langfun 0.0.2.dev20240422__py3-none-any.whl → 0.0.2.dev20240423__py3-none-any.whl

langfun/core/component.py

@@ -210,6 +210,12 @@ def get_contextual_override(var_name: str) -> ContextualOverride | None:
   return _get_scoped_value(_global_tls, _CONTEXT_OVERRIDES, var_name)
 
 
+def all_contextual_values() -> dict[str, Any]:
+  """Returns all contextual values provided from `lf.context` in scope."""
+  overrides = getattr(_global_tls, _CONTEXT_OVERRIDES, {})
+  return {k: v.value for k, v in overrides.items()}
+
+
 @contextlib.contextmanager
 def _contextual_scope(
     tls: threading.local, tls_key, **variables

langfun/core/component_test.py

@@ -84,6 +84,7 @@ class ComponentContextTest(unittest.TestCase):
           lf.get_contextual_override('y'),
           lf.ContextualOverride(3, cascade=False, override_attrs=False),
       )
+      self.assertEqual(lf.all_contextual_values(), dict(x=3, y=3, z=3))
 
       # Member attributes take precedence over `lf.context`.
       self.assertEqual(a1.x, 1)

langfun/core/language_model.py

@@ -24,6 +24,9 @@ from langfun.core import console
 from langfun.core import message as message_lib
 import pyglove as pg
 
+TOKENS_PER_REQUEST = 250  # Estimated num tokens for a single request
+DEFAULT_MAX_CONCURRENCY = 1  # Use this as max concurrency if no RPM or TPM data
+
 
 class LMSample(pg.Object):
   """Response candidate."""
@@ -604,3 +607,14 @@ class LanguageModel(component.Component):
             f'score: {r.score}',
             color='blue',
         )
+
+  def rate_to_max_concurrency(
+      self, requests_per_min: float = 0, tokens_per_min: float = 0
+  ) -> int:
+    """Converts a rate to a max concurrency."""
+    if tokens_per_min > 0:
+      return max(int(tokens_per_min / TOKENS_PER_REQUEST / 60), 1)
+    elif requests_per_min > 0:
+      return max(int(requests_per_min / 60), 1)  # Max concurrency can't be zero
+    else:
+      return DEFAULT_MAX_CONCURRENCY  # Default of 1

langfun/core/language_model_test.py

@@ -394,6 +394,38 @@ class LanguageModelTest(unittest.TestCase):
     with self.assertRaises(NotImplementedError):
       MockModel().score('hi', ['1', '2'])
 
+  def test_rate_to_max_concurrency_no_rpm_no_tpm(self) -> None:
+    lm = MockModel()
+    self.assertEqual(
+        lm_lib.DEFAULT_MAX_CONCURRENCY,
+        lm.rate_to_max_concurrency(requests_per_min=0, tokens_per_min=0),
+    )
+    self.assertEqual(
+        lm_lib.DEFAULT_MAX_CONCURRENCY,
+        lm.rate_to_max_concurrency(requests_per_min=-1, tokens_per_min=-1),
+    )
+
+  def test_rate_to_max_concurrency_only_rpm_specified_uses_rpm(self) -> None:
+    lm = MockModel()
+    test_rpm = 1e4
+    self.assertEqual(
+        lm.rate_to_max_concurrency(requests_per_min=test_rpm),
+        int(test_rpm / 60)
+    )
+
+  def test_rate_to_max_concurrency_tpm_specified_uses_tpm(self) -> None:
+    lm = MockModel()
+    test_tpm = 1e7
+    self.assertEqual(
+        lm.rate_to_max_concurrency(requests_per_min=1, tokens_per_min=test_tpm),
+        int(test_tpm / lm_lib.TOKENS_PER_REQUEST / 60),
+    )
+
+  def test_rate_to_max_concurrency_small_rate_returns_one(self) -> None:
+    lm = MockModel()
+    self.assertEqual(lm.rate_to_max_concurrency(requests_per_min=1), 1)
+    self.assertEqual(lm.rate_to_max_concurrency(tokens_per_min=1), 1)
+
 
 if __name__ == '__main__':
   unittest.main()

langfun/core/llms/anthropic.py

@@ -26,12 +26,15 @@ import requests
 
 SUPPORTED_MODELS_AND_SETTINGS = {
     # See https://docs.anthropic.com/claude/docs/models-overview
-    'claude-3-opus-20240229': pg.Dict(max_tokens=4096, max_concurrency=16),
-    'claude-3-sonnet-20240229': pg.Dict(max_tokens=4096, max_concurrency=16),
-    'claude-3-haiku-20240307': pg.Dict(max_tokens=4096, max_concurrency=16),
-    'claude-2.1': pg.Dict(max_tokens=4096, max_concurrency=16),
-    'claude-2.0': pg.Dict(max_tokens=4096, max_concurrency=16),
-    'claude-instant-1.2': pg.Dict(max_tokens=4096, max_concurrency=16),
+    # Rate limits from https://docs.anthropic.com/claude/reference/rate-limits
+    #     RPM/TPM for Claude-2.1, Claude-2.0, and Claude-Instant-1.2 estimated
+    #     as RPM/TPM of the largest-available model (Claude-3-Opus).
+    'claude-3-opus-20240229': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
+    'claude-3-sonnet-20240229': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
+    'claude-3-haiku-20240307': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
+    'claude-2.1': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
+    'claude-2.0': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
+    'claude-instant-1.2': pg.Dict(max_tokens=4096, rpm=4000, tpm=400000),
 }
 
