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

test/registered/spec/eagle/test_eagle_dp_attention.py

@@ -57,7 +57,7 @@ class TestEAGLE3EngineDPAttention(CustomTestCase):
             "triton" if is_in_amd_ci() else "fa3",
             "--mem-fraction-static",
             "0.75",
-            "--cuda-graph-max-bs",
+            "--cuda-graph-max-bs-decode",
             "64",
         ]
         with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1):

test/registered/spec/eagle/test_spec_eagle_fa3.py

@@ -44,7 +44,7 @@ class TestEagleLlama2Fa3Page256(
     attention_backend = "fa3"
     page_size = 256
     chunked_prefill_size = 4096  # must be divisible by page_size (256)
-    cuda_graph_max_bs = 5
+    cuda_graph_max_bs_decode = 5
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 

test/registered/spec/eagle/test_spec_eagle_stress.py

@@ -48,7 +48,7 @@ class TestEagle3Topk16V2Retract(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
     spec_topk = 16
     spec_tokens = 64
     disable_overlap = False
-    cuda_graph_max_bs = 5
+    cuda_graph_max_bs_decode = 5
     max_running_requests = 64
     gsm8k_accept_len_thres = 2.4
     extra_args = ("--max-total-tokens", 4500)  # small KV to trigger retract

test/registered/spec/eagle/test_spec_eagle_topk.py

@@ -28,7 +28,7 @@ class TestEagle3Topk16(Eagle3Base, SpecCorrectnessKit, SpecAccuracyKit, SpecLogp
     spec_topk = 16
     spec_tokens = 64
     disable_overlap = True  # synchronous baseline; SpecV2 subclass flips overlap on
-    cuda_graph_max_bs = 5
+    cuda_graph_max_bs_decode = 5
     acc_length_thres = 3.1
     batch_accept_len_thres = 1.75
     gsm8k_accept_len_thres = 2.4  # EAGLE3 topk16 gsm8k accept ~2.48
@@ -69,7 +69,7 @@ class TestEagleLlama3TokenMap(EagleLlama2Base, SpecAccuracyKit):
     draft_model = "lmsys/sglang-EAGLE-LLaMA3-Instruct-8B"
     spec_topk = 4
     spec_tokens = 8
-    cuda_graph_max_bs = 5
+    cuda_graph_max_bs_decode = 5
     gsm8k_accept_len_thres = 2.5  # FR-Spec token map lowers accept (~2.57)
     extra_args = (
         "--speculative-token-map",

test/registered/spec/eagle/test_spec_eagle_topk_page.py

@@ -26,7 +26,7 @@ class TestEagle3Page64Topk8(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
     spec_topk = 8
     spec_tokens = 32
     disable_overlap = False
-    cuda_graph_max_bs = 5
+    cuda_graph_max_bs_decode = 5
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 

test/registered/spec/eagle/test_spec_eagle_triton.py

@@ -32,7 +32,7 @@ class TestEagle3Triton(
 
     attention_backend = "triton"
     max_running_requests = 64
-    cuda_graph_max_bs = 64
+    cuda_graph_max_bs_decode = 64
     gsm8k_num_examples = 1000
     gsm8k_check_accept_len = False
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)