Cutlass dsl 4.5 bump

docker/install/install_python_packages.sh

@@ -19,6 +19,8 @@
 set -e
 set -u
 
+pip3 install --upgrade "setuptools>=77" "pip>=24"
+
 # Accept CUDA version as parameter (e.g., cu126, cu128, cu129)
 CUDA_VERSION=${1:-cu128}
 
@@ -32,7 +34,7 @@ pip3 install responses pytest scipy build cuda-python nvshmem4py-cu12
 if [[ "$CUDA_VERSION" == *"cu13"* ]]; then
   pip3 install --upgrade cuda-python==13.0
   pip3 install --upgrade nvidia-cudnn-cu13
-  pip3 install --upgrade "nvidia-cutlass-dsl[cu13]>=4.4.2"
+  pip3 install --upgrade "nvidia-cutlass-dsl[cu13]>=4.5.0"
 else
   pip3 install --upgrade cuda-python==12.*
   pip3 install --upgrade nvidia-cudnn-cu12

pyproject.toml

@@ -14,18 +14,18 @@
 
 [project]
 name = "flashinfer-python"
+dynamic = ["version", "dependencies"]
 description = "FlashInfer: Kernel Library for LLM Serving"
 requires-python = ">=3.10,<4.0"
 authors = [{ name = "FlashInfer team" }]
 license = "Apache-2.0"
 readme = "README.md"
 urls = { Homepage = "https://github.com/flashinfer-ai/flashinfer" }
-dynamic = ["dependencies", "version"]
 license-files = ["LICENSE", "LICENSE*.txt"]
 
 [project.optional-dependencies]
-cu12 = ["nvidia-cutlass-dsl>=4.4.2"]
-cu13 = ["nvidia-cutlass-dsl[cu13]>=4.4.2"]
+cu12 = ["nvidia-cutlass-dsl>=4.5.0"]
+cu13 = ["nvidia-cutlass-dsl[cu13]>=4.5.0"]
 
 [project.scripts]
 flashinfer = "flashinfer.__main__:cli"

requirements.txt

@@ -5,7 +5,6 @@ einops
 ninja
 numpy
 nvidia-cudnn-frontend>=1.13.0
-nvidia-cutlass-dsl>=4.4.2
 nvidia-ml-py
 packaging>=24.2
 requests

scripts/test_utils.sh

@@ -165,6 +165,17 @@ install_and_verify() {
         # Install precompiled kernels if enabled
         install_precompiled_kernels
 
+        # Sync dependencies from the branch's requirements.txt
+        pip install -r requirements.txt
+
+        # Install nvidia-cutlass-dsl with the correct CUDA extra to avoid
+        # version skew between libs-base and libs-cu13.
+        if [[ "${CUDA_VERSION}" == *"cu13"* ]] || [[ "${CUDA_VERSION}" == "13."* ]]; then
+            pip install --upgrade "nvidia-cutlass-dsl[cu13]>=4.5.0"
+        else
+            pip install --upgrade "nvidia-cutlass-dsl>=4.5.0"
+        fi
+
         # Install local python sources
         pip install -e . -v --no-deps
         echo ""

tests/gemm/test_mm_fp4.py

@@ -37,8 +37,8 @@ def _test_mm_fp4(
     if backend == "b12x":
         if not use_128x4_sf_layout:
             pytest.skip("b12x backend only supports 128x4 SF layout")
-        if compute_capability[0] != 12 or compute_capability[1] != 0:
-            pytest.skip("b12x backend only supports SM120 GPUs.")
+        if compute_capability[0] != 12:
+            pytest.skip("b12x backend only supports SM120/SM121 GPUs.")
         if not use_nvfp4:
             pytest.skip("b12x backend only supports NVFP4 (sf_vec_size=16).")
         if torch.version.cuda and int(torch.version.cuda.split(".")[0]) < 13:

tests/moe/test_b12x_fused_moe.py

@@ -65,14 +65,6 @@ def _cuda_13_or_newer():
         return False
 
 
-def _is_sm121():
-    """Check whether the current device reports compute capability SM121."""
-    if not torch.cuda.is_available():
-        return False
-    props = torch.cuda.get_device_properties(0)
-    return props.major == 12 and props.minor == 1
-
-
 # Skip decorators
 cute_dsl_available = pytest.mark.skipif(
     not is_cute_dsl_available(), reason="CuteDSL not available"
@@ -85,10 +77,6 @@ def _is_sm121():
     not _cuda_13_or_newer(),
     reason="b12x fused MoE requires CUDA 13 or later",
 )
-not_sm121 = pytest.mark.skipif(
-    _is_sm121(),
-    reason="b12x fused MoE is not supported on SM121",
-)
 
 
 # =============================================================================
@@ -532,7 +520,6 @@ def create_relu2_moe_tensors(
 @cute_dsl_available
 @sm120_required
 @cuda_13_required
-@not_sm121
 class TestB12xFunctional:
     """Tests for the functional API: b12x_fused_moe."""
 
@@ -613,7 +600,6 @@ def test_numerical_accuracy(
 @cute_dsl_available
 @sm120_required
 @cuda_13_required
-@not_sm121
 class TestB12xWrapper:
     """Tests for the wrapper API: B12xMoEWrapper."""
 
@@ -795,7 +781,6 @@ def test_wrapper_cuda_graph(self, num_tokens: int, num_experts: int):
 @cute_dsl_available
 @sm120_required
 @cuda_13_required
-@not_sm121
 class TestB12xApiConsistency:
     """Tests verifying consistency between b12x functional and wrapper APIs."""
 
@@ -874,7 +859,6 @@ def test_functional_vs_wrapper_output(self):
 @cute_dsl_available
 @sm120_required
 @cuda_13_required
-@not_sm121
 class TestMicroKernel:
     """Tests for the micro kernel path (routed_rows <= 20-40).
 
@@ -1076,7 +1060,6 @@ def test_micro_single_token_unique_path(self):
 @cute_dsl_available
 @sm120_required
 @cuda_13_required
-@not_sm121
 class TestRelu2Activation:
     """Tests for ReLU2 activation (non-gated, Nemotron-Super)."""