Skip to content

🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.

[fix] Keep streamed tool calls out of response content

cyanseek86aad5062149
diff --git a/scripts/serve_openai_api.py b/scripts/serve_openai_api.py
index 09e7fc1..7073ecc 100644
--- a/scripts/serve_openai_api.py
+++ b/scripts/serve_openai_api.py
@@ -129,9 +129,11 @@ def generate_stream_response(messages, temperature, top_p, max_tokens, tools=Non
Thread(target=_generate).start()
- full_text = ""
- emitted = 0
+ pending = ""
+ in_tool_call = False
+ tool_count = 0
thinking_ended = not bool(open_thinking)
+ strip_newlines = False
while True:
text = queue.get()
@@ -140,36 +142,58 @@ def generate_stream_response(messages, temperature, top_p, max_tokens, tools=Non
if isinstance(text, dict):
yield json.dumps(text, ensure_ascii=False)
continue
- full_text += text
+ pending += text
- if not thinking_ended:
- pos = full_text.find('</think>')
- if pos >= 0:
- thinking_ended = True
- new_r = full_text[emitted:pos]
- if new_r:
- yield json.dumps({"choices": [{"delta": {"reasoning_content": new_r}}]}, ensure_ascii=False)
- emitted = pos + len('</think>')
- after = full_text[emitted:].lstrip('\n')
- emitted = len(full_text) - len(after)
- if after:
- yield json.dumps({"choices": [{"delta": {"content": after}}]}, ensure_ascii=False)
- emitted = len(full_text)
+ while pending:
+ if strip_newlines:
+ pending = pending.lstrip('\n')
+ if not pending:
+ break
+ strip_newlines = False
+ if in_tool_call:
+ end = pending.find('</tool_call>')
+ if end < 0:
+ break
+ end += len('</tool_call>')
+ block, pending = pending[:end], pending[end:]
+ _, _, tool_calls = parse_response(block)
+ if tool_calls:
+ call = tool_calls[0]
+ call['index'] = tool_count
+ call['id'] = f"call_{int(time.time())}_{tool_count}"
+ tool_count += 1
+ yield json.dumps({"choices": [{"delta": {"tool_calls": [call]}}]}, ensure_ascii=False)
+ else:
+ yield json.dumps({"choices": [{"delta": {"content": block}}]}, ensure_ascii=False)
+ in_tool_call = False
+ continue
+
+ marker = '<tool_call>' if thinking_ended else '</think>'
+ field = 'content' if thinking_ended else 'reasoning_content'
+ pos = pending.find(marker)
+ if pos < 0:
+ # 保留可能跨 chunk 的标记前缀,其余文本立即发送。
+ keep = next((n for n in range(len(marker) - 1, 0, -1)
+ if pending.endswith(marker[:n])), 0)
+ end = len(pending) - keep
+ if end:
+ yield json.dumps({"choices": [{"delta": {field: pending[:end]}}]}, ensure_ascii=False)
+ pending = pending[end:]
+ break
+ if pos:
+ yield json.dumps({"choices": [{"delta": {field: pending[:pos]}}]}, ensure_ascii=False)
+ pending = pending[pos:]
+ if thinking_ended:
+ in_tool_call = True
else:
- new_r = full_text[emitted:]
- if new_r:
- yield json.dumps({"choices": [{"delta": {"reasoning_content": new_r}}]}, ensure_ascii=False)
- emitted = len(full_text)
- else:
- new_c = full_text[emitted:]
- if new_c:
- yield json.dumps({"choices": [{"delta": {"content": new_c}}]}, ensure_ascii=False)
- emitted = len(full_text)
-
- _, _, tool_calls = parse_response(full_text)
- if tool_calls:
- yield json.dumps({"choices": [{"delta": {"tool_calls": tool_calls}}]}, ensure_ascii=False)
- yield json.dumps({"choices": [{"delta": {}, "finish_reason": "tool_calls" if tool_calls else "stop"}]}, ensure_ascii=False)
+ pending = pending[len(marker):]
+ thinking_ended = True
+ strip_newlines = True
+
+ if pending:
+ field = 'content' if thinking_ended else 'reasoning_content'
+ yield json.dumps({"choices": [{"delta": {field: pending}}]}, ensure_ascii=False)
