Build package for CUDA 12.3; deprecate 12.1

Some CUTLASS kernels for SM90a are only available since CUDA 12.3:
https://github.com/NVIDIA/cutlass/blob/bbe579a9e3beb6ea6626d9227ec32d0dae119a49/include/cutlass/arch/mma_sm90.h#L67-L69

.github/workflows/wheel_manylinux_nightly.yaml

@@ -20,15 +20,15 @@ jobs:
         # matrix of build configs
         config:
           - gpu: 'none'
-            image: 'mlcaidev/package-cpu:b6b04f1'
-          - gpu: 'cuda-12.1'
-            image: 'mlcaidev/package-cu121:b6b04f1'
+            image: 'mlcaidev/package-cpu:254d630'
           - gpu: 'cuda-12.2'
-            image: 'mlcaidev/package-cu122:b6b04f1'
+            image: 'mlcaidev/package-cu122:254d630'
+          - gpu: 'cuda-12.3'
+            image: 'mlcaidev/package-cu123:254d630'
           - gpu: 'rocm-6.1'
-            image: 'mlcaidev/package-rocm61:b6b04f1'
+            image: 'mlcaidev/package-rocm61:254d630'
           - gpu: 'rocm-6.2'
-            image: 'mlcaidev/package-rocm62:b6b04f1'
+            image: 'mlcaidev/package-rocm62:254d630'
         exclude:
           - pkg_kind: ${{ github.event_name == 'schedule' && 'stable' || '' }}
 

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-6.1 rocm-6.2}"
+	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 cuda-12.3 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-6.1" "rocm-6.2")
+GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "cuda-12.3" "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-6.1" "rocm-6.2"}'
+	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "cuda-12.3" "rocm-6.1" "rocm-6.2"}'
 	exit -1
 fi
 

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-6.1 rocm-6.2}"
+	echo -e "--gpu {none cuda-11.7 cuda-11.8 cuda-12.1 cuda-12.2 cuda-12.3 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-6.1" "rocm-6.2")
+GPU_OPTIONS=("none" "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "cuda-12.3" "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-6.1" "rocm-6.2"}'
+	echo 'GPU version can only be {"none", "cuda-11.7" "cuda-11.8" "cuda-12.1" "cuda-12.2" "cuda-12.3" "rocm-6.1" "rocm-6.2"}'
 	exit -1
 fi
 

scripts/sync_package.py

@@ -167,6 +167,7 @@ def main():
             "cuda-11.8",
             "cuda-12.1",
             "cuda-12.2",
+            "cuda-12.3",
             "rocm-6.1",
             "rocm-6.2",
         ],