Fixes for ONNX and FP4 in 7.2

.azuredevops/rocm-ci.yml

@@ -1,13 +1,19 @@
+variables:
+- group: common
+- template: /.azuredevops/variables-global.yml@pipelines_repo
+
+parameters:
+- name: pipelinesRepoRef
+  type: string
+  default: refs/heads/develop
+
 resources:
   repositories:
   - repository: pipelines_repo
     type: github
     endpoint: ROCm
     name: ROCm/ROCm
-
-variables:
-- group: common
-- template: /.azuredevops/variables-global.yml@pipelines_repo
+    ref: ${{ parameters.pipelinesRepoRef }}
 
 trigger:
   batch: true

CMakeLists.txt

@@ -24,7 +24,7 @@
 cmake_minimum_required(VERSION 3.15 FATAL_ERROR)
 
 if("${CMAKE_SOURCE_DIR}" STREQUAL "${CMAKE_BINARY_DIR}")
     message(FATAL_ERROR "The binary and source directroy cannot be the same")
 endif()
 
 # Setup valid strings for build type
@@ -300,6 +300,8 @@
         # Disable because of too many FPs
         arithOperationsOnVoidPointer
         definePrefix:*test/include/test.hpp
+        definePrefix:*src/targets/gpu/kernels/include/migraphx/kernels/test.hpp
+        UseNamedLogicOperator:*src/targets/gpu/kernels/include/migraphx/kernels/debug.hpp
         ctuOneDefinitionRuleViolation:*test/*
         useSmartPointer:*src/api/api.cpp
         useSmartPointer:*make_shared_array.hpp

docs/dev/tools.rst

@@ -55,11 +55,11 @@ Provides additional debug information related to data. Use for debugging purpose
 
 Example output:
 
-.. image:: ../data/roctx1.jpg
+.. image:: ../data/roctx1.png
 
-Hotspot kerel timing information:
+Hotspot kernel timing information:
 
-.. image:: ../data/roctx2.jpg
+.. image:: ../data/roctx2.png
 
 The output provides ``SUM``, ``MIN``, ``MAX`` and ``COUNT`` information for each kernel executed for a given model. It also
 provides the average total time. The following three files are provided for reference:

docs/reference/MIGraphX-dev-env-vars.rst

@@ -160,6 +160,14 @@ Model performance tunable variables change the compilation behavior of a model.
 
       | Default: Split-k performance configurations are turned off.
 
+  * - | ``MIGRAPHX_FLASH_DECODING_NUM_SPLITS``
+      | Turns on flash decoding for attention fusion and sets the number of splits along the key-value sequence dimension.
+
+    - | ``0``: Flash decoding is turned off (i.e., number of splits is 0).
+      | ``N`` (where N > 1): Enables flash decoding with N splits along the key-value sequence dimension. For example, ``2`` enables flash decoding with 2 splits, ``4`` with 4 splits, etc.
+
+      | Default: flash decoding is turned off.
+
   * - | ``MIGRAPHX_DISABLE_FP16_INSTANCENORM_CONVERT``
       | When set, FP16 is not converted to FP32 in the ``InstanceNormalization`` ONNX operator. 
 

docs/sphinx/requirements.in

@@ -1,2 +1,2 @@
-rocm-docs-core==1.29.0
+rocm-docs-core==1.30.0
 sphinx-collapse

docs/sphinx/requirements.txt

@@ -211,7 +211,7 @@ requests==2.32.4
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==1.29.0
+rocm-docs-core==1.30.0
     # via -r requirements.in
 rpds-py==0.22.3
     # via

examples/onnxruntime/resnet50/resnet50.py

@@ -1,7 +1,7 @@
 #####################################################################################
 # The MIT License (MIT)
 #
-# Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+# Copyright (c) 2015-2025 Advanced Micro Devices, Inc. All rights reserved.
 #
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -219,7 +219,7 @@ def main():
         "resnet50.onnx",  # where to save the model (can be a file or file-like object)
         export_params=
         True,  # store the trained parameter weights inside the model file
-        opset_version=12,  # the ONNX version to export the model to
+        opset_version=18,  # the ONNX version to export the model to
         do_constant_folding=
         True,  # whether to execute constant folding for optimization
         input_names=['input'],  # the model's input names

src/CMakeLists.txt

@@ -190,6 +190,7 @@ register_migraphx_ops(
     ceil
     clip
     concat
+    concat_past_present
     contiguous
     convert
     convolution
@@ -212,8 +213,8 @@ register_migraphx_ops(
     gather
     gathernd
     get_tuple_elem
+    gqa_rotary_embedding
     greater
-    group_query_attention
     group
     gru
     identity