langfun 0.0.2.dev20240429__py3-none-any.whl → 0.1.2.dev202501150804__py3-none-any.whl

langfun/core/llms/openai_compatible.py

@@ -0,0 +1,179 @@
+# Copyright 2025 The Langfun Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Base for OpenAI compatible models (including OpenAI)."""
+
+from typing import Annotated, Any
+
+import langfun.core as lf
+from langfun.core import modalities as lf_modalities
+from langfun.core.llms import rest
+import pyglove as pg
+
+
+@lf.use_init_args(['api_endpoint', 'model'])
+class OpenAICompatible(rest.REST):
+  """Base for OpenAI compatible models."""
+
+  model: Annotated[
+      str, 'The name of the model to use.',
+  ] = ''
+
+  multimodal: Annotated[
+      bool, 'Whether this model has multimodal support.'
+  ] = False
+
+  @property
+  def headers(self) -> dict[str, Any]:
+    return {
+        'Content-Type': 'application/json'
+    }
+
+  def _request_args(
+      self, options: lf.LMSamplingOptions) -> dict[str, Any]:
+    """Returns a dict as request arguments."""
+    # Reference:
+    # https://platform.openai.com/docs/api-reference/completions/create
+    # NOTE(daiyip): options.top_k is not applicable.
+    args = dict(
+        n=options.n,
+        top_logprobs=options.top_logprobs,
+    )
+    if self.model:
+      args['model'] = self.model
+    if options.logprobs:
+      args['logprobs'] = options.logprobs
+    if options.temperature is not None:
+      args['temperature'] = options.temperature
+    if options.max_tokens is not None:
+      args['max_completion_tokens'] = options.max_tokens
+    if options.top_p is not None:
+      args['top_p'] = options.top_p
+    if options.stop:
+      args['stop'] = options.stop
+    if options.random_seed is not None:
+      args['seed'] = options.random_seed
+    return args
+
+  def _content_from_message(self, message: lf.Message) -> list[dict[str, Any]]:
+    """Returns a OpenAI content object from a Langfun message."""
+    content = []
+    for chunk in message.chunk():
+      if isinstance(chunk, str):
+        item = dict(type='text', text=chunk)
+      elif isinstance(chunk, lf_modalities.Image) and self.multimodal:
+        item = dict(type='image_url', image_url=dict(url=chunk.embeddable_uri))
+      else:
+        raise ValueError(f'Unsupported modality: {chunk!r}.')
+      content.append(item)
+    return content
+
+  def request(
+      self,
+      prompt: lf.Message,
+      sampling_options: lf.LMSamplingOptions
+  ) -> dict[str, Any]:
+    """Returns the JSON input for a message."""
+    request_args = self._request_args(sampling_options)
+
+    # Users could use `metadata_json_schema` to pass additional
+    # request arguments.
+    json_schema = prompt.metadata.get('json_schema')
+    if json_schema is not None:
+      if not isinstance(json_schema, dict):
+        raise ValueError(
+            f'`json_schema` must be a dict, got {json_schema!r}.'
+        )
+      if 'title' not in json_schema:
+        raise ValueError(
+            f'The root of `json_schema` must have a `title` field, '
+            f'got {json_schema!r}.'
+        )
+      request_args.update(
+          response_format=dict(
+              type='json_schema',
+              json_schema=dict(
+                  schema=json_schema,
+                  name=json_schema['title'],
+                  strict=True,
+              )
+          )
+      )
+      prompt.metadata.formatted_text = (
+          prompt.text
+          + '\n\n [RESPONSE FORMAT (not part of prompt)]\n'
+          + pg.to_json_str(request_args['response_format'], json_indent=2)
+      )
+
+    # Prepare messages.
+    messages = []
+    # Users could use `metadata_system_message` to pass system message.
+    system_message = prompt.metadata.get('system_message')
+    if system_message:
+      system_message = lf.SystemMessage.from_value(system_message)
+      messages.append(
+          dict(role='system',
+               content=self._content_from_message(system_message))
+      )
+    messages.append(
+        dict(role='user', content=self._content_from_message(prompt))
+    )
+    request = dict()
+    request.update(request_args)
+    request['messages'] = messages
+    return request
+
+  def _parse_choice(self, choice: dict[str, Any]) -> lf.LMSample:
+    # Reference:
+    # https://platform.openai.com/docs/api-reference/chat/object
+    logprobs = None
+    choice_logprobs = choice.get('logprobs')
+    if choice_logprobs:
+      logprobs = [
+          (
+              t['token'],
+              t['logprob'],
+              [(tt['token'], tt['logprob']) for tt in t['top_logprobs']],
+          )
+          for t in choice_logprobs['content']
+      ]
+    return lf.LMSample(
+        choice['message']['content'],
+        score=0.0,
+        logprobs=logprobs,
+    )
+
+  def result(self, json: dict[str, Any]) -> lf.LMSamplingResult:
+    """Returns a LMSamplingResult from a JSON response."""
+    usage = json['usage']
+    return lf.LMSamplingResult(
+        samples=[self._parse_choice(choice) for choice in json['choices']],
+        usage=lf.LMSamplingUsage(
+            prompt_tokens=usage['prompt_tokens'],
+            completion_tokens=usage['completion_tokens'],
+            total_tokens=usage['total_tokens'],
+            estimated_cost=self.estimate_cost(
+                num_input_tokens=usage['prompt_tokens'],
+                num_output_tokens=usage['completion_tokens'],
+            )
+        ),
+    )
+
+  def estimate_cost(
+      self,
+      num_input_tokens: int,
+      num_output_tokens: int
+  ) -> float | None:
+    """Estimate the cost based on usage."""
+    del num_input_tokens, num_output_tokens
+    return None

langfun/core/llms/openai_compatible_test.py

@@ -0,0 +1,495 @@
+# Copyright 2023 The Langfun Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for OpenAI models."""
+
+from typing import Any
+import unittest
+from unittest import mock
+
+import langfun.core as lf
+from langfun.core import modalities as lf_modalities
+from langfun.core.llms import openai_compatible
+import pyglove as pg
+import requests
+
+
+def mock_chat_completion_request(url: str, json: dict[str, Any], **kwargs):
+  del url, kwargs
+  messages = json['messages']
+  if len(messages) > 1:
+    system_message = f' system={messages[0]["content"]}'
+  else:
+    system_message = ''
+
+  if 'response_format' in json:
+    response_format = f' format={json["response_format"]["type"]}'
+  else:
+    response_format = ''
+
+  choices = []
+  for k in range(json['n']):
+    if json.get('logprobs'):
+      logprobs = dict(
+          content=[
+              dict(
+                  token='chosen_token',
+                  logprob=0.5,
+                  top_logprobs=[
+                      dict(
+                          token=f'alternative_token_{i + 1}',
+                          logprob=0.1
+                      ) for i in range(3)
+                  ]
+              )
+          ]
+      )
+    else:
+      logprobs = None
+
+    choices.append(dict(
+        message=dict(
+            content=(
+                f'Sample {k} for message.{system_message}{response_format}'
+            )
+        ),
+        logprobs=logprobs,
+    ))
+  response = requests.Response()
+  response.status_code = 200
+  response._content = pg.to_json_str(
+      dict(
+          choices=choices,
+          usage=lf.LMSamplingUsage(
+              prompt_tokens=100,
+              completion_tokens=100,
+              total_tokens=200,
+          ),
+      )
+  ).encode()
+  return response
+
+
+def mock_chat_completion_request_vision(
+    url: str, json: dict[str, Any], **kwargs
+):
+  del url, kwargs
+  choices = []
+  urls = [
+      c['image_url']['url']
+      for c in json['messages'][0]['content'] if c['type'] == 'image_url'
+  ]
+  for k in range(json['n']):
+    choices.append(pg.Dict(
+        message=pg.Dict(
+            content=f'Sample {k} for message: {"".join(urls)}'
+        ),
+        logprobs=None,
+    ))
+  response = requests.Response()
+  response.status_code = 200
+  response._content = pg.to_json_str(
+      dict(
+          choices=choices,
+          usage=lf.LMSamplingUsage(
+              prompt_tokens=100,
+              completion_tokens=100,
+              total_tokens=200,
+          ),
+      )
+  ).encode()
+  return response
+
+
+class OpenAIComptibleTest(unittest.TestCase):
+  """Tests for OpenAI compatible language model."""
