier, and a cd_regclass option on mfma/mfma_scaled (#11470, #11657, #11149, #11932, #11792)cdna5 is the canonical Gluon module for gfx1250, and gfx1250 remains as an alias (#11183, #11213)NVMMASharedLayout or TensorMemoryLinearLayout arguments no longer reuse the wrong compiled body. AMD scaled_dot rejects scale layouts that gave wrong results (#11094, #11201, #11443)triton_kernels.matmul supports unscaled FP8 E4M3 with FP16 or BF16, and DequantScaleRoundingMode.ROUND_NEAREST for NVFP4 with E4M3 scales (#11652, #11668)RaggedTensorMetadata object to BlackwellActMXScaleLayout and a_ragged_metadata, or they raise AssertionError. With gather_indx, the scales must already be in gathered row order (#11928)triton_kernels matmul adds a clc constraint for SM100+ persistent kernels (#11176)topk supports non-power-of-two widths and widths above 32 (#11445)pl.allocate_event, pl.start_event, and pl.end_event to time asynchronous work across loops, functions, and warps (#11122, #11142)rocprofiler backend supports PC sampling and HIP graph profiling, and no longer deadlocks or aborts at exit (#11040, #10506, #10989, #10999, #11009, #11008, #11559, #11824)cp312-abi3 wheel, and source builds enable it with TRITON_STABLE_ABI (#12006, #10958)nanobind==2.10.2), which slightly lowers launch overhead (#10283)TRITON_EXT_ENABLED=1, the wheel includes Triton's C++ headers and TableGen definitions (#10847)do_bench* return units, bf16x3/bf16x6 tl.dot precision, and type promotion for tl.sum, division, and modulo (#10729, #10798, #10893, #10968, #10883)RELEASE.md documents the release-branch process and ties Triton feature releases to PyTorch minor releases (#11713, #10932)enable_fp_fusion. Set TRITON_DEFAULT_FP_FUSION=0 to adopt the 3.10 default now, or =1 to keep fusion and silence the warning. Kernels whose numerics depend on fusion should call tl.fma explicitly. The new TRITON_FORCE_DISABLE_FP_FUSION=1 disables implicit fusion even for kernels that request it (#11270, #12387, #12433)can_reorder=True on tl.cat, tl.reshape, and tl.ravel: The argument still works in 3.9 but warns once per process. Triton 3.10 will remove it, and these operations will always preserve element order. Drop can_reorder=True now, and use dim= with tl.cat (#11293, #12393, #12433)tl.softmax(keep_dims=...): keep_dims has no effect, because softmax always keeps the input shape. Passing a value warns; remove the argument (#11430)amd.gfx1250 module: amd.gfx1250 re-exports amd.cdna5. Import cdna5 in new code (#11183)tl.make_block_ptr raises NotImplementedError, tl.advance was removed, and tl.load/tl.store no longer accept boundary_check or padding_option. Switch to tensor descriptors (tl.make_tensor_descriptor or the host-side TensorDescriptor), which provide bounds checking and zero padding (#10833)cache_modifier and eviction_policy on tl.load/tl.store (and tensor.store), volatile on tl.load, and keep_dims/ieee_rounding on tl.softmax are now keyword-only. Passing them positionally fails to compile (#10833, #11418, #11430)/ and % return float32 when an operand is a float narrower than 32 bits, including FP8 with Python scalars. Cast the result to keep the old dtype (#10883, #11121)tl.minimum, tl.maximum, comparisons, tl.clamp, tl.fma, or binary math functions, so tl.maximum(x_fp16, 0.0) returns float16. tl.minimum, tl.maximum, tl.clamp, tl.cumsum, tl.cumprod, and tl.min/tl.max with return_indices=True no longer upcast bfloat16 to float32. To keep float32, cast the inputs or pass dtype=tl.float32 to cumsum/cumprod. tl.cumprod gains dtype and, like cumsum, now upcasts integers narrower than 32 bits by default (#11472, #11913, #11473)x.exp() and x.sqrt() apply the same dtype checks as the tl.* functions and fail to compile on fp16 inputs. Cast first, for example x.to(tl.float32).exp() (#11218)triton.autotune raises ValueError if key, reset_to_zero, or restore_value names something that is not a kernel argument. Indexing with more : slices than the tensor has dimensions fails to compile. tl.make_tensor_descriptor rejects constant strides that are not 16-byte aligned, which previously gave wrong results (#11602, #10429, #10931)tl.pointer_type(..., address_space=...) takes a string ("global", "descriptor", or "constant"), and integers no longer work. tl.const pointer arguments use "constant". In IR, use named forms such as !tt.ptr<T, "descriptor">; integer forms such as !tt.ptr<T, 0> are rejected (#11385, #11294)>=3.11,<3.16) and no longer ships Python 3.10 wheels (#12006)gfx101x, gfx103x). Use Triton 3.8 for those GPUs (#12042)TRITON_PTXAS_BLACKWELL_PATH. TRITON_PTXAS_PATH applies only below sm_90, so move any Hopper override (#12035)PYBIND11_SYSPATH is no longer used, and out-of-tree code built