 
@@ -81,6 +84,7 @@ class Anthropic(lf.LanguageModel):
     super()._on_bound()
     self._api_key = None
     self.__dict__.pop('_api_initialized', None)
+    self.__dict__.pop('_session', None)
 
   @functools.cached_property
   def _api_initialized(self):
@@ -93,6 +97,17 @@ class Anthropic(lf.LanguageModel):
     self._api_key = api_key
     return True
 
+  @functools.cached_property
+  def _session(self) -> requests.Session:
+    assert self._api_initialized
+    s = requests.Session()
+    s.headers.update({
+        'x-api-key': self._api_key,
+        'anthropic-version': _ANTHROPIC_API_VERSION,
+        'content-type': 'application/json',
+    })
+    return s
+
   @property
   def model_id(self) -> str:
     """Returns a string to identify the model."""
@@ -100,7 +115,11 @@ class Anthropic(lf.LanguageModel):
 
   @property
   def max_concurrency(self) -> int:
-    return SUPPORTED_MODELS_AND_SETTINGS[self.model].max_concurrency
+    rpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('rpm', 0)
+    tpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('tpm', 0)
+    return self.rate_to_max_concurrency(
+        requests_per_min=rpm, tokens_per_min=tpm
+    )
 
   def _sample(self, prompts: list[lf.Message]) -> list[lf.LMSamplingResult]:
     assert self._api_initialized
@@ -165,8 +184,8 @@ class Anthropic(lf.LanguageModel):
   def _parse_response(self, response: requests.Response) -> lf.LMSamplingResult:
     """Parses Anthropic's response."""
     # NOTE(daiyip): Refer https://docs.anthropic.com/claude/reference/errors
-    output = response.json()
     if response.status_code == 200:
+      output = response.json()
       message = self._message_from_content(output['content'])
       input_tokens = output['usage']['input_tokens']
       output_tokens = output['usage']['output_tokens']
@@ -181,12 +200,11 @@ class Anthropic(lf.LanguageModel):
     else:
       if response.status_code == 429:
         error_cls = RateLimitError
-      elif response.status_code == 529:
+      elif response.status_code in (502, 529):
         error_cls = OverloadedError
       else:
         error_cls = AnthropicError
-      error = output['error']
-      raise error_cls(f'{error["type"]}: {error["message"]}')
+      raise error_cls(f'{response.status_code}: {response.content}')
 
   def _sample_single(self, prompt: lf.Message) -> lf.LMSamplingResult:
     request = dict()
@@ -198,17 +216,13 @@ class Anthropic(lf.LanguageModel):
             ]
         )
     )
-    response = requests.post(
-        _ANTHROPIC_MESSAGE_API_ENDPOINT,
-        json=request,
-        headers={
-            'x-api-key': self._api_key,
-            'anthropic-version': _ANTHROPIC_API_VERSION,
-            'content-type': 'application/json',
-        },
-        timeout=self.timeout,
-    )
-    return self._parse_response(response)
+    try:
+      response = self._session.post(
+          _ANTHROPIC_MESSAGE_API_ENDPOINT, json=request, timeout=self.timeout,
+      )
+      return self._parse_response(response)
+    except ConnectionError as e:
+      raise OverloadedError(str(e)) from e
 
 
 class Claude3(Anthropic):

langfun/core/llms/anthropic_test.py

@@ -98,20 +98,20 @@ def mock_requests_post_error(status_code, error_type, error_message):
   return _mock_requests
 
 
-class AuthropicTest(unittest.TestCase):
+class AnthropicTest(unittest.TestCase):
 
   def test_basics(self):
     self.assertEqual(
         anthropic.Claude3Haiku().model_id, 'claude-3-haiku-20240307'
     )
-    self.assertEqual(anthropic.Claude3Haiku().max_concurrency, 16)
+    self.assertGreater(anthropic.Claude3Haiku().max_concurrency, 0)
 
   def test_api_key(self):
     lm = anthropic.Claude3Haiku()
     with self.assertRaisesRegex(ValueError, 'Please specify `api_key`'):
       lm('hi')
 