+
+  def test_request_args(self):
+    self.assertEqual(
+        openai_compatible.OpenAICompatible(
+            api_endpoint='https://test-server',
+            model='test-model'
+        )._request_args(
+            lf.LMSamplingOptions(
+                temperature=1.0, stop=['\n'], n=1, random_seed=123
+            )
+        ),
+        dict(
+            model='test-model',
+            top_logprobs=None,
+            n=1,
+            temperature=1.0,
+            stop=['\n'],
+            seed=123,
+        ),
+    )
+
+  def test_call_chat_completion(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model',
+      )
+      self.assertEqual(
+          lm('hello', sampling_options=lf.LMSamplingOptions(n=2)),
+          'Sample 0 for message.',
+      )
+
+  def test_call_chat_completion_with_logprobs(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model',
+      )
+      results = lm.sample(['hello'], logprobs=True)
+      for result in results:
+        result.rebind(
+            {path: 0 for path in pg.query(result, '.*total_call_interval')},
+            skip_notification=True,
+        )
+      self.assertEqual(len(results), 1)
+      self.assertEqual(
+          results[0],
+          lf.LMSamplingResult(
+              [
+                  lf.LMSample(
+                      response=lf.AIMessage(
+                          text='Sample 0 for message.',
+                          metadata={
+                              'score': 0.0,
+                              'logprobs': [(
+                                  'chosen_token',
+                                  0.5,
+                                  [
+                                      ('alternative_token_1', 0.1),
+                                      ('alternative_token_2', 0.1),
+                                      ('alternative_token_3', 0.1),
+                                  ],
+                              )],
+                              'is_cached': False,
+                              'usage': lf.LMSamplingUsage(
+                                  prompt_tokens=100,
+                                  completion_tokens=100,
+                                  total_tokens=200,
+                                  estimated_cost=None,
+                              ),
+                          },
+                          tags=['lm-response'],
+                      ),
+                      logprobs=[(
+                          'chosen_token',
+                          0.5,
+                          [
+                              ('alternative_token_1', 0.1),
+                              ('alternative_token_2', 0.1),
+                              ('alternative_token_3', 0.1),
+                          ],
+                      )],
+                  )
+              ],
+              usage=lf.LMSamplingUsage(
+                  prompt_tokens=100,
+                  completion_tokens=100,
+                  total_tokens=200,
+                  estimated_cost=None,
+              ),
+          ),
+      )
+
+  def test_call_chat_completion_vision(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request_vision
+      lm_1 = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server',
+          model='test-model1',
+          multimodal=True
+      )
+      lm_2 = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server',
+          model='test-model2',
+          multimodal=True
+      )
+      for lm in (lm_1, lm_2):
+        self.assertEqual(
+            lm(
+                lf.UserMessage(
+                    'hello <<[[image]]>>',
+                    image=lf_modalities.Image.from_uri('https://fake/image')
+                ),
+                sampling_options=lf.LMSamplingOptions(n=2)
+            ),
+            'Sample 0 for message: https://fake/image',
+        )
+    lm_3 = openai_compatible.OpenAICompatible(
+        api_endpoint='https://test-server', model='test-model3'
+    )
+    with self.assertRaisesRegex(ValueError, 'Unsupported modality'):
+      lm_3(
+          lf.UserMessage(
+              'hello <<[[image]]>>',
+              image=lf_modalities.Image.from_uri('https://fake/image')
+          ),
+      )
+
+  def test_sample_chat_completion(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model'
+      )
+      results = lm.sample(
+          ['hello', 'bye'], sampling_options=lf.LMSamplingOptions(n=3)
+      )
+      for result in results:
+        result.rebind(
+            {path: 0 for path in pg.query(result, '.*total_call_interval')},
