[Infra] Enable CUDA forward compat in devcontainers.

.devcontainer/cccl-entrypoint.sh

@@ -8,6 +8,16 @@ if [[ -n "${GITHUB_ACTIONS:-}" ]]; then
     echo "::group::Initializing devcontainer..."
 fi
 
+# The CUDA forward compatibility (cuda-compat) libraries are only supported on
+# data center GPUs; activating them on other GPUs breaks CUDA. GPU jobs set
+# DATA_CENTER_GPU ('1' or '0') via the CI matrix, based on the runner's GPU.
+# Local users with a data center GPU can opt in with:
+#   .devcontainer/launch.sh --env DATA_CENTER_GPU=1
+if [[ "${DATA_CENTER_GPU:-0}" = "1" ]]; then
+    setup-cuda-compat;
+else
+    echo "DATA_CENTER_GPU is '${DATA_CENTER_GPU:-0}', skipping cuda-compat configuration.";
+fi
 devcontainer-utils-post-create-command;
 devcontainer-utils-init-git;
 devcontainer-utils-post-attach-command;

.github/actions/workflow-build/build-workflow.py

@@ -281,6 +281,12 @@ def get_gpu(gpu_string):
     if "gpu_count" not in result:
         result["gpu_count"] = 1
 
+    # dc (data center GPU) defaults to False. Used to decide whether the CUDA
+    # forward compatibility (cuda-compat) libraries may be activated in the
+    # container; they are only supported on data center GPUs.
+    if "dc" not in result:
+        result["dc"] = False
+
     # runner_id defaults to the GPU key. Used in runner label templates as {gpu_id}.
     if "runner_id" not in result:
         result["runner_id"] = gpu_string
@@ -597,7 +603,21 @@ def generate_dispatch_job_image(matrix_job, job_type):
 
 
 def generate_dispatch_job_environment(matrix_job, job_type):
-    return json.dumps(matrix_job.get("environment") or [])
+    environment = list(matrix_job.get("environment") or [])
+
+    # Inject the GPU's data center classification into every GPU job. The
+    # devcontainer entrypoint uses this to decide whether to run the
+    # setup-cuda-compat script, since the CUDA forward compatibility
+    # (cuda-compat) libraries are only supported on data center GPUs.
+    # Explicitly emit both cases ('1' and '0') so mock jobs reproduce CI.
+    # Skip if the tag already set the variable to allow manual overrides.
+    if get_job_type_info(job_type)["gpu"] and not any(
+        env.startswith("DATA_CENTER_GPU=") for env in environment
+    ):
+        gpu = get_gpu(matrix_job["gpu"])
+        environment.append(f"DATA_CENTER_GPU={'1' if gpu['dc'] else '0'}")
+
+    return json.dumps(environment)
 
 
 def generate_dispatch_job_command(matrix_job, job_type):

