ByteIR v1.9.3.0

What's Changed

• [torch-frontend] support lowering to byteir.resize, byteir.l2_norm
• [byteir] support byre serialization on cuda byre mlir
• [byteir] support decomposition of byteir.softmax, byteir.arg_max, byteir.addn
• [byteir] support merge two modules by mapping
• [brt] add BRTBackend
• [*] reduce wheel package size

New Contributors

• @yyp0 made their first contribution in #428

Full Changelog: v1.9.1.0...v1.9.3.0

ByteIR v1.9.1.0

What's Changed

• Enhance gemm codegen by @Xinyu302
• Enhance byteir_backend of torch.compile
• Support optimize contiguous slice as alias like op
• Refactor cat pipeline and executor
• Support 1xTF32 on GPU backend

New Contributors

• @JackyChenYi made their first contribution in #308
• @jianwenyyy made their first contribution in #358

Full Changelog: v1.8.4.0...v1.9.1.0

ByteIR 1.8.4.0

What's Changed

• More codegen fusion on mhlo, like slice-concat fusion and insert_slice fusion.
• More graph optimization.
• Enhance reduction codegen.
• Refactor byteir.compile.
• Add dynamo backend in torch_frontend.byteir_backend.
• E2E on NCCL backend.

Full Changelog: v1.8.1.0...v1.8.4.0

ByteIR 1.8.1.0

• Design ccl dialect to be lowered from torch communication ops.
• Release Torch Frontend v1.1.1, which split torchscript and dynamo/fx support.
• Support emitting stablehlo bytecode from frontends
• Compiler support accepting stablehlo bytecode.

ByteIR 1.7.1.0

• Update Torch-Frontend to 1.0.3
• Update ByteIR Compiler & Runtime to 1.7.1.0
• Support Serialization on Byre Dialect
• Init for NCCL Backend

ByteIR 1.4.0

[Release] Official Release ByteIR 1.4.0 (#71)

* [Sync] internal a6ef5f00...22d6dee6
* [AIT] Disabled hw info log, added error handling
* [CAT] Adjusted layout support
* [compiler/doc] Added codegen doc
* [frontend/torch] Added demo code, added more fx pattern for llm, fixed einsum, updated to 23b72244b1e1eaa0511cece8535b32810c1d0d7a
* [GPU] Added reduction codegen for PTX
* [Mhlo] Fixed transpose movedown, Added canonicalizer for gather with iota
* [Release] released 1.4.0 package
* [Runtime] Supported non-splat value in FillOp, added dropout support for flashV2
* [Util] Fixed bugs