against Triton's binding headers must port to nanobind. AMD code generation ignores LLVM_SYSPATH; set TRITON_AMD_CODEGEN_PATH to a prebuilt libtriton_amd_codegen library (#10283, #11485)async_copy_global_to_shared, async_copy_global_to_shared_im2col, and async_copy_shared_to_global raise an error. Use async_load, async_load_im2col, and async_store (#11112)cluster.arrive, cluster.wait, and mbarrier.fence_init_release_cluster. Use cluster.barrier(); the compiler inserts the mbarrier-init fence (#10960)_reinterpret on shared- and tensor-memory descriptors is renamed to reinterpret (#11131)MBarrierLayout without subviews, apart from indexing along the buffering dimension (#11501)TRITON_PLUGIN_PATHS; use extend_with/extend_dialects_with. make install and the CMake install components were removed; build the wheel with TRITON_EXT_ENABLED=1. Plugin passes and custom-op builders get add_/create_ prefixes automatically (#10775, #11157, #10847, #10737, #10754)tt.load, tt.store, tt.descriptor_load, nvws.descriptor_load, the async and TMA global-to-local copies, and the AMD buffer and async-copy ops replaced their Note truncated.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Aggregate types: @triton.aggregate and @gluon.aggregate are now public APIs. Aggregates support inherited fields, default values, generated constructo
@triton.aggregate and @gluon.aggregate are now public APIs. Aggregates support inherited fields, default values, generated constructors, immutable instances, and aggregate_replace() (#10095, #9572)tl.topk: Added a descending argument. Set descending=False to return the smallest values (#9355)tl.dot_scaled (#10311)tl.fdiv(..., ieee_rounding=True) now emits IEEE-rounded division, and floating-point atomic_min now returns a value-typed result (#10074, #10485)argmin, argmax, minimum, maximum, and clamp now match compiled behavior more closely when inputs contain NaNs (#10298, #10333, #10699)tma.store_wait now accepts a read_only argument. The default remains True; use read_only=False when the store must reach global memory before a release operation (#10415, #10419)RemoveLayoutConversions (#10646, #11029)tl.expect_zero (#9337, #9455, #9714, #10112, #10330, #10461)cdna5 as an alias for the gfx1250 target (#11383)gfx906 (#9628)gfx120x) targets (#10390, #10185)num_threads, num_warps, and smid to tl.extra.hip, and added clz and popc to the HIP libdevice (#9604, #10651)add4, sub4, mul4, and fma4 (#11084)GenericLinearEncodingAttr for supported swizzled and non-injective layouts, including local load/store lowering and permutation-matrix inference (#9765, #10122, #10515)GluonJITFunction (#9863)async_load and async_store names; the previous async_copy_* names remain available as aliases (#10083)triton_kernels adds an FP64 matmul path (#9393, #10634)Layout.storage_shape() for querying a layout's physical storage shape without materializing a conversion (#10554)do_bench_proton and do_bench_cudagraph_proton to reduce benchmark bias from CPU launch overhead and an unflushed L2 cache (#10149)roctracer to rocprofiler-sdk (#9704)tcgen05 code generation. CI now runs the full test suite (#10344, #10345, #10027)setup.py to CMake, added progress and resume support, and avoided repeated CUDA tool downloads (#9458, #9722, #10418, #10540)kernel_unload_hook and treated incomplete cache entries as misses (#9444, #9542, #10411)map_elementwise, tl.expect_zero, out-of-range float-to-integer casts, control-flow scoping, load semantics, and compiler-hint semantics (#10695, #10692, #10678, #10119, #10356)!tt.tensordesc<tensor<...>> must use the new !tt.tensordesc<..., #layout> form (#9851, #9984)tt.make_tensor_ptr and tt.advance IR operations were removed. Out-of-tree MLIR using those operations must migrate to regular pointer operations or tensor descriptors (#9668)buffer.load() should omit that argument; use buffer.get_reg_layout() when the layout is needed separately (#9594)float2 module was replaced by native packed operations such as add2, sub2, mul2, and fma2. Call these operations directly from the Blackwell module instead of importing blackwell.float2 (#11002, #11084)triton.tools.triton_to_gluon_translater to triton.tools.triton_to_gluon_translator. Update imports to use the corrected spelling (#9570)tl.dot output type: When a non-FP32 accumulator is supplied and out_dtype is omitted, the output now defaults to the accumulator dtype instead of tl.float32. Set out_dtype=tl.float32 explicitly to preserve the previous behavior (#10353)TRITON_PREFER_TMEM_16x256_LAYOUT environment variable (#10664)This release includes contributions from engineers at:
Special thanks to all contributors who submitted bug reports, feature requests, and code improvements!
Triton 3.7.1 is a patch release on top of 3.7.0. It fixes the following 2 regressions and contains no new features or API changes.
Triton 3.7.1 is a patch release on top of 3.7.0. It fixes the following 2 regressions and contains no new features or API changes.
Deprecation Warning for `make_block_ptr`: Emitted a deprecation warning when make_block_ptr is used
tl.squeeze / tl.unsqueeze: Added tl.squeeze and tl.unsqueeze operations to the standard library (#8924)constexpr values from JIT-compiled code (#8785)get_int_attr for Out-of-Tree Walk: Added get_int_attr to Operation to support out-of-tree IR walks (#8892)preload: Added optional device argument to preload and guardrails for cross-target preload (#8951, #8952, #9234)tl.cat(can_reorder=False): Added a non-reordering variant of tl.cat with broadcast support (#9312, #9163)desc.shape for FP4 Padded: Fixed desc.shape values for fp4-padded tensor descriptors (#9012)constexpr_functions (#8876)must_use_result for Methods: Fixed must_use_result check for methods (#8902)_semantic Default to None: Defaulted _semantic parameter to None (#8909)make_tensor_descriptor Error Typo: Fixed typo in make_tensor_descriptor error message (#8912)tl.cat Determinism: Made tl.cat deterministic via permute+reshape+join, then reverted (#9312, #8854, #8878)make_block_ptr: Emitted a deprecation warning when make_block_ptr is used (#9667)inspect.signature for builtins, lazily computed tuple type names, avoided find_paths_if and inspect.getclosurevars, removed outdated catch_warnings blocks — all to reduce JIT overhead (#8843, #8844, #8846, #8845, #8881)tcgen05.mma + Multicast: Added multicast support for tcgen05.mma (#9071)tcgen05.mma Verifier & Errors: Throw a clear error instead of miscompiling very large tcgen05.mma along N (#8915)LowerAref, per-partition asyncOp storage, explicit captures to WarpSpecializePartitionsOp, skip InsertTmemAref when WS isn't used (#8978, #9007, #9023, #9133, #9212)RegionBranchInterface: Made WarpSpecializePartitionsOp implement RegionBranchInterface (#8799)aref.get Filtering: aref.get creation now filters results not in the scheduled loop (#9114)tt.scan Layout Fixes: Fixed tt.scan with broadcasted layouts and additional scan layout issues (#9185, #9189)assert or print are no longer pipelined (#9055)wgmma wait(0) to first use of the accumulator (#9021, #9179)ext; improved robustness of ext slice rematerialization (#9194, #9019)AxisInfo for add/sub; reland of unvisited-operand handling (#9297, #8758)async_cp: Pick better layouts for small async_cp (#9183)getConvertBackwardSlice (#8291)memdesc_slice in Membar; extended membar with third-party ops via traits; AMD-aware membarFilter (#8755, #8798, #9265)ReduceOp lowering, later reverted on release branch (#9192, #9214)kReg smem Padding: Separated additive kReg shared-memory padding contribution (#9286)tcgen05.mma + multicast support and Generalized Encodings: continued generalization of TMEM and shared-memory layouts (#9071)SwizzledShared Layout, uniform hint on ttg.warp_id, and CGAEncoding rename (#9286, #9073, #8850, #9040, #9125)dotCanBeProperlyAsync when wgmma is not yielded by the loop and an associated infinite loop (#9274, #9282)FuncOpToLLVM Refactors: Moved handleArgPtrDatatype to Utility.h; support for LLVM struct/array types in DITypeAttr (#9120, #9124)JITFunction in preload: Support JITFunction in preload (#8794)3.7 is heavy on gfx1250 (RDNA4) maturation, warp specialization on AMD, Tensor Data Movement (TDM), and a new warp-pipeline path.