+ yield json.dumps({"choices": [{"delta": {}, "finish_reason": "tool_calls" if tool_count else "stop"}]}, ensure_ascii=False)
except Exception as e:
yield json.dumps({"error": str(e)})
diff --git a/tests/test_stream_tool_calls.py b/tests/test_stream_tool_calls.py
new file mode 100644
index 0000000..e51332f
--- /dev/null
+++ b/tests/test_stream_tool_calls.py
@@ -0,0 +1,147 @@
+"""python -m unittest discover -s tests -p test_stream_tool_calls.py -v"""
+import json
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+import torch
+from fastapi.testclient import TestClient
+from openai import OpenAI
+from transformers import AutoTokenizer
+
+from model.model_minimind import MiniMindConfig, MiniMindForCausalLM
+from scripts import serve_openai_api as server
+
+
+TOOL = '<tool_call>{"name":"weather","arguments":{"city":"杭州"}}</tool_call>'
+TOOLS = [{'type': 'function', 'function': {'name': 'weather', 'parameters': {
+ 'type': 'object', 'properties': {'city': {'type': 'string'}}, 'required': ['city']}}}]
+
+
+class ScriptedModel:
+ def __init__(self, chunks, tokenize=False):
+ self.chunks, self.tokenize = chunks, tokenize
+
+ def generate(self, input_ids, streamer, **kwargs):
+ if self.tokenize:
+ streamer.put(input_ids.cpu())
+ for token in streamer.tokenizer.encode(''.join(self.chunks), add_special_tokens=False):
+ streamer.put(torch.tensor([token]))
+ streamer.end()
+ else:
+ for chunk in self.chunks:
+ streamer.on_finalized_text(chunk)
+ streamer.on_finalized_text('', stream_end=True)
+
+
+class StreamToolTests(unittest.TestCase):
+ @classmethod
+ def setUpClass(cls):
+ cls.tokenizer = AutoTokenizer.from_pretrained(
+ Path(__file__).resolve().parents[1] / 'model', local_files_only=True)
+
+ def setUp(self):
+ globals_patch = patch.multiple(server, tokenizer=self.tokenizer, device='cpu',
+ model=ScriptedModel([]), create=True)
+ globals_patch.start()
+ self.addCleanup(globals_patch.stop)
+ self.http = TestClient(server.app)
+ self.addCleanup(self.http.close)
+ self.request = {'model': 'minimind', 'messages': [{'role': 'user', 'content': 'Hi'}]}
+
+ def frames(self, chunks, thinking=False):
+ server.model = ScriptedModel(chunks)
+ response = self.http.post('/v1/chat/completions', json={
+ **self.request, 'stream': True, 'open_thinking': thinking})
+ self.assertEqual(response.status_code, 200)
+ self.assertIn('text/event-stream', response.headers['content-type'])
+ return [json.loads(line[6:]) for line in response.text.splitlines()
+ if line.startswith('data: ')]
+
+ def assert_output(self, frames, content, count, reasoning=''):
+ self.assertFalse(any('error' in frame for frame in frames), frames)
+ deltas = [frame['choices'][0]['delta'] for frame in frames]
+ self.assertEqual(''.join(d.get('content', '') for d in deltas), content)
+ self.assertEqual(''.join(d.get('reasoning_content', '') for d in deltas), reasoning)
+ calls = [call for d in deltas for call in d.get('tool_calls', [])]
+ self.assertEqual(len(calls), count)
+ self.assertEqual([call['index'] for call in calls], list(range(count)))
+ self.assertEqual(len({call['id'] for call in calls}), count)
+ for call in calls:
+ self.assertEqual(call['function']['name'], 'weather')
+ self.assertEqual(json.loads(call['function']['arguments']), {'city': '杭州'})
+ self.assertEqual(frames[-1]['choices'][0]['finish_reason'], 'tool_calls' if count else 'stop')
+
+ def test_split_markers_and_multiple_calls(self):
+ text = 'Before ' + TOOL + ' between ' + TOOL + ' after'