-    with mock.patch('requests.post') as mock_request:
+    with mock.patch('requests.Session.post') as mock_request:
       mock_request.side_effect = mock_requests_post
 
       lm = anthropic.Claude3Haiku(api_key='fake key')
@@ -123,7 +123,7 @@ class AuthropicTest(unittest.TestCase):
       del os.environ['ANTHROPIC_API_KEY']
 
   def test_call(self):
-    with mock.patch('requests.post') as mock_request:
+    with mock.patch('requests.Session.post') as mock_request:
       mock_request.side_effect = mock_requests_post
       lm = anthropic.Claude3Haiku(api_key='fake_key')
       response = lm('hello', temperature=0.0, top_k=0.1, top_p=0.2, stop=['\n'])
@@ -140,7 +140,7 @@ class AuthropicTest(unittest.TestCase):
       self.assertIsNotNone(response.usage.total_tokens, 3)
 
   def test_mm_call(self):
-    with mock.patch('requests.post') as mock_mm_request:
+    with mock.patch('requests.Session.post') as mock_mm_request:
       mock_mm_request.side_effect = mock_mm_requests_post
       lm = anthropic.Claude3Haiku(api_key='fake_key')
       response = lm(lf_modalities.Image.from_bytes(image_content), lm=lm)
@@ -152,13 +152,13 @@ class AuthropicTest(unittest.TestCase):
         (529, 'service_unavailable', 'Service unavailable.'),
         (500, 'bad_request', 'Bad request.'),
     ]:
-      with mock.patch('requests.post') as mock_mm_request:
+      with mock.patch('requests.Session.post') as mock_mm_request:
         mock_mm_request.side_effect = mock_requests_post_error(
             status_code, error_type, error_message
         )
         lm = anthropic.Claude3Haiku(api_key='fake_key')
         with self.assertRaisesRegex(
-            Exception, f'{error_type}: {error_message}'
+            Exception, f'.*{status_code}: .*{error_message}'
         ):
           lm('hello', lm=lm, max_attempts=1)
 

langfun/core/llms/groq.py

@@ -78,6 +78,7 @@ class Groq(lf.LanguageModel):
     super()._on_bound()
     self._api_key = None
     self.__dict__.pop('_api_initialized', None)
+    self.__dict__.pop('_session', None)
 
   @functools.cached_property
   def _api_initialized(self):
@@ -85,11 +86,21 @@ class Groq(lf.LanguageModel):
     if not api_key:
       raise ValueError(
           'Please specify `api_key` during `__init__` or set environment '
-          'variable `GROQ_API_KEY` with your Anthropic API key.'
+          'variable `GROQ_API_KEY` with your Groq API key.'
       )
     self._api_key = api_key
     return True
 
+  @functools.cached_property
+  def _session(self) -> requests.Session:
+    assert self._api_initialized
+    s = requests.Session()
+    s.headers.update({
+        'Authorization': f'Bearer {self._api_key}',
+        'Content-Type': 'application/json',
+    })
+    return s
+
   @property
   def model_id(self) -> str:
     """Returns a string to identify the model."""
@@ -119,7 +130,7 @@ class Groq(lf.LanguageModel):
     return args
 
   def _content_from_message(self, prompt: lf.Message) -> list[dict[str, Any]]:
-    """Converts an message to Anthropic's content protocol (list of dicts)."""
+    """Converts an message to Groq's content protocol (list of dicts)."""
     # Refer: https://platform.openai.com/docs/api-reference/chat/create
     content = []
     for chunk in prompt.chunk():
@@ -138,7 +149,7 @@ class Groq(lf.LanguageModel):
     return content
 
   def _message_from_choice(self, choice: dict[str, Any]) -> lf.Message:
-    """Converts Anthropic's content protocol to message."""
+    """Converts Groq's content protocol to message."""
     # Refer: https://platform.openai.com/docs/api-reference/chat/create
     content = choice['message']['content']
     if isinstance(content, str):
@@ -148,10 +159,10 @@ class Groq(lf.LanguageModel):
     )
 
   def _parse_response(self, response: requests.Response) -> lf.LMSamplingResult:
-    """Parses Anthropic's response."""
+    """Parses Groq's response."""
     # Refer: https://platform.openai.com/docs/api-reference/chat/object
-    output = response.json()
     if response.status_code == 200:
+      output = response.json()
       samples = [
           lf.LMSample(self._message_from_choice(choice), score=0.0)
           for choice in output['choices']
@@ -169,12 +180,11 @@ class Groq(lf.LanguageModel):
       # https://platform.openai.com/docs/guides/error-codes/api-errors
       if response.status_code == 429:
         error_cls = RateLimitError
-      elif response.status_code in (500, 503):
+      elif response.status_code in (500, 502, 503):
         error_cls = OverloadedError
       else:
         error_cls = GroqError
-      error = output['error']
-      raise error_cls(f'{error["type"]}: {error["message"]}')
+      raise error_cls(f'{response.status_code}: {response.content}')
 
