langfun 0.0.2.dev20240330__py3-none-any.whl → 0.0.2.dev20240511__py3-none-any.whl

langfun/core/llms/openai.py

@@ -26,65 +26,55 @@ from openai import openai_object
 import pyglove as pg
 
 
-class Usage(pg.Object):
-  """Usage information per completion."""
-
-  prompt_tokens: int
-  completion_tokens: int
-  total_tokens: int
-
-
-class LMSamplingResult(lf.LMSamplingResult):
-  """LMSamplingResult with usage information."""
-
-  usage: Usage | None = None
-
-
-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-preview', 1),  # GPT-4 Turbo.
-    ('gpt-4-0125-preview', 1),  # GPT-4 Turbo
-    ('gpt-4-1106-preview', 1),  # GPT-4 Turbo
-    ('gpt-4-vision-preview', 1),  # GPT-4 Turbo with Vision.
-    # 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'])
@@ -93,7 +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'
@@ -145,7 +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):
@@ -163,8 +157,6 @@ class OpenAI(lf.LanguageModel):
     # NOTE(daiyip): options.top_k is not applicable.
     args = dict(
         n=options.n,
-        temperature=options.temperature,
-        max_tokens=options.max_tokens,
         stream=False,
         timeout=self.timeout,
         logprobs=options.logprobs,
@@ -173,13 +165,17 @@ class OpenAI(lf.LanguageModel):
     # Completion and ChatCompletion uses different parameter name for model.
     args['model' if self.is_chat_model else 'engine'] = self.model
 
+    if options.temperature is not None:
+      args['temperature'] = options.temperature
+    if options.max_tokens is not None:
+      args['max_tokens'] = options.max_tokens
     if options.top_p is not None:
       args['top_p'] = options.top_p
     if options.stop:
       args['stop'] = options.stop
     return args
 
-  def _sample(self, prompts: list[lf.Message]) -> list[LMSamplingResult]:
+  def _sample(self, prompts: list[lf.Message]) -> list[lf.LMSamplingResult]:
     assert self._api_initialized
     if self.is_chat_model:
       return self._chat_complete_batch(prompts)
@@ -187,7 +183,8 @@ class OpenAI(lf.LanguageModel):
       return self._complete_batch(prompts)
 
   def _complete_batch(
-      self, prompts: list[lf.Message]) -> list[LMSamplingResult]:
+      self, prompts: list[lf.Message]
+  ) -> list[lf.LMSamplingResult]:
 
     def _open_ai_completion(prompts):
       response = openai.Completion.create(
@@ -202,13 +199,13 @@ class OpenAI(lf.LanguageModel):
             lf.LMSample(choice.text.strip(), score=choice.logprobs or 0.0)
         )
 
-      usage = Usage(
+      usage = lf.LMSamplingUsage(
           prompt_tokens=response.usage.prompt_tokens,
           completion_tokens=response.usage.completion_tokens,
           total_tokens=response.usage.total_tokens,
       )
       return [
-          LMSamplingResult(
+          lf.LMSamplingResult(
               samples_by_index[index], usage=usage if index == 0 else None
           )
           for index in sorted(samples_by_index.keys())
@@ -220,20 +217,25 @@ class OpenAI(lf.LanguageModel):
         retry_on_errors=(
             openai_error.ServiceUnavailableError,
             openai_error.RateLimitError,
+            # Handling transient OpenAI server error (code 500). Check out
+            # https://platform.openai.com/docs/guides/error-codes/error-codes
+            (openai_error.APIError,
+             '.*The server had an error processing your request'),
         ),
     )[0]
 
   def _chat_complete_batch(
       self, prompts: list[lf.Message]
-  ) -> list[LMSamplingResult]:
+  ) -> list[lf.LMSamplingResult]:
     def _open_ai_chat_completion(prompt: lf.Message):
       if self.multimodal:
         content = []
         for chunk in prompt.chunk():
           if isinstance(chunk, str):
             item = dict(type='text', text=chunk)
-          elif isinstance(chunk, lf_modalities.Image) and chunk.uri:
-            item = dict(type='image_url', image_url=chunk.uri)
+          elif isinstance(chunk, lf_modalities.Image):
+            uri = chunk.uri or chunk.content_uri
+            item = dict(type='image_url', image_url=dict(url=uri))
           else:
             raise ValueError(f'Unsupported modality object: {chunk!r}.')
           content.append(item)
@@ -266,9 +268,9 @@ class OpenAI(lf.LanguageModel):
             )
         )
 
-      return LMSamplingResult(
+      return lf.LMSamplingResult(
           samples=samples,
-          usage=Usage(
+          usage=lf.LMSamplingUsage(
               prompt_tokens=response.usage.prompt_tokens,
               completion_tokens=response.usage.completion_tokens,
               total_tokens=response.usage.total_tokens,
@@ -291,26 +293,43 @@ class Gpt4(OpenAI):
 
 
 class Gpt4Turbo(Gpt4):
-  """GPT-4 Turbo with 128K context window size. Knowledge up to 4-2023."""
-  model = 'gpt-4-turbo-preview'
+  """GPT-4 Turbo with 128K context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo'
+  multimodal = True
 
 
-class Gpt4TurboVision(Gpt4Turbo):
-  """GPT-4 Turbo with vision."""
-  model = 'gpt-4-vision-preview'
+class Gpt4Turbo_20240409(Gpt4Turbo):  # pylint:disable=invalid-name
+  """GPT-4 Turbo with 128K context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo-2024-04-09'
   multimodal = True
 
 
-class Gpt4Turbo_0125(Gpt4Turbo):   # pylint:disable=invalid-name
-  """GPT-4 Turbo with 128K context window size. Knowledge up to 4-2023."""
+class Gpt4TurboPreview(Gpt4):
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Dec. 2023."""
+  model = 'gpt-4-turbo-preview'
+
+
+class Gpt4TurboPreview_0125(Gpt4TurboPreview):  # pylint: disable=invalid-name
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Dec. 2023."""
   model = 'gpt-4-0125-preview'
 
 
-class Gpt4Turbo_1106(Gpt4Turbo):   # pylint:disable=invalid-name
-  """GPT-4 Turbo @20231106. 128K context window. Knowledge up to 4-2023."""
+class Gpt4TurboPreview_1106(Gpt4TurboPreview):  # pylint: disable=invalid-name
+  """GPT-4 Turbo Preview with 128k context window. Knowledge up to Apr. 2023."""
   model = 'gpt-4-1106-preview'
 
 
+class Gpt4VisionPreview(Gpt4):
+  """GPT-4 Turbo vision preview. 128k context window. Knowledge to Apr. 2023."""
+  model = 'gpt-4-vision-preview'
+  multimodal = True
+
+
+class Gpt4VisionPreview_1106(Gpt4):  # pylint: disable=invalid-name
+  """GPT-4 Turbo vision preview. 128k context window. Knowledge to Apr. 2023."""
+  model = 'gpt-4-1106-vision-preview'
+
+
 class Gpt4_0613(Gpt4):    # pylint:disable=invalid-name
   """GPT-4 @20230613. 8K context window. Knowledge up to 9-2021."""
   model = 'gpt-4-0613'