feat(cutedsl): add CuTeDSL backend

.github/workflows/ci.yml

@@ -370,8 +370,27 @@ jobs:
             pytest --verbose --color=yes --durations=0 --showlocals --cache-clear
           )
           "${PYTEST[@]}" --maxfail=3 --numprocesses=4 \
+            --ignore=./python/jit/test_tilelang_jit_cutedsl.py \
             ./python
 
+      # CuTeDSL JIT tests require GEMM v1 (must be set before importing tilelang).
+      # Run them in a dedicated step to avoid changing the default GEMM selection
+      # (and to keep the rest of the CUDA tests on GEMM v2).
+      - name: Run CuTeDSL JIT tests (GEMM v1) with Python ${{ matrix.python-version }} (${{ matrix.runner.toolkit }})
+        id: cutedsl-tests
+        if: contains(matrix.runner.toolkit, 'CUDA')
+        env:
+          TILELANG_USE_GEMM_V1: "1"
+        run: |
+          cd testing
+          PYTEST=(
+            uv run --no-project -m --
+            pytest --verbose --color=yes --durations=0 --showlocals --cache-clear
+          )
+          # Avoid xdist contention on a single GPU by running this file in one worker.
+          "${PYTEST[@]}" --maxfail=3 --numprocesses=1 \
+            ./python/jit/test_tilelang_jit_cutedsl.py
+
       # AMD ROCm tests
       - name: Run ROCm tests with Python ${{ matrix.python-version }} (${{ matrix.runner.toolkit }})
         id: rocm-tests

CMakeLists.txt

@@ -215,7 +215,11 @@ elseif(USE_CUDA)
     src/runtime/runtime.cc
     src/target/ptx.cc
     src/target/codegen_cuda.cc
+    src/target/codegen_py.cc
+    src/target/codegen_utils.cc
+    src/target/codegen_cutedsl.cc
     src/target/rt_mod_cuda.cc
+    src/target/rt_mod_cutedsl.cc
   )
   list(APPEND TILE_LANG_SRCS ${TILE_LANG_CUDA_SRCS})
 

maint/scripts/run_local_ci_test.sh

@@ -14,7 +14,13 @@ cd examples
 python -m pytest -n 4 . --verbose --color=yes --durations=0 --showlocals --cache-clear
 cd ..
 
-# Run pytest in parallel (4 workers) for all tests in the testing/python directory
+# Run pytest in parallel (4 workers) for all tests in the testing/python directory.
+# IMPORTANT: CuTeDSL backend currently requires GEMM v1 (TILELANG_USE_GEMM_V1=1).
+# Do NOT export it globally here, or you'll silently change the default GEMM selection
+# for unrelated tests. Run the CuTeDSL JIT tests in a separate pytest invocation.
 cd testing/python
-python -m pytest -n 4 . --verbose --color=yes --durations=0 --showlocals --cache-clear
+python -m pytest -n 4 . --ignore=jit/test_tilelang_jit_cutedsl.py --verbose --color=yes --durations=0 --showlocals --cache-clear
+
+# CuTeDSL JIT tests (isolate env + avoid xdist contention on a single GPU)
+TILELANG_USE_GEMM_V1=1 python -m pytest -n 1 jit/test_tilelang_jit_cutedsl.py --verbose --color=yes --durations=0 --showlocals --cache-clear
 cd ..

requirements-test-cuda.txt

@@ -7,3 +7,5 @@
 # CUDA specific requirements
 flash-attn==2.5.8
 cuda-python==12.9.4
+# CuTeDSL (CUTLASS Python DSL with CuTe support)
+nvidia-cutlass-dsl>=4.3.1