langfun 0.1.2.dev202501080804__py3-none-any.whl → 0.1.2.dev202501240804__py3-none-any.whl

langfun/core/llms/vertexai.py

@@ -1,4 +1,4 @@
-# Copyright 2023 The Langfun Authors
+# 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.
@@ -15,10 +15,13 @@
 
 import functools
 import os
-from typing import Annotated, Any
+from typing import Annotated, Any, Literal
 
 import langfun.core as lf
+from langfun.core.llms import anthropic
 from langfun.core.llms import gemini
+from langfun.core.llms import openai_compatible
+from langfun.core.llms import rest
 import pyglove as pg
 
 try:
@@ -36,10 +39,21 @@ except ImportError:
   Credentials = Any
 
 
-@lf.use_init_args(['model'])
-@pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
-class VertexAI(gemini.Gemini):
-  """Language model served on VertexAI with REST API."""
+@pg.use_init_args(['api_endpoint'])
+class VertexAI(rest.REST):
+  """Base class for VertexAI models.
+
+  This class handles the authentication of vertex AI models. Subclasses
+  should implement `request` and `result` methods, as well as the `api_endpoint`
+  property. Or let users to provide them as __init__ arguments.
+
+  Please check out VertexAIGemini in `gemini.py` as an example.
+  """
+
+  model: Annotated[
+      str | None,
+      'Model ID.'
+  ] = None
 
   project: Annotated[
       str | None,
@@ -95,7 +109,7 @@ class VertexAI(gemini.Gemini):
     credentials = self.credentials
     if credentials is None:
       # Use default credentials.
-      credentials = google_auth.default(
+      credentials, _ = google_auth.default(
           scopes=['https://www.googleapis.com/auth/cloud-platform']
       )
     self._credentials = credentials
@@ -114,6 +128,17 @@ class VertexAI(gemini.Gemini):
     s.headers.update(self.headers or {})
     return s
 
+
+#
+# Gemini models served by Vertex AI.
+#
+
+
+@pg.use_init_args(['model'])
+@pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
+class VertexAIGemini(VertexAI, gemini.Gemini):
+  """Gemini models served by Vertex AI.."""
+
   @property
   def api_endpoint(self) -> str:
     assert self._api_initialized
@@ -124,7 +149,7 @@ class VertexAI(gemini.Gemini):
     )
 
 
-class VertexAIGeminiFlash2_0ThinkingExp_20241219(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiFlash2_0ThinkingExp_20241219(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini Flash 2.0 Thinking model launched on 12/19/2024."""
 
   api_version = 'v1alpha'
@@ -132,61 +157,405 @@ class VertexAIGeminiFlash2_0ThinkingExp_20241219(VertexAI):  # pylint: disable=i
   timeout = None
 
 
-class VertexAIGeminiFlash2_0Exp(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiFlash2_0Exp(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 2.0 Flash model."""
 
   model = 'gemini-2.0-flash-exp'
 
 
-class VertexAIGeminiExp_20241206(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiExp_20241206(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini Experimental model launched on 12/06/2024."""
 
   model = 'gemini-exp-1206'
 
 
-class VertexAIGeminiExp_20241114(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiExp_20241114(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini Experimental model launched on 11/14/2024."""
 
   model = 'gemini-exp-1114'
 
 
-class VertexAIGeminiPro1_5(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiPro1_5(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Pro model."""
 
   model = 'gemini-1.5-pro-latest'
 
 
-class VertexAIGeminiPro1_5_002(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiPro1_5_002(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Pro model."""
 
   model = 'gemini-1.5-pro-002'
 
 
-class VertexAIGeminiPro1_5_001(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiPro1_5_001(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Pro model."""
 
   model = 'gemini-1.5-pro-001'
 
 
-class VertexAIGeminiFlash1_5(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiFlash1_5(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Flash model."""
 
   model = 'gemini-1.5-flash'
 
 
-class VertexAIGeminiFlash1_5_002(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiFlash1_5_002(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Flash model."""
 
   model = 'gemini-1.5-flash-002'
 
 
-class VertexAIGeminiFlash1_5_001(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiFlash1_5_001(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.5 Flash model."""
 
   model = 'gemini-1.5-flash-001'
 
 
-class VertexAIGeminiPro1(VertexAI):  # pylint: disable=invalid-name
+class VertexAIGeminiPro1(VertexAIGemini):  # pylint: disable=invalid-name
   """Vertex AI Gemini 1.0 Pro model."""
 
