ci: add 4-GPU mi35x runner and rebalance off the saturated 8-GPU pool (#28745)

Co-authored-by: michaelzhang-ai <michaelzhang-ai@users.noreply.github.com>

test/registered/amd/test_deepseek_r1_mxfp4_4gpu.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
+register_amd_ci(est_time=3600, suite="stage-c-test-4-gpu-amd-mi35x")
 
 DEEPSEEK_R1_MODEL_PATH = "amd/DeepSeek-R1-MXFP4-Preview"
 SERVER_LAUNCH_TIMEOUT = 1800
@@ -30,12 +30,12 @@ class TestDeepseekR1MXFP4(CustomTestCase):
 
         # Workaround: AITER custom all-gather corrupts CUDA-graph IPC buffer
         # registration and triggers a decode-time "Memory access fault" on
-        # MI35x TP=8. Disable until the AITER-side fix lands (see PR body).
+        # MI35x. Disable until the AITER-side fix lands (see PR body).
         envs.SGLANG_USE_AITER_AG.set(False)
 
         other_args = [
             "--tp",
-            "8",
+            "4",
             "--chunked-prefill-size",
             "131072",
             "--model-loader-extra-config",
@@ -90,7 +90,8 @@ class TestDeepseekR1MXFP4(CustomTestCase):
             write_github_step_summary(
                 f"### test_bs_1_speed (deepseek-r1-mxfp4)\n" f"{speed=:.2f} token/s\n"
             )
-        self.assertGreater(speed, 75)
+        # Report-only: the previous >75 tok/s gate was calibrated for TP=8.
+        # Decode throughput at TP=4 differs; re-calibrate before re-enabling.
 
 
 class TestDeepseekR1MXFP4MTP(CustomTestCase):
@@ -105,7 +106,7 @@ class TestDeepseekR1MXFP4MTP(CustomTestCase):
 
         other_args = [
             "--tp",
-            "8",
+            "4",
             "--chunked-prefill-size",
             "131072",
             "--speculative-algorithm",
@@ -175,7 +176,8 @@ class TestDeepseekR1MXFP4MTP(CustomTestCase):
                 f"{speed=:.2f} token/s\n"
             )
             self.assertGreater(acc_length, 2.04)
-            self.assertGreater(speed, 150)
+            # Report-only: the previous >150 tok/s gate was calibrated for TP=8.
+            # Decode throughput at TP=4 differs; re-calibrate before re-enabling.
 
 
 if __name__ == "__main__":

test/run_suite.py

@@ -41,6 +41,7 @@ PER_COMMIT_SUITES = {
         "jit-kernel-unit-test-amd",
         "sgl-kernel-unit-test-2-gpu-amd",
         "stage-c-test-4-gpu-amd",
+        "stage-c-test-4-gpu-amd-mi35x",
         "stage-c-test-large-8-gpu-amd",
         "stage-c-test-large-8-gpu-amd-mi35x",
         # extra-a: label-gated PR opt-in suites in pr-test-amd-extra.yml