build: CUTLASS DSL 4.7.1 to match B12X

ci/cuda-versions.json

@@ -1,9 +1,9 @@
 {
   "dependency_policy": {
     "nvidia-cutlass-dsl": {
-      "provider_build_specifier": "==4.6.2",
-      "cuda_extra_specifier": "==4.6.2",
-      "ci_image_specifier": "==4.6.2",
+      "provider_build_specifier": "==4.7.1",
+      "cuda_extra_specifier": "==4.7.1",
+      "ci_image_specifier": "==4.7.1",
       "cuda_major_extras": {
         "13": ["cu13"]
       }

ci/lil_wheels/Dockerfile

@@ -21,6 +21,11 @@ ENV PATH=/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/
     BUILD_NCCL_EP=0 \
     BUILD_NIXL_EP=0
 
+# The builder ships CUTLASS DSL 4.6.2; B12X and the runtime image use 4.7.1.
+COPY ci/lil_wheels/cutlass-dsl.lock /tmp/cutlass-dsl.lock
+RUN --mount=type=cache,id=lil-flashinfer-pip-cu134-py312,target=/root/.cache/pip,sharing=locked \
+    python -m pip install --no-deps --require-hashes -r /tmp/cutlass-dsl.lock
+
 RUN python - <<'PY'
 import importlib.metadata
 import torch
@@ -40,8 +45,8 @@ expected = {
     "numpy": "2.1.0",
     "nvidia-ml-py": "13.610.43",
     "apache-tvm-ffi": "0.1.13.post3",
-    "nvidia-cutlass-dsl": "4.6.2",
-    "nvidia-cutlass-dsl-libs-cu13": "4.6.2",
+    "nvidia-cutlass-dsl": "4.7.1",
+    "nvidia-cutlass-dsl-libs-cu13": "4.7.1",
 }
 for package, version in expected.items():
     assert importlib.metadata.version(package) == version, package

ci/lil_wheels/cutlass-dsl.lock

@@ -0,0 +1,45 @@
+# CUTLASS DSL 4.7.1 for the wheel builder, which ships 4.6.2 (NGC PyTorch 26.08).
+# Hashes match blackwell-llm-docker's runtime lock (qwen38-runtime.lock).
+nvidia-cutlass-dsl==4.7.1 \
+    --hash=sha256:765850c74868a0b5a02fefbfc1b380842b64e5f7a1c8b1655e322a3b14c01726
+nvidia-cutlass-dsl-libs-base==4.7.1 \
+    --hash=sha256:06658e691eb1047c83be48ba501bf73f2ea7ecdd8c1ee5f9f26db587908e43ed \
+    --hash=sha256:0776506f666e1c6834a8a9195ef20299733ae013d234726829565be5d8509942 \
+    --hash=sha256:094579fb7700336a8d7e532c154777e99f1bbd5f54ec82cb9537a3584add9fe8 \
+    --hash=sha256:115608a7a514309e5a8448c6421a8597217c58760a3cadf2b781b37dd7c4d821 \
+    --hash=sha256:298e73ad0f1b0484e4b8438e9daae38732a6cadc9f4023475e5b0c2868985aba \
+    --hash=sha256:36192b4d0f72c4f47b39def20f4fda71058005617b7204395570768827524982 \
+    --hash=sha256:431e3648b84ad4a81a8d3b1fa7e445326cb5a524e64f6c22b4d191a76455bdf9 \
+    --hash=sha256:61097c5c355e1e42f8531efd4e8ead0ba61ae1ab6b0b9acfcbf2ca665b3dae5a \
+    --hash=sha256:93880e666f17513d5878f3b2ee76e9177a0acd59673dd27cb2c34dbdf5fe89f0 \
+    --hash=sha256:c2b67dacabfc905e4d132784fe4848647f40b0b97b1cdde3ce4a482642536b84 \
+    --hash=sha256:d01420777ac1ff22ee0c7842b6779fa58271b37c9a57cbd0bfff3eea932d35c7 \
+    --hash=sha256:e86c55a5010a7edfc067648a75cb29d0c5bd67d75210168cf507b0d0e4d15ba7
+nvidia-cutlass-dsl-libs-core==4.7.1 \
+    --hash=sha256:c65662abc9ab18fd997c57547895736928e559f4624bbe2aa1aa91418ea014ca
+nvidia-cutlass-dsl-libs-cu12==4.7.1 \
+    --hash=sha256:06f53df4ada14043b459993a6dc469d0505ad9a144d101b5f40461a58a8e80e3 \
+    --hash=sha256:10f98e415188e582a0159ed1bb375f4f18e5652bac286621e4362c0ad51ee845 \
+    --hash=sha256:49d556f2be188d75810fa8c547ad15dba40685138ac27d5441844182ac4955a9 \
+    --hash=sha256:5c1087e1d514b4afcaea9296e5f19c8e678aba93ee485e56f681fa7169e7a861 \
+    --hash=sha256:77684a86191195b2d6eb5f7088bb9bf25e45db58ceb44a9c02ffb0787bb95fb9 \
+    --hash=sha256:94e6da2629a3182f5b5810af22021d0e36d1ccf323bb8c93e71f1267ca7f82a4 \
+    --hash=sha256:b34a933cde1e5cf2df1f3cfd78efbd5484ea33dd0c97be201c057e168af79c10 \
+    --hash=sha256:c8a0f93a1cde41404696b3194dd1f589c6f3a7b426bcabe97845f8fe96f949f5 \
+    --hash=sha256:d3708b78b9b70dc1b29c9a48a7030d14a1a8c17b52f3f2d088db3d73c603bec8 \
+    --hash=sha256:f37fcbd94aad66b016dcea483db50ddb0fcaab7130c538df20ef29ddeeb8ba85 \
+    --hash=sha256:f57a3955c17b7dcd5472aa47c46513e5f3914e1325fc308afca1b9063a8f897b \
+    --hash=sha256:f5ab528aeb4548261fab7f393f2d119909f792c7165c986e98479ce2345a125d
+nvidia-cutlass-dsl-libs-cu13==4.7.1 \
+    --hash=sha256:20c14e16125cb8e36bc9947a9f0a57a660f44651365337b911c84abde4967f5f \
+    --hash=sha256:2f86fe8d1a7ef510b853a3800cbf324401a12fdc770cccf4ac5e20a9d0c7d1bd \
+    --hash=sha256:3aa56ee073da8584c84a3eac71dee32e6cc5ae6be958f8b589ca6dae276f299f \
+    --hash=sha256:425676733c1a1751e5fd4fc4ddd4ec6fbe6760d631ce12ed287dd63cf8c3646e \
+    --hash=sha256:4474ca7a9ea274f2977ee6fca7618c7611fcbad25bcdf756770458c8ad6a84a2 \
+    --hash=sha256:575e9558dfeca44402f53fe3d07e8cdfdc8cc4e847706213a5cc1bd7953cbdf1 \
+    --hash=sha256:67d363a9603cebdb847698bc6fc18bc14ba94f5c0f45c20b1138a316e068c7be \
+    --hash=sha256:7aaf4380a338ae36c53512f66b89470de6ee1f04f38163a89cd25e6dee92fe87 \
+    --hash=sha256:813ee19f8c8b6d39284a24ae99dd9a9b496b33c99270b2cf9548e79f49178405 \
+    --hash=sha256:990baa4dc8b6b0c7710afc89411e8eb05099a4c9bd01162bcbd365bd876edf59 \
+    --hash=sha256:a1d23cef44675e63825c3f5d773e7b4228802ca155eab61c0bc5e551caf35ad2 \
+    --hash=sha256:e675bf04dc6c91f7c4d5540699b7a031f2b8b44cd2687b32c20892f112f7e0d9