   def _sample(self, prompts: list[lf.Message]) -> list[lf.LMSamplingResult]:
     assert self._api_initialized
@@ -194,16 +204,15 @@ class Groq(lf.LanguageModel):
             ]
         )
     )
-    response = requests.post(
-        _CHAT_COMPLETE_API_ENDPOINT,
-        json=request,
-        headers={
-            'Authorization': f'Bearer {self._api_key}',
-            'Content-Type': 'application/json',
-        },
-        timeout=self.timeout,
-    )
-    return self._parse_response(response)
+    try:
+      response = self._session.post(
+          _CHAT_COMPLETE_API_ENDPOINT,
+          json=request,
+          timeout=self.timeout,
+      )
+      return self._parse_response(response)
+    except ConnectionError as e:
+      raise OverloadedError(str(e)) from e
 
 
 class GroqLlama3_8B(Groq):  # pylint: disable=invalid-name

langfun/core/llms/groq_test.py

@@ -107,7 +107,7 @@ class AuthropicTest(unittest.TestCase):
     with self.assertRaisesRegex(ValueError, 'Please specify `api_key`'):
       lm('hi')
 
-    with mock.patch('requests.post') as mock_request:
+    with mock.patch('requests.Session.post') as mock_request:
       mock_request.side_effect = mock_requests_post
 
       lm = groq.GroqMistral_8x7B(api_key='fake key')
@@ -119,7 +119,7 @@ class AuthropicTest(unittest.TestCase):
       del os.environ['GROQ_API_KEY']
 
   def test_call(self):
-    with mock.patch('requests.post') as mock_request:
+    with mock.patch('requests.Session.post') as mock_request:
       mock_request.side_effect = mock_requests_post
       lm = groq.GroqLlama3_70B(api_key='fake_key')
       response = lm(
@@ -143,7 +143,7 @@ class AuthropicTest(unittest.TestCase):
       self.assertIsNotNone(response.usage.total_tokens, 3)
 
   def test_mm_call(self):
-    with mock.patch('requests.post') as mock_mm_request:
+    with mock.patch('requests.Session.post') as mock_mm_request:
       mock_mm_request.side_effect = mock_mm_requests_post
       lm = groq.GroqLlama3_70B(multimodal=True, api_key='fake_key')
       response = lm(lf_modalities.Image.from_uri('https://fake/image.jpg'))
@@ -155,13 +155,13 @@ class AuthropicTest(unittest.TestCase):
         (503, 'service_unavailable', 'Service unavailable.'),
         (500, 'bad_request', 'Bad request.'),
     ]:
-      with mock.patch('requests.post') as mock_mm_request:
+      with mock.patch('requests.Session.post') as mock_mm_request:
         mock_mm_request.side_effect = mock_requests_post_error(
             status_code, error_type, error_message
         )
         lm = groq.GroqLlama3_70B(api_key='fake_key')
         with self.assertRaisesRegex(
-            Exception, f'{error_type}: {error_message}'
+            Exception, f'{status_code}:.*{error_type}'
         ):
           lm('hello', lm=lm, max_attempts=1)
 

