langfun 0.1.2.dev202501170804__py3-none-any.whl → 0.1.2.dev202501180803__py3-none-any.whl

langfun/core/llms/__init__.py

@@ -27,11 +27,15 @@ from langfun.core.llms.fake import StaticSequence
 # Compositional models.
 from langfun.core.llms.compositional import RandomChoice
 
-# REST-based models.
+# Base models by request/response protocol.
 from langfun.core.llms.rest import REST
+from langfun.core.llms.openai_compatible import OpenAICompatible
+from langfun.core.llms.gemini import Gemini
+from langfun.core.llms.anthropic import Anthropic
 
-# VertexAI-based models.
+# Base models by serving platforms.
 from langfun.core.llms.vertexai import VertexAI
+from langfun.core.llms.groq import Groq
 
 # Gemini models.
 from langfun.core.llms.google_genai import GenAI
@@ -60,9 +64,6 @@ from langfun.core.llms.vertexai import VertexAIGeminiFlash1_5_002
 from langfun.core.llms.vertexai import VertexAIGeminiFlash1_5_001
 from langfun.core.llms.vertexai import VertexAIGeminiPro1
 
-# Base for OpenAI-compatible models.
-from langfun.core.llms.openai_compatible import OpenAICompatible
-
 # OpenAI models.
 from langfun.core.llms.openai import OpenAI
 
@@ -119,7 +120,6 @@ from langfun.core.llms.openai import Gpt3Ada
 
 # Anthropic models.
 
-from langfun.core.llms.anthropic import Anthropic
 from langfun.core.llms.anthropic import Claude35Sonnet
 from langfun.core.llms.anthropic import Claude35Sonnet20241022
 from langfun.core.llms.anthropic import Claude35Sonnet20240620
@@ -135,7 +135,17 @@ from langfun.core.llms.vertexai import VertexAIClaude3_Opus_20240229
 
 # Misc open source models.
 
-from langfun.core.llms.groq import Groq
+# Gemma models.
+from langfun.core.llms.groq import GroqGemma2_9B_IT
+from langfun.core.llms.groq import GroqGemma_7B_IT
+
+# Llama models.
+from langfun.core.llms.vertexai import VertexAILlama
+from langfun.core.llms.vertexai import VertexAILlama3_2_90B
+from langfun.core.llms.vertexai import VertexAILlama3_1_405B
+from langfun.core.llms.vertexai import VertexAILlama3_1_70B
+from langfun.core.llms.vertexai import VertexAILlama3_1_8B
+
 from langfun.core.llms.groq import GroqLlama3_2_3B
 from langfun.core.llms.groq import GroqLlama3_2_1B
 from langfun.core.llms.groq import GroqLlama3_1_70B
@@ -143,18 +153,28 @@ from langfun.core.llms.groq import GroqLlama3_1_8B
 from langfun.core.llms.groq import GroqLlama3_70B
 from langfun.core.llms.groq import GroqLlama3_8B
 from langfun.core.llms.groq import GroqLlama2_70B
+
+# Mistral models.
+from langfun.core.llms.vertexai import VertexAIMistral
+from langfun.core.llms.vertexai import VertexAIMistralLarge_20241121
+from langfun.core.llms.vertexai import VertexAIMistralLarge_20240724
+from langfun.core.llms.vertexai import VertexAIMistralNemo_20240724
+from langfun.core.llms.vertexai import VertexAICodestral_20250113
+from langfun.core.llms.vertexai import VertexAICodestral_20240529
+
 from langfun.core.llms.groq import GroqMistral_8x7B
-from langfun.core.llms.groq import GroqGemma2_9B_IT
-from langfun.core.llms.groq import GroqGemma_7B_IT
+
+# DeepSeek models.
+from langfun.core.llms.deepseek import DeepSeek
+from langfun.core.llms.deepseek import DeepSeekChat
+
+# Whisper models.
 from langfun.core.llms.groq import GroqWhisper_Large_v3
 from langfun.core.llms.groq import GroqWhisper_Large_v3Turbo
 
 # LLaMA C++ models.
 from langfun.core.llms.llama_cpp import LlamaCppRemote
 
-# DeepSeek models.
-from langfun.core.llms.deepseek import DeepSeek
-from langfun.core.llms.deepseek import DeepSeekChat
 
 # Placeholder for Google-internal imports.
 

langfun/core/llms/gemini.py

@@ -380,7 +380,7 @@ class Gemini(rest.REST):
     return (
         cost_per_1m_input_tokens * num_input_tokens
         + cost_per_1m_output_tokens * num_output_tokens
-    ) / 1000_1000
+    ) / 1000_000
 
   @property
   def model_id(self) -> str:

langfun/core/llms/vertexai.py

@@ -20,6 +20,7 @@ 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
 
@@ -108,7 +109,7 @@ class VertexAI(rest.REST):
     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
@@ -281,3 +282,280 @@ class VertexAIClaude3_5_Haiku_20241022(VertexAIAnthropic):
   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-0.1.2.dev202501170804.dist-info → langfun-0.1.2.dev202501180803.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: langfun
-Version: 0.1.2.dev202501170804
+Version: 0.1.2.dev202501180803
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.1.2.dev202501170804.dist-info → langfun-0.1.2.dev202501180803.dist-info}/RECORD

@@ -71,7 +71,7 @@ langfun/core/eval/v2/reporting.py,sha256=QOp5jX761Esvi5w_UIRLDqPY_XRO6ru02-DOrdq
 langfun/core/eval/v2/reporting_test.py,sha256=UmYSAQvD3AIXsSyWQ-WD2uLtEISYpmBeoKY5u5Qwc8E,5696
 langfun/core/eval/v2/runners.py,sha256=DKEmSlGXjOXKWFdBhTpLy7tMsBHZHd1Brl3hWIngsSQ,15931
 langfun/core/eval/v2/runners_test.py,sha256=A37fKK2MvAVTiShsg_laluJzJ9AuAQn52k7HPbfD0Ks,11666
-langfun/core/llms/__init__.py,sha256=wA6t_E3peTYTjsW6uOHnOs9wjQ_Tj1WYlhVVLk2Sjcg,6867
+langfun/core/llms/__init__.py,sha256=50mJagAgkIhMwhOyHxGq_O5st4HhpnE-okeYzc7GU6c,7667
 langfun/core/llms/anthropic.py,sha256=z_DWDpR1VKNzv6wq-9CXLzWdqCDXRKuVFacJNpgBqAs,10826
 langfun/core/llms/anthropic_test.py,sha256=zZ2eSP8hhVv-RDSWxT7wX-NS5DfGfQmCjS9P0pusAHM,6556
 langfun/core/llms/compositional.py,sha256=csW_FLlgL-tpeyCOTVvfUQkMa_zCN5Y2I-YbSNuK27U,2872
@@ -80,7 +80,7 @@ langfun/core/llms/deepseek.py,sha256=Y7DlLUWrukbPVyBMesppd-m75Q-PxD0b3KnMKaoY_8I
 langfun/core/llms/deepseek_test.py,sha256=dS72i52bwMpCN4dJDvpJI59AnNChpwxS5eYYFrhGh90,1843
 langfun/core/llms/fake.py,sha256=gCHBYBLvBCsC78HI1hpoqXCS-p1FMTgY1P1qh_sGBPk,3070
 langfun/core/llms/fake_test.py,sha256=2h13qkwEz_JR0mtUDPxdAhQo7MueXaFSwsD2DIRDW9g,7653
-langfun/core/llms/gemini.py,sha256=tfM4vrt0WnvnrxRhWXZWh7Gp8dYYfMnSbi9uOstkSak,17399
+langfun/core/llms/gemini.py,sha256=itwTCmQHRjwSjt7_UzFfaat23gyRL-El4qmJrg-OGVA,17398
 langfun/core/llms/gemini_test.py,sha256=2ERhYWCJwnfDTQbCaZHFuB1TdWJFrOBS7yyCBInIdQk,6129
 langfun/core/llms/google_genai.py,sha256=85Vmx5QmsziON03PRsFQINSu5NF6pAAuFFhUdDteWGc,3662
 langfun/core/llms/google_genai_test.py,sha256=JZf_cbQ4GGGpwiQCLjFJn7V4jxBBqgZhIx91AzbGKVo,1250
@@ -94,7 +94,7 @@ langfun/core/llms/openai_compatible_test.py,sha256=0uFYhCiuHo2Wrlgj16-GRG6rW8P6E
 langfun/core/llms/openai_test.py,sha256=m85YjGCvWvV5ZYagjC0FqI0FcqyCEVCbUUs8Wm3iUrc,2475
 langfun/core/llms/rest.py,sha256=sWbYUV8S3SuOg9giq7xwD-xDRfaF7NP_ig7bI52-Rj4,3442
 langfun/core/llms/rest_test.py,sha256=zWGiI08f9gXsoQPJS9TlX1zD2uQLrJUB-1VpAJXRHfs,3475
-langfun/core/llms/vertexai.py,sha256=JV9iHsCM3Ee-4nE1ENNkTXIYGxjCHxrEeir175YpCM8,7869
+langfun/core/llms/vertexai.py,sha256=SVvLTqQZ6Ha8wZh3azkh4g3O838CpNkuP3XlgIrLMKo,15751
 langfun/core/llms/vertexai_test.py,sha256=6eLQOyeL5iGZOIWb39sFcf1TgYD_6TBGYdMO4UIvhf4,3333
 langfun/core/llms/cache/__init__.py,sha256=QAo3InUMDM_YpteNnVCSejI4zOsnjSMWKJKzkb3VY64,993
 langfun/core/llms/cache/base.py,sha256=rt3zwmyw0y9jsSGW-ZbV1vAfLxQ7_3AVk0l2EySlse4,3918
@@ -146,8 +146,8 @@ 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=Ot__1P1M8oJfoTp-M9-PQ6HUXqZKyMwvZ5f7yQ3yfyM,2326
-langfun-0.1.2.dev202501170804.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
-langfun-0.1.2.dev202501170804.dist-info/METADATA,sha256=X3MDNl6D6StuwltvUclYhE20uKNQ2x8lY3CkPggJyI4,8172
-langfun-0.1.2.dev202501170804.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-langfun-0.1.2.dev202501170804.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
-langfun-0.1.2.dev202501170804.dist-info/RECORD,,
+langfun-0.1.2.dev202501180803.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
+langfun-0.1.2.dev202501180803.dist-info/METADATA,sha256=W9jkpCCOZx-Tl8sNz3y1IdVZNG48qcjs21airG2TTI0,8172
+langfun-0.1.2.dev202501180803.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+langfun-0.1.2.dev202501180803.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
+langfun-0.1.2.dev202501180803.dist-info/RECORD,,