langfun 0.0.2.dev20240420__py3-none-any.whl → 0.0.2.dev20240421__py3-none-any.whl

langfun/core/llms/__init__.py

@@ -66,6 +66,13 @@ from langfun.core.llms.anthropic import Claude3Opus
 from langfun.core.llms.anthropic import Claude3Sonnet
 from langfun.core.llms.anthropic import Claude3Haiku
 
+from langfun.core.llms.groq import Groq
+from langfun.core.llms.groq import GroqLlama3_70B
+from langfun.core.llms.groq import GroqLlama3_8B
+from langfun.core.llms.groq import GroqLlama2_70B
+from langfun.core.llms.groq import GroqMistral_8x7B
+from langfun.core.llms.groq import GroqGemma7B_IT
+
 
 # LLaMA C++ models.
 from langfun.core.llms.llama_cpp import LlamaCppRemote

langfun/core/llms/groq.py

@@ -0,0 +1,251 @@
+# Copyright 2024 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.
+"""Language models from Groq."""
+
+import functools
+import os
+from typing import Annotated, Any
+
+import langfun.core as lf
+from langfun.core import modalities as lf_modalities
+import pyglove as pg
+import requests
+
+
+SUPPORTED_MODELS_AND_SETTINGS = {
+    # Refer https://console.groq.com/docs/models
+    'llama3-8b-8192': pg.Dict(max_tokens=8192, max_concurrency=16),
+    'llama3-70b-8192': pg.Dict(max_tokens=8192, max_concurrency=16),
+    'llama2-70b-4096': pg.Dict(max_tokens=4096, max_concurrency=16),
+    'mixtral-8x7b-32768': pg.Dict(max_tokens=32768, max_concurrency=16),
+    'gemma-7b-it': pg.Dict(max_tokens=8192, max_concurrency=16),
+}
+
+
+class GroqError(Exception):  # pylint: disable=g-bad-exception-name
+  """Base class for Groq errors."""
+
+
+class RateLimitError(GroqError):
+  """Error for rate limit reached."""
+
+
+class OverloadedError(GroqError):
+  """Groq's server is temporarily overloaded."""
+
+
+_CHAT_COMPLETE_API_ENDPOINT = 'https://api.groq.com/openai/v1/chat/completions'
+
+
+@lf.use_init_args(['model'])
+class Groq(lf.LanguageModel):
+  """Groq LLMs through REST APIs (OpenAI compatible).
+
+  See https://platform.openai.com/docs/api-reference/chat
+  """
+
+  model: pg.typing.Annotated[
+      pg.typing.Enum(
+          pg.MISSING_VALUE, list(SUPPORTED_MODELS_AND_SETTINGS.keys())
+      ),
+      'The name of the model to use.',
+  ]
+
+  multimodal: Annotated[bool, 'Whether this model has multimodal support.'] = (
+      False
+  )
+
+  api_key: Annotated[
+      str | None,
+      (
+          'API key. If None, the key will be read from environment variable '
+          "'GROQ_API_KEY'."
+      ),
+  ] = None
+
+  def _on_bound(self):
+    super()._on_bound()
+    self._api_key = None
+    self.__dict__.pop('_api_initialized', None)
+
+  @functools.cached_property
+  def _api_initialized(self):
+    api_key = self.api_key or os.environ.get('GROQ_API_KEY', None)
+    if not api_key:
+      raise ValueError(
+          'Please specify `api_key` during `__init__` or set environment '
+          'variable `GROQ_API_KEY` with your Anthropic API key.'
+      )
+    self._api_key = api_key
+    return True
+
+  @property
+  def model_id(self) -> str:
+    """Returns a string to identify the model."""
+    return self.model
+
+  @property
+  def max_concurrency(self) -> int:
+    return SUPPORTED_MODELS_AND_SETTINGS[self.model].max_concurrency
+
+  def _get_request_args(self, options: lf.LMSamplingOptions) -> dict[str, Any]:
+    """Returns a dict as request arguments."""
+    # `logprobs` and `top_logprobs` flags are not supported on Groq yet.
+    args = dict(
+        model=self.model,
+        n=options.n,
+        stream=False,
+    )
+
+    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 _content_from_message(self, prompt: lf.Message) -> list[dict[str, Any]]:
+    """Converts an message to Anthropic's content protocol (list of dicts)."""
+    # Refer: https://platform.openai.com/docs/api-reference/chat/create
+    content = []
+    for chunk in prompt.chunk():
+      if isinstance(chunk, str):
+        item = dict(type='text', text=chunk)
+      elif (
+          self.multimodal
+          and isinstance(chunk, lf_modalities.Image)
+          and chunk.uri
+      ):
+        # NOTE(daiyip): Groq only support image URL.
+        item = dict(type='image_url', image_url=chunk.uri)
+      else:
+        raise ValueError(f'Unsupported modality object: {chunk!r}.')
+      content.append(item)
+    return content
+
+  def _message_from_choice(self, choice: dict[str, Any]) -> lf.Message:
+    """Converts Anthropic's content protocol to message."""
+    # Refer: https://platform.openai.com/docs/api-reference/chat/create
+    content = choice['message']['content']
+    if isinstance(content, str):
+      return lf.AIMessage(content)
+    return lf.AIMessage.from_chunks(
+        [x['text'] for x in content if x['type'] == 'text']
+    )
+
+  def _parse_response(self, response: requests.Response) -> lf.LMSamplingResult:
+    """Parses Anthropic's response."""
+    # Refer: https://platform.openai.com/docs/api-reference/chat/object
+    output = response.json()
+    if response.status_code == 200:
+      samples = [
+          lf.LMSample(self._message_from_choice(choice), score=0.0)
+          for choice in output['choices']
+      ]
+      usage = output['usage']
+      return lf.LMSamplingResult(
+          samples,
+          usage=lf.LMSamplingUsage(
+              prompt_tokens=usage['prompt_tokens'],
+              completion_tokens=usage['completion_tokens'],
+              total_tokens=usage['total_tokens'],
+          ),
+      )
+    else:
+      # 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):
+        error_cls = OverloadedError
+      else:
+        error_cls = GroqError
+      error = output['error']
+      raise error_cls(f'{error["type"]}: {error["message"]}')
+
+  def _sample(self, prompts: list[lf.Message]) -> list[lf.LMSamplingResult]:
+    assert self._api_initialized
+    return self._parallel_execute_with_currency_control(
+        self._sample_single,
+        prompts,
+        retry_on_errors=(RateLimitError, OverloadedError),
+    )
+
+  def _sample_single(self, prompt: lf.Message) -> lf.LMSamplingResult:
+    request = dict()
+    request.update(self._get_request_args(self.sampling_options))
+    request.update(
+        dict(
+            messages=[
+                dict(role='user', content=self._content_from_message(prompt))
+            ]
+        )
+    )
+    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)
+
+
+class GroqLlama3_8B(Groq):  # pylint: disable=invalid-name
+  """Llama3-8B with 8K context window.
+
+  See: https://huggingface.co/meta-llama/Meta-Llama-3-8B
+  """
+
+  model = 'llama3-8b-8192'
+
+
+class GroqLlama3_70B(Groq):  # pylint: disable=invalid-name
+  """Llama3-70B with 8K context window.
+
+  See: https://huggingface.co/meta-llama/Meta-Llama-3-70B
+  """
+
+  model = 'llama3-70b-8192'
+
+
+class GroqLlama2_70B(Groq):  # pylint: disable=invalid-name
+  """Llama2-70B with 4K context window.
+
+  See: https://huggingface.co/meta-llama/Llama-2-70b
+  """
+
+  model = 'llama2-70b-4096'
+
+
+class GroqMistral_8x7B(Groq):  # pylint: disable=invalid-name
+  """Mixtral 8x7B with 32K context window.
+
+  See: https://huggingface.co/meta-llama/Llama-2-70b
+  """
+
+  model = 'mixtral-8x7b-32768'
+
+
+class GroqGemma7B_IT(Groq):  # pylint: disable=invalid-name
+  """Gemma 7B with 8K context window.
+
+  See: https://huggingface.co/google/gemma-1.1-7b-it
+  """
+
+  model = 'gemma-7b-it'

