From 54989b1fd025d40c51774f47acf866e17a3fc3fd Mon Sep 17 00:00:00 2001
From: Michael <13900043+michaelzhang-ai@users.noreply.github.com>
Date: Fri, 12 Jun 2026 15:12:56 -0700
Subject: [PATCH] [AMD] ci: add label-gated extra-a tier (kv_canary +
 mock_model unit tests) (#27822)

---
 .github/workflows/pr-test-amd-extra.yml       | 174 ++++++++++++++++++
 .github/workflows/pr-test-amd-rocm720.yml     |  20 ++
 .github/workflows/pr-test-amd.yml             |  19 ++
 .../kv_canary/test_self_unit_buffer_alloc.py  |   3 +-
 .../kv_canary/test_self_unit_endpoint.py      |   3 +-
 .../kv_canary/test_self_unit_future_tensor.py |   3 +-
 .../kv_canary/test_self_unit_perturb.py       |   3 +-
 .../kv_canary/test_self_unit_plan_input.py    |   3 +-
 .../kv_canary/test_self_unit_pool_patcher.py  |   3 +-
 .../test_self_unit_pool_patcher_utils.py      |   3 +-
 .../kv_canary/test_self_unit_radix_walker.py  |   3 +-
 ..._unit_req_to_expected_token_ids_manager.py |   3 +-
 .../kv_canary/test_self_unit_runner_health.py |   3 +-
 .../test_self_unit_runner_per_forward.py      |   3 +-
 .../test_self_unit_runner_swa_divergence.py   |   3 +-
 .../kv_canary/test_self_unit_runner_sweep.py  |   3 +-
 .../test_self_unit_sweep_plan_builder.py      |   3 +-
 .../kv_canary/test_self_unit_token_oracle.py  |   3 +-
 .../kv_canary/test_self_unit_violation.py     |   3 +-
 .../test_self_unit_canary_mock_wiring.py      |   3 +-
 .../mock_model/test_self_unit_install.py      |   3 +-
 .../mock_model/test_self_unit_oracle.py       |   3 +-
 .../test_self_unit_oracle_torch_vs_ref.py     |   3 +-
 .../test_self_unit_sampler_hookpoint.py       |   3 +-
 test/run_suite.py                             |   8 +
 25 files changed, 263 insertions(+), 21 deletions(-)
 create mode 100644 .github/workflows/pr-test-amd-extra.yml

