@@ -4,6 +4,71 @@ Full documentation for MIGraphX is available at
44[ https://rocmdocs.amd.com/projects/AMDMIGraphX/en/latest/ ] ( https://rocmdocs.amd.com/projects/AMDMIGraphX/en/latest/ ) .
55
66
7+ ## MIGraphX 2.14 for ROCm 7.1.0
8+
9+ ### Added
10+
11+ * Added Python 3.13 support.
12+ * Added PyTorch wheels to the Dockerfile.
13+ * Added Python API for returning serialized bytes.
14+ * Added ` fixed_pad ` operator for padding dynamic shapes to the maximum static shape.
15+ * Added matcher to upcast base ` Softmax ` operations.
16+ * Added support for the ` convolution_backwards ` operator through rocMLIR.
17+ * Added ` LSE ` output to attention fusion.
18+ * Added flags to ` EnableControlFlowGuard ` due to BinSkim errors.
19+ * Added new environment variable documentation and reorganized structure.
20+ * Added ` stash_type ` attribute for ` LayerNorm ` and expanded test coverage.
21+ * Added operator builders (phase 2).
22+ * Added ` MIGRAPHX_GPU_HIP_FLAGS ` to allow extra HIP compile flags.
23+
24+ ### Changed
25+
26+ * Updated C API to include ` current() ` caller information in error reporting.
27+ * Updated documentation dependencies:
28+ * ** rocm-docs-core** bumped from 1.21.1 → 1.25.0 across releases.
29+ * ** Doxygen** updated to 1.14.0.
30+ * ** urllib3** updated from 2.2.2 → 2.5.0.
31+ * Updated ` src/CMakeLists.txt ` to support ` msgpack ` 6.x (` msgpack-cxx ` ).
32+ * Updated model zoo test generator to fix test issues and add summary logging.
33+ * Updated ` rocMLIR ` and ` ONNXRuntime ` mainline references across commits.
34+ * Updated module sorting algorithm for improved reliability.
35+ * Restricted FP8 quantization to ` dot ` and ` convolution ` operators.
36+ * Moved ONNX Runtime launcher script into MIGraphX and updated build scripts.
37+ * Simplified ONNX ` Resize ` operator parser for correctness and maintainability.
38+ * Updated ` any_ptr ` assertion to avoid failure on default HIP stream.
39+ * Print kernel and module information on compile failure.
40+
41+ ### Fixed
42+
43+ * Fixed error in ` MIGRAPHX_GPU_COMPILE_PARALLEL ` documentation (#4337 ).
44+ * Fixed rocMLIR ` rewrite_reduce ` issue (#4218 ).
45+ * Fixed bug with ` invert_permutation ` on GPU (#4194 ).
46+ * Fixed compile error when ` MIOPEN ` is disabled (missing ` std ` includes) (#4281 ).
47+ * Fixed ONNX ` Resize ` parsing when input and output shapes are identical (#4133 , #4161 ).
48+ * Fixed issue with MHA in attention refactor (#4152 ).
49+ * Fixed synchronization issue from upstream ONNX Runtime (#4189 ).
50+ * Fixed spelling error in “Contiguous” (#4287 ).
51+ * Fixed tidy complaint about duplicate header (#4245 ).
52+ * Fixed ` reshape ` , ` transpose ` , and ` broadcast ` rewrites between pointwise and reduce operators (#3978 ).
53+ * Fixed extraneous include file in HIPRTC-based compilation (#4130 ).
54+ * Fixed CI Perl dependency issue for SLES builds (#4254 ).
55+ * Fixed compiler warnings for ROCm 7.0 of `` error: unknown warning option '-Wnrvo' `` (#4192 ).
56+
57+ ### Optimized
58+
59+ * Reduced nested visits in reference operators to improve compile time.
60+ * Avoided dynamic memory allocation during kernel launches.
61+ * Removed redundant NOP instructions for GFX11/12 platforms.
62+ * Improved ` Graphviz ` output (node color and layout updates).
63+ * Optimized interdependency checking during compilation.
64+ * Skip hipBLASLt solutions requiring workspace size larger than 128 MB for efficient memory utilization.
65+
66+ ### Removed
67+
68+ * Removed Perl dependency from SLES builds.
69+ * Removed redundant includes and unused internal dependencies.
70+
71+
772## MIGraphX 2.13 for ROCm 7.0.0
873
974### Added
0 commit comments