   model = 'gemini-1.0-pro'
+
+
+#
+# Anthropic models on Vertex AI.
+#
+
+
+@pg.use_init_args(['model'])
+@pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
+class VertexAIAnthropic(VertexAI, anthropic.Anthropic):
+  """Anthropic models on VertexAI."""
+
+  location: Annotated[
+      Literal['us-east5', 'europe-west1'],
+      'GCP location with Anthropic models hosted.'
+  ] = 'us-east5'
+
+  api_version = 'vertex-2023-10-16'
+
+  @property
+  def headers(self):
+    return {
+        'Content-Type': 'application/json; charset=utf-8',
+    }
+
+  @property
+  def api_endpoint(self) -> str:
+    return (
+        f'https://{self.location}-aiplatform.googleapis.com/v1/projects/'
+        f'{self._project}/locations/{self.location}/publishers/anthropic/'
+        f'models/{self.model}:streamRawPredict'
+    )
+
+  def request(
+      self,
+      prompt: lf.Message,
+      sampling_options: lf.LMSamplingOptions
+  ):
+    request = super().request(prompt, sampling_options)
+    request['anthropic_version'] = self.api_version
+    del request['model']
+    return request
+
+
+# pylint: disable=invalid-name
+
+
+class VertexAIClaude3_Opus_20240229(VertexAIAnthropic):
+  """Anthropic's Claude 3 Opus model on VertexAI."""
+  model = 'claude-3-opus@20240229'
+
+
+class VertexAIClaude3_5_Sonnet_20241022(VertexAIAnthropic):
+  """Anthropic's Claude 3.5 Sonnet model on VertexAI."""
+  model = 'claude-3-5-sonnet-v2@20241022'
+
+
+class VertexAIClaude3_5_Sonnet_20240620(VertexAIAnthropic):
+  """Anthropic's Claude 3.5 Sonnet model on VertexAI."""
+  model = 'claude-3-5-sonnet@20240620'
+
+
+class VertexAIClaude3_5_Haiku_20241022(VertexAIAnthropic):
+  """Anthropic's Claude 3.5 Haiku model on VertexAI."""
+  model = 'claude-3-5-haiku@20241022'
+
+# pylint: enable=invalid-name
+
+#
+# Llama models on Vertex AI.
+# pylint: disable=line-too-long
+# Pricing: https://cloud.google.com/vertex-ai/generative-ai/pricing?_gl=1*ukuk6u*_ga*MjEzMjc4NjM2My4xNzMzODg4OTg3*_ga_WH2QY8WWF5*MTczNzEzNDU1Mi4xMjQuMS4xNzM3MTM0NzczLjU5LjAuMA..#meta-models
+# pylint: enable=line-too-long
+
+LLAMA_MODELS = {
+    'llama-3.2-90b-vision-instruct-maas': pg.Dict(
+        latest_update='2024-09-25',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # Free during preview.
+        cost_per_1m_input_tokens=None,
+        cost_per_1m_output_tokens=None,
+    ),
+    'llama-3.1-405b-instruct-maas': pg.Dict(
+        latest_update='2024-09-25',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=5,
+        cost_per_1m_output_tokens=16,
+    ),
+    'llama-3.1-70b-instruct-maas': pg.Dict(
+        latest_update='2024-09-25',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # Free during preview.
+        cost_per_1m_input_tokens=None,
+        cost_per_1m_output_tokens=None,
+    ),
+    'llama-3.1-8b-instruct-maas': pg.Dict(
+        latest_update='2024-09-25',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # Free during preview.
+        cost_per_1m_input_tokens=None,
+        cost_per_1m_output_tokens=None,
+    )
+}
+
+
+@pg.use_init_args(['model'])
+@pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
+class VertexAILlama(VertexAI, openai_compatible.OpenAICompatible):
+  """Llama models on VertexAI."""
+
+  model: pg.typing.Annotated[
+      pg.typing.Enum(pg.MISSING_VALUE, list(LLAMA_MODELS.keys())),
+      'Llama model ID.',
+  ]
+
+  locations: Annotated[
+      Literal['us-central1'],
+      (
+          'GCP locations with Llama models hosted. '
+          'See https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/llama#regions-quotas'
+      )
+  ] = 'us-central1'
+
+  @property
+  def api_endpoint(self) -> str:
+    assert self._api_initialized
+    return (
+        f'https://{self._location}-aiplatform.googleapis.com/v1beta1/projects/'
+        f'{self._project}/locations/{self._location}/endpoints/'
+        f'openapi/chat/completions'
+    )
+
+  def request(
+      self,
+      prompt: lf.Message,
+      sampling_options: lf.LMSamplingOptions
+  ):
+    request = super().request(prompt, sampling_options)
+    request['model'] = f'meta/{self.model}'
+    return request
+
+  @property
+  def max_concurrency(self) -> int:
+    rpm = LLAMA_MODELS[self.model].get('rpm', 0)
+    tpm = LLAMA_MODELS[self.model].get('tpm', 0)
+    return self.rate_to_max_concurrency(
+        requests_per_min=rpm, tokens_per_min=tpm
+    )
+
+  def estimate_cost(
+      self,
+      num_input_tokens: int,
+      num_output_tokens: int
+  ) -> float | None:
+    """Estimate the cost based on usage."""
+    cost_per_1m_input_tokens = LLAMA_MODELS[self.model].get(
+        'cost_per_1m_input_tokens', None
+    )
+    cost_per_1m_output_tokens = LLAMA_MODELS[self.model].get(
+        'cost_per_1m_output_tokens', None
+    )
