rocm-10.1

What's Changed

• Fuse expert SilU heads (MoE) into batched GEMM via fuse_horizontal by @TedThemistokleous in #5087
• Refactor rnn ops to op builders by @pfultz2 in #4606
• Add a migraphx code review skill by @pfultz2 in #5111
• dyn_slice operator for symbolics and data-dependent operators by @CharlieL7 in #5112
• Jenkins: include tools/requirements-py.txt in docker image hash by @causten in #5158
• 8/11 bump rocmlir by @causten in #5129
• [ROCM-26610] Replace tags with SHAs by @eddieliao in #5166
• Bump rocm-docs-core from 1.38.0 to 1.40.0 in /docs/sphinx by @dependabot[bot] in #5176
• add missing MLIR wrapper stubs for MLIR-disabled builds by @nkotakal0107 in #5154
• windows-gpu: free disk space before ROCm install by @causten in #5185
• Fix dynamic-shape parse failure in Softplus and Softsign by @zhihuidu-amd in #5136
• Release protobuf state on ONNX DLL unload by @ivarusic-amd in #5157
• Rewrite broadcast->layout by @pfultz2 in #5141
• Set copilot code review level to xhigh by @pfultz2 in #5191
• Duplicate Lambda Parameter from Repeated Input by @causten in #5130
• Horizontally fuse unsqueeze by @pfultz2 in #5180
• Jenkins: create ccache dirs on host before container launch to fix root ownership by @causten in #5188
• Dont fuse multi-output literal by @pfultz2 in #5178
• [AIRADSW-850] Fix simplify_algebra conv fusion regression after rewrite_convolution by @urpetkov-amd in #5167
• Fuse concat of reshapes whose inputs differ along the concat axis by @pfultz2 in #5181
• Enhance symbolic onnx parsing via sym_eval by @shivadbhavsar in #5148
• [AIMIGRAPHX-1242] Use ROCm Extras prefix for TheRock packages by @kentqian in #5134
• CI: update migraphx-reports repo location to AMD-ROCm-Internal by @causten in #5159
• Add promote_storage_type pass to compute storage-only types in float by @pfultz2 in #5138
• Problem cache follow-on: pluggable backends, layered cache paths, compile-options API, and offline tooling by @danieyan-amd in #5117
• CI: report.yaml Generate Performance Report workflow by @causten in #5199
• CI: remove cross-org workflows; harden performance.yaml by @causten in #5209
• Layout convolution weights as yxck by @pfultz2 in #5149
• Add rewrite broadcasts for fuse_reduce by @pfultz2 in #5146
• Emit appropriate msg when no mxr files are dumped by @ahsan-ca in #5164
• Changelog fixes by @CharlieL7 in #5200
• Prevent AI making short unit test helper functions by @CharlieL7 in #5206
• Onnxruntime Weekly Sync 2026-08-28 by @github-actions[bot] in #5214
• AIMIGRAPHX-317 g+g heuristic added to apply by @bdevorem in #4439
• Install clangd by @pfultz2 in #5163
• Add executable path flags and --all option to tools/format.py by @pfultz2 in #5218
• Format by @pfultz2 in #5222
• Bump rocm-docs-core from 1.40.0 to 1.40.2 in /docs/sphinx by @dependabot[bot] in #5220
• Bump CI to ROCm 10.0 by @causten in #5216
• Fix ROCm dependency discovery after cget/rbuild toolchain setup by @kentqian in #5224
• Split pointwise ops over concat inputs to enable fused_concat fusion by @pfultz2 in #5182
• Format python files by @pfultz2 in #5228
• Bump version of migraphx by @causten in #5225
• Insert match::opaque into deep matchers to fix large symbol names on windows and debug gcc builds by @pfultz2 in #5124
• Update MIGraphX EP shim to default to the AMD pci vendor ID by @TedThemistokleous in #5221
• Simplify concat same broadcast by @pfultz2 in #5179
• Leaky relu using max by @pfultz2 in #5005
• Skip literals that are not scalars or iotas when fusing attention by @ahsan-ca in #5211
• Remove concat->reshapes->slice by @pfultz2 in #5183
• Extend reduce vectorization ot use up 8 vector lanes. by @pfultz2 in #5230
• Fix seg fault when splitting non-fusible MLIR modules by @justinrosner in #5187
• Add block batch reduce algorithm by @pfultz2 in #5177
• Parse TopK with dyn slice & ONNX node UID expansion by @CharlieL7 in #5150
• Update CHANGELOG.md with versioned 2.16 and 2.17 sections by @causten in #5229
