move runs_on + rdma into runner_configs.yml (#25264)

.github/workflows/_pr-test-check-changes.yml

@@ -29,8 +29,8 @@ on:
         value: ${{ jobs.run.outputs.partitions }}
       partition_model_sha:
         value: ${{ jobs.run.outputs.partition_model_sha }}
-      b200_runner:
+      runs_on_map:
-        value: ${{ jobs.run.outputs.b200_runner }}
+        value: ${{ jobs.run.outputs.runs_on_map }}
       enable_retry:
         value: ${{ jobs.run.outputs.enable_retry }}
       continue_on_error:
@@ -47,7 +47,7 @@ jobs:
       multimodal_gen: ${{ steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }}
       partitions: ${{ steps.partitions.outputs.partitions }}
       partition_model_sha: ${{ steps.partition-model-sha.outputs.sha }}
-      b200_runner: ${{ steps.set-runner.outputs.b200_runner }}
+      runs_on_map: ${{ steps.runner-map.outputs.runs_on_map }}
       enable_retry: ${{ steps.set-retry.outputs.enable_retry }}
       continue_on_error: ${{ steps.set-continue-on-error.outputs.continue_on_error }}
     steps:
@@ -174,6 +174,13 @@ jobs:
             echo "b200_runner=4-gpu-b200" >> $GITHUB_OUTPUT
           fi
 
+      - name: Build runs_on_map (resolves $b200_runner sentinel)
+        id: runner-map
+        run: |
+          python3 scripts/ci/runner_configs.py --map \
+            '${{ steps.set-runner.outputs.b200_runner }}' \
+            >> "$GITHUB_OUTPUT"
+
       - name: Enable retry for CI
         id: set-retry
         run: |

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

@@ -15,15 +15,11 @@ on:
         type: string
         required: true
       runner_config:
-        description: 'Key in scripts/ci/runner_configs.yml (install script / artifact version / install timeout).'
+        description: 'Key in scripts/ci/runner_configs.yml. Resolves install script, artifact version, install timeout, runs-on label, and rdma_devices.'
-        type: string
-        required: true
-      runs_on:
-        description: 'GHA runner label. B200 stages pass needs.check-changes.outputs.b200_runner for dynamic selection.'
         type: string
         required: true
       check_changes:
-        description: 'toJson(needs.check-changes.outputs). Read via fromJson(...).main_package / sgl_kernel / continue_on_error etc.'
+        description: 'toJson(needs.check-changes.outputs). Read via fromJson(...).main_package / sgl_kernel / continue_on_error / runs_on_map etc.'
         type: string
         required: true
       caller_inputs:
@@ -79,11 +75,11 @@ jobs:
       always() &&
       ((github.event_name == 'schedule' || fromJson(inputs.caller_inputs).test_parallel_dispatch == true) || (!failure() && !cancelled())) &&
       (fromJson(inputs.check_changes).main_package == 'true' || fromJson(inputs.check_changes).sgl_kernel == 'true')
-    runs-on: ${{ inputs.runs_on }}
+    # runs-on resolved from runs_on_map; check-changes already substituted
+    # $b200_runner (see runner_configs.py --map). rdma_devices is exported
+    # below in a setup step via $GITHUB_ENV.
+    runs-on: ${{ fromJson(fromJson(inputs.check_changes).runs_on_map)[inputs.runner_config] }}
     timeout-minutes: 240
-    env:
-      # Only stage-c-test-8-gpu-h20 needs the RDMA device list.
-      SGLANG_CI_RDMA_ALL_DEVICES: ${{ inputs.runner_config == '8-gpu-h20' && 'mlx5_1,mlx5_2,mlx5_3,mlx5_4' || '' }}
     strategy:
       fail-fast: false
       max-parallel: ${{ fromJson(inputs.partitions)[inputs.self_name].max_parallel }}
@@ -98,6 +94,9 @@ jobs:
         id: rc
         run: python3 scripts/ci/runner_configs.py '${{ inputs.runner_config }}' >> "$GITHUB_OUTPUT"
 
+      - name: Export rdma_devices to job env
+        run: echo "SGLANG_CI_RDMA_ALL_DEVICES=${{ steps.rc.outputs.rdma_devices || '' }}" >> "$GITHUB_ENV"
+
       - uses: ./.github/actions/check-stage-health
 
