Repository navigation
What's Changed
- [TorchToLinalg] Use the end pad for the avg_pool window-end clamp by @ziereis in #4756
- [TorchToTosa] Fix avg_pool1d dilation decode by @sahas3 in #4797
- [FxImporter] Fix shared and nested get_attr imports by @catcor01 in #4770
- [FxImporter] Import dense tensor literals from byte buffers by @catcor01 in #4787
- [TorchToLinalg] Choose the zero point extension from its Torch dtype by @ziereis in #4754
- [TorchOnnxToTorch] Apply Gemm alpha without C by @sahas3 in #4798
- release: checkout before downloading artifacts by @j2kun in #4800
- [OnnxImporter] Normalize ai.onnx to the default domain by @GodHforever in #4775
- [TorchOnnxToTorch] QuantizeLinear / DequantizeLinear emit quantized_decomposed ops by @juanigp in #4782
- [Torch] Fold scalar floating-point cumsum by @gokcan-karakus in #4805
- [TorchToTosa] Do not reshape FP8 linear by @Christoffer-JL in #4812
- [TorchToLinalg] Add first-class PT2E per-token quantize/dequantize/choose_qparams ops by @guikorol in #4765
- [Stablehlo] Fix dynamic-shape crash in AtenClampTensorOp lowering by @alexxony in #4789
- [Torch] Fix scalar infinite-order vector norm decomposition by @gokcan-karakus in #4807
- [TorchToTosa] Preserve int64 lowering by @davsva01 in #4783
- [TorchToTosa] Lower static 1D replication padding by @gokcan-karakus in #4806
- [Torch] Decompose scalar gather through rank-one tensors by @gokcan-karakus in #4808
- [Torch] Decompose scalar max.dim by @gokcan-karakus in #4809
- [RefBackend] Pass contiguous inputs with identity-layout strides by @felixtensor in #4815
- [TorchToLinalg] Fix adaptive pooling non-finite option by @Lallapallooza in #4821
- [LLVM] Integrate LLVM at 2fc577681ca1c896597252941593da9b61e0960b by @jackv1ck in #4802
- [TorchToStablehlo] Add lowering for aten.diagonal by @shashank-mcw in #4829
- [OnnxToTorch] add lowering for com.microsoft.MatMulIntegerToFloat by @Manewing in #4830
- [OnnxToTorch] add lowering for com.microsoft.FastGelu by @Manewing in #4835
- [ONNXToTorch] Fix ORT e2e tests for MatMulIntegerToFloat by @guikorol in #4844
- [Python] Make the torch-mlir package prefix configurable by @sgjzfzzf in #4781
- [TorchToArith] Add lowering for aten.not by @Jeoffrey-Alex in #4851
- [TorchOnnxToTorch] Migration to PT2E for QLinearAdd, QLinearMul, QLinearLeakyRelu, QLinearSigmoid, QLinearConcat, QLinearGlobalAveragePoo, QLinearAveragePool by @guikorol in #4833
- [TMTensor] Bufferize
insoperands at the op that reads them by @nmalimban in #4820 - [OnnxToTorch] Support up to four outputs in lowering of SkipSimplifiedLayerNormalization op by @riestmo-nxp in #4846
- [TorchToLinalg] Avoid over-padding pooling inputs when ceil_mode=true by @sahas3 in #4790
- [TorchToLinalg] Add first-class torchao affine quantization support by @roberto-laudani in #4843
- [TorchToLinalg] Compute quantize/dequantize payloads in f32 by @ziereis in #4836
- [TorchToTosa] Support avg_pool2d divisor_override by @gokcan-karakus in #4824
- [TorchToTosa] Update LIT tests to new TOSA by @guikorol in #4857
- [TorchOnnxToTorch] Lower com.microsoft.MatMulNBits to PT2E QDQ by @ziereis in #4837
- [TorchToTosa] Support max_pool2d with indices by @gokcan-karakus in #4823
- [LLVM] Integrate LLVM at 0223b87de6dd35a12844bc9f92575e3217626dd5 by @jackv1ck in #4859
New Contributors
- @GodHforever made their first contribution in #4775
- @guikorol made their first contribution in #4765
- @alexxony made their first contribution in #4789
- @felixtensor made their first contribution in #4815
- @shashank-mcw made their first contribution in #4829
- @Jeoffrey-Alex made their first contribution in #4851
Full Changelog: v20261001...v20261101