rocm-7.0.1

What's Changed

• Add E2E test for the OCP Fp8 fused kernel with QuantizeLinear and DeQuantizeLinear by @umangyadav in #1747
• [TOSA] Set accType to Float16 for the Fp8 types by @umangyadav in #1745
• Remove scheduling barrier hack for LDS barrier lowering by @dhernandez0 in #1749
• Fixes for group conv emit-key by @dhernandez0 in #1748
• Fix performance for non-standard layouts by @dhernandez0 in #1741
• [6.4]Fix bug when both A and B are broadcasted (FoldBroadcast pass) by @dhernandez0 in #1744
• [TOSA] Fix accType for the Quant Convolutions as well by @umangyadav in #1752
• [6.4] Update gfx12 target in AmdArchDB by @TedThemistokleous in #1746
• Add Fp8 to quick-tuning by @djramic in #1753
• Add bf16 to tuning runner by @djramic in #1739
• Enable output swizzle for multiple outputs by @dhernandez0 in #1750
• Use AddDim for unit input dimensions to help getMaxVectorization() by @dhernandez0 in #1755
• [DO NOT SQUASH] Enable atomic add bf16 reduction and split-k for Navi4x by @dhernandez0 in #1732
• Enable bf16 atomic add for gfx950 by @dhernandez0 in #1734
• Add test from SWDEV-518130 by @dhernandez0 in #1757
• [6.4]fix compilation with HIP SDK 6.3 for Windows by @apwojcik in #1742
• Add lookup for more layouts in PerfRunner and Add an option for verifying each perfConfig with tuningRunner by @umangyadav in #1758
• Rocmlir tuning driver datatype fix by @dorde-antic in #1761
• [CI] Added gfx942 architecture to the 'Tune MLIR kernels' stage by @stefankoncarevic in #1733
• Fix dependency graph creation in RockPipeline and not generate loops with negative iterations by @umangyadav in #1760
• Fix GlobalLoad 4b lowering by @dhernandez0 in #1764
• Improve performance of quantizelinear for int4 by @dhernandez0 in #1706
• Add fp8 convolution to the tuning runner by @djramic in #1738
• Introduce perfConfig V3 with param to select different schedule by @umangyadav in #1767
• Support for causal attention and more strict checks for KV-Cache by @dhernandez0 in #1770
• Fix generateMlirDriverCommandLine for attention in perfRunner by @dhernandez0 in #1773
• Remove hasValidChip() from ConvGenerator by @dorde-antic in #1771
• Use MLIR based kernels for verification in MIGraphX stage by @umangyadav in #1766
• [DO NOT SQUASH] March LLVM upstream merge by @dhernandez0 in #1763
• Add requirements.txt file and modify Dockerfile by @dorde-antic in #1776
• Add checks for uid and devices by @causten in #1777
• Fix Dockerfile URL for requirements.txt by @stefankoncarevic in #1778
• Adjust Dockerfile for Separate hip-python Installation by @stefankoncarevic in #1781
• Skip unsupported datatypes in perfRunner by @djramic in #1780
• Fix initialization for split-k by @dhernandez0 in #1784
• Use hip-python API instead of rocm_agent_enumerator by @dorde-antic in #1762
• Recover split-k fusion tests removed in last upstream merge by @dhernandez0 in #1785
• Add hip-python to requirements.txt and update LLVM version by @dhernandez0 in #1787
• Fix split-k fusion when there are two or more consecutive linalg.genericops by @dhernandez0 in #1782
• Remove Machine Names Due to Security Team Advisory by @stefankoncarevic in #1788
• Remove fp8 check on nightly CI. by @stefankoncarevic in #1789
• [DO NOT SQUASH] upstream merge for sprint 48 by @dhernandez0 in #1786
• Move requirements.txt -> pip_requirements.txt due to issues with cget by @dhernandez0 in #1792
• Python script for testing metrics and plotting correlations by @dorde-antic in #1769
• Fix attention bugs (swap thread and iter when Q LDS is bypassed and bf16 tests) by @dhernandez0 in #1797
• Sort Dimensions based on Layout in case of input fusion by @umangyadav in #1793
• Fix kernel generation when kernelRepeats are more than 1 by @umangyadav in #1799
• Workaround issue 1802 by @dhernandez0 in #1800
• Add Gemm+Elementwise+Gemm support by @dhernandez0 in #1774
• Add dependencies for rocprofv3 by @djramic in #1801
• Remove perfTest from Jenkins by @dhernandez0 in #1803
• Add Tier1 model configs to rocMLIR by @dorde-antic in #1794
• GEMM+GEMM migraphx integration by @dhernandez0 in #1791
• Fix for issue 1802 workaround by @dhernandez0 in #1806
• Update MI300 quick-tuning list by @mirza-halilcevic in #1765
• gemm+gemm: extend allowed types by @dhernandez0 in #1795
• Bump Dockerfiles to rocm-6.4 by @dorde-antic in #1808
• Disable code coverage on nightly and weekly CI, and expand it to run on WMMA by @mirza-halilcevic in #1813
• Fix grep ROCM_VERSION in Docker image build by @djramic in #1814
• Add GEMM scheduleV2 by @umangyadav in #1772
• Modify Tier1 models tuning problems by @dorde-antic in #1810
• Prepare Jenkinsfiles for rocm-6.4 by @dorde-antic in #1809
• Remove unused files by @dhernandez0 in #1804
• Update AmdArchDb.cpp with gfx950 target info by @mirza-halilcevic in #1802
• Add pybind11 to pip_requirements.txt by @mirza-halilcevic in #1816
• Use migraphx.greater instead of migraphx.greater_or_equal by @dhernandez0 in #1827
• Change rounding mode for FP32 to Fp16 truncation by @umangyadav in #1833