       - uses: ./.github/actions/check-maintenance

.github/workflows/pr-test-extra.yml

@@ -107,7 +107,6 @@ jobs:
     with:
       self_name: extra-a-test-1-gpu-small
       runner_config: 1-gpu-small
-      runs_on: 1-gpu-5090
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -121,7 +120,6 @@ jobs:
     with:
       self_name: extra-a-test-1-gpu-large
       runner_config: 1-gpu-large
-      runs_on: 1-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -136,7 +134,6 @@ jobs:
     with:
       self_name: extra-a-test-2-gpu-large
       runner_config: 2-gpu-large
-      runs_on: 2-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -151,7 +148,6 @@ jobs:
     with:
       self_name: extra-b-test-4-gpu-h100
       runner_config: 4-gpu-h100
-      runs_on: 4-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -165,7 +161,6 @@ jobs:
     with:
       self_name: extra-b-test-4-gpu-b200
       runner_config: 4-gpu-b200
-      runs_on: ${{ needs.check-changes.outputs.b200_runner }}
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -180,7 +175,6 @@ jobs:
     with:
       self_name: extra-b-test-8-gpu-h200
       runner_config: 8-gpu-h200
-      runs_on: 8-gpu-h200
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -194,7 +188,6 @@ jobs:
     with:
       self_name: extra-b-test-deepep-8-gpu-h200
       runner_config: deepep-8-gpu-h200
-      runs_on: 8-gpu-h200-deepep
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}

.github/workflows/pr-test-jit-kernel.yml

@@ -9,7 +9,10 @@ on:
       sgl_kernel:
         required: true
         type: string
-      b200_runner:
+      runner_config:
+        required: true
+        type: string
+      runs_on_map:
         required: true
         type: string
       git_ref:
@@ -157,7 +160,7 @@ jobs:
     if: |
       github.event_name != 'schedule' &&
       inputs.test_parallel_dispatch != 'true'
-    runs-on: ${{ inputs.b200_runner }}
+    runs-on: ${{ fromJson(inputs.runs_on_map)[inputs.runner_config] }}
     timeout-minutes: 240
     steps:
       - uses: actions/checkout@v4

.github/workflows/pr-test-multimodal-gen.yml

@@ -9,7 +9,10 @@ on:
       sgl_kernel:
         required: true
         type: string
-      b200_runner:
+      runner_config:
+        required: true
+        type: string
+      runs_on_map:
         required: true
         type: string
       continue_on_error:
@@ -264,7 +267,7 @@ jobs:
     if: |
       ((github.event_name == 'schedule' || inputs.test_parallel_dispatch == 'true') || (inputs.caller_needs_failure != 'true' && !cancelled())) &&
       inputs.multimodal_gen == 'true'
-    runs-on: ${{ inputs.b200_runner }}
+    runs-on: ${{ fromJson(inputs.runs_on_map)[inputs.runner_config] }}
     timeout-minutes: 240
     steps:
       - name: Checkout code

.github/workflows/pr-test-sgl-kernel.yml

@@ -6,7 +6,10 @@ on:
       sgl_kernel:
         required: true
         type: string
-      b200_runner:
+      runner_config:
+        required: true
+        type: string
+      runs_on_map:
         required: true
         type: string
       git_ref:
@@ -110,7 +113,7 @@ jobs:
           echo "All benchmark tests completed!"
 
   sgl-kernel-b200-test:
-    runs-on: ${{ inputs.b200_runner }}
+    runs-on: ${{ fromJson(inputs.runs_on_map)[inputs.runner_config] }}
     timeout-minutes: 240
     steps:
       - uses: actions/checkout@v4

