diff --git a/.agents/skills/model-microbenchmark/SKILL.md b/.agents/skills/model-microbenchmark/SKILL.md
index 7e48325dcb..7f18c2bb87 100644
--- a/.agents/skills/model-microbenchmark/SKILL.md
+++ b/.agents/skills/model-microbenchmark/SKILL.md
@@ -31,6 +31,9 @@ Requests: `[N*]Q[@K][c][:D]`, comma-separated, per DP rank.
 | `2048@10240` | prefill after an 8192-token prefix-cache hit |
 | `2048@10240c` | chunked prefill: the prefix came from earlier chunks |
 
+Verify cases require `--speculative-config` with `num_speculative_tokens`
+at least as large as the case's draft count; insufficient capacity is rejected.
+
 Hit vs chunk only differs on models with SWA bounded replay (DSV4.1): a hit
 also recomputes the last window (128 tokens) of the prefix.
 
diff --git a/tests/benchmarks/test_forward_cli.py b/tests/benchmarks/test_forward_cli.py
index 34f8e0c355..d44c6b9466 100644
--- a/tests/benchmarks/test_forward_cli.py
+++ b/tests/benchmarks/test_forward_cli.py
@@ -9,6 +9,7 @@ import numpy as np
 import pytest
 
 from vllm.benchmarks.forward.batch import build_batch
+from vllm.benchmarks.forward.main import run_case
 from vllm.benchmarks.forward.shapes import jitter_kv_lens, parse_case
 from vllm.platforms import current_platform
 
@@ -84,6 +85,18 @@ def test_jitter_kv_lens():
         jitter_kv_lens(case, 0, 1000, 0)
 
 
+@pytest.mark.parametrize("capacity", [0, 2])
+def test_bench_forward_rejects_oversized_verify(capacity):
+    """Reject drafts beyond the runner's buffer before launching CUDA kernels."""
+    runner = SimpleNamespace(num_speculative_steps=capacity)
+    case = parse_case("verify=2*4@64:3")
+    with pytest.raises(
+        ValueError,
+        match=rf"Case 'verify' needs 3 draft tokens.*supports {capacity}",
+    ):
+        run_case(runner, case, SimpleNamespace())
+
+
 @pytest.mark.benchmark
 # TODO(woosuk): Support ROCm and re-enable this test there.
 @pytest.mark.skipif(
diff --git a/vllm/benchmarks/forward/main.py b/vllm/benchmarks/forward/main.py
index ff709562a8..d7841c641e 100644
--- a/vllm/benchmarks/forward/main.py
+++ b/vllm/benchmarks/forward/main.py
@@ -120,6 +120,14 @@ def _make_engine_args(args: argparse.Namespace, cases: list[BenchCase]):
 
 
 def run_case(runner, case: BenchCase, args: argparse.Namespace) -> dict[str, Any]:
+    required_drafts = max((r.num_draft_tokens for r in case.requests), default=0)
+    if required_drafts > runner.num_speculative_steps:
+        raise ValueError(
+            f"Case {case.name!r} needs {required_drafts} draft tokens per request, "
+            f"but the runner supports {runner.num_speculative_steps}. "
+            "Set --speculative-config with a sufficient num_speculative_tokens."
+        )
+
     from vllm.benchmarks.forward.batch import build_batch
     from vllm.benchmarks.forward.capture import ForwardBench
     from vllm.benchmarks.forward.kv_init import touched_kv_views, zero_kv_blocks