langfun/core/llms/groq_test.py

@@ -0,0 +1,170 @@
+# 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 Groq models."""
+
+import os
+from typing import Any
+import unittest
+from unittest import mock
+from langfun.core import modalities as lf_modalities
+from langfun.core.llms import groq
+import pyglove as pg
+import requests
+
+
+def mock_requests_post(url: str, json: dict[str, Any], **kwargs):
+  del url, kwargs
+
+  response = requests.Response()
+  response.status_code = 200
+  response._content = pg.to_json_str({
+      'choices': [{
+          'message': {
+              'content': [{
+                  'type': 'text',
+                  'text': (
+                      f'hello with temperature={json.get("temperature")}, '
+                      f'top_p={json.get("top_p")}, '
+                      f'max_tokens={json.get("max_tokens")}, '
+                      f'stop={json.get("stop")}.'
+                  ),
+              }],
+          }
+      }],
+      'usage': {
+          'prompt_tokens': 2,
+          'completion_tokens': 1,
+          'total_tokens': 3,
+      },
+  }).encode()
+  return response
+
+
+def mock_mm_requests_post(url: str, json: dict[str, Any], **kwargs):
+  del url, kwargs
+  v = json['messages'][0]['content'][0]
+  image = lf_modalities.Image.from_uri(v['image_url'])
+
+  response = requests.Response()
+  response.status_code = 200
+  response._content = pg.to_json_str({
+      'choices': [
+          {
+              'message': {
+                  'content': [{
+                      'type': 'text',
+                      'text': image.uri,
+                  }],
+              }
+          }
+      ],
+      'usage': {
+          'prompt_tokens': 2,
+          'completion_tokens': 1,
+          'total_tokens': 3,
+      },
+  }).encode()
+  return response
+
+
+def mock_requests_post_error(status_code, error_type, error_message):
+  def _mock_requests(url: str, json: dict[str, Any], **kwargs):
+    del url, json, kwargs
+    response = requests.Response()
+    response.status_code = status_code
+    response._content = pg.to_json_str(
+        {
+            'error': {
+                'type': error_type,
+                'message': error_message,
+            }
+        }
+    ).encode()
+    return response
+
+  return _mock_requests
+
+
+class AuthropicTest(unittest.TestCase):
+
+  def test_basics(self):
+    self.assertEqual(groq.GroqMistral_8x7B().model_id, 'mixtral-8x7b-32768')
+    self.assertEqual(groq.GroqMistral_8x7B().max_concurrency, 16)
+
+  def test_api_key(self):
+    lm = groq.GroqMistral_8x7B()
+    with self.assertRaisesRegex(ValueError, 'Please specify `api_key`'):
+      lm('hi')
+
+    with mock.patch('requests.post') as mock_request:
+      mock_request.side_effect = mock_requests_post
+
+      lm = groq.GroqMistral_8x7B(api_key='fake key')
+      self.assertRegex(lm('hi').text, 'hello.*')
+
+      os.environ['GROQ_API_KEY'] = 'abc'
+      lm = groq.GroqMistral_8x7B()
+      self.assertRegex(lm('hi').text, 'hello.*')
+      del os.environ['GROQ_API_KEY']
+
+  def test_call(self):
+    with mock.patch('requests.post') as mock_request:
+      mock_request.side_effect = mock_requests_post
+      lm = groq.GroqLlama3_70B(api_key='fake_key')
+      response = lm(
+          'hello',
+          temperature=0.0,
+          max_tokens=1024,
+          top_k=0.1,
+          top_p=0.2,
+          stop=['\n'],
+      )
+      self.assertEqual(
+          response.text,
+          (
+              'hello with temperature=0.0, top_p=0.2, '
+              "max_tokens=1024, stop=['\\n']."
+          ),
+      )
+      self.assertIsNotNone(response.usage)
+      self.assertIsNotNone(response.usage.prompt_tokens, 2)
+      self.assertIsNotNone(response.usage.completion_tokens, 1)
+      self.assertIsNotNone(response.usage.total_tokens, 3)
+
+  def test_mm_call(self):
+    with mock.patch('requests.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'))
+      self.assertEqual(response.text, 'https://fake/image.jpg')
+
+  def test_call_errors(self):
+    for status_code, error_type, error_message in [
+        (429, 'rate_limit', 'Rate limit exceeded.'),
+        (503, 'service_unavailable', 'Service unavailable.'),
+        (500, 'bad_request', 'Bad request.'),
+    ]:
+      with mock.patch('requests.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}'
+        ):
+          lm('hello', lm=lm, max_attempts=1)
+
+
+if __name__ == '__main__':
+  unittest.main()

{langfun-0.0.2.dev20240420.dist-info → langfun-0.0.2.dev20240421.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langfun
-Version: 0.0.2.dev20240420
+Version: 0.0.2.dev20240421
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.0.2.dev20240420.dist-info → langfun-0.0.2.dev20240421.dist-info}/RECORD

@@ -46,13 +46,15 @@ langfun/core/eval/matching.py,sha256=aqNlYrlav7YmsB7rUlsdfoi1RLA5CYqn2RGPxRlPc78
 langfun/core/eval/matching_test.py,sha256=FFHYD7IDuKe5RMjkx74ksukiwUhO5a_SS340JaIPMws,4898
 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=c_9lVKzFjnxHKgRjY_dUiJzBmW1jWALy3mtYv0uMyl0,2953
+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/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/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
@@ -99,8 +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.dev20240420.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
-langfun-0.0.2.dev20240420.dist-info/METADATA,sha256=R4bRp7OO2PSjDyKe48YvIbMptLTkeqesP98ZxJ17woc,3405
-langfun-0.0.2.dev20240420.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-langfun-0.0.2.dev20240420.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
-langfun-0.0.2.dev20240420.dist-info/RECORD,,
+langfun-0.0.2.dev20240421.dist-info/LICENSE,sha256=WNHhf_5RCaeuKWyq_K39vmp9F28LxKsB4SpomwSZ2L0,11357
+langfun-0.0.2.dev20240421.dist-info/METADATA,sha256=-43mHV_1OLXkg277zBFm6nRAcaS8VMfePW1QzyrIc2o,3405
+langfun-0.0.2.dev20240421.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+langfun-0.0.2.dev20240421.dist-info/top_level.txt,sha256=RhlEkHxs1qtzmmtWSwYoLVJAc1YrbPtxQ52uh8Z9VvY,8
+langfun-0.0.2.dev20240421.dist-info/RECORD,,