.github/workflows/pr-test.yml

@@ -209,8 +209,9 @@ jobs:
       needs.check-changes.outputs.sgl_kernel == 'true'
     uses: ./.github/workflows/pr-test-sgl-kernel.yml
     with:
+      runner_config: 4-gpu-b200
+      runs_on_map: ${{ needs.check-changes.outputs.runs_on_map }}
       sgl_kernel: ${{ needs.check-changes.outputs.sgl_kernel }}
-      b200_runner: ${{ needs.check-changes.outputs.b200_runner }}
       git_ref: ${{ inputs.git_ref || '' }}
       skip_stage_health_check: ${{ inputs.skip_stage_health_check == true }}
     secrets: inherit
@@ -227,9 +228,10 @@ jobs:
       needs.check-changes.outputs.jit_kernel == 'true'
     uses: ./.github/workflows/pr-test-jit-kernel.yml
     with:
+      runner_config: 4-gpu-b200
+      runs_on_map: ${{ needs.check-changes.outputs.runs_on_map }}
       jit_kernel: ${{ needs.check-changes.outputs.jit_kernel }}
       sgl_kernel: ${{ needs.check-changes.outputs.sgl_kernel }}
-      b200_runner: ${{ needs.check-changes.outputs.b200_runner }}
       git_ref: ${{ inputs.git_ref || '' }}
       test_parallel_dispatch: ${{ inputs.test_parallel_dispatch == true && 'true' || 'false' }}
       skip_stage_health_check: ${{ inputs.skip_stage_health_check == true }}
@@ -245,7 +247,6 @@ jobs:
     with:
       self_name: stage-a-test-1-gpu-small
       runner_config: 1-gpu-small
-      runs_on: 1-gpu-5090
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -325,7 +326,6 @@ jobs:
     with:
       self_name: stage-b-test-1-gpu-small
       runner_config: 1-gpu-small
-      runs_on: 1-gpu-5090
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -340,7 +340,6 @@ jobs:
     with:
       self_name: stage-b-test-1-gpu-large
       runner_config: 1-gpu-large
-      runs_on: 1-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -355,7 +354,6 @@ jobs:
     with:
       self_name: stage-b-test-2-gpu-large
       runner_config: 2-gpu-large
-      runs_on: 2-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -369,7 +367,6 @@ jobs:
     with:
       self_name: stage-b-test-4-gpu-b200
       runner_config: 4-gpu-b200
-      runs_on: ${{ needs.check-changes.outputs.b200_runner }}
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -386,9 +383,10 @@ jobs:
       needs.check-changes.outputs.multimodal_gen == 'true'
     uses: ./.github/workflows/pr-test-multimodal-gen.yml
     with:
+      runner_config: 4-gpu-b200
+      runs_on_map: ${{ needs.check-changes.outputs.runs_on_map }}
       multimodal_gen: ${{ needs.check-changes.outputs.multimodal_gen }}
       sgl_kernel: ${{ needs.check-changes.outputs.sgl_kernel }}
-      b200_runner: ${{ needs.check-changes.outputs.b200_runner }}
       continue_on_error: ${{ needs.check-changes.outputs.continue_on_error }}
       git_ref: ${{ inputs.git_ref || '' }}
       test_parallel_dispatch: ${{ inputs.test_parallel_dispatch == true && 'true' || 'false' }}
@@ -403,7 +401,6 @@ jobs:
     with:
       self_name: stage-c-test-4-gpu-h100
       runner_config: 4-gpu-h100
-      runs_on: 4-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -417,7 +414,6 @@ jobs:
     with:
       self_name: stage-c-test-8-gpu-h200
       runner_config: 8-gpu-h200
-      runs_on: 8-gpu-h200
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -438,7 +434,6 @@ jobs:
     with:
       self_name: stage-c-test-8-gpu-h20
       runner_config: 8-gpu-h20
-      runs_on: 8-gpu-h20
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -452,7 +447,6 @@ jobs:
     with:
       self_name: stage-c-test-deepep-4-gpu-h100
       runner_config: deepep-4-gpu-h100
-      runs_on: 4-gpu-h100
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -468,7 +462,6 @@ jobs:
     with:
       self_name: stage-c-test-4-gpu-b200
       runner_config: 4-gpu-b200
-      runs_on: ${{ needs.check-changes.outputs.b200_runner }}
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -483,7 +476,6 @@ jobs:
     with:
       self_name: stage-c-test-dsv4-4-gpu-b200
       runner_config: dsv4-4-gpu-b200
-      runs_on: ${{ needs.check-changes.outputs.b200_runner }}
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -498,7 +490,6 @@ jobs:
     with:
       self_name: stage-c-test-dsv4-8-gpu-h200
       runner_config: dsv4-8-gpu-h200
-      runs_on: 8-gpu-h200
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}

scripts/ci/runner_configs.py

@@ -1,14 +1,24 @@
-"""Emit a runner_config's setup details (install / artifact_version /
+"""Emit runner_config setup for GitHub Actions $GITHUB_OUTPUT.
-install_timeout) in $GITHUB_OUTPUT format. Reads scripts/ci/runner_configs.yml.
+
-Called by .github/workflows/_pr-test-stage.yml.
+runner_configs.py <runner_config>
+    Per-field `key=value` lines (install / artifact_version /
+    install_timeout / rdma_devices). `runs_on` is intentionally omitted —
+    it carries the `$b200_runner` sentinel and is resolved via --map.
+    Called per stage by _pr-test-stage.yml.
+
+runner_configs.py --map <b200_runner_label>
+    `runs_on_map={json}` — flat dict {runner_config: runs_on}, with
+    `$b200_runner` substituted. Called once by _pr-test-check-changes.yml.
 """
 
+import json
 import os
 import sys
 
 import yaml
 
 _YAML_PATH = os.path.join(os.path.dirname(__file__), "runner_configs.yml")
+_B200_SENTINEL = "$b200_runner"
 
 
 def load() -> dict:
@@ -16,12 +26,35 @@ def load() -> dict:
         return yaml.safe_load(f)["runner_configs"]
 
