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

python/sglang/test/ascend/vlm_utils.py

@@ -17,7 +17,7 @@ class TestVLMModels(CustomTestCase):
     mmmu_accuracy = 0.00
     other_args = [
         "--trust-remote-code",
-        "--cuda-graph-max-bs",
+        "--cuda-graph-max-bs-decode",
         "32",
         "--enable-multimodal",
         "--mem-fraction-static",

python/sglang/test/doc_patch.py

@@ -52,7 +52,7 @@ def launch_server_cmd(command: str, host: str = "0.0.0.0", port: int = None):
         f"--max-running-requests {DEFAULT_MAX_RUNNING_REQUESTS} "
         f"--max-total-tokens {DEFAULT_MAX_TOTAL_TOKENS} "
         f"--disable-cuda-graph "
-        f"--cuda-graph-max-bs 4"
+        f"--cuda-graph-max-bs-decode 4"
     )
 
     full_command = f"{command} --port {port} {extra_flags}"

python/sglang/test/gpt_oss_common.py

@@ -40,7 +40,7 @@ class BaseTestGptOss(CustomTestCase):
         }[(model_variant, quantization)]
 
         if model_variant == "20b":
-            other_args += ["--cuda-graph-max-bs", "600"]
+            other_args += ["--cuda-graph-max-bs-decode", "600"]
         # Respect SGLANG_USE_AITER if already set, otherwise default to "0" for HIP
         if _is_hip and "SGLANG_USE_AITER" not in os.environ:
             os.environ["SGLANG_USE_AITER"] = "0"

python/sglang/test/kits/mmmu_vlm_kit.py

@@ -367,7 +367,7 @@ class MMMUMultiModelTestBase(CustomTestCase):
                 api_key=self.api_key,
                 other_args=[
                     "--trust-remote-code",
-                    "--cuda-graph-max-bs",
+                    "--cuda-graph-max-bs-decode",
                     "64",
                     "--enable-multimodal",
                     "--mem-fraction-static",

python/sglang/test/kv_canary/consts.py

@@ -28,7 +28,7 @@ DSV4_POOL_SERVER_ARGS: Final[tuple[str, ...]] = (
     "--enable-dp-attention",
     "--moe-a2a-backend",
     "deepep",
-    "--cuda-graph-max-bs",
+    "--cuda-graph-max-bs-decode",
     "128",
     "--max-running-requests",
     "256",

python/sglang/test/server_fixtures/hybrid_attn_backend_fixture.py

@@ -28,7 +28,7 @@ GSM_DATASET_PATH = None
 # Default server arguments shared across all hybrid-attn-backend tests
 DEFAULT_HYBRID_ATTN_SERVER_ARGS = [
     "--trust-remote-code",
-    "--cuda-graph-max-bs",
+    "--cuda-graph-max-bs-decode",
     "8",
     "--prefill-attention-backend",
     "fa3",

python/sglang/test/server_fixtures/mmmu_fixture.py

@@ -44,7 +44,7 @@ class MMMUServerBase(CustomTestCase):
         # Build server args with MMMU-specific settings
         server_args = [
             "--trust-remote-code",
-            "--cuda-graph-max-bs",
+            "--cuda-graph-max-bs-decode",
             "64",
             "--enable-multimodal",
             "--mem-fraction-static",

python/sglang/test/server_fixtures/ngram_fixture.py

@@ -26,7 +26,7 @@ from sglang.test.test_utils import (
 
 DEFAULT_NGRAM_SERVER_ARGS = [
     "--trust-remote-code",
-    "--cuda-graph-max-bs",
+    "--cuda-graph-max-bs-decode",
     "8",
     "--speculative-algorithm",
     "NGRAM",

python/sglang/test/server_fixtures/spec_eagle_fixture.py

@@ -64,7 +64,7 @@ class SpecEagleServerBase(CustomTestCase):
     # bf16 rather than fp16: fp16 activations can overflow (-> Inf -> NaN) on
     # degenerate draft branches in verify and trip the CI NaN asserts.
     dtype = "bfloat16"
-    cuda_graph_max_bs = None
+    cuda_graph_max_bs_decode = None
     trust_remote_code = True
     # Launch with --enable-return-hidden-states so SpecHiddenStatesKit can probe
     # per-request hidden states; per-request gated, so other requests don't pay.
@@ -109,8 +109,8 @@ class SpecEagleServerBase(CustomTestCase):
             args.append("--trust-remote-code")
         if cls.enable_return_hidden_states:
             args.append("--enable-return-hidden-states")
-        if cls.cuda_graph_max_bs is not None:
-            args += ["--cuda-graph-max-bs", str(cls.cuda_graph_max_bs)]
+        if cls.cuda_graph_max_bs_decode is not None:
+            args += ["--cuda-graph-max-bs-decode", str(cls.cuda_graph_max_bs_decode)]
         args += [str(a) for a in cls.extra_args]
         return args
 

python/sglang/test/server_fixtures/standalone_fixture.py

@@ -49,7 +49,7 @@ class StandaloneServerBase:
         assert cls.attention_backend, f"{cls.__name__} must set `attention_backend`"
         args = [
             "--trust-remote-code",
-            "--cuda-graph-max-bs",
+            "--cuda-graph-max-bs-decode",
             "8",
             "--speculative-algorithm",
             "STANDALONE",

python/sglang/test/test_deterministic_utils.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
 DEFAULT_MODEL = "Qwen/Qwen3-8B"
 COMMON_SERVER_ARGS = [
     "--trust-remote-code",
-    "--cuda-graph-max-bs",
+    "--cuda-graph-max-bs-decode",
     "32",
     "--enable-deterministic-inference",
 ]