1 Release published by 1 person

• v0.3.1

  published Jun 26, 2025

37 Pull requests merged by 10 people

• Use onnx_ir common passes

  #2420 merged Jun 26, 2025

• Update VERSION to 0.3.2

  #2421 merged Jun 26, 2025

• Create torch_2_8 apis

  #2419 merged Jun 26, 2025

• Bump onnx ir requirement to 0.1.3

  #2418 merged Jun 26, 2025

• Add sphinx inter link to onnx_ir

  #2415 merged Jun 25, 2025

• Update VERSION to 0.3.1

  #2414 merged Jun 24, 2025

• Add phi2/phi4 test cases for mha/gqa fusion

  #2409 merged Jun 23, 2025

• Fix proto handling in version converter

  #2411 merged Jun 23, 2025

• Support dynamic shapes for aten_unfold

  #2407 merged Jun 20, 2025

• [optimizer] Replace value.nbytes with value.size

  #2399 merged Jun 19, 2025

• Move _c_api_utils.py to version_converter package

  #2401 merged Jun 19, 2025

• Fix fusion ordering for partial rotary embedding

  #2402 merged Jun 19, 2025

• Move gemm_to_matmul_add rule to ort fusion rules

  #2398 merged Jun 19, 2025

• Updates to the rewriter tutorial

  #2397 merged Jun 19, 2025

• [rewriter] Decouple llama rule sets and make API explicit

  #2388 merged Jun 18, 2025

• [torchlib] Implement aten__upsample_bicubic2d_aa and aten__upsample_bilinear2d_aa functions

  #2383 merged Jun 17, 2025

• fix: check for rank of bias in bias-gelu fusion🐛

  #2393 merged Jun 16, 2025

• Fixes to MHA fusion

  #2380 merged Jun 14, 2025

• [Rewriter]: Add ∘ MatMul -> Gemm

  #2356 merged Jun 14, 2025

• fix: pattern match gelu from contrib and onnx ops🐛

  #2364 merged Jun 14, 2025

• [torchlib] Unregister aten::max.other

  #2377 merged Jun 13, 2025

• Fusion extensions to improve GQA fusion

  #2374 merged Jun 13, 2025

• [CI] Fix execnet.gateway_base.DumpError: can't serialize <class 'numpy.int64'>

  #2379 merged Jun 13, 2025

• chore(deps): bump onnx-weekly from 1.19.0.dev20250419 to 1.19.0.dev20250602 in /requirements/ci

  #2376 merged Jun 12, 2025

• Add missing converter for _local_scalar_dense

  #2367 merged Jun 11, 2025

• Re-enable fused matmul rules

  #2370 merged Jun 10, 2025

• Cast-cast elimination

  #2368 merged Jun 8, 2025

• Update autocast.py to fix attribute creation error

  #2365 merged Jun 6, 2025

• Fix fused matmul check/rewrite functions

  #2331 merged Jun 6, 2025

• docs: cleanup documentation for function-based rewrites📄

  #2359 merged Jun 5, 2025

• Always fold the Transpose node in the constant folder

  #2355 merged Jun 5, 2025

• Test SDPA fusion via MHA

  #2366 merged Jun 5, 2025

• Enable CSE in optimizer

  #2361 merged Jun 4, 2025

• Require onnx-ir 0.1.1

  #2360 merged Jun 3, 2025

• SDPA fusion cleanup

  #2352 merged Jun 3, 2025

• Fix test names for pytest

  #2358 merged Jun 2, 2025

• Support common subexpression elimination pass (CSE)

  #2304 merged May 30, 2025

20 Pull requests opened by 9 people

• [Rewriter] Add optimizer to fold Pad operators into Conv

  #2363 opened Jun 4, 2025

• [DRAFT] Merge metadata props when fusing

  #2375 opened Jun 12, 2025

• Fix rewrite rules to handle attribute references (RefAttr) properly

  #2381 opened Jun 14, 2025

• [torchlib] Implement torch.ops.prims.broadcast_in_dim.default

  #2382 opened Jun 14, 2025

• [torchlib] Fix aten.embedding_bag.padding_idx to support None values

  #2384 opened Jun 14, 2025

• Fix padding_idx=None handling in aten_embedding_bag_padding_idx

  #2385 opened Jun 14, 2025

• Fix ast.Expression deprecation warning by using proper constructor

  #2386 opened Jun 14, 2025

• [torchlib] Implement quantize_per_channel and dequantize_per_channel

  #2390 opened Jun 14, 2025

• Clean up rewriter code: improve efficiency, finish TODOs, and enhance documentation

  #2392 opened Jun 14, 2025

• Implement E-graph based pattern matching for efficient and robust rewriting

  #2395 opened Jun 16, 2025

• fix: handling of default attrs in SimplifiedLayerNormalization + LayerNormalization🐛

