Skip to content

Release v3.14.0

Latest

Choose a tag to compare

@devtbi devtbi released this 16 Nov 10:15
5332c68

IREE v3.14.0 Release Notes

Release Candidate: iree-3.14.0rc20261115
PRs: 109 since v3.13.0
VMFB Bytecode Version: 17.0 (unchanged from v3.13.0)
HAL Module Version: 0.7 (unchanged from v3.13.0)

Breaking changes

  • iree_linalg_ext.sort takes its inputs as ins(). The op used to sort its outs() operands in place. Inputs now go in ins() and destination tensors in outs(), like other LinalgExt ops. Update hand-written IR and custom lowerings that create the op. #24658
  • iree_linalg_ext.sort is no longer stable by default. The O(n²) bubble sort lowering was replaced with faster algorithms. Without the new is_stable(true) attribute, elements that the comparator does not order end up in unspecified order. The Torch input conversion sets is_stable(true), and the StableHLO conversion forwards is_stable from stablehlo.sort. #25262
  • Tracy is updated to 0.15. Use a Tracy 0.15 profiler, or the iree-tracy-capture from this release, to capture traces. IREE_TRACE_MESSAGE levels map to Tracy message severities, and the exported TracyClient CMake target now enables TRACY_ENABLE. #25241
  • The iree-global-opt-convert-batch-matmul-to-matmul pass was removed. Broadcasts now fold into the indexing maps of batch_matmul (iree-global-opt-fold-broadcast-into-batch-matmul), and dispatch creation collapses the result into a plain matmul. #25114

Notable changes

Compiler

  • Quantization:
    • Quantized convolutions convert to integer math. im2col is now one generic rewrite for any convolution (grouped, dilated, 1-D/3-D, dynamic and integer), and it runs after the QDQ conversion. #25011, #25099
    • Constant 4-bit and 2-bit quantized weights are packed into i4/i2 storage. #25145
    • Broadcast and collapsed quantized weights reach the integer QDQ rewrite. #25114, #25197, #25086
    • Torch: torchao quantize_affine/dequantize_affine convert to LinalgExt, and float zero points of PT2E and torchao ops that are zero or integral are converted. #25104, #25112, #25136
  • Data tiling:
    • Experimental early data tiling during global optimization with --iree-opt-experimental-early-data-tiling. Layouts propagate through the surrounding tensor graph. #24969, #24968
    • Named contractions with user indexing maps are collapsed like generic ones, so data tiling no longer depends on the named or generic form. #25113
  • New LinalgExt ops group_matmul and group_mmt4d, the counterpart of PyTorch's grouped_mm, for Mixture-of-Experts models. #24909
  • CPU:
    • RISC-V V: f16 (vfmacc, vfwmacc, Zvfhmin) and bf16 (vfwmaccbf16) inner_tiled MMA. #25000, #25120
    • Packed f32 layouts for Wasm SIMD128. #25199
    • math.roundeven stays a native instruction on AArch64, x86 with SSE4.1 and riscv64 with D, instead of a software sequence. #25219
    • Invariant reductions are no longer fused into reduction tile loops, so tiled softmax no longer recomputes the row maximum for each tile. #25201
  • Reproducible .vmfb files: LLVMCPU no longer writes uninitialized source location lines, and multi-device modules get deterministic allocator-select global names. #25239, #25243
  • Compiler plugins can handle StableHLO composites and custom calls before input conversion, and C and Python API callers can activate plugins through the session. #25067, #25226
  • Torch: ONNX reductions with a runtime axes tensor compile. #25041
  • WebGPU: fixes for client API detection, wide elementwise vectors, read-only bindings and executable export names. #25051, #25053, #25215, #25195
  • Correctness fixes:
    • Attention normalization used the wrong operand order, which gave wrong results for a division. #25170
    • GPU integer reductions used wrong identities for narrow and unsigned types. #25176
    • Mutable globals that share a splat initializer are no longer updated through the same resource. #25030
    • VM branches no longer leak a reference, and merged scf.index_switch ops keep their selectors. #24781, #25037
    • Executable canonicalization and CSE in the Flow pipeline actually run. #25160

Runtime

  • Python:
    • from_dlpack_capsule takes memory_type and memory_access. DLPack imports validate shapes and offsets and keep the producer alive, and buffers that legacy DLPack cannot describe, such as read-only ones, are not exported. #25227
    • HAL enums support __index__, so they work as integer arguments with nanobind 2.15. #25163, #25223
    • Memory leaks on error paths are fixed. #25014
  • VM: the bytecode verifier checks register ordinals against the function's declared register counts. #25088
  • HAL: drivers handle a device group being reassigned consistently. #25174
  • WebGPU and WebAssembly: wasm64 support in the WebGPU bridge, and fixes for resource lifetimes in command buffers, staging buffer mappings, copies within one buffer and inline completion draining. #25054, #25075, #25070, #25068, #25069, #25193, #25245

Developer tools

  • Release candidates have build provenance attestations. Verify a file with gh attestation verify <file> --repo iree-org/iree. Stable releases keep them for the iree-dist tarballs, which are shipped unchanged with their .sigstore.json bundles. The iree-base-compiler and iree-base-runtime wheels change when they are promoted to the final version, so they have no attestation yet; attesting them is in progress in #25204. #25202, #25220
  • Bazel: HAL drivers from outside IREE (like IREE_EXTERNAL_HAL_DRIVERS in CMake), Android detection through @platforms//os:android, and include paths when IREE is an external repository. #25238, #25107, #25101
  • CMake: builds against an installed LLVM/MLIR SDK honor tool overrides and no longer need FileCheck with IREE_BUILD_TESTS=OFF. #25167, #25146
  • AMDGPU device bitcode comes from the official ROCm distribution, pinned in build_tools/third_party/amdgpu_device_libs/rocm_package.json. #25133

New contributors

Full changelog

Commit history: v3.13.0...v3.14.0

Release notes were AI-generated and edited.