[CI] Re-enable GB300

.github/workflows/_pr-test-stage.yml

@@ -107,11 +107,8 @@ jobs:
     # caller that declares its runners as a matrix cannot, since job-level `if`
     # sees github/needs/vars/inputs but not `matrix`. Callers that never declare
     # the input read back null and match every runner_config.
-    # Temporarily skip the unavailable GB300 runner before GitHub tries to
-    # allocate it. Covers both callers: pr-test and nightly.
     if: |
       always() &&
-      inputs.runner_config != '4-gpu-gb300' &&
       (fromJson(inputs.caller_inputs).runner_filter == null ||
        fromJson(inputs.caller_inputs).runner_filter == '' ||
        fromJson(inputs.caller_inputs).runner_filter == 'all' ||

.github/workflows/nightly-test-nvidia.yml

@@ -23,7 +23,7 @@ on:
           - '2-gpu-large'
           - '4-gpu-h100'
           - '4-gpu-b200'
-          # - '4-gpu-gb300'  # Temporarily disabled while the GB300 runner is unavailable.
+          - '4-gpu-gb300'
           - '8-gpu-h200'
           - '8-gpu-b200'
           - 'diffusion'

.github/workflows/release-whl-deepgemm.yml

@@ -120,11 +120,10 @@ jobs:
           # - arch_label: sm120
           #   runner: 1-gpu-5090
           #   wheel_arch: x86_64
-          # Temporarily disabled while the GB300 runner is unavailable; the aarch64 wheel still publishes untested.
-          # - arch_label: sm100-aarch64
-          #   runner: 4-gpu-gb300
-          #   wheel_arch: aarch64
-          #   timeout: 180
+          - arch_label: sm100-aarch64
+            runner: 4-gpu-gb300
+            wheel_arch: aarch64
+            timeout: 180
     runs-on: ${{ matrix.runner }}
     timeout-minutes: ${{ matrix.timeout }}
     steps:

.github/workflows/rerun-test.yml

@@ -119,8 +119,7 @@ jobs:
   rerun-test-cuda:
     name: rerun-test-cuda (partition ${{ matrix.partition }})
     needs: prepare-partitions
-    # Temporarily reject reruns targeting the unavailable GB300 runner.
-    if: inputs.mode == 'cuda' && inputs.runs_on != '4-gpu-gb300'
+    if: inputs.mode == 'cuda'
     strategy:
       fail-fast: false
       matrix: ${{ fromJson(needs.prepare-partitions.outputs.matrix) }}
@@ -163,10 +162,6 @@ jobs:
 
       - name: Install dependencies
         timeout-minutes: ${{ fromJson(inputs.install_timeout) }}
-        env:
-          # Public-read-only PAT: the checkout above leaves no job token for the
-          # install script's github.com clones. Unset -> anonymous.
-          GH_TOKEN: ${{ secrets.GH_PAT_FOR_CI_CLONE }}
         run: |
           if [[ "${{ inputs.runs_on }}" == "1-gpu-5090" ]]; then
             source /etc/profile.d/sglang-ci.sh
@@ -294,8 +289,6 @@ jobs:
 
       - name: Install dependencies (diffusion)
         timeout-minutes: 20
-        env:
-          GH_TOKEN: ${{ secrets.GH_PAT_FOR_CI_CLONE }}
         run: bash scripts/ci/cuda/ci_install_dependency.sh diffusion
 
       - name: Run test

scripts/ci/utils/slash_command_handler.py

@@ -24,10 +24,6 @@
 # (single path component under that dir, no whitespace/operators).
 _ALLOWED_INSTALL_SCRIPT = re.compile(r"^scripts/ci/cuda/[\w.-]+\.sh$")
 
-# Temporarily disabled while their runners are unavailable. rerun-test.yml
-# skips these too; rejecting here gives a clear reply instead of a skipped run.
-_DISABLED_RUNNER_CONFIGS = {"4-gpu-gb300"}
-
 # Configuration
 PERMISSIONS_FILE_PATH = ".github/CI_PERMISSIONS.json"
 TEST_GROUPS_FILE_PATH = "scripts/ci/rerun_test_groups.json"
@@ -996,12 +992,6 @@ def _resolve_runner_config(rc, full_path, suite):
             f"— not in scripts/ci/runner_configs.yml.\n\n"
             f"Known runner_configs: {known}",
         )
-    if rc in _DISABLED_RUNNER_CONFIGS:
-        return _dispatch_err(
-            suite,
-            f"runner_config `{rc}` in `{full_path}` is temporarily disabled "
-            f"while its runners are unavailable.",
-        )
     install_script = cfg["install"]
     if not _ALLOWED_INSTALL_SCRIPT.match(install_script):
         return _dispatch_err(