 
-if __name__ == "__main__":
+def _emit_single(rc: str) -> None:
-    if len(sys.argv) != 2:
+    # runs_on goes through --map (resolves $b200_runner). Suppress it here so a
-        sys.exit("usage: runner_configs.py <runner_config>")
+    # consumer can't accidentally read the raw sentinel value.
-    rc = sys.argv[1]
+    cfg = load().get(rc)
-    config = load().get(rc)
+    if cfg is None:
-    if config is None:
         sys.exit(f"unknown runner_config: {rc!r}")
-    for key, value in config.items():
+    for key, value in cfg.items():
+        if key == "runs_on":
+            continue
         print(f"{key}={value}")
+
+
+def _emit_map(b200_runner: str) -> None:
+    runs_on = {
+        name: (b200_runner if cfg.get("runs_on") == _B200_SENTINEL else cfg["runs_on"])
+        for name, cfg in load().items()
+    }
+    print(f"runs_on_map={json.dumps(runs_on, separators=(',', ':'))}")
+
+
+if __name__ == "__main__":
+    args = sys.argv[1:]
+    if len(args) == 1:
+        _emit_single(args[0])
+    elif len(args) == 2 and args[0] == "--map":
+        _emit_map(args[1])
+    else:
+        sys.exit(
+            "usage:\n"
+            "  runner_configs.py <runner_config>\n"
+            "  runner_configs.py --map <b200_runner_label>"
+        )

scripts/ci/runner_configs.yml

@@ -3,9 +3,16 @@
 # scripts/ci/runner_configs.py (CLI wrapper), which is in turn called by
 # .github/workflows/_pr-test-stage.yml.
 #
-# Each runner_config carries install script, actions/download-artifact major
+# Each runner_config carries:
-# version, and install-step wall-clock cap (minutes, enforced via
+#   - install: install script path
-# `timeout-minutes:` on the install step in _pr-test-stage.yml).
+#   - artifact_version: actions/download-artifact major version
+#   - install_timeout: install-step wall-clock cap (minutes), enforced via
+#     `timeout-minutes:` on the install step in _pr-test-stage.yml
+#   - runs_on: GHA runner label for the stage's `runs-on:`. The literal
+#     `$b200_runner` is substituted at workflow-load time with the dynamic
+#     b200 runner tag from check-changes (see runner_configs.py --map).
+#   - rdma_devices (optional): exported as SGLANG_CI_RDMA_ALL_DEVICES env
+#     to the stage job; absent means unset.
 
 _anchors:
   default_install: &default scripts/ci/cuda/ci_install_dependency.sh
@@ -13,14 +20,14 @@ _anchors:
   dsv4_install:    &dsv4    scripts/ci/cuda/ci_install_dsv4_dep.sh
 
 runner_configs:
-  1-gpu-small:       { install: *default, artifact_version: v4, install_timeout: "20" }
+  1-gpu-small:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 1-gpu-5090 }
-  1-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20" }
+  1-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 1-gpu-h100 }
-  2-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20" }
+  2-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 2-gpu-h100 }
-  4-gpu-b200:        { install: *default, artifact_version: v6, install_timeout: "20" }
+  4-gpu-b200:        { install: *default, artifact_version: v6, install_timeout: "20", runs_on: $b200_runner }
-  4-gpu-h100:        { install: *default, artifact_version: v4, install_timeout: "20" }
+  4-gpu-h100:        { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 4-gpu-h100 }
-  8-gpu-h200:        { install: *default, artifact_version: v4, install_timeout: "20" }
+  8-gpu-h200:        { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
-  8-gpu-h20:         { install: *deepep,  artifact_version: v4, install_timeout: "20" }
+  8-gpu-h20:         { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h20, rdma_devices: "mlx5_1,mlx5_2,mlx5_3,mlx5_4" }
-  deepep-4-gpu-h100: { install: *deepep,  artifact_version: v4, install_timeout: "20" }
+  deepep-4-gpu-h100: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 4-gpu-h100 }
-  deepep-8-gpu-h200: { install: *deepep,  artifact_version: v4, install_timeout: "20" }
+  deepep-8-gpu-h200: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200-deepep }
-  dsv4-4-gpu-b200:   { install: *dsv4,    artifact_version: v6, install_timeout: "30" }
+  dsv4-4-gpu-b200:   { install: *dsv4,    artifact_version: v6, install_timeout: "30", runs_on: $b200_runner }
-  dsv4-8-gpu-h200:   { install: *dsv4,    artifact_version: v4, install_timeout: "30" }
+  dsv4-8-gpu-h200:   { install: *dsv4,    artifact_version: v4, install_timeout: "30", runs_on: 8-gpu-h200 }