[misc] Use --cuda-graph-max-bs-decode in tests, examples, and docs (#29591)

test/registered/backends/test_deepseek_r1_fp8_trtllm_backend.py

@@ -31,7 +31,7 @@ class TestDeepseekR1Fp8Flashinfer(CustomTestCase):
             "8192",
             "--mem-fraction-static",
             "0.9",
-            "--cuda-graph-max-bs",
+            "--cuda-graph-max-bs-decode",
             "128",
             "--max-prefill-tokens",
             "8192",

test/registered/backends/test_flashinfer_trtllm_gen_attn_backend.py

@@ -28,7 +28,7 @@ class TestFlashinferTrtllmGenAttnBackend(CustomTestCase):
             other_args=[
                 "--attention-backend",
                 "trtllm_mha",
-                "--cuda-graph-max-bs",
+                "--cuda-graph-max-bs-decode",
                 "512",
                 "--tp-size",
                 "4",

test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py

@@ -78,7 +78,7 @@ class FlashinferTrtllmGenMoeBackendBF16Base:
                 "triton",
                 "--moe-runner-backend",
                 cls.backend,
-                "--cuda-graph-max-bs",
+                "--cuda-graph-max-bs-decode",
                 "512",
                 "--tp-size",
                 "4",
@@ -261,7 +261,7 @@ class FlashinferTrtllmGenMoeBackendNvFp4OnlineBase:
                 "triton",
                 "--moe-runner-backend",
                 cls.backend,
-                "--cuda-graph-max-bs",
+                "--cuda-graph-max-bs-decode",
                 "128",
                 "--tp-size",
                 "4",

test/registered/backends/test_torch_compile.py

@@ -32,7 +32,7 @@ class TestTorchCompile(CustomTestCase, MMLUMixin):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--enable-torch-compile", "--cuda-graph-max-bs", "4"],
+            other_args=["--enable-torch-compile", "--cuda-graph-max-bs-decode", "4"],
         )
 
     @classmethod