AMD Developer Challenge 2025: Distributed Inference

README.md

@@ -2,7 +2,9 @@
 
 🏆 Grand Prize Winner Project for AMD Developer Challenge 2025
 
+- [2026/02] All2All, GEMM-RS, and AG-GEMM kernels are now open-sourced for [AMD Developer Challenge 2025: Distributed Inference](https://amdchallenge2025.datamonsters.com/)
 ### News 🔥
+
 - [2025/11] [Easily Build and Share ROCm Kernels with Hugging Face](https://huggingface.co/blog/build-rocm-kernels) – Discover how to easily build and share RadeonFlow_Kernels using [Hugging Face's kernel-builder](https://github.com/huggingface/kernel-builder).
 
 - [2025/06] [AMD Developer Cloud](https://www.amd.com/en/developer/resources/cloud-access/amd-developer-cloud.html) now provides free AMD Instinct Accelerators, you can try out our project using their MI300X.

dist-infer/.clang-format

@@ -0,0 +1,3 @@
+IndentWidth: 4
+AlignAfterOpenBracket: BlockIndent
+PackConstructorInitializers: CurrentLine

dist-infer/.clangd

@@ -0,0 +1,3 @@
+CompileFlags:
+  Add: ["-xhip", "--rocm-path=/opt/rocm"]
+  Remove: ["-x*"]

dist-infer/.gitignore

@@ -0,0 +1,11 @@
+logs/
+build/
+.cache/
+.vscode-server/
+.popcorn.yaml
+*.cu
+*.hip
+*.lock
+submission.py
+torch-build/
+__pycache__

dist-infer/ag-gemm/CMakeLists.txt

@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.21)
+cmake_policy(VERSION 3.21.3...3.27)
+set(PROJECT_NAME "ag_gemm")
+project(${PROJECT_NAME} LANGUAGES HIP CXX)
+
+find_package(Python3 REQUIRED COMPONENTS Development Interpreter)
+find_package(Torch CONFIG REQUIRED)
+find_package(HIP CONFIG REQUIRED)
+
+# required for python binding
+find_library(TORCH_PYTHON_LIBRARY torch_python PATH ${TORCH_INSTALL_PREFIX}/lib)
+
+add_library(${PROJECT_NAME} SHARED perf_gemm.cc)
+set_source_files_properties(ag_gemm.cpp PROPERTIES LANGUAGE HIP)
+target_link_libraries(${PROJECT_NAME} PRIVATE ${TORCH_PYTHON_LIBRARY} ${TORCH_LIBRARIES} hip::device Python3::Python)
+set_target_properties(${PROJECT_NAME} PROPERTIES PREFIX "")
+target_compile_features(${PROJECT_NAME} PRIVATE cxx_std_20)
+
+# for host compile
+target_compile_definitions(${PROJECT_NAME} PRIVATE -D__${CMAKE_HIP_ARCHITECTURES}__)
+target_compile_options(${PROJECT_NAME} PRIVATE -save-temps)

dist-infer/ag-gemm/Makefile

@@ -0,0 +1,39 @@
+.PHONY: all config build
+
+TARGET ?= gfx942
+
+BUILD_TYPE ?= RelWithDebInfo
+BUILD_DIR ?= build
+
+PYTHON_DIR ?= $(shell python -c "import site; print(site.getsitepackages()[0])")
+
+all: config build submit
+
+config:
+	PYTORCH_ROCM_ARCH=$(TARGET) cmake -B $(BUILD_DIR) . \
+		-DCMAKE_BUILD_TYPE=$(BUILD_TYPE) \
+		-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
+		-DCMAKE_PREFIX_PATH="/opt/rocm;$(PYTHON_DIR)" \
+		-DCMAKE_HIP_ARCHITECTURES=$(TARGET) \
+		-DGPU_TARGETS=$(TARGET) \
+		-DAMDGPU_TARGETS=$(TARGET) \
+		-G Ninja
+
+build:
+	cmake --build $(BUILD_DIR) -j8
+
+test: build
+	PYTHONPATH=$(PYTHONPATH):$(realpath tools):$(realpath $(BUILD_DIR)) python tools/smoke_test.py
+
+clean:
+	rm -r $(BUILD_DIR)
+
+local:
+	python submit.py local_test
+	POPCORN_GPUS=2 POPCORN_FD=2 python eval.py benchmark benchmark.txt
+
+submit:
+	python submit.py
+
+dis:
+	roc-obj -d build/ag_gemm.so