langfun/core/llms/openai.py

@@ -26,54 +26,55 @@ from openai import openai_object
 import pyglove as pg
 
 
-SUPPORTED_MODELS_AND_SETTINGS = [
-    # Model name, max concurrent requests.
-    # The concurrent requests is estimated by TPM/RPM from
-    # https://platform.openai.com/account/limits
-    # GPT-4 Turbo models.
-    ('gpt-4-turbo', 8),  # GPT-4 Turbo with Vision
-    ('gpt-4-turbo-2024-04-09', 8),  # GPT-4-Turbo with Vision, 04/09/2024
-    ('gpt-4-turbo-preview', 8),  # GPT-4 Turbo Preview
-    ('gpt-4-0125-preview', 8),  # GPT-4 Turbo Preview, 01/25/2024
-    ('gpt-4-1106-preview', 8),  # GPT-4 Turbo Preview, 11/06/2023
-    ('gpt-4-vision-preview', 8),  # GPT-4 Turbo Vision Preview.
-    ('gpt-4-1106-vision-preview', 8),  # GPT-4 Turbo Vision Preview, 11/06/2023
-    # GPT-4 models.
-    ('gpt-4', 4),
-    ('gpt-4-0613', 4),
-    ('gpt-4-0314', 4),
-    ('gpt-4-32k', 4),
-    ('gpt-4-32k-0613', 4),
-    ('gpt-4-32k-0314', 4),
-    # GPT-3.5 Turbo models.
-    ('gpt-3.5-turbo', 16),
-    ('gpt-3.5-turbo-0125', 16),
-    ('gpt-3.5-turbo-1106', 16),
-    ('gpt-3.5-turbo-0613', 16),
-    ('gpt-3.5-turbo-0301', 16),
-    ('gpt-3.5-turbo-16k', 16),
-    ('gpt-3.5-turbo-16k-0613', 16),
-    ('gpt-3.5-turbo-16k-0301', 16),
-    # GPT-3.5 models.
-    ('text-davinci-003', 8),  # GPT-3.5, trained with RHLF.
-    ('text-davinci-002', 4),  # Trained with SFT but no RHLF.
-    ('code-davinci-002', 4),
-    # GPT-3 instruction-tuned models.
-    ('text-curie-001', 4),
-    ('text-babbage-001', 4),
-    ('text-ada-001', 4),
-    ('davinci', 4),
-    ('curie', 4),
-    ('babbage', 4),
-    ('ada', 4),
-    # GPT-3 base models without instruction tuning.
-    ('babbage-002', 4),
-    ('davinci-002', 4),
-]
-
-
-# Model concurreny setting.
-_MODEL_CONCURRENCY = {m[0]: m[1] for m in SUPPORTED_MODELS_AND_SETTINGS}
+# From https://platform.openai.com/settings/organization/limits
+_DEFAULT_TPM = 250000
+_DEFAULT_RPM = 3000
+
+SUPPORTED_MODELS_AND_SETTINGS = {
+    # Models from https://platform.openai.com/docs/models
+    # RPM is from https://platform.openai.com/docs/guides/rate-limits
+    # GPT-4-Turbo models
+    'gpt-4-turbo': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-turbo-2024-04-09': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-turbo-preview': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-0125-preview': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-1106-preview': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-vision-preview': pg.Dict(rpm=10000, tpm=1500000),
+    'gpt-4-1106-vision-preview': pg.Dict(
+        rpm=10000, tpm=1500000
+    ),
+    # GPT-4 models
+    'gpt-4': pg.Dict(rpm=10000, tpm=300000),
+    'gpt-4-0613': pg.Dict(rpm=10000, tpm=300000),
+    'gpt-4-0314': pg.Dict(rpm=10000, tpm=300000),
+    'gpt-4-32k': pg.Dict(rpm=10000, tpm=300000),
+    'gpt-4-32k-0613': pg.Dict(rpm=10000, tpm=300000),
+    'gpt-4-32k-0314': pg.Dict(rpm=10000, tpm=300000),
+    # GPT-3.5-Turbo models
+    'gpt-3.5-turbo': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-0125': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-1106': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-0613': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-0301': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-16k': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-16k-0613': pg.Dict(rpm=10000, tpm=2000000),
+    'gpt-3.5-turbo-16k-0301': pg.Dict(rpm=10000, tpm=2000000),
+    # GPT-3.5 models
+    'text-davinci-003': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'text-davinci-002': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'code-davinci-002': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    # GPT-3 instruction-tuned models
+    'text-curie-001': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'text-babbage-001': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'text-ada-001': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'davinci': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'curie': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'babbage': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'ada': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    # GPT-3 base models
+    'babbage-002': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+    'davinci-002': pg.Dict(rpm=_DEFAULT_RPM, tpm=_DEFAULT_TPM),
+}
 
 
 @lf.use_init_args(['model'])
@@ -82,7 +83,7 @@ class OpenAI(lf.LanguageModel):
 
   model: pg.typing.Annotated[
       pg.typing.Enum(
-          pg.MISSING_VALUE, [m[0] for m in SUPPORTED_MODELS_AND_SETTINGS]
+          pg.MISSING_VALUE, list(SUPPORTED_MODELS_AND_SETTINGS.keys())
       ),
       'The name of the model to use.',
   ] = 'gpt-3.5-turbo'
@@ -134,7 +135,11 @@ class OpenAI(lf.LanguageModel):
 
   @property
   def max_concurrency(self) -> int:
-    return _MODEL_CONCURRENCY[self.model]
+    rpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('rpm', 0)
+    tpm = SUPPORTED_MODELS_AND_SETTINGS[self.model].get('tpm', 0)
+    return self.rate_to_max_concurrency(
+        requests_per_min=rpm, tokens_per_min=tpm
+    )
 
   @classmethod
   def dir(cls):

langfun/core/llms/openai_test.py

@@ -11,7 +11,7 @@
 # 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."""
+"""Tests for OpenAI models."""
 
 import unittest
 from unittest import mock
@@ -85,7 +85,7 @@ def mock_chat_completion_query_vision(messages, *, n=1, **kwargs):
   )
 
 
-class OpenaiTest(unittest.TestCase):
+class OpenAITest(unittest.TestCase):
   """Tests for OpenAI language model."""
 