+            skip_notification=True,
+        )
+
+    self.assertEqual(len(results), 2)
+    self.assertEqual(
+        results[0],
+        lf.LMSamplingResult(
+            [
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 0 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 1 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 2 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+            ],
+            usage=lf.LMSamplingUsage(
+                prompt_tokens=100, completion_tokens=100, total_tokens=200,
+                estimated_cost=None,
+            ),
+        ),
+    )
+    self.assertEqual(
+        results[1],
+        lf.LMSamplingResult(
+            [
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 0 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 1 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 2 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=33,
+                            completion_tokens=33,
+                            total_tokens=66,
+                            estimated_cost=None,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+            ],
+            usage=lf.LMSamplingUsage(
+                prompt_tokens=100, completion_tokens=100, total_tokens=200,
+                estimated_cost=None,
+            ),
+        ),
+    )
+
+  def test_sample_with_contextual_options(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model'
+      )
+      with lf.use_settings(sampling_options=lf.LMSamplingOptions(n=2)):
+        results = lm.sample(['hello'])
+      for result in results:
+        result.rebind(
+            {path: 0 for path in pg.query(result, '.*total_call_interval')},
+            skip_notification=True,
+        )
+
+    self.assertEqual(len(results), 1)
+    self.assertEqual(
+        results[0],
+        lf.LMSamplingResult(
+            [
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 0 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=50,
+                            completion_tokens=50,
+                            total_tokens=100,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+                lf.LMSample(
+                    lf.AIMessage(
+                        'Sample 1 for message.',
+                        score=0.0,
+                        logprobs=None,
+                        is_cached=False,
+                        usage=lf.LMSamplingUsage(
+                            prompt_tokens=50,
+                            completion_tokens=50,
+                            total_tokens=100,
+                        ),
+                        tags=[lf.Message.TAG_LM_RESPONSE],
+                    ),
+                    score=0.0,
+                    logprobs=None,
+                ),
+            ],
+            usage=lf.LMSamplingUsage(
+                prompt_tokens=100, completion_tokens=100, total_tokens=200
+            ),
+        )
+    )
+
+  def test_call_with_system_message(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model'
+      )
+      self.assertEqual(
+          lm(
+              lf.UserMessage(
+                  'hello',
+                  system_message='hi',
+              ),
+              sampling_options=lf.LMSamplingOptions(n=2)
+          ),
+          '''Sample 0 for message. system=[{'type': 'text', 'text': 'hi'}]''',
+      )
+
+  def test_call_with_json_schema(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_chat_completion_request
+      lm = openai_compatible.OpenAICompatible(
+          api_endpoint='https://test-server', model='test-model'
+      )
+      self.assertEqual(
+          lm(
+              lf.UserMessage(
+                  'hello',
+                  json_schema={
+                      'type': 'object',
+                      'properties': {
+                          'name': {'type': 'string'},
+                      },
+                      'required': ['name'],
+                      'title': 'Person',
+                  }
+              ),
+              sampling_options=lf.LMSamplingOptions(n=2)
+          ),
+          'Sample 0 for message. format=json_schema',
+      )
+
+    # Test bad json schema.
+    with self.assertRaisesRegex(ValueError, '`json_schema` must be a dict'):
+      lm(lf.UserMessage('hello', json_schema='foo'))
+
+    with self.assertRaisesRegex(
+        ValueError, 'The root of `json_schema` must have a `title` field'
+    ):
+      lm(lf.UserMessage('hello', json_schema={}))
+
+
+if __name__ == '__main__':
+  unittest.main()