ci/lil_wheels/runtime.lock

@@ -5,7 +5,7 @@ cuda.version=13.4
 pytorch.runtime-version=2.14.0a0+4fdf77b940.nv26.08
 pytorch.version=2.14.0a0+4fdf77b940.nv26.8.63802676
 pytorch.commit=4fdf77b940a7e61d854b9be527884e38f32d335a
-cutlass-dsl.version=4.6.2
+cutlass-dsl.version=4.7.1
 pip.version=26.2.1
 setuptools.version=81.0.0
 packaging.version=26.2

pyproject.toml

@@ -25,11 +25,11 @@ urls = { Homepage = "https://github.com/flashinfer-ai/flashinfer" }
 license-files = ["LICENSE", "LICENSE*.txt"]
 
 [project.optional-dependencies]
-# The b12x snapshot requires the 4.6.2 compiler and matching runtime libraries.
+# The b12x snapshot requires the 4.7.1 compiler and matching runtime libraries.
 # Keep both CUDA extras installable with those base pins; selecting an extra
 # must not upgrade the compiler independently of the snapshot.
-cu12 = ["nvidia-cutlass-dsl==4.6.2"]
-cu13 = ["nvidia-cutlass-dsl[cu13]==4.6.2"]
+cu12 = ["nvidia-cutlass-dsl==4.7.1"]
+cu13 = ["nvidia-cutlass-dsl[cu13]==4.7.1"]
 # DEPRECATED alias, kept so existing `pip install ".[nvep]"` commands keep
 # working. The moe_ep runtime deps (cuda-python, nccl4py) are now part of the
 # BASE dependencies (requirements.txt) and the NIXL-EP submodule build runs by

requirements.txt

@@ -30,11 +30,11 @@ ninja
 numpy
 nvidia-cudnn-frontend>=1.25.0
 # Preserve the b12x snapshot's compiler and complete matching library set.
-nvidia-cutlass-dsl==4.6.2
-nvidia-cutlass-dsl-libs-base==4.6.2
-nvidia-cutlass-dsl-libs-core==4.6.2
-nvidia-cutlass-dsl-libs-cu12==4.6.2
-nvidia-cutlass-dsl-libs-cu13==4.6.2
+nvidia-cutlass-dsl==4.7.1
+nvidia-cutlass-dsl-libs-base==4.7.1
+nvidia-cutlass-dsl-libs-core==4.7.1
+nvidia-cutlass-dsl-libs-cu12==4.7.1
+nvidia-cutlass-dsl-libs-cu13==4.7.1
 nvidia-ml-py
 packaging>=24.2
 requests