[CI] Bound the CUDA graph capture range in test launches and lift the spec fixture's admission cap (#33776)

test/registered/spec/eagle/test_spec_eagle_stress.py

@@ -30,7 +30,6 @@ class TestEagle3Perf(Eagle3Base, SpecPerfKit):
 class TestEagleLlama2Retract(EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit):
     """Retract under a small KV budget; must not leak."""
 
-    max_running_requests = 64
     extra_args = ("--max-total-tokens", 4500)  # small KV to trigger retract
     env_overrides = (
         (envs.SGLANG_TEST_RETRACT, True),
@@ -46,7 +45,6 @@ class TestEagle3Topk16V2Retract(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
     spec_tokens = 64
     disable_overlap = False
     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
     env_overrides = (

test/registered/spec/eagle/test_spec_eagle_topk.py

@@ -40,6 +40,9 @@ class TestEagle3Topk16(
     spec_tokens = 64
     disable_overlap = False
     enable_return_hidden_states = True
+    # Verify materializes bs * spec_tokens fp32 logit rows: 262MB here, but
+    # 2.1GB at the fixture's 64 -- OOM on a 32GB card.
+    max_running_requests = 8
     cuda_graph_max_bs_decode = 5
     acc_length_thres = 3.1
     batch_accept_len_thres = 1.75

test/registered/spec/eagle/test_spec_eagle_topk_page.py

@@ -31,6 +31,8 @@ class TestEagle3Page4Topk8(Eagle3Base, SpecAccuracyKit, SpecLogprobKit, SpecFeat
     # Preset accept-length values are topk=1 numbers -- loose for a topk=8
     # tree; tighten once CI reports the actuals.
     gsm8k_accept_len_thres = 2.0
+    # Same fp32 logits budget as the topk=16 classes (bs * spec_tokens * vocab).
+    max_running_requests = 16
     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

@@ -32,8 +32,6 @@ class TestEagle3Triton(
     """Overlap scheduler on triton (kits listed in bases)."""
 
     attention_backend = "triton"
-    max_running_requests = 64
-    cuda_graph_max_bs_decode = 64
     gsm8k_num_examples = 200
     gsm8k_check_accept_len = False
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)