Skip to content

Official Inference Code of EAGLE-4

fix(core): reject a stop_check workload shorter than concurrency x waves

kevin9327573cb7cfa673
diff --git a/tests/test_stop_check.py b/tests/test_stop_check.py
new file mode 100644
index 0000000..6c037e8
--- /dev/null
+++ b/tests/test_stop_check.py
@@ -0,0 +1,43 @@
+# Copyright 2026 Eagle AI Corp.
+# SPDX-License-Identifier: LicenseRef-Eagle-4-Source-Available
+
+"""tools/stop_check.py refuses a workload file too short for the requested batches before sending anything."""
+
+import importlib.util
+import json
+from pathlib import Path
+import sys
+import types
+
+import pytest
+
+
+class Tokenizer:
+ eos_token_id = 2
+
+ def encode(self, text, add_special_tokens=False):
+ return [10] if text == "\n\n" else [ord(c) for c in text]
+
+ def apply_chat_template(self, messages, **kwargs):
+ return messages[0]["content"]
+
+
+def load(monkeypatch):
+ transformers = types.ModuleType("transformers")
+ transformers.AutoTokenizer = types.SimpleNamespace(from_pretrained=lambda path: Tokenizer())
+ monkeypatch.setitem(sys.modules, "transformers", transformers)
+ monkeypatch.setitem(sys.modules, "requests", types.ModuleType("requests"))
+ spec = importlib.util.spec_from_file_location("stop_check", Path(__file__).parents[1] / "tools/stop_check.py")
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module
+
+
+def test_short_data_is_a_usage_error(tmp_path, monkeypatch):
+ module = load(monkeypatch)
+ data = tmp_path / "question_multiturn.jsonl"
+ data.write_text("".join(json.dumps({"turns": [f"question {i}"]}) + "\n" for i in range(3)))
+ argv = ["--api", "sglang", "--port", "1", "--data", str(data), "--tokenizer", str(tmp_path)]
+ with pytest.raises(SystemExit, match="has 3 conversations.*needs 32"):
+ module.main([*argv, "--out", str(tmp_path / "out.json")])
+ assert not (tmp_path / "out.json").exists()
diff --git a/tools/stop_check.py b/tools/stop_check.py
index bf269f1..a9c7764 100644
--- a/tools/stop_check.py
+++ b/tools/stop_check.py
@@ -164,6 +164,11 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
base = f"http://127.0.0.1:{args.port}"
model = requests.get(f"{base}/v1/models", timeout=60).json()["data"][0]["id"] if args.api != "sglang" else None
turns = first_turns(args.data, args.concurrency * args.waves)
+ if len(turns) < args.concurrency * args.waves:
+ raise SystemExit(
+ f"{args.data} has {len(turns)} conversations; --concurrency {args.concurrency} x --waves {args.waves} "
+ f"needs {args.concurrency * args.waves}"
+ )
prompts = [
tokenizer.encode(
tokenizer.apply_chat_template(