ttg.warp_id and AMD Conversions (#8659)v_permlane16_swap: Enabled for convert_layout and reduceOp on GFX1250 (#8724)AMDWmmaEncodingAttr, scalar-pointer cluster-load avoidance (#9342, #9340, #9129)AMDWMMALayout Rank Consistency (#9127)i8xi8xi32 v3, missing f64.16x16x4.f64, and clamp operand on WMMA int intrinsic (#9267, #9271, #9291, #9359)AsyncWait in UpdateAsyncWaitCount (#9352)dim > 2 (#8994)AsyncCopy by default for gfx950 and gfx1250 — later reverted on release/3.7.x (#9445, #9087)v_perm for convert_layout (#9014)ReorderInstructions: Removed sinkSecondLoad, sinkDotConversion, and moveUpTranspose optimizations (#9119, #9139, #9204, #9229)ReorderInstructions with MoveUpPrologueLoads (#9328)UpdateAsyncWaitCount: Support single-block execute regions (#9126)OptimizeLDSUsage Removal (#8282)finite/isfinited, rint, clampf via v_med3: libdevice and codegen additions (#9097, #9166, #9256)BlockPingpong Improvements: Debug messages and dot-dominates-predecessors fix (#8804, #9027)kWidth mandatory for WMMA v3 (#8783)copysign Replacement: Replaced LLVM copysign intrinsic (#8789)CanonicalizePointers: Support MakeTensorDescOp in CanonicalizePointers (#9228)PartitionedSharedEncodingAttr: Introduced and reverted (#9314, #9367)scf.if Combining: Added PrepareIfCombining pass (#9253)SinkLayoutConversions Pass (#9168)addOccurrence for proper LLVM-option disabling; ScopedNoAliasAAWrapperPass in MIR swap pipeline (#8711, #9311, #9309)atomic_cas Fixes: Wrong struct index for atomic-CAS pattern, ignored sem/scope, and atomic-CAS for non-int types (#8867, #9042, #9116)uniformSum Crash: Fixed null uniformSum in CanonicalizePointers (#8991)RangeAnalysis tripCount: Fixed trip-count calculation (#9383, #9944)BlockPingpong for non-MFMA dot (#9618, #9948)CanonicalizePointers Different Bases (#9541, #9950)tcgen05 MMA on sm110 (Jetson Thor) (#9160)tcgen05.ld.red on sm103: Implemented in Gluon (#9151)NVIDIA::canSkipBarSync: Resurrected (#9246)AsyncTMACopyGlobalToLocalOp, tensor-descriptor support, fix for tma load, and driver support (#9202, #9225, #9303, #9305)tt.split/join in WS Data Partition: Hopper WS support for tt.split/tt.join (#456, #9147)w_scale Mask: Fixed Hopper mask (#8974)get_view(): Added get_view() for Gluon layouts (#9270)PaddedSharedLayouts (#9336)to_linear_layout: Allow TM layouts in to_linear_layout for printing (#8682)SharedLinearEncoding: Continued lowering generalization (carry-over from 3.6 with backend updates).requires_persistent (#9198)Tensor.clone: Briefly added clone for triton_kernels.tensor.Tensor, then reverted (#9178, #9208)mxfp4→bf16 Conversion via mul.bf16x2 (#8967)ex2.approx.ftz for swiglu (#8801, #8905, #9164)distributed.py / bench_utils.py: Extracted common code from bench_mlp.py and distributed.py (#8866)num_stages Adjustment: For bf16/fp16 × mxfp (#8773)reduce_forward Metadata: Improved performance (#9068)p_matmul Asserts & Fixes (#9376)symm_mem_pool by Argument (#9092, #9155)deactivate / get_data Overhead Reduction: Especially for CUDA-graph profiling; exposed get_data_msgpack (#9030)get_data API: Export profile data directly in Python (#8928)clear_data API: Remove pre-deactivation data (#8971)finalize Cleanup: Clean up context source after teardown (#9069)GlobalScratchAllocOp Deprecation: Deprecated Proton's own op in favor of TritonGPU's, with a custom backend (#8976)TRITON_ENABLE_HW_TRACE in CuptiProfiler (#9324)fresh_knobs Default Behavior (#9184)tl.dot BF16xN Nondeterminism (#8818)DOCKER_API_VERSION (release/3.7.x) (#10244)assert_close: Propagate err_msg to numpy (#9170)test_line_info_ir_source Flake Fix (#9161)CMAKE_LIBRARY_OUTPUT_DIRECTORY: Fixed build