langfun 0.1.2.dev202512150805__tar.gz → 0.1.2.dev202512260805__tar.gz

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langfun
-Version: 0.1.2.dev202512150805
+Version: 0.1.2.dev202512260805
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/agentic/action.py

@@ -288,6 +288,8 @@ class Action(pg.Object):
 
     with session.track_action(self, max_execution_time=max_execution_time):
       try:
+        # Early terminate the action if the execution time is exceeded.
+        session.check_execution_time()
         result = self.call(session=session, **kwargs)
         self._invocation.end(result)
       except BaseException as e:
@@ -2005,9 +2007,6 @@ class Session(pg.Object, pg.views.html.HtmlTreeView.Extension):
           'signal the start and end of the session.'
       )
 
-    # Early terminate the action if the execution time is exceeded.
-    self.check_execution_time()
-
     invocation = ActionInvocation(
         pg.maybe_ref(action),
         max_execution_time=self._child_max_execution_time(max_execution_time)

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/eval/v2/checkpointing.py

@@ -122,7 +122,7 @@ class Checkpointer(experiment_lib.Plugin):
       example: Example,
   ) -> None:
     """Saves the example to the checkpoint file."""
-    if example.newly_processed:
+    if example.newly_processed or runner.current_run.force_recompute_metrics:
       self._save_example(runner, experiment, example)
 
   def _load_experiment(
@@ -173,9 +173,9 @@ class Checkpointer(experiment_lib.Plugin):
         output_ckpt_file = current_run.output_path_for(
             experiment, os.path.basename(ckpt_file)
         )
-        if ckpt_file != output_ckpt_file and any(
-            e for e in loaded_examples if not e.has_error
-        ):
+        if (not runner.current_run.force_recompute_metrics
+            and ckpt_file != output_ckpt_file
+            and any(e for e in loaded_examples if not e.has_error)):
           # Write the error-free warm-start examples to the output checkpoint
           # file.
           with SequenceWriter(output_ckpt_file) as writer:
@@ -373,7 +373,6 @@ class BulkCheckpointer(Checkpointer):
       )
       if pg.io.path_exists(input_ckpt_file):
         return [input_ckpt_file]
-    print('CCC', experiment.hash, [])
     return []
 
   def on_experiment_complete(

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/eval/v2/experiment.py

@@ -376,23 +376,24 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
   def run(
       self,
       root_dir: str,
-      id: str | None = None,   # pylint: disable=redefined-builtin
+      id: str | None = None,  # pylint: disable=redefined-builtin
       *,
       runner: str = 'parallel',
       warm_start_from: str | None = None,
-      filter: Callable[['Experiment'], bool] | None = None,   # pylint: disable=redefined-builtin
+      filter: Callable[['Experiment'], bool] | None = None,  # pylint: disable=redefined-builtin
       example_ids: list[int] | None = None,
       shuffle_inputs: bool = False,
       raise_if_has_error: bool = False,
       reevaluate_upon_previous_errors: bool = True,
       reprocess: bool | list[int] = False,
+      force_recompute_metrics: bool = False,
       generate_example_html: Literal['new', 'all', 'no'] | list[int] = 'new',
       process_timeout: int | None = None,
       use_cache: Literal['global', 'per_dataset', 'no'] = 'per_dataset',
       note: str | None = None,
       tags: list[str] | None = None,
       plugins: list['Plugin'] | None = None,
-      **kwargs
+      **kwargs,
   ) -> 'Run':
     """Runs the experiment.
 
