Internal change

langfun/core/llms/__init__.py

@@ -35,8 +35,12 @@
 from langfun.core.llms.openai import OpenAI
 
 from langfun.core.llms.openai import Gpt4Turbo
-from langfun.core.llms.openai import Gpt4Turbo_0125
-from langfun.core.llms.openai import Gpt4TurboVision
+from langfun.core.llms.openai import Gpt4Turbo_20240409
+from langfun.core.llms.openai import Gpt4TurboPreview
+from langfun.core.llms.openai import Gpt4TurboPreview_0125
+from langfun.core.llms.openai import Gpt4TurboPreview_1106
+from langfun.core.llms.openai import Gpt4VisionPreview
+from langfun.core.llms.openai import Gpt4VisionPreview_1106
 from langfun.core.llms.openai import Gpt4
 from langfun.core.llms.openai import Gpt4_0613
 from langfun.core.llms.openai import Gpt4_32K

langfun/core/llms/anthropic.py

@@ -29,6 +29,9 @@
     'claude-3-opus-20240229': pg.Dict(max_tokens=4096, max_concurrency=1),
     'claude-3-sonnet-20240229': pg.Dict(max_tokens=4096, max_concurrency=4),
     '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),
 }
 
 
@@ -210,6 +213,7 @@ def _sample_single(self, prompt: lf.Message) -> lf.LMSamplingResult:
 
 class Claude3(Anthropic):
   """Base class for Claude 3 models. 200K input tokens and 4K output tokens."""
+  multimodal = True
 
 
 class Claude3Opus(Claude3):
@@ -228,3 +232,18 @@ class Claude3Haiku(Claude3):
   """Anthropic's most compact model."""
 
   model = 'claude-3-haiku-20240307'
+
+
+class Claude2(Anthropic):
+  """Predecessor to Claude 3 with 100K context window.."""
+  model = 'claude-2.0'
+
+
+class Claude21(Anthropic):
+  """Updated Claude 2 model with improved accuracy and 200K context window."""
+  model = 'claude-2.1'
+
+
+class ClaudeInstant(Anthropic):
+  """Cheapest small and fast model, 100K context window."""
+  model = 'claude-instant-1.2'

langfun/core/llms/openai.py

@@ -31,10 +31,13 @@
     # 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-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),
@@ -284,26 +287,43 @@ 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'

langfun/core/llms/openai_test.py

@@ -157,17 +157,19 @@ def test_call_chat_completion(self):
   def test_call_chat_completion_vision(self):
     with mock.patch('openai.ChatCompletion.create') as mock_chat_completion:
       mock_chat_completion.side_effect = mock_chat_completion_query_vision
-      lm = openai.Gpt4TurboVision(api_key='test_key')
-      self.assertEqual(
-          lm(
-              lf.UserMessage(
-                  'hello {{image}}',
-                  image=lf_modalities.Image.from_uri('https://fake/image')
-              ),
-              sampling_options=lf.LMSamplingOptions(n=2)
-          ),
-          'Sample 0 for message: https://fake/image',
-      )
+      lm_1 = openai.Gpt4Turbo(api_key='test_key')
+      lm_2 = openai.Gpt4VisionPreview(api_key='test_key')
+      for lm in (lm_1, lm_2):
+        self.assertEqual(
+            lm(
+                lf.UserMessage(
+                    'hello {{image}}',
+                    image=lf_modalities.Image.from_uri('https://fake/image')
+                ),
+                sampling_options=lf.LMSamplingOptions(n=2)
+            ),
+            'Sample 0 for message: https://fake/image',
+        )
 
   def test_sample_completion(self):
     with mock.patch('openai.Completion.create') as mock_completion: