ROCm 6.1 package

mlc-ai · Aug 23, 2024 · 4b13131 · 4b13131
1 parent f4fed53
commit 4b13131
Show file tree

Hide file tree

Showing 5 changed files with 42 additions and 37 deletions.
diff --git a/.github/workflows/wheel_manylinux_nightly.yaml b/.github/workflows/wheel_manylinux_nightly.yaml
@@ -20,29 +20,21 @@ jobs:
         # matrix of build configs
         config:
           - gpu: 'none'
-            image: 'mlcaidev/package-cpu:7a83f97'
+            image: 'mlcaidev/package-cpu:16b1781'
           - gpu: 'cuda-12.1'
-            image: 'mlcaidev/package-cu121:7a83f97'
+            image: 'mlcaidev/package-cu121:16b1781'
           - gpu: 'cuda-12.2'
-            image: 'mlcaidev/package-cu122:7a83f97'
-          # - gpu: 'rocm-5.6'
-          #   image: 'mlcaidev/package-rocm56:7a83f97'
-          # - gpu: 'rocm-5.7'
-          #   image: 'mlcaidev/package-rocm57:7a83f97'
+            image: 'mlcaidev/package-cu122:16b1781'
+          - gpu: 'rocm-6.1'
+            image: 'mlcaidev/package-rocm61:16b1781'
+          - gpu: 'rocm-6.2'
+            image: 'mlcaidev/package-rocm62:16b1781'
         exclude:
           - pkg_kind: ${{ github.event_name == 'schedule' && 'stable' || '' }}
 
-    runs-on: ubuntu-latest
+    runs-on: [self-hosted, Linux, X64]
 
     steps:
-    - name: Reclaim disk space
-      run: |
-        df -h
-        sudo rm -rf /usr/local/lib/android
-        sudo rm -rf /opt/hostedtoolcache/CodeQL
-        sudo rm -rf /usr/local/.ghcup
-        df -h
-
     - uses: actions/checkout@v3
       with:
         submodules: 'recursive'
@@ -56,24 +48,32 @@ jobs:
         git clone https://github.com/mlc-ai/mlc-llm mlc-llm --recursive
     - name: Sync MLC AI Package
       run: |
-        python scripts/sync_package.py --gpu ${{ matrix.config.gpu }} --package tvm --package-name ${{ matrix.pkg_kind == 'nightly' && 'mlc-ai-nightly' || 'mlc-ai' }} --revision origin/mlc ${{ matrix.pkg_kind == 'nightly' && '--skip-checkout' || '' }} --skip-conda
+        python3 scripts/sync_package.py --gpu ${{ matrix.config.gpu }} --package tvm --package-name ${{ matrix.pkg_kind == 'nightly' && 'mlc-ai-nightly' || 'mlc-ai' }} --revision origin/mlc ${{ matrix.pkg_kind == 'nightly' && '--skip-checkout' || '' }} --skip-conda
     - name: Sync MLC LLM Package
       if: matrix.pkg_kind == 'nightly'
       run: |
-        python scripts/sync_package.py --gpu ${{ matrix.config.gpu }} --package mlc-llm --package-name ${{ matrix.pkg_kind == 'nightly' && 'mlc-llm-nightly' || 'mlc-llm' }} --revision origin/main ${{ matrix.pkg_kind == 'nightly' && '--skip-checkout' || '' }} --skip-conda
+        python3 scripts/sync_package.py --gpu ${{ matrix.config.gpu }} --package mlc-llm --package-name ${{ matrix.pkg_kind == 'nightly' && 'mlc-llm-nightly' || 'mlc-llm' }} --revision origin/main ${{ matrix.pkg_kind == 'nightly' && '--skip-checkout' || '' }} --skip-conda
     - name: Build TVM Unity
+      uses: gacts/run-and-post-run@v1
       env:
         IMAGE: ${{ matrix.config.image }}
         GPU: ${{ matrix.config.gpu }}
-      run: |
-        docker/bash.sh --no-gpu $IMAGE ./scripts/build_mlc_ai_wheel_manylinux.sh --gpu $GPU
+      with:
+        run: |
+          docker/bash.sh --no-gpu $IMAGE ./scripts/build_mlc_ai_wheel_manylinux.sh --gpu $GPU
+        post: |
+          docker/bash.sh --no-gpu $IMAGE ./scripts/cleanup_workspace.sh
     - name: Build MLC-LLM
       if: matrix.pkg_kind == 'nightly'
+      uses: gacts/run-and-post-run@v1
       env:
         IMAGE: ${{ matrix.config.image }}
         GPU: ${{ matrix.config.gpu }}
-      run: |
-        docker/bash.sh --no-gpu $IMAGE ./scripts/build_mlc_llm_wheel_manylinux.sh --gpu $GPU
+      with:
+        run: |
+          docker/bash.sh --no-gpu $IMAGE ./scripts/build_mlc_llm_wheel_manylinux.sh --gpu $GPU
+        post: |
+          docker/bash.sh --no-gpu $IMAGE ./scripts/cleanup_workspace.sh
     - name: Wheel-Deploy
       if: github.ref == 'refs/heads/main'
       uses: softprops/action-gh-release@v1

diff --git a/scripts/build_mlc_ai_wheel_manylinux.sh b/scripts/build_mlc_ai_wheel_manylinux.sh
@@ -6,7 +6,7 @@ source /opt/rh/gcc-toolset-11/enable # GCC-11 is the hightest GCC version compat
 function usage() {
 	echo "Usage: $0 [--gpu GPU-VERSION]"
 	echo
-	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 rocm-5.6 rocm-5.7}"
+	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 rocm-6.1 rocm-6.2}"
 	echo -e "\tSpecify the GPU version (CUDA/ROCm) in the TVM (default: none)."
 }
 
@@ -44,7 +44,7 @@ function audit_mlc_ai_wheel() {
 TVM_PYTHON_DIR="/workspace/tvm/python"
 PYTHON_VERSIONS_CPU=("3.7" "3.8" "3.9" "3.10" "3.11" "3.12")
 PYTHON_VERSIONS_GPU=("3.7" "3.8" "3.9" "3.10" "3.11" "3.12")
-GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-5.6" "rocm-5.7")
+GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-6.1" "rocm-6.2")
 GPU="none"
 
 while [[ $# -gt 0 ]]; do
@@ -71,7 +71,7 @@ done
 if ! in_array "${GPU}" "${GPU_OPTIONS[*]}"; then
 	echo "Invalid GPU option: ${GPU}"
 	echo
-	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-5.6" "rocm-5.7"}'
+	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-6.1" "rocm-6.2"}'
 	exit -1
 fi
 
@@ -86,7 +86,7 @@ fi
 AUDITWHEEL_OPTS="--plat ${AUDITWHEEL_PLAT} -w repaired_wheels/"
 AUDITWHEEL_OPTS="--exclude libtinfo ${AUDITWHEEL_OPTS}"
 if [[ ${GPU} == rocm* ]]; then
-	AUDITWHEEL_OPTS="--exclude libamdhip64 --exclude libhsa-runtime64 --exclude librocm_smi64 --exclude librccl ${AUDITWHEEL_OPTS}"
+	AUDITWHEEL_OPTS="--exclude libamdhip64 --exclude libhsa-runtime64 --exclude librocm_smi64 --exclude librccl --exclude libhipblas --exclude libhipblaslt ${AUDITWHEEL_OPTS}"
 elif [[ ${GPU} == cuda* ]]; then
 	AUDITWHEEL_OPTS="--exclude libcuda --exclude libcudart --exclude libnvrtc --exclude libcublas --exclude libcublasLt ${AUDITWHEEL_OPTS}"
 fi
@@ -105,9 +105,10 @@ elif [[ ${GPU} == cuda* ]]; then
 fi
 
 if [[ ${GPU} == rocm* ]]; then
-	echo set\(USE_LLVM \"/opt/rocm/llvm/bin/llvm-config --ignore-libllvm --link-static\"\) >>config.cmake
+	echo set\(USE_LLVM \"llvm-config --ignore-libllvm --link-static\"\) >>config.cmake
 	echo set\(USE_ROCM ON\) >>config.cmake
-	echo set\(USE_RCCL /opt/rocm/rccl/ \) >>config.cmake
+	echo set\(USE_HIPBLAS ON\) >>config.cmake
+	echo set\(USE_RCCL /opt/rocm/\) >>config.cmake
 elif [[ ${GPU} == cuda* ]]; then
 	echo set\(USE_LLVM \"llvm-config --ignore-libllvm --link-static\"\) >>config.cmake
 	echo set\(USE_CUDA ON\) >>config.cmake
@@ -135,7 +136,7 @@ git config --global --add safe.directory /workspace/tvm
 mkdir -p build
 cd build
 cmake ..
-make -j$(nproc)
+make -j12
 find . -type d -name 'CMakeFiles' -exec rm -rf {} +
 