@@ -445,6 +446,8 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
         meaning that existing checkpoints will be ignored. If a list of
         example IDs, it indicates that only the specified examples will be
         reprocessed.
+      force_recompute_metrics: If True, it will recompute the metrics for all
+        examples, even if the previous checkpoints have metric metadata.
       generate_example_html: Among 'new', 'all', 'no' or a list of example IDs.
         If 'new', generate HTML files for all newly processed examples, and
           keep/copy existing HTML files for unchanged examples.
@@ -481,6 +484,7 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
             raise_if_has_error=raise_if_has_error,
             reevaluate_upon_previous_errors=reevaluate_upon_previous_errors,
             reprocess=reprocess,
+            force_recompute_metrics=force_recompute_metrics,
             generate_example_html=generate_example_html,
             use_cache=use_cache,
             process_timeout=process_timeout,
@@ -884,6 +888,14 @@ class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
       )
   ] = True
 
+  force_recompute_metrics: Annotated[
+      bool,
+      (
+          'If True, force recompute the metrics even if metric metadata is '
+          'already present from previous checkpoint.'
+      )
+  ] = False
+
   note: Annotated[
       str | None,
       'The user note for the current run.'
@@ -1003,7 +1015,7 @@ class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
     load_metadata_ids = set()
     if isinstance(self.generate_example_html, list):
       load_metadata_ids = set(self.generate_example_html)
-    elif self.generate_example_html == 'all':
+    elif self.generate_example_html == 'all' or self.force_recompute_metrics:
       load_metadata_ids = self.examples_to_evaluate(experiment)
     load_metadata_ids -= self.examples_to_reprocess(experiment)
     return load_metadata_ids

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/eval/v2/progress.py

@@ -227,7 +227,11 @@ class Progress(pg.Object, pg.views.HtmlTreeView.Extension):
 
   def merge_from(self, other: 'Progress') -> None:
     """Merges the progress from another progress."""
-    with pg.notify_on_change(False), pg.allow_writable_accessors(True):
+    with (
+        self._lock,
+        pg.notify_on_change(False),
+        pg.allow_writable_accessors(True)
+    ):
       if other.start_time is not None and (
           self.start_time is None or self.start_time > other.start_time):
         self.start_time = other.start_time
@@ -268,16 +272,17 @@ class Progress(pg.Object, pg.views.HtmlTreeView.Extension):
         stop_time=self.stop_time_str,
     )
     if self.execution_summary:
-      time_info['execution'] = pg.Dict(
-          {
-              k: pg.Dict(
-                  num_started=v.num_started,
-                  num_ended=v.num_ended,
-                  num_failed=v.num_failed,
-                  avg_duration=round(v.avg_duration, 2),
-              ) for k, v in self.execution_summary.breakdown.items()
-          }
-      )
+      with self._lock:
+        time_info['execution'] = pg.Dict(
+            {
+                k: pg.Dict(
+                    num_started=v.num_started,
+                    num_ended=v.num_ended,
+                    num_failed=v.num_failed,
+                    avg_duration=round(v.avg_duration, 2),
+                ) for k, v in self.execution_summary.breakdown.items()
+            }
+        )
     return pg.format(time_info, verbose=False)
 
   def _html_tree_view(

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/eval/v2/reporting.py

@@ -99,6 +99,7 @@ class ExampleHtmlGenerator(experiment_lib.Plugin):
 
     generate_example_html = current_run.generate_example_html
     if (generate_example_html == 'all'
+        or runner.current_run.force_recompute_metrics
         or (generate_example_html == 'new' and example.newly_processed)
         or (isinstance(generate_example_html, list)
             and example.id in generate_example_html)):

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/eval/v2/runners/base.py

@@ -396,6 +396,7 @@ class RunnerBase(Runner):
         item,
         raise_if_has_error=self.current_run.raise_if_has_error,
         reevaluate_upon_previous_errors=self.current_run.reevaluate_upon_previous_errors,
+        force_recompute_metrics=self.current_run.force_recompute_metrics,
     )
     self.on_example_complete(evaluation, item)
     return item

