v3.7.2

What's Changed

New Hardware Support

• Add NVL archs by @anmyachev in #6995

New Features

• Add Python 3.15 support by @anmyachev in #7125
• Enable FpSanitizer pass for Intel XPU backend by @etiotto in #6469
• Add DescriptorPrefetchOp to software pipelines by @anthonycanino in #6344
• Support rank > 2 tensors in descriptor load matmul loop pipelining by @etiotto in #6423
• Support rank > 2 tensors in 2D block I/O lowering by @etiotto in #6418
• Support PAD_NAN for tensor descriptor block IO loads by @whitneywhtsang in #6548
• Enable has_bfloat16_arithmetic by @whitneywhtsang in #6546
• Enable vectorized gather/scatter for descriptor load/store by @etiotto in #6504
• Lower block pointers to tensor descriptors in Python frontend by @whitneywhtsang in #6666
• FlashAttention backward pass for SYCL-TLA provider by @mieshkiwrk in #6481
• Use SPV_INTEL_predicated_io extension by @whitneywhtsang in #6816
• Re-enable sanitize-overflow by @mieshkiwrk in #6582
• Switch to static launcher by @anmyachev in #6650

Performance Improvements

• Add HoistLayoutConversions pass with liveness-based GRF budget heuristic by @etiotto in #6460
• Elementwise prefetch: software pipelining for non-dot loads (Xe3P+) by @etiotto in #6712
• Use LoopAwareCSE in make_ttgir to reduce register pressure by @etiotto in #6648
• Widen elementwise store vectorization to 256 bits (Xe3P) by @etiotto in #6727
• Widen tt.load/store vectorization; introduce WidenLoadStoreEncoding pass by @etiotto in #6787
• Rewrite contiguous descriptor_gather to 2D block loads by @etiotto in #6584
• Rewrite contiguous descriptor_scatter to descriptor_store by @etiotto in #6604
• Support 1D strided stores via reshape to 2D block IO by @etiotto in #6531
• Support 1D strided load to 2D block load reshape by @etiotto in #6747
• Add iterative backward rematerialization to RemoveLayoutConversions by @etiotto in #6380
• Support rank > 2 descriptor loads in transpose fusion (OptimizeDotOperands) by @etiotto in #6421
• Use twisted grid order in FlashAttention kernel for better performance by @chengjunlu in #6750
• Replace ballot-based histogram algorithm with atomic-per-element by @etiotto in #6621
• Cache results from get_current_target call by @anmyachev in #6585

Bug Fixes

• Fix transpose load using wrong base width and height by @chengjunlu in #6433
• Fix crash on broadcast mask in RemoveMasks InvariantMaskValidator by @wdziurdz in #6477
• Fix rank-reducing tensor descriptor load/store lowering by @whitneywhtsang in #6490
• Fix masked div/rem undefined behavior exploitation by LLVM O3 by @etiotto in #6675
• Fix dominance violation in RemoveLayoutConversions hoistConvertOnTopOfExtOrBroadcast by @wdziurdz in #6681
• Fix dead hoistConvertDotOperand for DPAS encoding by @etiotto in #6688
• Fix backward remat conflict detection in RemoveLayoutConversions by @etiotto in #6701
• Fix adjust base width bug in block IO by @chengjunlu in #6734
• Fix float32 reduction accuracy regression in ReduceOpToLLVM by @whitneywhtsang in #6919
• Fix std::optional<bool> misuse in IntegerRangeAnalysis getTripCount causing incorrect cmpi folding by @whitneywhtsang in #6925
• Workaround reduce accuracy issue in TIMM model with upstream ReduceOp combine order by @chengjunlu in #6667
• Fix vLLM startup segfault: initialize Level Zero early by @lslusarczyk in #6767
• Guard sqrt_cr/divide_cr builtins behind SPV_INTEL_rounded_divide_sqrt capability by @whitneywhtsang in #6777
• Trigger 256-GRF retry when zebin is degenerate (LTS2 IGC silent failure) by @dev-tomek in #6958
• Don't raise an error when .ze_info section is missing from zebin by @dev-tomek in #6921
• Skip block IO when surface pitch exceeds 24-bit hardware limit by @dev-tomek in #7032
• Fall back instead of asserting on column_major descriptor load/store that can't transpose by @dev-tomek in #7202
• Widen sub-byte bit ops forbidden by SPV_INTEL_int4 by @wdziurdz in #6927
• Use -cl-fp32-correctly-rounded-divide-sqrt for LTS drivers by @anmyachev in #7006

New Contributors

• @Roshan1299 made their first contribution in #6471
• @exolyr made their first contribution in #6560
• @lslusarczyk made their first contribution in #6767

Requirements:

• Python 3.10, 3.11, 3.12, 3.13, 3.14, 3.15.
• Agama rolling from KOBUK (1260) or LTS2 release.
• Intel® Deep Learning Essentials 2026.0

Note

triton-3.7.2+xpu-py3-none-any.whl is a temporary metadata-only compatibility shim. It does not contain any Triton implementation and exists solely to improve pip dependency resolution in environments where triton and triton-xpu may be installed simultaneously.

Full Changelog: v3.7.1...v3.7.2