diff --git a/python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py b/python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py
index 2bd30e03d..475038f54 100644
--- a/python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py
+++ b/python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py
@@ -50,24 +50,6 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-# Measured immediately before prefill graph construction, after model weights,
-# KV cache, and the eager runner's static buffers have been allocated.  Below
-# this budget, compiling/capturing a multi-bucket prefill graph is likely to
-# OOM or make no forward progress for large models.  Keep an explicitly chosen
-# backend untouched: an operator may intentionally trade KV capacity for it.
-_MIN_AUTO_PREFILL_CUDA_GRAPH_FREE_MEMORY_GB = 4.0
-
-
-def should_skip_auto_prefill_cuda_graph_for_memory(
-    available_memory_gb: float,
-    cuda_graph_config_locked: set[tuple[str, str]],
-) -> bool:
-    """Return whether an auto-selected prefill graph lacks capture headroom."""
-    return (
-        (Phase.PREFILL, "backend") not in cuda_graph_config_locked
-        and available_memory_gb < _MIN_AUTO_PREFILL_CUDA_GRAPH_FREE_MEMORY_GB
-    )
-
 
 class GraphCapture(msgspec.Struct, frozen=True, kw_only=True):
     runner: Optional[BaseRunner]
@@ -396,20 +378,6 @@ def capture_prefill_graph(
 
     tic = time.perf_counter()
     before_mem = get_available_gpu_memory(model_runner.device, model_runner.gpu_id)
-    if should_skip_auto_prefill_cuda_graph_for_memory(
-        before_mem,
-        getattr(model_runner.server_args, "_cuda_graph_config_locked", set()),
-    ):
-        logger.warning(
-            "Disabling auto-selected prefill CUDA graph: only %.2f GiB is free "
-            "after model/KV/eager-buffer allocation; at least %.2f GiB is "
-            "required for capture. Set an explicit prefill CUDA graph backend "
-            "to override this safety gate.",
-            before_mem,
-            _MIN_AUTO_PREFILL_CUDA_GRAPH_FREE_MEMORY_GB,
-        )
-        return result(eager_runner)
-
     role = "draft" if model_runner.is_draft_worker else "target"
     capture_name = f"{role} prefill"
     logger.info(
diff --git a/test/registered/unit/model_executor/model_runner_components/test_cuda_graph_setup.py b/test/registered/unit/model_executor/model_runner_components/test_cuda_graph_setup.py
index 0632fe9c3..7e6a2ef1c 100644
--- a/test/registered/unit/model_executor/model_runner_components/test_cuda_graph_setup.py
+++ b/test/registered/unit/model_executor/model_runner_components/test_cuda_graph_setup.py
@@ -3,28 +3,15 @@ from types import SimpleNamespace
 
 import pytest
 
-from sglang.srt.model_executor.cuda_graph_config import Phase
 from sglang.srt.model_executor.model_runner_components import cuda_graph_setup
 from sglang.srt.model_executor.model_runner_components.cuda_graph_setup import (
     capture_decode_graph,
-    should_skip_auto_prefill_cuda_graph_for_memory,
 )
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 
 
-def test_auto_prefill_cuda_graph_memory_gate():
-    assert should_skip_auto_prefill_cuda_graph_for_memory(3.99, set())
-    assert not should_skip_auto_prefill_cuda_graph_for_memory(4.0, set())
-
-
-def test_explicit_prefill_backend_bypasses_memory_gate():
-    assert not should_skip_auto_prefill_cuda_graph_for_memory(
-        0.0, {(Phase.PREFILL, "backend")}
-    )
-
-
 def test_model_runner_can_override_decode_graph_runner(monkeypatch):
     class CustomGraphRunner:
         def __init__(self, model_runner):
diff --git a/test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py b/test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py
index e52b08e0c..81baa61b8 100644
--- a/test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py
+++ b/test/registered/unit/model_executor/test_prefill_cuda_graph_runner.py
@@ -61,6 +61,54 @@ class _FakeKVIndexKernel:
 
 
 class TestPrefillCudaGraphRunnerChunkedPrefix(CustomTestCase):
+    def test_low_free_memory_still_captures_prefill_graph(self):
+        eager_runner = object()
+        prefill_runner = object()
+        model_runner = SimpleNamespace(
+            device="cuda",
+            gpu_id=0,
+            is_draft_worker=False,
+            spec_algorithm=SimpleNamespace(is_eagle=lambda: False),
+            server_args=SimpleNamespace(
+                enable_lora=False,
+                cuda_graph_config=SimpleNamespace(
+                    prefill=SimpleNamespace(bs=[1], backend=Backend.BREAKABLE)
+                ),
+            ),
+            model=SimpleNamespace(),
+            model_config=SimpleNamespace(context_len=8192, num_hidden_layers=1),
+            req_to_token_pool=SimpleNamespace(size=1),
+        )
+        language_model = SimpleNamespace(layers=[object()])
+
+        with (
+            patch.object(graph_setup, "check_cuda_graph_backend", return_value=False),
+            patch.object(
+                graph_setup, "resolve_language_model", return_value=language_model
+            ),
+            patch.object(
+                graph_setup,
+                "compute_attention_and_moe_layers",
+                return_value=([object()], [], [], [], []),
+            ),
+            patch.object(
+                graph_setup,
+                "get_available_gpu_memory",
+                side_effect=[3.99, 3.5],
+            ),
+            patch.object(
+                graph_setup,
+                "PrefillCudaGraphRunner",
+                return_value=prefill_runner,
+            ),
+        ):
+            capture = capture_prefill_graph(
+                model_runner=model_runner,
+                eager_runner=eager_runner,
+            )
+
+        self.assertIs(capture.runner, prefill_runner)
+
     def test_eagle_target_tc_piecewise_skips_last_mode_capture(self):
         eager_runner = object()
         model_runner = SimpleNamespace(