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/llms/__init__.py

@@ -43,6 +43,7 @@ from langfun.core.llms.azure_openai import AzureOpenAI
 # Gemini models.
 from langfun.core.llms.google_genai import GenAI
 from langfun.core.llms.google_genai import Gemini3ProPreview
+from langfun.core.llms.google_genai import Gemini3FlashPreview
 from langfun.core.llms.google_genai import Gemini25Pro
 from langfun.core.llms.google_genai import Gemini25Flash
 from langfun.core.llms.google_genai import Gemini25ProPreview_20250605
@@ -94,6 +95,7 @@ from langfun.core.llms.vertexai import VertexAIGemini25Flash
 from langfun.core.llms.vertexai import VertexAIGemini25FlashImagePreview
 from langfun.core.llms.vertexai import VertexAIGemini3ProPreview
 from langfun.core.llms.vertexai import VertexAIGemini3ProImagePreview
+from langfun.core.llms.vertexai import VertexAIGemini3FlashPreview
 
 # For backward compatibility.
 GeminiPro1_5 = Gemini15Pro
@@ -160,6 +162,7 @@ from langfun.core.llms.openai import Gpt35
 from langfun.core.llms.anthropic import Claude45
 from langfun.core.llms.anthropic import Claude45Haiku_20251001
 from langfun.core.llms.anthropic import Claude45Sonnet_20250929
+from langfun.core.llms.anthropic import Claude45Opus_20251101
 from langfun.core.llms.anthropic import Claude4
 from langfun.core.llms.anthropic import Claude4Sonnet_20250514
 from langfun.core.llms.anthropic import Claude4Opus_20250514
@@ -179,6 +182,7 @@ from langfun.core.llms.anthropic import Claude3Haiku_20240307
 from langfun.core.llms.vertexai import VertexAIAnthropic
 from langfun.core.llms.vertexai import VertexAIClaude45Haiku_20251001
 from langfun.core.llms.vertexai import VertexAIClaude45Sonnet_20250929
+from langfun.core.llms.vertexai import VertexAIClaude45Opus_20251101
 from langfun.core.llms.vertexai import VertexAIClaude4Opus_20250514
 from langfun.core.llms.vertexai import VertexAIClaude4Sonnet_20250514
 from langfun.core.llms.vertexai import VertexAIClaude37Sonnet_20250219

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/llms/anthropic.py

@@ -113,6 +113,32 @@ SUPPORTED_MODELS = [
             max_output_tokens_per_minute=400_000,
         ),
     ),
+    AnthropicModelInfo(
+        model_id='claude-opus-4-5-20251101',
+        provider='Anthropic',
+        in_service=True,
+        description='Claude 4.5 Opus model (11/01/2025).',
+        release_date=datetime.datetime(2025, 11, 1),
+        input_modalities=(
+            AnthropicModelInfo.INPUT_IMAGE_TYPES
+            + AnthropicModelInfo.INPUT_DOC_TYPES
+        ),
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=200_000,
+            max_output_tokens=64_000,
+        ),
+        pricing=lf.ModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.5,
+            cost_per_1m_input_tokens=5,
+            cost_per_1m_output_tokens=25,
+        ),
+        rate_limits=AnthropicModelInfo.RateLimits(
+            # Tier 4 rate limits
+            max_requests_per_minute=2000,
+            max_input_tokens_per_minute=1_000_000,
+            max_output_tokens_per_minute=400_000,
+        ),
+    ),
     AnthropicModelInfo(
         model_id='claude-4-opus-20250514',
         provider='Anthropic',
@@ -300,6 +326,32 @@ SUPPORTED_MODELS = [
             max_output_tokens_per_minute=0,
         ),
     ),
+    AnthropicModelInfo(
+        model_id='claude-opus-4-5@20251101',
+        alias_for='claude-opus-4-5-20251101',
+        provider='VertexAI',
+        in_service=True,
+        description='Claude 4.5 Opus model served on VertexAI (11/01/2025).',
+        release_date=datetime.datetime(2025, 11, 1),
+        input_modalities=(
+            AnthropicModelInfo.INPUT_IMAGE_TYPES
+            + AnthropicModelInfo.INPUT_DOC_TYPES
+        ),
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=200_000,
+            max_output_tokens=64_000,
+        ),
+        pricing=lf.ModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.5,
+            cost_per_1m_input_tokens=5,
+            cost_per_1m_output_tokens=25,
+        ),
+        rate_limits=AnthropicModelInfo.RateLimits(
+            max_requests_per_minute=100,
+            max_input_tokens_per_minute=1_000_000,
+            max_output_tokens_per_minute=80_000,
+        ),
+    ),
     AnthropicModelInfo(
         model_id='claude-opus-4@20250514',
         alias_for='claude-opus-4-20250514',
@@ -834,6 +886,13 @@ class Claude45Sonnet_20250929(Claude45):
   model = 'claude-sonnet-4-5-20250929'
 