diff --git a/.github/workflows/pr-test-amd-extra.yml b/.github/workflows/pr-test-amd-extra.yml
new file mode 100644
index 000000000..ff86d6dd4
--- /dev/null
+++ b/.github/workflows/pr-test-amd-extra.yml
@@ -0,0 +1,174 @@
+name: PR Test Extra (AMD)
+# Label-gated AMD extra CI workflow — the AMD mirror of pr-test-extra.yml.
+#
+# Adds AMD runtime to a PR only when the author opts in: the PR must carry
+# BOTH `run-ci` (basic-CI prerequisite) and `run-ci-extra` (explicit opt-in).
+# The label check happens at runtime in pr-gate.yml via a live
+# `gh pr view`-style fetch, so reruns after adding the labels (e.g. via a
+# slash command) pick up the new label set — a workflow-level `if` would read
+# the frozen event payload, which never updates on rerun. The job graph also
+# runs unconditionally on workflow_dispatch / workflow_call so it can be
+# triggered manually or chained from the AMD scheduler.
+#
+# Stage: extra-a (1-gpu-small-amd). The job mirrors the container bring-up of
+# pr-test-amd.yml and dispatches `run_suite.py --hw amd --suite
+# extra-a-test-1-gpu-small-amd`. Only the mock-model / kv_canary *unit* tests
+# are onboarded so far; the canary *e2e* tests (which would land in
+# 1-/2-gpu-large) need the canary JIT kernel ported to ROCm first, so those
+# suites are intentionally not registered for AMD yet.
+
+on:
+  pull_request:
+    # `labeled` lets the workflow re-fire when `run-ci-extra` (or `run-ci`)
+    # is added after the latest push. See call-gate.if for the matching guard
+    # that prevents unrelated label additions from dispatching a full run.
+    types: [opened, synchronize, reopened, labeled]
+  workflow_dispatch:
+    inputs:
+      runner_arch:
+        description: 'AMD runner pool to dispatch GPU jobs to'
+        required: false
+        type: choice
+        default: mi325
+        options:
+          - mi300
+          - mi325
+      rocm_version:
+        description: 'ROCm container variant (empty = Dockerfile default; rocm720 = ROCm 7.2.0)'
+        required: false
+        type: choice
+        default: ''
+        options:
+          - ''
+          - rocm720
+      aiter_ref:
+        description: 'Override AITER commit (optional, leave empty to use Dockerfile default)'
+        required: false
+        type: string
+        default: ''
+      continue_on_error:
+        description: 'Continue on error (do not fail the workflow on test failures)'
+        required: false
+        type: boolean
+        default: false
+  workflow_call:
+    inputs:
+      ref:
+        description: 'Git ref (branch, tag, or SHA) to test. If not provided, uses the default branch.'
+        required: false
+        type: string
+        default: ''
+      rocm_version:
+        description: 'ROCm container variant (empty = Dockerfile default; rocm720 = ROCm 7.2.0)'
+        required: false
+        type: string
+        default: ''
+      aiter_ref:
+        description: 'Override AITER commit (optional, leave empty to use Dockerfile default)'
+        required: false
+        type: string
+        default: ''
+      continue_on_error:
+        description: 'Continue on error (do not fail the workflow on test failures)'
+        required: false
+        type: boolean
+        default: false
+
+env:
+  AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
+  DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
+  DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
+
+concurrency:
+  group: pr-test-amd-extra-${{ github.event_name }}-${{ github.head_ref || github.ref_name || 'default' }}-${{ inputs.ref || 'all' }}
+  cancel-in-progress: ${{ github.event_name != 'workflow_call' }}
+
+permissions:
+  actions: write
+  contents: read
+  issues: read
+  pull-requests: read
+
+jobs:
+  # =============================================== PR Gate ====================================================
+  # Runtime live-fetch label gate (mirrors pr-test-extra.yml's call-gate):
+  # requires both `run-ci` and `run-ci-extra`. A failure here cascades to
+  # every test job via `needs`, so a PR without the labels ends in one red
+  # ~30s gate job plus a row of skipped jobs instead of consuming AMD runners.
+  #
+  # The job-level `if` only filters the `labeled` event type so that adding an
+  # unrelated label doesn't dispatch a full run; the actual label-presence
+  # gate is enforced at runtime inside pr-gate.yml.
+  call-gate:
+    if: |
+      github.event_name != 'pull_request' ||
+      github.event.action != 'labeled' ||
+      github.event.label.name == 'run-ci' ||
+      github.event.label.name == 'run-ci-extra'
+    uses: ./.github/workflows/pr-gate.yml
+    with:
+      require-run-ci: true
+      require-run-ci-extra: true
+    secrets: inherit
+
+  # =============================================== extra-a (1-gpu-small) ===============================================
+  # Single unpartitioned job: the 21 onboarded unit tests total ~233s, so the
+  # expensive per-job setup (container bring-up + sgl-kernel ROCm build + dep
+  # install, several minutes) dominates. Partitioning would multiply that
+  # setup across scarce AMD GPUs to shave only a couple minutes of test time,
+  # so one GPU running the whole suite sequentially is the better trade.
+  extra-a-test-1-gpu-small-amd:
+    name: ${{ format('extra-a-test-1-gpu-small-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi325') }}
+    needs: [call-gate]
+    if: ${{ !cancelled() && needs.call-gate.result == 'success' }}
+    runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - name: Start CI container
+        # `rocm_version` (e.g. rocm720) selects an alternate ROCm container; empty uses the Dockerfile default.
+        run: bash scripts/ci/amd/amd_ci_start_container.sh ${{ inputs.rocm_version && format('--rocm-version {0}', inputs.rocm_version) || '' }}
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Run test
+        timeout-minutes: 45
+        run: |
+          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite extra-a-test-1-gpu-small-amd --timeout-per-file 1800 ${{ inputs.continue_on_error == true && '--continue-on-error' || '' }}
+
+  # =============================================== aggregator ====================================================
+  # Single fan-in job so branch protection / notifications depend on one job
+  # rather than every matrix leg. Fails if any dependent failed or was
+  # cancelled; `skipped` (e.g. PR without the opt-in labels) is not a failure.
+  pr-test-amd-extra-finish:
+    needs:
+      [
+        call-gate,
+        extra-a-test-1-gpu-small-amd,
+      ]
+    if: always()
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check all dependent job statuses
+        run: |
+          json_needs='${{ toJson(needs) }}'
+          job_names=$(echo "$json_needs" | jq -r 'keys_unsorted[]')
+          for job in $job_names; do
+            result=$(echo "$json_needs" | jq -r --arg j "$job" '.[$j].result')
+            echo "$job: $result"
+            if [[ "$result" == "failure" || "$result" == "cancelled" ]]; then
+              echo "The above jobs failed."
+              exit 1
+            fi
+          done
+          echo "All jobs completed successfully"
+          exit 0
diff --git a/.github/workflows/pr-test-amd-rocm720.yml b/.github/workflows/pr-test-amd-rocm720.yml
index 482d0b440..0348212f9 100644
--- a/.github/workflows/pr-test-amd-rocm720.yml
+++ b/.github/workflows/pr-test-amd-rocm720.yml
@@ -203,6 +203,26 @@ jobs:
               - "python/pyproject_rocm.toml"
               - "python/pyproject_other.toml"
 