  #2396 opened Jun 17, 2025

• Implement aten::feature_dropout

  #2404 opened Jun 19, 2025

• [docs] Document rewriter pattern options

  #2406 opened Jun 19, 2025

• Support gqa in aten spda

  #2408 opened Jun 20, 2025

• [Rewriter]: fuse successive Relu/Clip nodes

  #2410 opened Jun 22, 2025

• Add support for onnx fusions

  #2412 opened Jun 23, 2025

• chore(deps): bump ruff from 0.11.4 to 0.12.0 in /requirements/lintrunner

  #2413 opened Jun 24, 2025

• [pass] Enable DeduplicateInitializersPass

  #2416 opened Jun 24, 2025

• Eliminate unnecessary ScatterND

  #2422 opened Jun 28, 2025

• Add stacktrace in metadata_props for rewriter created nodes

  #2423 opened Jun 28, 2025

27 Issues closed by 4 people

• ATen ops list

  #258 closed Jun 23, 2025

• [torchlib] Support dynamic shapes for aten_unfold

  #2309 closed Jun 20, 2025

• Move _c_api_utils to version_converter

  #2400 closed Jun 19, 2025

• [rewriter] Decouple llama rule sets

  #2128 closed Jun 18, 2025

• [Feature request] Implement logging to improve pattern mismatching debug

  #1366 closed Jun 17, 2025

• Potential improvement to masked_scatter op

  #2113 closed Jun 14, 2025

• Debug passes

  #2197 closed Jun 14, 2025

• [torchlib] `aten.min.other`

  #2146 closed Jun 14, 2025

• Domain of `Gelu` in `bias_gelu_rules` set

  #2362 closed Jun 14, 2025

• [Rewriter] Why doesn't it match?

  #1491 closed Jun 6, 2025

• [IR] More serialization support

  #2290 closed Jun 6, 2025

• onnxscript/ir code freeze

  #2351 closed Jun 6, 2025

• [IR] Implement graph editing APIs as convenience functions

  #2294 closed Jun 3, 2025

• IR Roadmap

  #1324 closed Jun 2, 2025

• [IR] Journaling system

  #1717 closed Jun 2, 2025

• [IR] Create a utility for merging models

  #1700 closed Jun 2, 2025

• [IR] Make the schemas module public

  #1929 closed Jun 2, 2025

• [IR] Pass manager follow up

  #1551 closed Jun 2, 2025

• [IR] Handle nodes with defined output names

  #1571 closed Jun 2, 2025

• [IR] Fix Value repr

  #1572 closed Jun 2, 2025

• [IR] Graph composition

  #2311 closed Jun 2, 2025

• [IR] Integrate symbolic shape inference

  #1383 closed Jun 2, 2025

• [IR] Support sparse tensor

  #1428 closed Jun 2, 2025

• [IR] Create a comprehensive checker

  #1431 closed Jun 2, 2025

• [IR] Common passes

  #1473 closed May 31, 2025

• [IR] Create a name-fix pass

  #1432 closed May 31, 2025

• CSE pass

  #2105 closed May 30, 2025

9 Issues opened by 4 people

• Attention fusion (SDPA/MHA) broken for BART decoder with/wo past🐛

  #2424 opened Jun 29, 2025

• [docs] Document rewriter options

  #2405 opened Jun 19, 2025

• Implement aten::feature_dropout

  #2403 opened Jun 19, 2025

• Explore the idea of pattern matching with egraph (copilot)

  #2394 opened Jun 16, 2025

• Clean up rewriter code

  #2391 opened Jun 14, 2025

• [torchlib] Implement quantize_per_channel and dequantize_per_channel

  #2389 opened Jun 14, 2025

• Handling of default attrs of `LayerNormalization` in `SkipLayerNormFusion`

  #2378 opened Jun 13, 2025

• Support matching variadic inputs

  #2373 opened Jun 11, 2025

• [torchlib] slice_scatter does not support start and end to be None

  #2372 opened Jun 11, 2025

10 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Add New Metadata and Pattern Features

  #2271 commented on Jun 2, 2025 • 7 new comments

• [torchlib] Make index_put dynamic

  #2263 commented on Jun 26, 2025 • 2 new comments

• Ensure rewrite rules handle attribute-refs

  #1442 commented on Jun 14, 2025 • 0 new comments

• [torchlib] Implement `torch.ops.prims.broadcast_in_dim.default`

  #2218 commented on Jun 14, 2025 • 0 new comments

• [torchlib] Implement various upsample functions

  #1159 commented on Jun 14, 2025 • 0 new comments

• [torchlib] Implement `aten.embedding_bag.padding_idx`

  #2219 commented on Jun 14, 2025 • 0 new comments

• [torchlib] `torch.ops.aten.embedding_bag.padding_idx` may see `padding_idx=None`

  #2145 commented on Jun 14, 2025 • 0 new comments

• Converter call to `ast.Expression` is invalid

  #2114 commented on Jun 14, 2025 • 0 new comments

• [torchlib] Improve/reimplement index_put

  #1749 commented on Jun 23, 2025 • 0 new comments

• Optimizer Suggestions for Common Graph Fusions

  #2301 commented on Jun 24, 2025 • 0 new comments