Blackwell Support Codegen

.github/workflows/documentation.yaml

@@ -11,7 +11,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@v4
       with:
         submodules: recursive
 

.github/workflows/update-relax.yaml

@@ -14,7 +14,7 @@ jobs:
 
     steps:
     - name: Checkout
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
       with:
         submodules: true
 

.github/workflows/windows-build.yaml

@@ -25,7 +25,7 @@ jobs:
     - uses: actions/checkout@v3
       with:
         submodules: 'recursive'
-    - uses: conda-incubator/setup-miniconda@v2
+    - uses: conda-incubator/setup-miniconda@v3
       with:
         activate-environment: mlc-llm-build
         channel-priority: strict

ci/jenkinsfile.groovy

@@ -18,12 +18,12 @@
 import org.jenkinsci.plugins.pipeline.modeldefinition.Utils
 
 run_cpu = "bash ci/bash.sh mlcaidev/ci-cpu:4d61e5d -e GPU cpu -e MLC_CI_SETUP_DEPS 1"
-run_cuda = "bash ci/bash.sh mlcaidev/ci-cu121:4d61e5d -e GPU cuda-12.1 -e MLC_CI_SETUP_DEPS 1"
-run_rocm = "bash ci/bash.sh mlcaidev/ci-rocm57:4d61e5d -e GPU rocm-5.7 -e MLC_CI_SETUP_DEPS 1"
+run_cuda = "bash ci/bash.sh mlcaidev/ci-cu128:4d61e5d -e GPU cuda-12.8 -e MLC_CI_SETUP_DEPS 1"
+run_rocm = "bash ci/bash.sh mlcaidev/ci-rocm63:4d61e5d -e GPU rocm-6.3 -e MLC_CI_SETUP_DEPS 1"
 
-pkg_cpu = "bash ci/bash.sh mlcaidev/package-rocm61:254d630 -e GPU cpu -e MLC_CI_SETUP_DEPS 1"
-pkg_cuda = "bash ci/bash.sh mlcaidev/package-cu122:254d630 -e GPU cuda-12.2 -e MLC_CI_SETUP_DEPS 1"
-pkg_rocm = "bash ci/bash.sh mlcaidev/package-rocm61:254d630 -e GPU rocm-6.1 -e MLC_CI_SETUP_DEPS 1"
+pkg_cpu = "bash ci/bash.sh mlcaidev/package-rocm62:254d630 -e GPU cpu -e MLC_CI_SETUP_DEPS 1"
+pkg_cuda = "bash ci/bash.sh mlcaidev/package-cu128:254d630 -e GPU cuda-12.8 -e MLC_CI_SETUP_DEPS 1"
+pkg_rocm = "bash ci/bash.sh mlcaidev/package-rocm63:254d630 -e GPU rocm-6.3 -e MLC_CI_SETUP_DEPS 1"
 
 
 def per_exec_ws(folder) {

ci/task/build_lib.sh

@@ -24,7 +24,7 @@ if [[ ${GPU} == rocm* ]]; then
 elif [[ ${GPU} == cuda* ]]; then
     echo set\(USE_VULKAN ON\) >>config.cmake
     echo set\(CMAKE_CUDA_COMPILER_LAUNCHER ccache\) >>config.cmake
-    echo set\(CMAKE_CUDA_ARCHITECTURES "80;90"\) >>config.cmake
+    echo set\(CMAKE_CUDA_ARCHITECTURES "80;90;100;120"\) >>config.cmake
     echo set\(CMAKE_CUDA_FLAGS \"\$\{CMAKE_CUDA_FLAGS\} -t $NUM_THREADS\"\) >>config.cmake
     echo set\(USE_CUDA ON\) >>config.cmake
     echo set\(USE_CUBLAS ON\) >>config.cmake

ci/task/test_model_compile.sh

@@ -9,11 +9,11 @@ pip install --force-reinstall wheels/*.whl
 
 if [[ ${GPU} == cuda* ]]; then
     TARGET=cuda
-    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-cu123
+    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-cu128
     export LD_LIBRARY_PATH=/usr/local/cuda/compat/:$LD_LIBRARY_PATH
 elif [[ ${GPU} == rocm* ]]; then
     TARGET=rocm
-    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-rocm57
+    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-rocm63
 elif [[ ${GPU} == metal ]]; then
     TARGET=metal
     pip install --pre -U --force-reinstall -f https://mlc.ai/wheels mlc-ai-nightly-cpu

ci/task/test_unittest.sh

@@ -8,7 +8,7 @@ if [[ -n ${MLC_CI_SETUP_DEPS:-} ]]; then
     # Install dependency
     pip install --force-reinstall wheels/*.whl
     pip install --quiet pytest
-    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-cu123
+    pip install --pre -U --no-index -f https://mlc.ai/wheels mlc-ai-nightly-cu128
     export LD_LIBRARY_PATH=/usr/local/cuda/compat/:$LD_LIBRARY_PATH
 fi
 

cmake/gen_cmake_config.py

@@ -83,12 +83,12 @@
     if use_flashInfer:
         while True:
             user_input = input("Enter your CUDA compute capability: ")
-            if user_input in ["80", "86", "89", "90"]:
+            if user_input in ["80", "86", "89", "90", "100", "120"]:
                 cmake_config_str += f"set(FLASHINFER_CUDA_ARCHITECTURES {user_input})\n"
                 cmake_config_str += f"set(CMAKE_CUDA_ARCHITECTURES {user_input})\n"
                 break
             else:
-                print(f"Invalid input: {user_input}. FlashInfer requires 80, 86, 89, or 90.")
+                print(f"Invalid input: {user_input}. FlashInfer requires 80, 86, 89, 90, 100 or 120")
 
     print("\nWriting the following configuration to config.cmake...")
     print(cmake_config_str)

docs/install/tvm.rst

@@ -53,6 +53,13 @@ A nightly prebuilt Python package of Apache TVM Unity is provided.
               conda activate your-environment
               python -m pip install --pre -U -f https://mlc.ai/wheels mlc-ai-nightly-cu123
 
+         .. tab:: CUDA 12.8
+
+            .. code-block:: bash
+
+              conda activate your-environment
+              python -m pip install --pre -U -f https://mlc.ai/wheels mlc-ai-nightly-cu128
+
          .. tab:: ROCm 6.1
 
             .. code-block:: bash
@@ -67,6 +74,13 @@ A nightly prebuilt Python package of Apache TVM Unity is provided.
               conda activate your-environment
               python -m pip install --pre -U -f https://mlc.ai/wheels mlc-ai-nightly-rocm62
 
+         .. tab:: ROCm 6.3
+
+            .. code-block:: bash
+
+              conda activate your-environment
+              python -m pip install --pre -U -f https://mlc.ai/wheels mlc-ai-nightly-rocm63
+
          .. tab:: Vulkan
 
             Supported in all Linux packages.