+ for chunks in ([text], list(text), ['Before <tool_', text[len('Before <tool_'):]]):
+ with self.subTest(chunks=len(chunks)):
+ self.assert_output(self.frames(chunks), 'Before between after', 2)
+
+ def test_thinking_boundary_and_tool_text_inside_reasoning(self):
+ reasoning = 'Consider ' + TOOL + '\n'
+ text = reasoning + '</think>\n\n' + TOOL + ' answer'
+ self.assert_output(self.frames(list(text), thinking=True), ' answer', 1, reasoning)
+
+ def test_invalid_and_unfinished_calls_preserve_text(self):
+ for text in ('plain < text', 'end <tool_', '<tool_call>{"name":',
+ '<tool_call>bad json</tool_call>', '<tool_call>[]</tool_call>'):
+ with self.subTest(text=text):
+ self.assert_output(self.frames(list(text)), text, 0)
+ invalid = '<tool_call>bad json</tool_call>'
+ self.assert_output(self.frames([invalid, TOOL]), invalid, 1)
+
+ def test_empty_and_unfinished_thinking(self):
+ self.assert_output(self.frames([]), '', 0)
+ self.assert_output(self.frames(['reason </thi'], thinking=True), '', 0, 'reason </thi')
+
+ def test_plain_content_and_tool_call_arrive_before_generation_ends(self):
+ def generate(input_ids, streamer, **kwargs):
+ streamer.on_finalized_text('first ')
+ self.assertTrue(continue_generation.wait(5))
+ streamer.on_finalized_text(TOOL)
+ self.assertTrue(finish_generation.wait(5))
+ streamer.on_finalized_text(' last', stream_end=True)
+
+ from threading import Event
+ continue_generation, finish_generation = Event(), Event()
+ server.model.generate = generate
+ stream = server.generate_stream_response(self.request['messages'], 0.7, 0.92, 128)
+ self.addCleanup(continue_generation.set)
+ self.addCleanup(finish_generation.set)
+ first = json.loads(next(stream))
+ if first['choices'][0]['delta'].get('role') == 'assistant':
+ first = json.loads(next(stream))
+ self.assertEqual(first['choices'][0]['delta']['content'], 'first ')
+ continue_generation.set()
+ call = json.loads(next(stream))
+ self.assertEqual(call['choices'][0]['delta']['tool_calls'][0]['function']['name'], 'weather')
+ finish_generation.set()
+ self.assert_output([first, call] + [json.loads(frame) for frame in stream], 'first last', 1)
+
+ def test_real_text_streamer_and_openai_client(self):
+ server.model = ScriptedModel(['Before ', TOOL, ' after'], tokenize=True)
+ with OpenAI(api_key='local-test', base_url='http://testserver/v1',
+ http_client=self.http, max_retries=0) as client:
+ chunks = list(client.chat.completions.create(**self.request, tools=TOOLS, stream=True))
+ self.assertEqual(''.join(c.choices[0].delta.content or '' for c in chunks), 'Before after')
+ calls = [call for c in chunks for call in c.choices[0].delta.tool_calls or []]
+ self.assertEqual(len(calls), 1)
+ self.assertEqual(calls[0].index, 0)
+ self.assertEqual(json.loads(calls[0].function.arguments), {'city': '杭州'})
+
+ def test_tiny_minimind_generation(self):
+ torch.manual_seed(42)
+ server.model = MiniMindForCausalLM(MiniMindConfig(
+ hidden_size=32, num_hidden_layers=1, num_attention_heads=4,
+ num_key_value_heads=2, intermediate_size=64,
+ vocab_size=len(self.tokenizer), max_position_embeddings=256)).eval()
+ response = self.http.post('/v1/chat/completions', json={
+ **self.request, 'stream': True, 'max_tokens': 4})
+ frames = [json.loads(line[6:]) for line in response.text.splitlines() if line.startswith('data: ')]
+ self.assertFalse(any('error' in frame for frame in frames), frames)
+ self.assertEqual(frames[-1]['choices'][0]['finish_reason'], 'stop')
+
+
+if __name__ == '__main__':
+ unittest.main()