• Implement with_attn_bias in AttentionConfiguration by @dorde-antic in #1834
• Add rocprofv3 to perfRunner by @djramic in #1779
• Fix rocm version in migraphx CI docker image by @djramic in #1837
• Upstream merge sprint 50 by @djramic in #1815
• [CI] Set 3600s test timeout and update LIT worker configuration by @stefankoncarevic in #1832
• Remove hardcoded value for render group id in Dockerfile by @umangyadav in #1839
• add back render group but do not assign GID by @umangyadav in #1843
• Causal attention by @dhernandez0 in #1829
• Correct rocprof invocation in fusion benchmarking path. by @stefankoncarevic in #1841
• conv+gemm support by @dhernandez0 in #1820
• Problem config for tier 1 models by @aarushjain29 in #1836
• conv+gemm migraphx integration by @dhernandez0 in #1823
• Separate new Tier1 tuning problems by @dorde-antic in #1849
• Disable test temporarily to pass CI by @umangyadav in #1850
• Implement GQA in AttentionConfiguration by @dorde-antic in #1847
• Correct layout map access in MLIROnlyConfig by @stefankoncarevic in #1855
• Add missing LDS barriers to attention by @dhernandez0 in #1853
• Causal masking: migraphx integration by @dhernandez0 in #1831
• Updated ATTN_TEST_PARAMETERS in reportUtils.py by @stefankoncarevic in #1858
• [CLONE] Add CI node checks and retries. Refactored the pipeline to resolve compilation errors and address incorrect syntax by @umangyadav in #1835
• Modify CI to use Tier1 and rotate through configs by @dorde-antic in #1840
• Allow retries for failing tests / Remove failing tests by @dorde-antic in #1819
• Print rocm version and permissions for /dev/dri /dev/kfd by @umangyadav in #1860
• Added fix for environment variables being overwritten in parallel and matrix runs and more diagnostic info by @leo-amd in #1867
• Remove deprecated use of find_package by @umangyadav in #1861
• Add i8 in DATA_TYPES_ATTENTION by @dorde-antic in #1838
• Fix docker arguments for Public CI by @umangyadav in #1868
• Adjust lit parallelism for MFMA architectures based on GPU type by @stefankoncarevic in #1865
• Remove ALLOW_RETRIES by @dhernandez0 in #1852
• Make number of LIT workers 8 by default by @dhernandez0 in #1871
• Make sure arch can't be empty in AmdArchDb by @dhernandez0 in #1862
• Added cleanWs to the post in internal rocMLIR pipeline by @leo-amd in #1872
• Added 5min timeout to the resetGPUs() func by @leo-amd in #1864
• Install rocMLIR requirements in MITuna venv by @dorde-antic in #1879
• Fix Azure builds by @umangyadav in #1884
• Fix output file path for rocprofv3 by @djramic in #1883
• Add missing space for perfRunner and some other minor issues by @dhernandez0 in #1885
• Attention: return LSE (log-sum-exp) by @dhernandez0 in #1882
• Refactor splitConfigFile to use sed-based slicing by @dorde-antic in #1870
• [DO NOT SQUASH] LLVM Upstream Merge Sprint 53 by @mirza-halilcevic in #1863
• Fix arch in Fusion test by @umangyadav in #1888
• Use migraphx-ci image for migraphx stage by @umangyadav in #1892
• Increase time out on resetGPUs to allow resetting more than 8 GPUs by @umangyadav in #1893
• Upstream merge Sprint 54 with Navi3x fixes by @umangyadav in #1891
• Define datatypes for attention dynamically based on chip type by @dorde-antic in #1894
• Define DATA_TYPES_ATTENTION if class AttentionConfiguration is reused outside by @dorde-antic in #1897
• Parameter Sweep for Attention by @dorde-antic in #1830
• [DO NOT SQUASH] Add double rate MFMA instruction for gfx950 by @umangyadav in #1896
• Attention LSE migraphx integration by @dhernandez0 in #1887
• Weekly CI: Tuning on Navi3x/Navi4x by @dorde-antic in #1890
• Allow user to choose softmax type by @dhernandez0 in #1900
• perfRunner: use regex for exact flag detection in config generation by @dorde-antic in #1903
• perfRunner: skip f32 attention kernels on Navi by @dorde-antic in #1902
• Upstream merge 55 by @umangyadav in #1899
• Add acc_type to MatMulOp by @djramic in #1880
• Refactor perfRunner: repr implementation in PerfConfiguration by @dorde-antic in #1901
• [DO NOT SQUASH] Attention: let migraphx decide softmax type and LSE return type by @dhernandez0 in #1904
• [BACKPORT] Allow softmax type conversion to happen before or after elementwise by @umangyadav in #1920
• [7.0][BACKPORT] Fix multi buffer test on gfx950 by @djramic in #1915
• [7.0][BACKPORT] Find first gemm index after fusing linalg.generic ops by @dhernandez0 in #1923
• [7.0][BACKPORT] Update minCU count for MI308 by @causten in #1928
• [BACKPORT] Introduce new quick tune lists based on Tier1 configs and separated b… by @umangyadav in #1938
• [BACKPORT] Cherry pick fix for const folding of immediate args by @umangyadav in #1941
• [Backport] Update tests to excludes unsupported tests on Navi2x (#1943) by @umangyadav in #1945
• [Backport] Add regularization for multiple linalgs in preSoftmaxBody by @umangyadav in #1958
• [Backport] Refactor and fix creation of ElementWise Region for Gemm+Gemm like ops by @umangyadav in #1965

New Contributors

• @TedThemistokleous made their first contribution in #1746

Full Changelog: rocm-6.4.3...rocm-7.0.1