[AMD] ci: add extra-a 1-gpu-large tier (fp8kv-triton, streaming-session, spec-standalone) (#28458)

test/registered/quant/test_fp8kv_triton.py

@@ -3,7 +3,7 @@ from types import SimpleNamespace
 from urllib.parse import urlparse
 
 from sglang.srt.utils import kill_process_tree
-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.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -13,6 +13,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=73, stage="extra-a", runner_config="1-gpu-large")
+register_amd_ci(est_time=94, suite="extra-a-test-1-gpu-large-amd")
 
 
 class TestFP8KVCacheTritonBackend(CustomTestCase):

test/registered/sessions/test_streaming_session_extra.py

@@ -1,6 +1,6 @@
 import unittest
 
-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.kits.streaming_session_kit import StreamingSessionKitMixin
 from sglang.test.server_fixtures.streaming_session_fixture import (
     StreamingSessionServerBase,
@@ -11,6 +11,7 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=691, stage="extra-a", runner_config="1-gpu-large")
+register_amd_ci(est_time=562, suite="extra-a-test-1-gpu-large-amd")
 
 
 class TestStreamingSessionRetractMixedChunk(

test/registered/spec/test_spec_standalone_extra.py

@@ -1,14 +1,22 @@
 import unittest
 
-from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.srt.utils import is_hip
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.server_fixtures.standalone_fixture import StandaloneServerBase
 from sglang.test.test_utils import CustomTestCase
 
 # Non-V2 standalone speculative decoding tests (FA3, Triton, FlashInfer
 # backends). Sibling V2 classes stay per-commit in test_spec_standalone.py.
 register_cuda_ci(est_time=406, stage="extra-a", runner_config="1-gpu-large")
+# AMD: fa3 / flashinfer attention backends are not built in the ROCm
+# sgl_kernel, so only the triton-backend class runs on ROCm (the fa3 and
+# flashinfer classes are skipped on ROCm below).
+register_amd_ci(est_time=103, suite="extra-a-test-1-gpu-large-amd")
+
+_AMD_SKIP_BACKEND = "fa3 / flashinfer attention backends are CUDA-only (not in the ROCm sgl_kernel build)"
 
 
+@unittest.skipIf(is_hip(), _AMD_SKIP_BACKEND)
 class TestStandaloneSpeculativeDecodingBase(StandaloneServerBase, CustomTestCase):
     attention_backend = "fa3"
     speculative_eagle_topk = 2
@@ -24,6 +32,7 @@ class TestStandaloneSpeculativeDecodingTriton(StandaloneServerBase, CustomTestCa
     enable_deterministic_inference = True
 
 
+@unittest.skipIf(is_hip(), _AMD_SKIP_BACKEND)
 class TestStandaloneSpeculativeDecodingFlashinfer(StandaloneServerBase, CustomTestCase):
     attention_backend = "flashinfer"
     speculative_eagle_topk = 2

test/run_suite.py

@@ -43,14 +43,18 @@ 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
+        # extra-a: label-gated PR opt-in suites 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.
+        # dispatch when the PR carries the `run-ci-extra` label). 1-gpu-small
+        # carries the mock-model / kv_canary *unit* tests; 1-gpu-large carries
+        # the subset of model e2e tests validated to pass on mi325 (quant
+        # fp8kv-triton, sessions streaming-session EAGLE3, spec standalone
+        # triton-backend variant). The rest of CUDA
+        # extra-a tests fail on ROCm (missing flash_attn.cute/flash_ops
+        # kernels, OOM, or accuracy regressions — e.g. gemma4-mtp-31b dips
+        # below the gsm8k floor on the topk=3 leg) and stay CUDA-only for now.
         "extra-a-test-1-gpu-small-amd",
+        "extra-a-test-1-gpu-large-amd",
     ],
     HWBackend.MUSA: [],
     HWBackend.CUDA: [