 
+# pylint: disable=invalid-name
+class Claude45Opus_20251101(Claude45):
+  """Claude 4.5 Opus model 20251101."""
+
+  model = 'claude-opus-4-5-20251101'
+
+
 class Claude4(Anthropic):
   """Base class for Claude 4 models."""
 

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/llms/gemini.py

@@ -200,6 +200,32 @@ SUPPORTED_MODELS = [
             max_tokens_per_minute=1_000_000,
         ),
     ),
+    # Gemini 3 Flash Preview
+    GeminiModelInfo(
+        model_id='gemini-3-flash-preview',
+        in_service=True,
+        provider=pg.oneof(['Google GenAI', 'VertexAI']),
+        model_type='instruction-tuned',
+        description=(
+            'Gemini 3 Flash Preview: High-efficiency, low-latency multimodal'
+            ' model optimized for agentic workflows.'
+        ),
+        release_date=datetime.datetime(2025, 12, 17),
+        input_modalities=GeminiModelInfo.ALL_SUPPORTED_INPUT_TYPES,
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=1_048_576,
+            max_output_tokens=65_536,
+        ),
+        pricing=GeminiModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.05,
+            cost_per_1m_input_tokens=0.50,
+            cost_per_1m_output_tokens=3.00,
+        ),
+        rate_limits=lf.ModelInfo.RateLimits(
+            max_requests_per_minute=2_000,
+            max_tokens_per_minute=4_000_000,
+        ),
+    ),
     # Gemini 2.5 Flash
     GeminiModelInfo(
         model_id='gemini-2.5-flash',

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/llms/google_genai.py

@@ -143,6 +143,12 @@ class Gemini3ProImagePreview(GenAI):
   response_modalities = ['TEXT', 'IMAGE']
 
 
+class Gemini3FlashPreview(GenAI):
+  """Gemini 3 Flash Preview model."""
+
+  model = 'gemini-3-flash-preview'
+
+
 class Gemini25FlashImagePreview(GenAI):
   """Gemini 2.5 Flash Image Preview model."""
   model = 'gemini-2.5-flash-image-preview'

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun/core/llms/vertexai.py

@@ -240,6 +240,13 @@ class VertexAIGemini3ProImagePreview(VertexAIGemini):  # pylint: disable=invalid
   response_modalities = ['TEXT', 'IMAGE']
 
 
+class VertexAIGemini3FlashPreview(VertexAIGemini):  # pylint: disable=invalid-name
+  """Gemini 3 Flash Preview model launched on 12/17/2025."""
+
+  model = 'gemini-3-flash-preview'
+  location = 'global'
+
+
 class VertexAIGemini25Pro(VertexAIGemini):  # pylint: disable=invalid-name
   """Gemini 2.5 Pro GA model launched on 06/17/2025."""
 
@@ -439,6 +446,12 @@ class VertexAIClaude45Sonnet_20250929(VertexAIAnthropic):
   model = 'claude-sonnet-4-5@20250929'
 
 
+class VertexAIClaude45Opus_20251101(VertexAIAnthropic):
+  """Anthropic's Claude 4.5 Opus model on VertexAI."""
+
+  model = 'claude-opus-4-5@20251101'
+
+
 class VertexAIClaude4Opus_20250514(VertexAIAnthropic):
   """Anthropic's Claude 4 Opus model on VertexAI."""
   model = 'claude-opus-4@20250514'

{langfun-0.1.2.dev202512150805 → langfun-0.1.2.dev202512260805}/langfun.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langfun
-Version: 0.1.2.dev202512150805
+Version: 0.1.2.dev202512260805
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors