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

python/sglang/test/server_fixtures/spec_eagle_fixture.py

@@ -58,8 +58,12 @@ class SpecEagleServerBase(CustomTestCase):
     attention_backend = "flashinfer"
     # Primary axis: False -> overlap scheduler; True -> synchronous (non-overlap).
     disable_overlap = False
-    mem_fraction_static = 0.85
-    max_running_requests = 8
+    # Leaves ~3.3GB on a 32GB card for the verify logits and activations at a
+    # cap of 64; higher OOMs, lower starves the KV pool into capping the batch.
+    mem_fraction_static = 0.80
+    # The eval kits drive 128 client threads, so a small cap just serializes them.
+    # Capture follows: capture_bs is clipped to req_to_token_pool.size (cap + 1).
+    max_running_requests = 64
     chunked_prefill_size = 128
     # bf16 rather than fp16: fp16 activations can overflow (-> Inf -> NaN) on
     # degenerate draft branches in verify and trip the CI NaN asserts.

python/sglang/test/test_utils.py

@@ -952,6 +952,15 @@ def popen_launch_server(
         other_args = list(other_args)
         other_args += ["--device", str(device)]
 
+    # Prefill dominates capture time: the bucket list runs to chunked_prefill_size
+    # (8192 on H100-class GPUs) and its largest buckets cost seconds each, while
+    # 97% of CI prefill batches are under 1024 tokens -- a server that serves one
+    # test file captures the rest and never replays it. Decode is left alone: its
+    # capture cost is per-phase, not per-bucket. Pass the flag to opt out.
+    prefill_flag = "--cuda-graph-max-bs-prefill"
+    if not any(str(arg).startswith(prefill_flag) for arg in other_args):
+        other_args = list(other_args) + [prefill_flag, "1024"]
+
     # CI-specific: Validate cache and enable offline mode if complete
     if env is None:
         env = os.environ.copy()