   def test_model_id(self):
@@ -98,7 +98,7 @@ class OpenaiTest(unittest.TestCase):
     )
 
   def test_max_concurrency(self):
-    self.assertEqual(openai.Gpt35(api_key='test_key').max_concurrency, 8)
+    self.assertGreater(openai.Gpt35(api_key='test_key').max_concurrency, 0)
 
   def test_get_request_args(self):
     self.assertEqual(

langfun/core/template.py

@@ -38,6 +38,10 @@ NO_TEMPLATE_DOCSTR_SIGN = 'THIS IS NOT A TEMPLATE'
 _TLS_RENDER_STACK = '_template_render_stack'
 _TLS_RENDER_RESULT_CACHE = '_template_render_result_cache'
 
+# The prefix for fields or contextual attributes to be treated as additional
+# metadata for rendered message.
+_ADDITIONAL_METADATA_PREFIX = 'metadata_'
+
 
 class Template(
     natural_language.NaturalLanguageFormattable,
@@ -303,19 +307,19 @@ class Template(
               with modality.format_modality_as_ref():
                 rendered_text = self._template.render(**inputs)
 
+            # Carry additional metadata.
+            metadata = self.additional_metadata()
+
         if self.clean:
           rendered_text = rendered_text.strip()
 
-        # Fill the variables for rendering the template as metadata.
-        message = message_cls(
-            text=rendered_text,
-            metadata={
-                k: pg.Ref(v)
-                for k, v in inputs.items()
-                if not inspect.ismethod(v)
-            },
+        metadata.update(
+            {k: pg.Ref(v) for k, v in inputs.items() if not inspect.ismethod(v)}
         )
 
+        # Fill the variables for rendering the template as metadata.
+        message = message_cls(text=rendered_text, metadata=metadata)
+
         # Tag input as rendered message.
         message.tag(message_lib.Message.TAG_RENDERED)
 
@@ -340,6 +344,20 @@ class Template(
       top = pg.object_utils.thread_local_pop(_TLS_RENDER_STACK)
       assert top is self, (top, self)
 
+  def additional_metadata(self) -> dict[str, Any]:
+    """Returns additional metadta to be carried in the rendered message."""
+    metadata = {}
+    # Carry metadata from `lf.context`.
+    for k, v in component.all_contextual_values().items():
+      if k.startswith(_ADDITIONAL_METADATA_PREFIX):
+        metadata[k.removeprefix(_ADDITIONAL_METADATA_PREFIX)] = v
+
+    # Carry metadata from fields.
+    for k, v in self.sym_init_args.items():
+      if k.startswith(_ADDITIONAL_METADATA_PREFIX):
+        metadata[k.removeprefix(_ADDITIONAL_METADATA_PREFIX)] = v
+    return metadata
+
   #
   # Implements `pg.typing.CustomTyping`.
   #

langfun/core/template_test.py

@@ -16,6 +16,7 @@ import inspect
 import unittest
 
 from langfun.core import component
+from langfun.core import message as message_lib
 from langfun.core import modality
 from langfun.core import subscription
 from langfun.core.template import Template
@@ -427,6 +428,14 @@ class RenderTest(unittest.TestCase):
     # Test len.
     self.assert_partial(Template('Hello {{len(x)}}'), 'Hello {{len(x)}}')
 
+  def test_additional_metadata(self):
+    t = Template('hi', metadata_weights=1.0, y=2)
+    self.assertEqual(t.render(), message_lib.UserMessage('hi', weights=1.0))
+
+    t = Template('hi')
+    with component.context(metadata_weights=1.0, y=2):
+      self.assertEqual(t.render(), message_lib.UserMessage('hi', weights=1.0))
+
 
 class TemplateRenderEventTest(unittest.TestCase):
 

{langfun-0.0.2.dev20240422.dist-info → langfun-0.0.2.dev20240423.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langfun
-Version: 0.0.2.dev20240422
+Version: 0.0.2.dev20240423
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.0.2.dev20240422.dist-info → langfun-0.0.2.dev20240423.dist-info}/RECORD

