[CI] Tune threshold (#5787)

test/srt/run_suite.py

@@ -82,10 +82,10 @@ suites = {
         TestFile("test_triton_moe_channel_fp8_kernel.py", 25),
     ],
     "per-commit-2-gpu": [
-        TestFile("models/lora/test_lora_tp.py", 300),
+        TestFile("models/lora/test_lora_tp.py", 150),
         TestFile("test_data_parallelism.py", 90),
-        TestFile("test_dp_attention.py", 90),
-        TestFile("test_mla_tp.py", 420),
+        TestFile("test_dp_attention.py", 150),
+        TestFile("test_mla_tp.py", 174),
         TestFile("test_moe_ep.py", 220),
         TestFile("test_patch_torch.py", 30),
         TestFile("test_update_weights_from_distributed.py", 100),

test/srt/test_bench_one_batch.py

@@ -51,7 +51,7 @@ class TestBenchOneBatch(CustomTestCase):
                 f"### test_torch_compile_tp2_bs1\n"
                 f"output_throughput : {output_throughput:.2f} token/s\n"
             )
-            self.assertGreater(output_throughput, 235)
+            self.assertGreater(output_throughput, 225)
 
 
 if __name__ == "__main__":

test/srt/test_torch_native_attention_backend.py

@@ -28,7 +28,7 @@ class TestTorchNativeAttnBackend(CustomTestCase):
 
         if is_in_ci():
             # Torch native backend is expected to be slower
-            assert output_throughput > 50, f"{output_throughput=}"
+            self.assertGreater(output_throughput, 40)
 
     def test_mmlu(self):
         model = DEFAULT_MODEL_NAME_FOR_TEST