• Bump tornado from 6.5.7 to 6.5.8 in /docs/sphinx by @dependabot[bot] in #5231
• [AIMIGRAPHX-1100] Add no-rebuild callback for verify by @eddieliao in #5067
• Bump misspell to v1.28 by @causten in #5242
• Disable subprocess spawning during hiprtc kernel compilation by @pnikolic-amd in #5189
• Add adaptive benchmarking to MIGraphX by @justinrosner in #5133
• Use Python3.11 in SLES docker by @ahsan-ca in #5247
• Add proper handling of prefill attributes from rocMLIR/rocmlirTriton by @justinrosner in #5249
• Fix MLIR benchmark fill for non-packed outputs by @umangyadav in #5252
• Fix unsafe MLIR split boundaries by @umangyadav in #5253
• [ROCM-28806] Make navi4x gemm_tune_invalid_sol_index guard version specific by @eddieliao in #5174
• [ROCM-26731] Exclude external PRs from CI GitHub action by @eddieliao in #5240
• Improve find_permutation with ambiguous layouts by @pfultz2 in #5153
• [AIMIGRAPHX-1279] Skip horizontal fusion if dots are dependent by @eddieliao in #5250
• Accuracy: Round to nearest, ties to even for generic_float by @CharlieL7 in #5193
• Work around runtime bug by using add instead of dot by @justinrosner in #5267
• Have reduce_sum, reduce_mean, and reduce_prod accumulate in wider floating point for fp8 by @CharlieL7 in #5160
• Fix matcher has_value tolerance for low precision types by @CharlieL7 in #5190
• Revert "[ROCM-26731] Exclude external PRs from CI GitHub action" by @causten in #5263
• Eliminate concat after reshapes by @pfultz2 in #5137
• Drop null problem-cache sentinels on load, keep in-run dedup by @danieyan-amd in #5204
• Fix Gather on empty (zero-element) tensor aborting at ONNX parse (#5246) by @itikhono in #5262
• Added support to set mlir defaults by @pnikolic-amd in #4651
• [AIRADSW-852] Fixing regression with slice-squeeze rewrite for scalar inputs by @urpetkov-amd in #5192
• Support nonpacked fill by @pfultz2 in #5275
• Add std C++ components to rocm namespace and add unit tests by @pfultz2 in #4752
• Fix tidy warnings in generic_float by @pfultz2 in #5288
• Marshall exception in simple_par_for by @pfultz2 in #5265
• Rewrite broadcast reshapes by @pfultz2 in #5273
• [AIRADSW-709] Remove division of batch size since does not scale with workgroups by @Zhaeong in #5285
• Have convolution_backwards error on indivisible groups and reference operator accumulate in double by @CharlieL7 in #5248
• Rewrite skinny gemms to mul+reduce_sum by @pfultz2 in #4811
• ensure standard shapes for lrn operator by @kahmed10 in #5277
• Fix conv concat fusion with different output channels by @urpetkov-amd in #5297
• Add support for HipGraph by @pfultz2 in #4956
• Avoid no-op concat rewrites in simplify_algebra by @justinrosner in #5300
• Use rocmlirtriton as the compiler backend by @causten in #5215
• Fix hipExtModuleLaunchKernel C linkage on Linux by @kentqian in #5298
• Disable flaky test_ck_gemm_softmax_gemm_0 GPU verify by @causten in #5305
• Fix out-of-order split concat rewrite by @CharlieL7 in #5303
• Fix horizontal fusion of dependent operations by partitioning key groups into independent subgroups by @pfultz2 in #5280
• Hiprtc compilation use streams instead of flie write/read by @pnikolic-amd in #5244
• [AIMIGRAPHX-1208] Run ONNX Model Zoo on CI by @eddieliao in #5243
• Refactor adaptive benchmarking by @pfultz2 in #5279
• Dont fuse multi outputs when the layouts are different by @pfultz2 in #5320
• Match the reduce_mean variance pattern through shape transforms by @pfultz2 in #5311
• Propagate broadcast in binary ops by @pfultz2 in #5291
• Limit reshape output fusion in mlir by @pfultz2 in #5283
• Reduce redundant work in adaptive GPU tuning benchmarks by @justinrosner in #5317
• Bump RMT to 0925 by @causten in #5321
• Set CHANGELOG for 10.1 by @causten in #5326
• Resolve model compilation performance for 10.1 ROCm release by @causten in #5331

New Contributors

• @nkotakal0107 made their first contribution in #5154
• @zhihuidu-amd made their first contribution in #5136

Full Changelog: rocm-10.0...rocm-10.1