@@ -1,15 +1,15 @@
 langfun/__init__.py,sha256=3iCC7F8XoRZ7Gvus11NT50e4KDOQJxIPn9a7TlLzuVI,1880
 langfun/core/__init__.py,sha256=6QEuXOZ9BXxm6TjpaMXuLwUBTYO3pkFDqn9QVBXyyPQ,4248
-langfun/core/component.py,sha256=VRPfDB_2jEnxcB3-HoiVjG4ID-SMenNPIsytb0uXMPg,9674
-langfun/core/component_test.py,sha256=VAPd6V_-odAe8rBvesW3ogYDd6OSqRq4FaPhfgOM4Zg,7949
+langfun/core/component.py,sha256=oxesbC0BoE_TbtxwW5x-BAZWxZyyJbuPiX5S38RqCv0,9909
+langfun/core/component_test.py,sha256=uR-_Sz_42Jxc5qzLIB-f5_pXmNwnC01Xlbv5NOQSeSU,8021
 langfun/core/concurrent.py,sha256=TRc49pJ3HQro2kb5FtcWkHjhBm8UcgE8RJybU5cU3-0,24537
 langfun/core/concurrent_test.py,sha256=mwFMZhDUdppnDr7vDSTwcbMHwrdsIoKJwRYNtl4ZWL4,15185
 langfun/core/console.py,sha256=bk5rNPNm9rMGW5YT2HixxU04p2umnoabn5SDz6Dqe88,2317
 langfun/core/console_test.py,sha256=5SYJdxpJGLgdSSQqqMPoA1X6jpsLD8rgcyk-EgI65oE,1077
 langfun/core/langfunc.py,sha256=WXdTc3QsmGD_n80KD9dFRr5MHpGZ9E_y_Rhtk4t9-3w,11852
 langfun/core/langfunc_test.py,sha256=sQaKuZpGGmG80GRifhbxkj7nfzQLJKj4Vuw5y1s1K3U,8378
-langfun/core/language_model.py,sha256=1_GO6oEm0wXnE7aRRLOdT-A4j_6YvRanS5oMgfobcIs,18331
-langfun/core/language_model_test.py,sha256=KvXXOr64TsSs3WkEALCLLZSlz09i7hBiHDOZ_8Eq8_o,13047
+langfun/core/language_model.py,sha256=mJfQ_Zqq9IyVyZUdYMQ1BPrpo4Gn8yxDJb_RghQFP_I,18911
+langfun/core/language_model_test.py,sha256=oWQjnyiJugSpHJKda-qLaSvmbm1sx_v-ZXrHvw_kNk4,14172
 langfun/core/memory.py,sha256=f-asN1F7Vehgdn_fK84v73GrEUOxRtaW934keutTKjk,2416
 langfun/core/message.py,sha256=QhvV9t5qaryPcruyxxcXi3gm9QDInkSldwTtK6sVJ3c,15734
 langfun/core/message_test.py,sha256=Z23pUM5vPnDrYkIIibe2KL73D5HKur_awI0ut_EQFQA,9501
@@ -21,8 +21,8 @@ langfun/core/sampling.py,sha256=vygWvgC8MFw0_AKNSmz-ywMXJYWf8cl0tI8QycvAmyI,5795
 langfun/core/sampling_test.py,sha256=U7PANpMsl9E_pa4_Y4FzesSjcwg-u-LKHGCWSgv-8FY,3663
 langfun/core/subscription.py,sha256=euawEuSZP-BHydaT-AQpfYFL0m5pWPGcW0upFhrojqc,10930
 langfun/core/subscription_test.py,sha256=Y4ZdbZEwm83YNZBxHff0QR4QUa4rdaNXA3_jfIcArBo,8717
-langfun/core/template.py,sha256=zVD8dAsXFfgF25aKh2WqSuCEHVqriCC-4tLbQqTMa2w,17662
-langfun/core/template_test.py,sha256=1hDdYfvXJVoslTUudh3WhxU7VnDSiIz6MkxPfmuHKAY,13572
+langfun/core/template.py,sha256=dr3tZCbXH2qWzigO_EFVHe0GDnnCu58Tru5Mvlzin4o,18447
+langfun/core/template_test.py,sha256=xty7PgdNhGpw7ZRZ6QGwhKZWG6dyRgI16Lg3p7IMLJg,13944
 langfun/core/text_formatting.py,sha256=ytjj7opnRJ6w-pkglL2CZUyfYDXLpNf65E42LBb31gc,5158
 langfun/core/text_formatting_test.py,sha256=nyKC6tn2L4hPJiqQHgxcbQsJJi4A4Nbj8FiO8iT6B80,1514
 langfun/core/coding/__init__.py,sha256=5utju_fwEsImaiftx4oXKl9FAM8p281k8-Esdh_-m1w,835
@@ -47,18 +47,18 @@ langfun/core/eval/matching_test.py,sha256=FFHYD7IDuKe5RMjkx74ksukiwUhO5a_SS340Ja
 langfun/core/eval/scoring.py,sha256=aKeanBJf1yO3Q9JEtgPWoiZk_3M_GiqwXVXX7x_g22w,6172
 langfun/core/eval/scoring_test.py,sha256=YH1cIxBWtfdKcAV9Fh10vLkV5J-gxk8b6nxW4Z2u5pk,4024
 langfun/core/llms/__init__.py,sha256=1bPg1QI8duOZCYINm-jWi094x0JtLmsk4KX60qIC_gs,3245