 UNICODE_WIDTH=32 # Dummy value, irrelevant for Python 3

diff --git a/scripts/build_mlc_llm_wheel_manylinux.sh b/scripts/build_mlc_llm_wheel_manylinux.sh
@@ -6,7 +6,7 @@ source /opt/rh/gcc-toolset-11/enable # GCC-11 is the hightest GCC version compat
 function usage() {
 	echo "Usage: $0 [--gpu GPU-VERSION]"
 	echo
-	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 rocm-5.6 rocm-5.7}"
+	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 rocm-6.1 rocm-6.2}"
 	echo -e "\tSpecify the GPU version (CUDA/ROCm) in the MLC-LLM (default: none)."
 }
 
@@ -44,7 +44,7 @@ function audit_mlc_llm_wheel() {
 MLC_LLM_PYTHON_DIR="/workspace/mlc-llm/python"
 PYTHON_VERSIONS_CPU=("3.7" "3.8" "3.9" "3.10" "3.11" "3.12")
 PYTHON_VERSIONS_GPU=("3.7" "3.8" "3.9" "3.10" "3.11" "3.12")
-GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-5.6" "rocm-5.7")
+GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-6.1" "rocm-6.2")
 GPU="none"
 
 while [[ $# -gt 0 ]]; do
@@ -71,7 +71,7 @@ done
 if ! in_array "${GPU}" "${GPU_OPTIONS[*]}"; then
 	echo "Invalid GPU option: ${GPU}"
 	echo
-	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-5.6" "rocm-5.7"}'
+	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "rocm-6.1" "rocm-6.2"}'
 	exit -1
 fi
 
@@ -86,7 +86,7 @@ fi
 AUDITWHEEL_OPTS="--plat ${AUDITWHEEL_PLAT} -w repaired_wheels/"
 AUDITWHEEL_OPTS="--exclude libtvm --exclude libtvm_runtime --exclude libvulkan ${AUDITWHEEL_OPTS}"
 if [[ ${GPU} == rocm* ]]; then
-	AUDITWHEEL_OPTS="--exclude libamdhip64 --exclude libhsa-runtime64 --exclude librocm_smi64 --exclude librccl ${AUDITWHEEL_OPTS}"
+	AUDITWHEEL_OPTS="--exclude libamdhip64 --exclude libhsa-runtime64 --exclude librocm_smi64 --exclude librccl --exclude libhipblas --exclude libhipblaslt ${AUDITWHEEL_OPTS}"
 elif [[ ${GPU} == cuda* ]]; then
 	AUDITWHEEL_OPTS="--exclude libcuda --exclude libcudart --exclude libnvrtc  --exclude libcublas --exclude libcublasLt ${AUDITWHEEL_OPTS}"
 fi
@@ -103,7 +103,8 @@ fi
 
 if [[ ${GPU} == rocm* ]]; then
 	echo set\(USE_ROCM ON\) >>config.cmake
-	echo set\(USE_RCCL /opt/rocm/rccl/ \) >>config.cmake
+	echo set\(USE_HIPBLAS ON\) >>config.cmake
+	echo set\(USE_RCCL /opt/rocm/\) >>config.cmake
 elif [[ ${GPU} == cuda* ]]; then
 	echo set\(USE_CUDA ON\) >>config.cmake
 	echo set\(USE_CUTLASS ON\) >>config.cmake
@@ -127,7 +128,7 @@ fi
 mkdir -p build
 cd build
 cmake ..
-make -j$(nproc)
+make -j12
 find . -type d -name 'CMakeFiles' -exec rm -rf {} +
 
 UNICODE_WIDTH=32 # Dummy value, irrelevant for Python 3

diff --git a/scripts/cleanup_workspace.sh b/scripts/cleanup_workspace.sh
@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+
+rm -rf /workspace/tvm/* /workspace/mlc-llm/*
diff --git a/scripts/sync_package.py b/scripts/sync_package.py
@@ -167,8 +167,8 @@ def main():
             "cuda-11.8",
             "cuda-12.1",
             "cuda-12.2",
-            "rocm-5.6",
-            "rocm-5.7",
+            "rocm-6.1",
+            "rocm-6.2",
         ],
         help="GPU (CUDA/ROCm) version to be linked to the resultant binaries,"
         "or none, to disable CUDA/ROCm. Defaults to none.",