[TRTLLM-13767][chore] upgrade CUTLASS DSL stack to 4.6.1

.claude/skills/kernel-cute-writing/references/api-core.md

@@ -126,8 +126,7 @@ cute.is_major(x, d)    # True if dimension d is stride-1
 ```python
 cute.make_tensor(ptr, layout)                  # From pointer + layout
 cute.make_identity_tensor(shape)               # Coordinate mapping tensor
-cute.make_rmem_tensor(layout)                  # Register memory tensor
-cute.make_fragment(shape, dtype)               # Register fragment
+cute.make_rmem_tensor(layout_or_shape, dtype)  # Register memory tensor
 ```
 
 ### Data Initialization

.claude/skills/kernel-cute-writing/references/concepts-tensors.md

@@ -81,8 +81,7 @@ mA = cute.make_tensor(ptr, layout=layout)
 
 ### Register/Fragment Tensors (Owning)
 ```python
-rmem = cute.make_rmem_tensor(layout)
-frag = cute.make_fragment(shape, dtype)
+rmem = cute.make_rmem_tensor(layout_or_shape, dtype)
 ```
 
 ## Accessing Tensors

constraints.txt

@@ -1,10 +1,18 @@
-# These vulnerabilities were inherited from the base image (pytorch:25.12-py3) and should be removed when the base image
-# is updated.
-# Upgrade base image nvidia-cutlass-dsl 4.3.5 to 4.4.2
-nvidia-cutlass-dsl>=4.4.2
+# setup.py appends these constraints to wheel install_requires, so every entry
+# must also be a valid TensorRT-LLM runtime dependency outside the base image.
+# Keep the base-image and wheel CuTe DSL stack aligned with requirements.txt.
+nvidia-cutlass-dsl==4.6.1
+# Keep the CuTe DSL and FlashAttention 4 runtime stack aligned with requirements.txt.
+apache-tvm-ffi==0.1.13.post2
+quack-kernels==0.6.2
+torch-c-dlpack-ext==0.1.3
+# CUTLASS DSL 4.6.1 libraries require protobuf 6.x.
+protobuf>=6.30.2,<7
 # The `nvidia-cutlass-dsl` package does not pin numpy at all, which can be problematic in certain CI
 # stages.
 numpy>=2.0.0,<2.4 # numba 0.63.1 requires numpy<2.4
+# The remaining vulnerabilities were inherited from the base image
+# (pytorch:25.12-py3) and can be removed after the base image is updated.
 # WAR against https://github.com/advisories/GHSA-qcq2-496w-v96p
 mistune>=3.3.0
 # WAR against https://github.com/advisories/GHSA-rch3-82jr-f9w9

docker/Dockerfile.multi

@@ -66,11 +66,13 @@ RUN --mount=type=bind,source=docker/common,target=/opt/docker/common \
 
 # Install constraints after install.sh so cleanup() doesn't delete the file mid-RUN
 COPY constraints.txt /tmp/constraints.txt
+# constraints.txt is also consumed as a pip constraint file, which forbids extras.
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip3 uninstall -y tornado black nbconvert pillow nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base numpy wandb || true && \
+    pip3 uninstall -y tornado black nbconvert pillow nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base \
+        nvidia-cutlass-dsl-libs-core nvidia-cutlass-dsl-libs-cu12 nvidia-cutlass-dsl-libs-cu13 numpy wandb || true && \
     # Remove any leftover namespace dirs or dist-info that pip missed
     rm -rf $(python3 -c "import site; print(site.getsitepackages()[0])")/nvidia_cutlass_dsl* && \
-    pip3 install -r /tmp/constraints.txt && \
+    pip3 install -r /tmp/constraints.txt "nvidia-cutlass-dsl[cu13]==4.6.1" && \
     rm /tmp/constraints.txt
 
 # Install UCX, NIXL, etcd

docker/common/install_fa4.sh

@@ -2,7 +2,7 @@
 
 set -ex
 
-FLASH_ATTN_4_VERSION="4.0.0b11"
+FLASH_ATTN_4_VERSION="4.0.0b19"
 
 if [ -n "${GITHUB_MIRROR}" ]; then
   export PIP_INDEX_URL="https://urm.nvidia.com/artifactory/api/pypi/pypi-remote/simple"

docs/source/installation/installation-guide.md

@@ -68,7 +68,8 @@ Before installing the latest version, uninstall any previous CUTLASS DSL install
 [CUTLASS DSL installation guide](https://docs.nvidia.com/cutlass/latest/media/docs/pythonDSL/quick_start.html#installation):
 
 ```bash
-pip3 uninstall nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base nvidia-cutlass-dsl-libs-cu13
+pip3 uninstall nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base \
+    nvidia-cutlass-dsl-libs-core nvidia-cutlass-dsl-libs-cu12 nvidia-cutlass-dsl-libs-cu13
 ```
 
 ```bash

jenkins/L0_Test.groovy

@@ -6325,7 +6325,11 @@ def launchTestJobs(pipeline, testFilter, globalVars)
                             trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get remove -y python3-pygments")
                             // Remove stale nvidia-cutlass-dsl from the base image to prevent namespace
                             // directory corruption when pip upgrades to the version required by tensorrt_llm.
-                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 uninstall -y nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base || true")
+                            trtllm_utils.llmExecStepWithRetry(
+                                pipeline,
+                                script: "pip3 uninstall -y nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base " +
+                                    "nvidia-cutlass-dsl-libs-core nvidia-cutlass-dsl-libs-cu12 " +
+                                    "nvidia-cutlass-dsl-libs-cu13 || true")
                             trtllm_utils.llmExecStepWithRetry(pipeline, script: 'rm -rf $(python3 -c "import site; print(site.getsitepackages()[0])")/nvidia_cutlass_dsl*')
                         }
                         trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get update && apt-get install -y python3-pip git rsync curl wget")

requirements.txt

@@ -76,17 +76,17 @@ xdsl>=0.59.0 # Optional: required for MLIR-based elementwise fusion in AutoDeplo
 tiktoken
 blobfile
 openai-harmony==0.0.4
-nvidia-cutlass-dsl[cu13]==4.5.0; python_version >= "3.10"
+nvidia-cutlass-dsl[cu13]==4.6.1; python_version >= "3.10" # required by Blackwell CuTe DSL kernels
 nvidia-matmul-heuristics==0.1.0.27; python_version >= "3.10" # analytic GEMM heuristics for CuTe DSL autotuner tactic pruning
-quack-kernels>=0.2.10; python_version >= "3.10" # required for MinimaxM3 MSA
+quack-kernels==0.6.2; python_version >= "3.10" # required for MinimaxM3 MSA and FlashAttention 4; pins CUTLASS DSL 4.6.1
 jinja2 # required for MinimaxM3 MSA
 plotly
 numexpr
 partial_json_parser
 mcp<2.0.0 # 2.0.0 removed mcp.server.fastmcp; scaffolding still uses the 1.x APIs
-apache-tvm-ffi==0.1.6 # used for reduce nvidia-cutlass-dsl host overhead
+apache-tvm-ffi==0.1.13.post2 # required by CUTLASS DSL 4.6.1's TVM-FFI provider
 torch-c-dlpack-ext==0.1.3 # used for reduce nvidia-cutlass-dsl host overhead, optional package for improved torch tensor calling perf
-flash-attn-4==4.0.0b11
+flash-attn-4==4.0.0b19
 mistral-common>=1.10.0
 torchao>=0.14.1,<0.16.0
 cuda-core