-langfun/core/llms/anthropic.py,sha256=p-tjttvithBg2b4tgxIS2F-Zk5AYAh5e-lW-8e1p4wc,7865
-langfun/core/llms/anthropic_test.py,sha256=OuLDxeiPRdqsfKILS0R6jJLTRs3-1KCIotPPr7IbIDU,5502
+langfun/core/llms/anthropic.py,sha256=7W9YdPN3SlAFhAIQlihMkrpo7tTY_4NvD0KIlCrqcsk,8505
+langfun/core/llms/anthropic_test.py,sha256=TMM30myyEhwF99Le4RvJEXOn8RYl0q1FRkt9Q9nl1jk,5540
 langfun/core/llms/fake.py,sha256=b-Xk5IPTbUt-elsyzd_i3n1tqzc_kgETXrEvgJruSMk,2824
 langfun/core/llms/fake_test.py,sha256=ZlDQgL41EX3eYTfBQNp2nB2LciqCmtoHgCsGvW4XhwI,4184
 langfun/core/llms/google_genai.py,sha256=n8zyJwh9UCTgb6-8LyvmjVNFGZQ4-zfzZ0ulkhHAnR8,8624
 langfun/core/llms/google_genai_test.py,sha256=_UcGTfl16-aDUlEWFC2W2F8y9jPUs53RBYA6MOCpGXw,7525
-langfun/core/llms/groq.py,sha256=ZULexLJoU_IJ6vjQimMsmv0xnCOTPGrJVkPLbjfqC5w,7600
-langfun/core/llms/groq_test.py,sha256=o95z76qwOwmsOxC2WhHJ4roFzxFRoVjkC7KETlfsVis,5250
+langfun/core/llms/groq.py,sha256=NaGItVL_pkOpqPpI4bPGU27xLFRoaeizZ49v2s-4ERs,7844
+langfun/core/llms/groq_test.py,sha256=M6GtlrsOvDun_j-sR8cPh4W_moHWZNSTiThu3kuwbbc,5281
 langfun/core/llms/llama_cpp.py,sha256=Y_KkMUf3Xfac49koMUtUslKl3h-HWp3-ntq7Jaa3bdo,2385
 langfun/core/llms/llama_cpp_test.py,sha256=ZxC6defGd_HX9SFRU9U4cJiQnBKundbOrchbXuC1Z2M,1683
-langfun/core/llms/openai.py,sha256=Z_pujF3B2QMzWBgOdV67DKAfZ8Wmyeb_6F9BkcGHyaE,12344
-langfun/core/llms/openai_test.py,sha256=S83nVUq1Za15-rq-tCGOZPGPGByVgk0YdamoO7gnNpw,8270
+langfun/core/llms/openai.py,sha256=06nPhmw0zIA5Zqv3eqsrZtYLHnKwW7N8yt3LlFUFVpI,13247
+langfun/core/llms/openai_test.py,sha256=Yt_W6k8YXpT3bs0JroARofCGmn_Uq3u61LmZxqWS2DQ,8272
 langfun/core/llms/cache/__init__.py,sha256=QAo3InUMDM_YpteNnVCSejI4zOsnjSMWKJKzkb3VY64,993
 langfun/core/llms/cache/base.py,sha256=cFfYvOIUae842pncqCAsRvqXCk2AnAsRYVx0mcIoAeY,3338
 langfun/core/llms/cache/in_memory.py,sha256=YfFyJEhLs73cUiB0ZfhMxYpdE8Iuxxw-dvMFwGHTSHw,4742
@@ -101,8 +101,8 @@ langfun/core/templates/demonstration.py,sha256=vCrgYubdZM5Umqcgp8NUVGXgr4P_c-fik
 langfun/core/templates/demonstration_test.py,sha256=SafcDQ0WgI7pw05EmPI2S4v1t3ABKzup8jReCljHeK4,2162
 langfun/core/templates/selfplay.py,sha256=yhgrJbiYwq47TgzThmHrDQTF4nDrTI09CWGhuQPNv-s,2273
 langfun/core/templates/selfplay_test.py,sha256=DYVrkk7uNKCqJGEHH31HssU2BPuMItU1vJLzfcXIlYg,2156
-langfun-0.0.2.dev20240422.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
-langfun-0.0.2.dev20240422.dist-info/METADATA,sha256=33ozyCNrWSLP3XvGJsjL4FjNwG9gTapSOP-aVnMH1hA,3405
-langfun-0.0.2.dev20240422.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-langfun-0.0.2.dev20240422.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
-langfun-0.0.2.dev20240422.dist-info/RECORD,,
+langfun-0.0.2.dev20240423.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
+langfun-0.0.2.dev20240423.dist-info/METADATA,sha256=tK11XSi0smD_bL0tVdbT_YsNXu-xs1KKYbTX2powidg,3405
+langfun-0.0.2.dev20240423.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+langfun-0.0.2.dev20240423.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
+langfun-0.0.2.dev20240423.dist-info/RECORD,,