ci/matrix.yaml

@@ -931,18 +931,24 @@ projects:
 #              Defaults to the GPU key. Useful when the key includes a suffix
 #              (e.g. 'h100_2gpu') but the runner label should use the base name
 #              (e.g. 'h100').
+# dc -> Whether the GPU is a data center product. Injected into all GPU jobs as
+#       the DATA_CENTER_GPU environment variable ('1' or '0'), which the
+#       devcontainer entrypoint uses to decide whether to run the
+#       setup-cuda-compat script. The CUDA forward compatibility (cuda-compat)
+#       libraries are only supported on data center GPUs and must not be
+#       activated on other GPUs.
 gpus:
-  t4:         { name: 'T4',          sm: 75,  runner: 't4-latest',         gpu_count: 1 } # 16 GB,  10 runners
-  rtx2080:    { name: 'RTX2080',     sm: 75,  runner: 'rtx2080-latest',    gpu_count: 1 } #  8 GB,  12 runners
-  rtxa6000:   { name: 'RTXA6000',    sm: 86,  runner: 'rtxa6000-latest',   gpu_count: 1 } # 48 GB,  12 runners
-  l4:         { name: 'L4',          sm: 89,  runner: 'l4-latest',         gpu_count: 1 } # 24 GB,  48 runners
-  rtx4090:    { name: 'RTX4090',     sm: 89,  runner: 'rtx4090-latest',    gpu_count: 1 } # 24 GB,  10 runners
-  h100:       { name: 'H100',        sm: 90,  runner: 'h100-latest',       gpu_count: 1 } # 80 GB,  16 runners
-  h100_2gpu:  { name: 'H100 2-GPU',  sm: 90,  runner: 'h100-latest',       gpu_count: 2, runner_id: 'h100' } # 2 x 80 GB
-  gb300:      { name: 'GB300',       sm: 103, runner: 'gb300-latest',      gpu_count: 1 }
+  t4:         { name: 'T4',          sm: 75,  runner: 't4-latest',         gpu_count: 1, dc: true  } # 16 GB,  10 runners
+  rtx2080:    { name: 'RTX2080',     sm: 75,  runner: 'rtx2080-latest',    gpu_count: 1, dc: false } #  8 GB,  12 runners
+  rtxa6000:   { name: 'RTXA6000',    sm: 86,  runner: 'rtxa6000-latest',   gpu_count: 1, dc: true } # 48 GB,  12 runners
+  l4:         { name: 'L4',          sm: 89,  runner: 'l4-latest',         gpu_count: 1, dc: true  } # 24 GB,  48 runners
+  rtx4090:    { name: 'RTX4090',     sm: 89,  runner: 'rtx4090-latest',    gpu_count: 1, dc: false } # 24 GB,  10 runners
+  h100:       { name: 'H100',        sm: 90,  runner: 'h100-latest',       gpu_count: 1, dc: true  } # 80 GB,  16 runners
+  h100_2gpu:  { name: 'H100 2-GPU',  sm: 90,  runner: 'h100-latest',       gpu_count: 2, runner_id: 'h100', dc: true } # 2 x 80 GB
+  gb300:      { name: 'GB300',       sm: 103, runner: 'gb300-latest',      gpu_count: 1, dc: true  }
   # Very small number of runners on loan from cuda-python while we wait for our order to arrive.
   # Limit jobs on these:
-  rtxpro6000: { name: 'RTXPRO6000',  sm: 120, runner: 'rtxpro6000-latest', gpu_count: 1 }
+  rtxpro6000: { name: 'RTXPRO6000',  sm: 120, runner: 'rtxpro6000-latest', gpu_count: 1, dc: true  }
 
 # Tags are used to define a `matrix job` in the workflow section.
 #

docs/infrastructure/ci/references/matrix_yaml.rst

@@ -37,7 +37,7 @@ Top-level keys
    * - ``projects``
      - Project definitions: supported standards, display name, and ``job_map`` expansions.
    * - ``gpus``
-     - GPU runner pools and their ``sm`` value.
+     - GPU runner pools and their ``sm`` value and ``dc`` (data center GPU) classification.
    * - ``tags``
      - The fields a job entry accepts, with ``required`` flags and defaults.
 
@@ -119,6 +119,10 @@ Computed internally, not user-specified:
 - ``gpu`` requirement, ``cuda_ext``, ``name``, and ``invoke`` script details - also taken from the ``jobs`` section.
 - ``force_producer_ctk`` — set in the ``jobs`` section to pin a producer build's CTK independent
   of the consumer's ``ctk`` tag. Used mainly for python packaging special cases.
+- ``DATA_CENTER_GPU`` — injected into the ``environment`` of every GPU job as ``1`` or ``0``, taken
+  from the selected GPU's ``dc`` field. The devcontainer entrypoint gates the CUDA forward
+  compatibility (``cuda-compat``) setup on this variable; forward compatibility is only
+  supported on data center GPUs.
 
 Annotated example
 ~~~~~~~~~~~~~~~~~