+  # =============================================== extra (scheduled) ====================================================
+  # ROCm 7.2 mirror of pr-test-amd.yml's `call-pr-test-amd-extra`: chain the
+  # label-gated AMD extra tier into this workflow's daily schedule, but in a
+  # ROCm 7.2 container (`rocm_version: rocm720`). On `schedule` (and
+  # run_all_tests dispatch) the extra suite runs on `main` without the
+  # `run-ci-extra` label (pr-gate.yml only enforces labels on pull_request
+  # events). Targeted dispatches (target_stage set) are excluded. Not wired
+  # into any finish aggregator so the base rocm720 run never depends on it.
+  call-pr-test-amd-extra-rocm720:
+    if: |
+      (github.event_name == 'schedule' || inputs.run_all_tests == true) &&
+      !(inputs.target_stage || inputs.target_stage_select)
+    uses: ./.github/workflows/pr-test-amd-extra.yml
+    with:
+      ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
+      rocm_version: rocm720
+      aiter_ref: ${{ inputs.aiter_ref }}
+      continue_on_error: true
+    secrets: inherit
+
   # =============================================== sgl-kernel ====================================================
   sgl-kernel-unit-test-amd-rocm720:
     needs: [check-changes]
diff --git a/.github/workflows/pr-test-amd.yml b/.github/workflows/pr-test-amd.yml
index e5991f614..23a38df39 100644
--- a/.github/workflows/pr-test-amd.yml
+++ b/.github/workflows/pr-test-amd.yml
@@ -187,6 +187,25 @@ jobs:
               - "python/pyproject_rocm.toml"
               - "python/pyproject_other.toml"
 