with empty directory (#8810)llvm_update_compile_flags Removal (#9167)LLVM_BUILD_SHARED_LIBS Canonicalization (#8933)TRITON_EXT_ENABLED for Wheels (#9935, #9959)nvidia-toolchain-version.json Update.TRITON_DEFAULT_BACKEND: Control driver.active via this env var (#9144)TRITON_PTXAS_BLACKWELL_PATH: Allow override of ptxas-blackwell binary (#8945)topk in Plugin Example: Increment index in plugin example (#9315)link.py (#9084)AxisInfo (#9266)topk Operation: Added to language documentation (#9345)warp_specialize Docs: Updated gl.warp_specialize docs (#8553)LinearLayout Output Matrix Comment: Doc fix (#9243)triton_kernels matmul refactor (BC-breaking): The matrix-multiplication refactor introduces a backwards-incompatible API surface. Downstream users of triton_kernels.matmul_* should review call sites (#8765)tcgen05.cp Lowering Generalization & tcgen05.mma Encoding Acceptance: Continued from 3.6, with new verifier behavior and stricter encoding checks.GlobalScratchAllocOp Deprecated: Replaced with TritonGPU's GlobalScratchAllocOp + custom backend. Out-of-tree consumers must migrate (#8976)make_block_ptr Deprecated: A deprecation warning is now emitted; users should migrate to tensor descriptors (#9667)This release includes contributions from engineers at:
Special thanks to all contributors who submitted bug reports, feature requests, and code improvements!
Un-deprecated min/max (#8734): Un-deprecated min/max on scalar tensors
tl.trans and tl.dot operationsdot_scaled operationsdot_scaled Handling (#8658): Fixed missing handling for None acc in dot_scaledtl.cdiv (#8669): Optimized tl.cdiv for common case of 32-bit divisorsast.Numtcgen05.cp Lowering (#8225): Implemented generic lowering for tcgen05.cpatomic_cas operationtt.LoadOptcgen05.ld/st genericallytcgen05.mma to accept SharedLinearEncodingAttrgl.warp_specialize API for better usabilitynum_ctas in Gluonttgl.get_num_warps metafunctiongather and its layout teststl functions into gluon and expose catasync_copy to Gluon for gfx1250x in triton_kernelssplit_k on m * nBitmatrixMetadata and RaggedTensorMetadata; deprecated triton_kernels.routingy_indx and uniform distributionreduction_n=2 to bench_mlp.pysplit_k > 1 with fused scatterHopperValue layoutgl.warp_specialize APIThis release includes contributions from engineers at:
Special thanks to all contributors who submitted bug reports, feature requests, and code improvements!
This release is meant to fix the following issue:
This release is meant to fix the following issue:
Fix sm103 (GB300) support broken by Triton 3.5.0 release (https://github.com/triton-lang/triton/pull/8045)
Warp Specialization Enhancements (#8005): Made warp specialization require at least 4 warps with proper error messaging to prevent compiler crashes
mask parameter to tl.device_assert for easier debugging with masked operationsconstexpr_function to support cache invalidation and capability checkstl.float16 and other FP types@builtin or @core.extern functions modify their argumentscore.extern_elementwiseConstant{Int|Float}Op type and value orderTargetLibraryInfoImplconvert_layout that:
ldmatrix/stmatrix and transpose versionsselect and shuffle instructionsPaddedSharedEncoding with non-default orderlowerLdSt pathassert_trivial flag for performance validationnumel and nbytes properties (#7507)map_elementwise (#7564)tensor.sum (#7617)splat returning auto encoding (#7490)make dev-install-llvm in READMETRITON_DEBUG at import time (#7767)This release includes contributions from engineers at:
Special thanks to all contributors who submitted bug reports, feature requests, and code improvements!
One column per quarter.