+    if cost_per_1m_output_tokens is None or cost_per_1m_input_tokens is None:
+      return None
+    return (
+        cost_per_1m_input_tokens * num_input_tokens
+        + cost_per_1m_output_tokens * num_output_tokens
+    ) / 1000_000
+
+
+# pylint: disable=invalid-name
+class VertexAILlama3_2_90B(VertexAILlama):
+  """Llama 3.2 90B vision instruct model on VertexAI."""
+
+  model = 'llama-3.2-90b-vision-instruct-maas'
+
+
+class VertexAILlama3_1_405B(VertexAILlama):
+  """Llama 3.1 405B vision instruct model on VertexAI."""
+
+  model = 'llama-3.1-405b-instruct-maas'
+
+
+class VertexAILlama3_1_70B(VertexAILlama):
+  """Llama 3.1 70B vision instruct model on VertexAI."""
+
+  model = 'llama-3.1-70b-instruct-maas'
+
+
+class VertexAILlama3_1_8B(VertexAILlama):
+  """Llama 3.1 8B vision instruct model on VertexAI."""
+
+  model = 'llama-3.1-8b-instruct-maas'
+# pylint: enable=invalid-name
+
+#
+# Mistral models on Vertex AI.
+# pylint: disable=line-too-long
+# Pricing: https://cloud.google.com/vertex-ai/generative-ai/pricing?_gl=1*ukuk6u*_ga*MjEzMjc4NjM2My4xNzMzODg4OTg3*_ga_WH2QY8WWF5*MTczNzEzNDU1Mi4xMjQuMS4xNzM3MTM0NzczLjU5LjAuMA..#mistral-models
+# pylint: enable=line-too-long
+
+
+MISTRAL_MODELS = {
+    'mistral-large-2411': pg.Dict(
+        latest_update='2024-11-21',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=2,
+        cost_per_1m_output_tokens=6,
+    ),
+    'mistral-large@2407': pg.Dict(
+        latest_update='2024-07-24',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=2,
+        cost_per_1m_output_tokens=6,
+    ),
+    'mistral-nemo@2407': pg.Dict(
+        latest_update='2024-07-24',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=0.15,
+        cost_per_1m_output_tokens=0.15,
+    ),
+    'codestral-2501': pg.Dict(
+        latest_update='2025-01-13',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=0.3,
+        cost_per_1m_output_tokens=0.9,
+    ),
+    'codestral@2405': pg.Dict(
+        latest_update='2024-05-29',
+        in_service=True,
+        rpm=0,
+        tpm=0,
+        # GA.
+        cost_per_1m_input_tokens=0.2,
+        cost_per_1m_output_tokens=0.6,
+    ),
+}
+
+
+@pg.use_init_args(['model'])
+@pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
+class VertexAIMistral(VertexAI, openai_compatible.OpenAICompatible):
+  """Mistral AI models on VertexAI."""
+
+  model: pg.typing.Annotated[
+      pg.typing.Enum(pg.MISSING_VALUE, list(MISTRAL_MODELS.keys())),
+      'Mistral model ID.',
+  ]
+
+  locations: Annotated[
+      Literal['us-central1', 'europe-west4'],
+      (
+          'GCP locations with Mistral models hosted. '
+          'See https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/mistral#regions-quotas'
+      )
+  ] = 'us-central1'
+
+  @property
+  def api_endpoint(self) -> str:
+    assert self._api_initialized
+    return (
+        f'https://{self._location}-aiplatform.googleapis.com/v1/projects/'
+        f'{self._project}/locations/{self._location}/publishers/mistralai/'
+        f'models/{self.model}:rawPredict'
+    )
+
+  @property
+  def max_concurrency(self) -> int:
+    rpm = MISTRAL_MODELS[self.model].get('rpm', 0)
+    tpm = MISTRAL_MODELS[self.model].get('tpm', 0)
+    return self.rate_to_max_concurrency(
+        requests_per_min=rpm, tokens_per_min=tpm
+    )
+
+  def estimate_cost(
+      self,
+      num_input_tokens: int,
+      num_output_tokens: int
+  ) -> float | None:
+    """Estimate the cost based on usage."""
+    cost_per_1m_input_tokens = MISTRAL_MODELS[self.model].get(
+        'cost_per_1m_input_tokens', None
+    )
+    cost_per_1m_output_tokens = MISTRAL_MODELS[self.model].get(
+        'cost_per_1m_output_tokens', None
+    )
+    if cost_per_1m_output_tokens is None or cost_per_1m_input_tokens is None:
+      return None
+    return (
+        cost_per_1m_input_tokens * num_input_tokens
+        + cost_per_1m_output_tokens * num_output_tokens
+    ) / 1000_000
+
+
+# pylint: disable=invalid-name
+class VertexAIMistralLarge_20241121(VertexAIMistral):
+  """Mistral Large model on VertexAI released on 2024/11/21."""
+
+  model = 'mistral-large-2411'
+
+
+class VertexAIMistralLarge_20240724(VertexAIMistral):
+  """Mistral Large model on VertexAI released on 2024/07/24."""
+
+  model = 'mistral-large@2407'
+
+
+class VertexAIMistralNemo_20240724(VertexAIMistral):
+  """Mistral Nemo model on VertexAI released on 2024/07/24."""
+
+  model = 'mistral-nemo@2407'
+
+
+class VertexAICodestral_20250113(VertexAIMistral):
+  """Mistral Nemo model on VertexAI released on 2024/07/24."""
+
+  model = 'codestral-2501'
+
+
+class VertexAICodestral_20240529(VertexAIMistral):
+  """Mistral Nemo model on VertexAI released on 2024/05/29."""
+
+  model = 'codestral@2405'
+# pylint: enable=invalid-name