+  # =============================================== extra (scheduled) ====================================================
+  # Chain the label-gated AMD extra tier into the scheduled run, mirroring
+  # pr-test.yml's `call-pr-test-extra`. On `schedule` (and run_all_tests
+  # dispatch) the extra suite runs on `main` without needing the
+  # `run-ci-extra` label (pr-gate.yml only enforces labels on pull_request
+  # events). Targeted /rerun-stage dispatches (target_stage set) are excluded.
+  # Not added to `pr-test-amd-finish` so the base AMD gate never depends on
+  # the opt-in extra suite.
+  call-pr-test-amd-extra:
+    if: |
+      (github.event_name == 'schedule' || inputs.run_all_tests == true) &&
+      !(inputs.target_stage || inputs.target_stage_select)
+    uses: ./.github/workflows/pr-test-amd-extra.yml
+    with:
+      ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
+      aiter_ref: ${{ inputs.aiter_ref }}
+      continue_on_error: true
+    secrets: inherit
+
   # =============================================== sgl-kernel ====================================================
   sgl-kernel-unit-test-amd:
     name: ${{ format('sgl-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
diff --git a/test/registered/kv_canary/test_self_unit_buffer_alloc.py b/test/registered/kv_canary/test_self_unit_buffer_alloc.py
index 97d5d8435..27c3cf55f 100644
--- a/test/registered/kv_canary/test_self_unit_buffer_alloc.py
+++ b/test/registered/kv_canary/test_self_unit_buffer_alloc.py
@@ -12,10 +12,11 @@ from sglang.srt.kv_canary.pool_patcher.buffer_alloc import (
     make_row_source,
     resolve_real_kv_read_bytes,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=10, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=10, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _config(mode: RealKvHashMode) -> CanaryConfig:
diff --git a/test/registered/kv_canary/test_self_unit_endpoint.py b/test/registered/kv_canary/test_self_unit_endpoint.py
index b301b27b5..8adcded2b 100644
--- a/test/registered/kv_canary/test_self_unit_endpoint.py
+++ b/test/registered/kv_canary/test_self_unit_endpoint.py
@@ -21,11 +21,12 @@ from sglang.srt.kv_canary.expected_inputs import ExpectedInputs
 from sglang.srt.kv_canary.state import (
     ViolationLog,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import DEFAULT_DEVICE
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=20, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=20, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_endpoint(*, device, kernel_kind=CanaryLaunchTag.HEAD_K_FULL, swa_lut=None):
diff --git a/test/registered/kv_canary/test_self_unit_future_tensor.py b/test/registered/kv_canary/test_self_unit_future_tensor.py
index e593cd6d0..a0256378f 100644
--- a/test/registered/kv_canary/test_self_unit_future_tensor.py
+++ b/test/registered/kv_canary/test_self_unit_future_tensor.py
@@ -6,10 +6,11 @@ from typing import cast
 import torch
 
 from sglang.srt.kv_canary.runner.future_tensor import FutureTensors
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=20, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=20, suite="extra-a-test-1-gpu-small-amd")
 
 
 class _FakeEvent:
diff --git a/test/registered/kv_canary/test_self_unit_perturb.py b/test/registered/kv_canary/test_self_unit_perturb.py
index 5f3d4ed62..8e9a240f7 100644
--- a/test/registered/kv_canary/test_self_unit_perturb.py
+++ b/test/registered/kv_canary/test_self_unit_perturb.py
@@ -27,7 +27,7 @@ from sglang.srt.kv_canary.perturb.utils import (
     flip_first_byte_in_source,
     pick_target_group,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import (
     DEFAULT_DEVICE,
     make_buffer_group,
@@ -41,6 +41,7 @@ if TYPE_CHECKING:
     from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 
 register_cuda_ci(est_time=10, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=10, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestParseTargetGroupKind(CustomTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_plan_input.py b/test/registered/kv_canary/test_self_unit_plan_input.py
index 937b7221b..7449b0947 100644
--- a/test/registered/kv_canary/test_self_unit_plan_input.py
+++ b/test/registered/kv_canary/test_self_unit_plan_input.py
@@ -6,7 +6,7 @@ from types import SimpleNamespace
 import torch
 
 from sglang.srt.kv_canary.plan_input import PlanInput
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import (
     DEFAULT_DEVICE,
     make_forward_batch,
@@ -14,6 +14,7 @@ from sglang.test.kv_canary.fixtures import (
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=30, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=30, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_static_plan_input(*, bs_capacity: int, device) -> PlanInput:
diff --git a/test/registered/kv_canary/test_self_unit_pool_patcher.py b/test/registered/kv_canary/test_self_unit_pool_patcher.py
index 859914721..19f0934e7 100644
--- a/test/registered/kv_canary/test_self_unit_pool_patcher.py
+++ b/test/registered/kv_canary/test_self_unit_pool_patcher.py
@@ -15,7 +15,7 @@ from sglang.jit_kernel.kv_canary.verify import (
 )
 from sglang.srt.kv_canary.buffer_group import PoolKind
 from sglang.srt.kv_canary.pool_patcher.api import attach_canary_buffers
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import (
     DEFAULT_DEVICE,
     make_base_config,
@@ -25,6 +25,7 @@ from sglang.test.kv_canary.fixtures import (
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=45, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=45, suite="extra-a-test-1-gpu-small-amd")
 
