Bump CuTe DSL to 4.6.0 (#31714)

python/pyproject.toml

@@ -30,7 +30,7 @@ dependencies = [
   "easydict",  # Required by remote model code (e.g. DeepSeek-OCR) loaded via trust_remote_code; validated by transformers 5.4+ check_imports
   "einops",
   "fastapi",
-  "flash-attn-4==4.0.0b15",
+  "flash-attn-4>=4.0.0b18",
   "flashinfer_python[cu13]==0.6.14", # keep it aligned with jit-cache version in Dockerfile
   "gguf",
   "helion==0.2.6",
@@ -45,7 +45,7 @@ dependencies = [
   "ninja",
   "numba==0.65.1",
   "numpy",
-  "nvidia-cutlass-dsl[cu13]==4.5.2",
+  "nvidia-cutlass-dsl[cu13]==4.6.0",
   "nvidia-mathdx==25.6.0",
   "nvidia-ml-py",
   "openai==2.6.1",
@@ -62,7 +62,7 @@ dependencies = [
   "pydantic",
   "python-multipart",
   "pyzmq>=25.1.2",
-  "quack-kernels>=0.4.1",
+  "quack-kernels>=0.6.1",
   "requests",
   "scipy",
   "sentencepiece",

python/sglang/kernels/ops/attention/cute_utils/_tcgen05.py

@@ -11,8 +11,8 @@ from cutlass.cutlass_dsl import dsl_user_op
 
 NVVM_CTA_GROUP_MAP = [
     None,
-    nvvm.Tcgen05GroupKind.CTA_1,
+    nvvm.CTAGroupKind.CTA_1,
-    nvvm.Tcgen05GroupKind.CTA_2,
+    nvvm.CTAGroupKind.CTA_2,
 ]
 LDST_MAP = {
     "32x32b": (nvvm.Tcgen05LdStShape.SHAPE_32X32B, 1),
@@ -134,7 +134,7 @@ def commit(mbar, cta_mask=None, cta_group: int = 1, *, loc=None, ip=None):
     mbar_llvm = mbar.to_llvm_ptr(loc=loc, ip=ip)
     group = NVVM_CTA_GROUP_MAP[cta_group]
     if cutlass.const_expr(cta_mask is not None):
-        nvvm.tcgen05_commit_arrive(
+        nvvm.tcgen05_commit(
             mbar_llvm,
             multicast_mask=cta_mask.ir_value(loc=loc, ip=ip),
             group=group,
@@ -142,7 +142,7 @@ def commit(mbar, cta_mask=None, cta_group: int = 1, *, loc=None, ip=None):
             ip=ip,
         )
     else:
-        nvvm.tcgen05_commit_arrive(mbar_llvm, group=group, loc=loc, ip=ip)
+        nvvm.tcgen05_commit(mbar_llvm, group=group, loc=loc, ip=ip)
 
 
 @dsl_user_op

python/sglang/multimodal_gen/test/test_utils.py

@@ -34,7 +34,7 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
-SGL_TEST_FILES_CI_DATA_REVISION = "d51ca9623e0bb27087da243a44c942fdda5aafe5"
+SGL_TEST_FILES_CI_DATA_REVISION = "320949ecc2587474a2f535229ffc8f47ed16ee51"
 
 if current_platform.is_npu():
     SGL_TEST_FILES_CI_DATA_REVISION = "6b62f4b6825c76a25fd2ba28248df68f2b400e65"

python/sglang/srt/utils/common.py

@@ -1820,7 +1820,7 @@ def suppress_noisy_warnings():
     cutlass_dsl_noisy = {
         (
             DeprecationWarning,
-            "Use explicit `struct.scalar.ptr` for pointer instead.",
+            "Using `struct.scalar` as pointer is deprecated.",
         ),
         (
             UserWarning,

scripts/ci/cuda/ci_install_dependency.sh

@@ -417,34 +417,6 @@ download_flashinfer_cache() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
-force_reinstall_cutlass_dsl_libs_cu13() {
-    # nvidia-cutlass-dsl[cu13] has additive PyPI extras: installing it pulls in
-    # both -libs-base and -libs-cu13. The two wheels ship intentionally-different
-    # content for the same paths (cutlass/_mlir/dialects/_gpu_ops_gen.py and
-    # cutlass/_mlir/_mlir_libs/_cutlass_ir.cpython-*.so) -- each Python wrapper
-    # is paired with a matching pybind11 .so. If install order leaves the .py
-    # from one wheel and the .so from the other, GPUModuleOp.__init__ raises
-    # TypeError: incompatible function arguments at kernel-compile time.
-    #
-    # Force-reinstall -libs-cu13 LAST so both files come from the same wheel
-    # (BOTH-cu13 state), eliminating the mismatch. The version is parsed from
-    # pyproject.toml so this stays in sync with whatever nvidia-cutlass-dsl
-    # version the project pins.
-    if [ "$CU_MAJOR" != "13" ]; then
-        return
-    fi
-
-    CUTLASS_DSL_VERSION=$(grep -Po -m1 'nvidia-cutlass-dsl(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' "${REPO_ROOT}/python/pyproject.toml" || echo "")
-    if [ -z "$CUTLASS_DSL_VERSION" ]; then
-        echo "WARNING: could not detect nvidia-cutlass-dsl version from pyproject.toml; skipping libs-cu13 force-reinstall"
-        return
-    fi
-
-    $PIP_CMD install --force-reinstall --no-deps "nvidia-cutlass-dsl-libs-cu13==${CUTLASS_DSL_VERSION}" $PIP_INSTALL_SUFFIX
-
-    mark_step_done "${FUNCNAME[0]}"
-}
-
 stabilize_flashinfer_jit_paths() {
     # In venv mode, FlashInfer JIT writes build.ninja with hardcoded -isystem
     # paths. Per-job venvs get unique paths, but the JIT cache is shared on the
@@ -624,7 +596,6 @@ main() {
     install_sglang_router
     install_flashinfer_cubin
     download_flashinfer_cache
-    force_reinstall_cutlass_dsl_libs_cu13
     stabilize_flashinfer_jit_paths
     install_extra_deps
     install_test_tools