[AMD] Restore AMD CI registrations dropped by #37436 (#38581)

test/registered/debug_utils/test_tensor_dump_forward_hook.py

@@ -14,11 +14,14 @@ from sglang.srt.layers.linear import LinearBase
 from sglang.srt.models.qwen2 import Qwen2MLP
 from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
 from sglang.srt.utils import add_prefix, get_device
-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.layer_ut_utils import init_single_process_dist
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
+# Backend-specific: the hook resolves get_device() and the matching distributed
+# backend, so only an AMD run exercises the HIP device and dump path.
+register_amd_ci(est_time=15, suite="stage-b-test-1-gpu-small-amd")
 
 TEST_HIDDEN_SIZE = 32
 

test/registered/observability/test_tracing.py

@@ -34,7 +34,7 @@ from sglang.srt.observability.trace import (
 )
 from sglang.srt.utils import kill_process_tree
 from sglang.srt.utils.network import get_zmq_socket
-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 (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -47,6 +47,10 @@ logger = logging.getLogger(__name__)
 
 # CI registration
 register_cuda_ci(est_time=113, stage="extra-a", runner_config="1-gpu-small")
+# Backend-specific: the span assertions require PREFILL_FORWARD/DECODE_FORWARD
+# to be emitted from the scheduler forward path, which ROCm reaches through its
+# own attention backend and graph replay.
+register_amd_ci(est_time=113, suite="stage-b-test-1-gpu-small-amd")
 
 
 # ============================================================================

test/registered/openai_server/function_call/test_anthropic_tool_use.py

@@ -17,6 +17,7 @@ import requests
 
 from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import (
+    register_amd_ci,
     register_cpu_ci,
     register_cuda_ci,
 )
@@ -29,6 +30,9 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=50, stage="base-b", runner_config="1-gpu-large")
+# Backend-specific: streaming tool-argument assembly is rebuilt from
+# incremental decode output, so ROCm chunk-boundary divergence surfaces here.
+register_amd_ci(est_time=140, suite="stage-b-test-1-gpu-small-amd")
 register_cpu_ci(est_time=54, suite="stage-b-test-cpu-intel")
 
 # System message to guide Llama3.2 to produce proper tool call format

test/registered/openai_server/function_call/test_openai_function_calling.py

@@ -5,7 +5,11 @@ import openai
 
 from sglang.srt.utils import is_npu, kill_process_tree
 from sglang.srt.utils.hf_transformers_utils import get_tokenizer
-from sglang.test.ci.ci_register import register_cuda_ci, register_npu_ci
+from sglang.test.ci.ci_register import (
+    register_amd_ci,
+    register_cuda_ci,
+    register_npu_ci,
+)
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -16,6 +20,9 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=100, stage="base-b", runner_config="1-gpu-large")
+# Backend-specific: the llama3 and pythonic parsers run on real decoded text,
+# so ROCm decode divergence surfaces as tool calls that no longer parse.
+register_amd_ci(est_time=73, suite="stage-b-test-1-gpu-small-amd")
 # Backend-specific: Ascend uses a local model mirror and its native
 # attention backend, while sharing the protocol assertions below.
 register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)

test/registered/reasoning/test_reasoning.py

@@ -4,7 +4,7 @@ import unittest
 import requests
 
 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.kits.reasoning_kit import (
     ReasoningTokenUsageMixin,
     SeparateReasoningMixin,
@@ -18,6 +18,10 @@ from sglang.test.test_utils import (
 )
 
 register_cuda_ci(est_time=129, stage="base-b", runner_config="1-gpu-large")
+# Backend-specific: the thinking-token and separate-reasoning assertions read
+# decoded output from a 30B MoE, so a ROCm fused-MoE or sampling regression
+# shows up here as malformed reasoning_content that CUDA cannot catch.
+register_amd_ci(est_time=200, suite="stage-b-test-1-gpu-small-amd")
 
 
 class TestEnableThinking(