langfun/core/llms/vertexai_test.py

@@ -17,6 +17,8 @@ import os
 import unittest
 from unittest import mock
 
+from google.auth import exceptions
+import langfun.core as lf
 from langfun.core.llms import vertexai
 
 
@@ -48,5 +50,55 @@ class VertexAITest(unittest.TestCase):
     del os.environ['VERTEXAI_LOCATION']
 
 
+class VertexAIAnthropicTest(unittest.TestCase):
+  """Tests for VertexAI Anthropic models."""
+
+  def test_basics(self):
+    with self.assertRaisesRegex(ValueError, 'Please specify `project`'):
+      lm = vertexai.VertexAIClaude3_5_Sonnet_20241022()
+      lm('hi')
+
+    model = vertexai.VertexAIClaude3_5_Sonnet_20241022(project='langfun')
+
+    # NOTE(daiyip): For OSS users, default credentials are not available unless
+    # users have already set up their GCP project. Therefore we ignore the
+    # exception here.
+    try:
+      model._initialize()
+    except exceptions.DefaultCredentialsError:
+      pass
+
+    self.assertEqual(
+        model.api_endpoint,
+        (
+            'https://us-east5-aiplatform.googleapis.com/v1/projects/'
+            'langfun/locations/us-east5/publishers/anthropic/'
+            'models/claude-3-5-sonnet-v2@20241022:streamRawPredict'
+        )
+    )
+    self.assertEqual(
+        model.headers,
+        {
+            'Content-Type': 'application/json; charset=utf-8',
+        },
+    )
+    request = model.request(
+        lf.UserMessage('hi'), lf.LMSamplingOptions(temperature=0.0),
+    )
+    self.assertEqual(
+        request,
+        {
+            'anthropic_version': 'vertex-2023-10-16',
+            'max_tokens': 8192,
+            'messages': [
+                {'content': [{'text': 'hi', 'type': 'text'}], 'role': 'user'}
+            ],
+            'stream': False,
+            'temperature': 0.0,
+            'top_k': 40,
+        },
+    )
+
+
 if __name__ == '__main__':
   unittest.main()