 
 class PoolPatcherHelper:
diff --git a/test/registered/kv_canary/test_self_unit_pool_patcher_utils.py b/test/registered/kv_canary/test_self_unit_pool_patcher_utils.py
index cfb1019dc..eefa62bab 100644
--- a/test/registered/kv_canary/test_self_unit_pool_patcher_utils.py
+++ b/test/registered/kv_canary/test_self_unit_pool_patcher_utils.py
@@ -3,10 +3,11 @@ from __future__ import annotations
 import unittest
 
 from sglang.srt.kv_canary.pool_patcher.utils import wrap_method
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=10, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=10, suite="extra-a-test-1-gpu-small-amd")
 
 
 class _FakeObj:
diff --git a/test/registered/kv_canary/test_self_unit_radix_walker.py b/test/registered/kv_canary/test_self_unit_radix_walker.py
index 6ef7dce93..4caef8b8c 100644
--- a/test/registered/kv_canary/test_self_unit_radix_walker.py
+++ b/test/registered/kv_canary/test_self_unit_radix_walker.py
@@ -6,11 +6,12 @@ import torch
 
 from sglang.srt.kv_canary.radix_cache_walker import walk_radix_cache_for_canary
 from sglang.srt.mem_cache.swa_radix_cache import SWARadixCache, TreeNode
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import DEFAULT_DEVICE, make_radix_cache
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=30, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=30, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestSelfUnitRadixWalker(CustomTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_req_to_expected_token_ids_manager.py b/test/registered/kv_canary/test_self_unit_req_to_expected_token_ids_manager.py
index 518c4fbb9..a11db78c5 100644
--- a/test/registered/kv_canary/test_self_unit_req_to_expected_token_ids_manager.py
+++ b/test/registered/kv_canary/test_self_unit_req_to_expected_token_ids_manager.py
@@ -10,11 +10,12 @@ from sglang.srt.kv_canary.req_to_expected_token_ids_manager import (
     compute_req_all_ids_info,
     populate_req_to_expected_token_ids,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import DEFAULT_DEVICE, make_forward_batch
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=15, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=15, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_req(*, origin: list[int], output: list[int]) -> SimpleNamespace:
diff --git a/test/registered/kv_canary/test_self_unit_runner_health.py b/test/registered/kv_canary/test_self_unit_runner_health.py
index 7255ba60c..b033a0e55 100644
--- a/test/registered/kv_canary/test_self_unit_runner_health.py
+++ b/test/registered/kv_canary/test_self_unit_runner_health.py
@@ -11,7 +11,7 @@ from sglang.srt.kv_canary.config import CanaryConfig
 from sglang.srt.kv_canary.runner import stats_logger as stats_logger_module
 from sglang.srt.kv_canary.runner.health_checker import KernelRunCounterHealthChecker
 from sglang.srt.kv_canary.state import CanaryDeviceState
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.runner_test_base import (
     CanaryManagerTestCase,
     make_config,
@@ -20,6 +20,7 @@ from sglang.test.kv_canary.runner_test_base import (
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=45, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=45, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestSelfUnitManagerHealth(CanaryManagerTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_runner_per_forward.py b/test/registered/kv_canary/test_self_unit_runner_per_forward.py
index de5ebcf32..d99ac6eb1 100644
--- a/test/registered/kv_canary/test_self_unit_runner_per_forward.py
+++ b/test/registered/kv_canary/test_self_unit_runner_per_forward.py
@@ -12,7 +12,7 @@ from sglang.srt.kv_canary import endpoint as endpoint_module
 from sglang.srt.kv_canary.expected_inputs import ExpectedInputs
 from sglang.srt.kv_canary.runner import kernel_launcher as kernel_launcher_module
 from sglang.srt.kv_canary.state import ViolationLog
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import make_buffer_group, make_forward_batch
 from sglang.test.kv_canary.runner_test_base import (
     CanaryManagerTestCase,
@@ -21,6 +21,7 @@ from sglang.test.kv_canary.runner_test_base import (
 )
 
 register_cuda_ci(est_time=45, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=45, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestManagerPerForward(CanaryManagerTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_runner_swa_divergence.py b/test/registered/kv_canary/test_self_unit_runner_swa_divergence.py
index 7d0ef0059..0e3dc80c7 100644
--- a/test/registered/kv_canary/test_self_unit_runner_swa_divergence.py
+++ b/test/registered/kv_canary/test_self_unit_runner_swa_divergence.py
@@ -15,12 +15,13 @@ from sglang.srt.kv_canary.runner.swa_divergence import (
     SwaDivergenceReporter,
     compute_swa_full_idx_divergence,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import make_buffer_group
 from sglang.test.kv_canary.runner_test_base import CanaryManagerTestCase, make_manager
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=45, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=45, suite="extra-a-test-1-gpu-small-amd")
 
 _DEVICE = torch.device("cuda")
 
diff --git a/test/registered/kv_canary/test_self_unit_runner_sweep.py b/test/registered/kv_canary/test_self_unit_runner_sweep.py
index a2d02e1c6..49e55f018 100644
--- a/test/registered/kv_canary/test_self_unit_runner_sweep.py
+++ b/test/registered/kv_canary/test_self_unit_runner_sweep.py
@@ -4,7 +4,7 @@ import unittest
 from unittest.mock import patch
 
 from sglang.srt.kv_canary import endpoint as endpoint_module
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import (
     make_forward_batch,
     make_radix_cache,
@@ -17,6 +17,7 @@ from sglang.test.kv_canary.runner_test_base import (
 )
 
 register_cuda_ci(est_time=45, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=45, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _run_one_cycle(manager, forward_batch) -> None:
diff --git a/test/registered/kv_canary/test_self_unit_sweep_plan_builder.py b/test/registered/kv_canary/test_self_unit_sweep_plan_builder.py
index b9d5c0731..2cab72e1f 100644
--- a/test/registered/kv_canary/test_self_unit_sweep_plan_builder.py
+++ b/test/registered/kv_canary/test_self_unit_sweep_plan_builder.py
@@ -5,7 +5,7 @@ import unittest
 import torch
 
 from sglang.srt.kv_canary.sweep_plan_builder import build_verify_plan_radix_sweep
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import (
     DEFAULT_DEVICE,
     make_radix_cache,
@@ -14,6 +14,7 @@ from sglang.test.kv_canary.fixtures import (
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=30, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=30, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestSelfUnitSweepPlanBuilder(CustomTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_token_oracle.py b/test/registered/kv_canary/test_self_unit_token_oracle.py
index f8866473e..fb0ebb3d4 100644
--- a/test/registered/kv_canary/test_self_unit_token_oracle.py
+++ b/test/registered/kv_canary/test_self_unit_token_oracle.py
@@ -9,11 +9,12 @@ from sglang.srt.kv_canary.expected_inputs import ExpectedInputs
 from sglang.srt.kv_canary.token_oracle.oracle import HashOracle
 from sglang.srt.kv_canary.token_oracle.oracle_manager import TokenOracleManager
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kv_canary.fixtures import DEFAULT_DEVICE
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=1, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=1, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestTokenOracleManager(CustomTestCase):
diff --git a/test/registered/kv_canary/test_self_unit_violation.py b/test/registered/kv_canary/test_self_unit_violation.py
index 6a9ff778a..7245586a0 100644
--- a/test/registered/kv_canary/test_self_unit_violation.py
+++ b/test/registered/kv_canary/test_self_unit_violation.py
@@ -15,10 +15,11 @@ from sglang.srt.kv_canary.runner.violation_reporter import (
     ViolationReporter,
     _format_violation,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=5, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=5, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_row(
diff --git a/test/registered/mock_model/test_self_unit_canary_mock_wiring.py b/test/registered/mock_model/test_self_unit_canary_mock_wiring.py
index 7916867d7..2b3fbd78e 100644
--- a/test/registered/mock_model/test_self_unit_canary_mock_wiring.py
+++ b/test/registered/mock_model/test_self_unit_canary_mock_wiring.py
@@ -12,11 +12,12 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardMode,
     _stable_hash_str_to_i64,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.mock_model.utils import mock_model_server_args, mock_model_server_env
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=60, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=60, suite="extra-a-test-1-gpu-small-amd")
 
 
 @dataclasses.dataclass
diff --git a/test/registered/mock_model/test_self_unit_install.py b/test/registered/mock_model/test_self_unit_install.py
index 28f655b52..0bcc35feb 100644
--- a/test/registered/mock_model/test_self_unit_install.py
+++ b/test/registered/mock_model/test_self_unit_install.py
@@ -9,10 +9,11 @@ os.environ["SGLANG_KV_CANARY_ENABLE_TOKEN_ORACLE"] = "1"
 from sglang.srt.kv_canary.token_oracle.install import install_token_oracle_from_env
 from sglang.srt.kv_canary.token_oracle.oracle import HashOracle
 from sglang.srt.layers.sampler import _CUSTOM_SAMPLER_FACTORIES
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=60, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=60, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_server_args(*, sampling_backend: str) -> SimpleNamespace:
diff --git a/test/registered/mock_model/test_self_unit_oracle.py b/test/registered/mock_model/test_self_unit_oracle.py
index fe7f02e6c..c28b723e2 100644
--- a/test/registered/mock_model/test_self_unit_oracle.py
+++ b/test/registered/mock_model/test_self_unit_oracle.py
@@ -10,10 +10,11 @@ from sglang.srt.kv_canary.token_oracle.oracle import (
     HashOracle,
     _splitmix64_tensor,
 )
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=60, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=60, suite="extra-a-test-1-gpu-small-amd")
 
 
 _U64_MASK: int = (1 << 64) - 1
diff --git a/test/registered/mock_model/test_self_unit_oracle_torch_vs_ref.py b/test/registered/mock_model/test_self_unit_oracle_torch_vs_ref.py
index 3043e66ec..c880a8d4f 100644
--- a/test/registered/mock_model/test_self_unit_oracle_torch_vs_ref.py
+++ b/test/registered/mock_model/test_self_unit_oracle_torch_vs_ref.py
@@ -7,10 +7,11 @@ import torch
 
 from sglang.jit_kernel.kv_canary.verify_ref import splitmix64
 from sglang.srt.kv_canary.token_oracle.oracle import HashOracle
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=30, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=30, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestHashOracleTorchVsRef(CustomTestCase):
diff --git a/test/registered/mock_model/test_self_unit_sampler_hookpoint.py b/test/registered/mock_model/test_self_unit_sampler_hookpoint.py
index 2d4e2c44a..c46c07954 100644
--- a/test/registered/mock_model/test_self_unit_sampler_hookpoint.py
+++ b/test/registered/mock_model/test_self_unit_sampler_hookpoint.py
@@ -18,10 +18,11 @@ from sglang.srt.kv_canary.token_oracle.oracle import HashOracle
 from sglang.srt.kv_canary.token_oracle.sampler import install_oracle_sampler
 from sglang.srt.layers.sampler import _CUSTOM_SAMPLER_FACTORIES
 from sglang.srt.server_args import SAMPLING_BACKEND_CHOICES
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=60, stage="extra-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=60, suite="extra-a-test-1-gpu-small-amd")
 
 
 class TestInstallOracleSampler(CustomTestCase):
diff --git a/test/run_suite.py b/test/run_suite.py
index d9c139e64..a7a0a951b 100644
--- a/test/run_suite.py
+++ b/test/run_suite.py
@@ -43,6 +43,14 @@ PER_COMMIT_SUITES = {
         "stage-c-test-4-gpu-amd",
         "stage-c-test-large-8-gpu-amd",
         "stage-c-test-large-8-gpu-amd-mi35x",
+        # extra-a: label-gated PR opt-in suite in pr-test-amd-extra.yml
+        # (mirror of CUDA extra-a; tests stay tagged per-commit but only
+        # dispatch when the PR carries the `run-ci-extra` label). Only the
+        # 1-gpu-small mock-model / kv_canary *unit* tests are onboarded so
+        # far; the canary *e2e* tests (1-/2-gpu-large) need the canary JIT
+        # kernel ported to ROCm first, so those suites are intentionally
+        # not yet registered for AMD.
+        "extra-a-test-1-gpu-small-amd",
     ],
     HWBackend.MUSA: [],
     HWBackend.CUDA: [