model-library 0.1.2__tar.gz → 0.1.4__tar.gz

{model_library-0.1.2 → model_library-0.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.2
+Version: 0.1.4
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT
@@ -15,14 +15,13 @@ Requires-Dist: backoff<3.0,>=2.2.1
 Requires-Dist: redis<7.0,>=6.2.0
 Requires-Dist: tiktoken==0.11.0
 Requires-Dist: pillow
-Requires-Dist: openai<2.0,>=1.97.1
+Requires-Dist: openai<3.0,>=2.0
 Requires-Dist: anthropic<1.0,>=0.57.1
-Requires-Dist: together<2.0,>=1.5.25
 Requires-Dist: mistralai<2.0,>=1.9.10
 Requires-Dist: xai-sdk<2.0,>=1.0.0
 Requires-Dist: ai21<5.0,>=4.0.3
 Requires-Dist: boto3<2.0,>=1.38.27
-Requires-Dist: google-genai[aiohttp]<2.0,>=1.48.0
+Requires-Dist: google-genai[aiohttp]>=1.51.0
 Requires-Dist: google-cloud-storage>=1.26.0
 Dynamic: license-file
 

{model_library-0.1.2 → model_library-0.1.4}/examples/prompt_caching.py

@@ -44,19 +44,11 @@ async def run(model: LLM) -> None:
 
     async def query_with_logging(tag: str, question: str) -> None:
         user_prompt = f"{task_spec}\n\nQUESTION: {question}"
-        result = await model.query(
+        await model.query(
             input=user_prompt,
             system_prompt=system_prefix,
             tools=tools,
         )
-        meta = result.metadata
-        model.logger.info(
-            "%s -> cache_write=%s cache_read=%s uncached_input=%s",
-            tag,
-            meta.cache_write_tokens,
-            meta.cache_read_tokens,
-            meta.in_tokens,
-        )
 
     await query_with_logging(
         "first_query",

{model_library-0.1.2 → model_library-0.1.4}/model_library/base/base.py

@@ -64,6 +64,7 @@ class LLMConfig(BaseModel):
     max_tokens: int = DEFAULT_MAX_TOKENS
     temperature: float | None = None
     top_p: float | None = None
+    top_k: int | None = None
     reasoning: bool = False
     reasoning_effort: str | None = None
     supports_images: bool = False
@@ -106,6 +107,7 @@ class LLM(ABC):
         self.max_tokens: int = config.max_tokens
         self.temperature: float | None = config.temperature
         self.top_p: float | None = config.top_p
+        self.top_k: int | None = config.top_k
 
         self.reasoning: bool = config.reasoning
         self.reasoning_effort: str | None = config.reasoning_effort
@@ -218,6 +220,10 @@ class LLM(ABC):
         Join input with history
         Log, Time, and Retry
         """
+
+        # verbose on debug
+        verbose = self.logger.isEnabledFor(logging.DEBUG)
+
         # format str input
         if isinstance(input, str):
             input = [TextInput(text=input)]
@@ -226,11 +232,11 @@ class LLM(ABC):
         input = [*files, *images, *input]
 
         # format input info
-        item_info = f"--- input ({len(input)}): {get_pretty_input_types(input)}\n"
+        item_info = (
+            f"--- input ({len(input)}): {get_pretty_input_types(input, verbose)}\n"
+        )
         if history:
-            item_info += (
-                f"--- history({len(history)}): {get_pretty_input_types(history)}\n"
-            )
+            item_info += f"--- history({len(history)}): {get_pretty_input_types(history, verbose)}\n"
 
         # format tool info
         tool_results = [t for t in input if isinstance(t, ToolResult)]
@@ -251,7 +257,7 @@ class LLM(ABC):
 
         # unique logger for the query
         query_id = uuid.uuid4().hex[:14]
-        query_logger = logging.getLogger(f"{self.logger.name}<query={query_id}>")
+        query_logger = self.logger.getChild(f"query={query_id}")
 
         query_logger.info(
             "Query started:\n" + item_info + tool_info + f"--- kwargs: {short_kwargs}\n"
@@ -277,6 +283,7 @@ class LLM(ABC):
         output.metadata.cost = await self._calculate_cost(output.metadata)
 
         query_logger.info(f"Query completed: {repr(output)}")
+        query_logger.debug(output.model_dump(exclude={"history", "raw"}))
 
         return output
 
@@ -316,7 +323,7 @@ class LLM(ABC):
             )
 
         # costs for long context
-        total_in = metadata.in_tokens + (metadata.cache_read_tokens or 0)
+        total_in = metadata.total_input_tokens
         if costs.context and total_in > costs.context.threshold:
             input_cost, output_cost = costs.context.get_costs(
                 input_cost,

{model_library-0.1.2 → model_library-0.1.4}/model_library/base/output.py

@@ -59,6 +59,33 @@ class QueryResultCost(BaseModel):
             )
         )
 
+    @computed_field
+    @property
+    def total_input(self) -> float:
+        return sum(
+            filter(
+                None,
+                [
+                    self.input,
+                    self.cache_read,
+                    self.cache_write,
+                ],
+            )
+        )
+
+    @computed_field
+    @property
+    def total_output(self) -> float:
+        return sum(
+            filter(
+                None,
+                [
+                    self.output,
+                    self.reasoning,
+                ],
+            )
+        )
+
     @override
     def __repr__(self):
         use_cents = self.total < 1
@@ -92,6 +119,33 @@ class QueryResultMetadata(BaseModel):
     def default_duration_seconds(self) -> float:
         return self.duration_seconds or 0
 
+    @computed_field
+    @property
+    def total_input_tokens(self) -> int:
+        return sum(
+            filter(
+                None,
+                [
+                    self.in_tokens,
+                    self.cache_read_tokens,
+                    self.cache_write_tokens,
+                ],
+            )
+        )
+
+    @computed_field
+    @property
+    def total_output_tokens(self) -> int:
+        return sum(
+            filter(
+                None,
+                [
+                    self.out_tokens,
+                    self.reasoning_tokens,
+                ],
+            )
+        )
+
     def __add__(self, other: "QueryResultMetadata") -> "QueryResultMetadata":
         return QueryResultMetadata(
             in_tokens=self.in_tokens + other.in_tokens,
@@ -107,6 +161,7 @@ class QueryResultMetadata(BaseModel):
             ),
             duration_seconds=self.default_duration_seconds
             + other.default_duration_seconds,
+            cost=self.cost,
         )
 
     @override

{model_library-0.1.2 → model_library-0.1.4}/model_library/base/utils.py

@@ -21,12 +21,13 @@ def sum_optional(a: int | None, b: int | None) -> int | None:
     return (a or 0) + (b or 0)
 
 
-def get_pretty_input_types(input: Sequence["InputItem"]) -> str:
+def get_pretty_input_types(input: Sequence["InputItem"], verbose: bool = False) -> str:
     # for logging
     def process_item(item: "InputItem"):
         match item:
             case TextInput():
-                return truncate_str(repr(item))
+                item_str = repr(item)
+                return item_str if verbose else truncate_str(item_str)
             case FileBase():  # FileInput
                 return repr(item)
             case ToolResult():

model_library-0.1.4/model_library/config/README.md

@@ -0,0 +1,169 @@
+# Model Library Configuration
+
+This directory contains YAML configuration files that define all available models in the model-proxy library.
+
+## Configuration Structure
+
+Each model configuration is organized into distinct sections:
+
+### Core Sections
+
+#### `properties`
+Model-specific technical characteristics and capabilities:
+- `context_window`: Maximum context window in tokens
+- `max_tokens`: Maximum output tokens the model can generate
+- `training_cutoff`: Training data cutoff date (string or null)
+- `reasoning_model`: Whether the model is a reasoning/thinking model
+
+```yaml
+properties:
+  context_window: 200_000
+  max_tokens: 32_000
+  training_cutoff: "2025-03"
+  reasoning_model: false
+```
+
+#### `supports`
+Feature support flags indicating model capabilities:
+- `images`: Supports image inputs
+- `videos`: Supports video inputs
+- `files`: Supports file inputs
+- `batch`: Supports batch requests
+- `temperature`: Supports temperature parameter
+- `tools`: Supports tool/function calling
+
+```yaml
+supports:
+  images: true
+  files: true
+  tools: true
+  batch: true
+  temperature: true
+  videos: false
+```
+
+#### `metadata`
+Vals platform-specific metadata for model availability and status:
+- `deprecated`: Model is deprecated and should not be used for new projects
+- `available_for_everyone`: Model is available to all users
+- `available_as_evaluator`: Model can be used as an evaluator
+- `ignored_for_cost`: Exclude from cost calculations
+
+```yaml
+metadata:
+  deprecated: false
+  available_for_everyone: true
+  available_as_evaluator: false
+  ignored_for_cost: false
+```
+
+#### Other Sections
+
+- `costs_per_million_token`: Pricing information (input, output, cache, batch, context)
+- `default_parameters`: Default parameter values (temperature, top_p, reasoning_effort)
+- `provider_properties`: Provider-specific configuration options
+- `alternative_keys`: Alternative model identifiers/aliases
+
+## Configuration Inheritance
+
+Configurations support hierarchical inheritance through `base-config` blocks:
+
+### 1. Provider-level base-config
+```yaml
+base-config:
+  company: Anthropic
+  open_source: false
+  supports:
+    images: true
+    tools: true
+  metadata:
+    available_for_everyone: true
+```
+
+### 2. Model-block base-config
+```yaml
+claude-4-models:
+  base-config:
+    supports:
+      temperature: true
+    default_parameters:
+      temperature: 1
+
+  anthropic/claude-opus-4-1-20250805:
+    # Inherits from both provider and block base-configs
+    properties:
+      context_window: 200_000
+      max_tokens: 32_000
+```
+
+### 3. Individual model overrides
+Models can override any inherited configuration:
+
+```yaml
+anthropic/claude-opus-4-1-20250805:
+  properties:
+    context_window: 200_000
+    max_tokens: 32_000
+  metadata:
+    available_for_everyone: false  # Override base-config
+```
+
+## Alternative Keys
+
+Models can define alternative identifiers that map to the same configuration:
+
+```yaml
+anthropic/claude-3-5-sonnet-20241022:
+  label: Claude 3.5 Sonnet Latest
+  properties:
+    context_window: 200_000
+    max_tokens: 8_192
+  alternative_keys:
+    - anthropic/claude-3-5-sonnet-latest
+    - anthropic/claude-3.5-sonnet-latest
+```
+
+Alternative keys can also override configuration:
+
+```yaml
+alternative_keys:
+  - anthropic/claude-opus-4-1-20250805-thinking:
+      properties:
+        reasoning_model: true
+```
+
+## Generating all_models.json
+
+After making changes to any YAML configuration file, regenerate the compiled configuration:
+
+```bash
+make config
+```
+
+This generates `all_models.json` which is used by the model registry at runtime.
+
+## Schema Validation
+
+The configuration is validated using Pydantic models defined in `register_models.py`:
+- `Properties` - Model properties
+- `Supports` - Feature support flags
+- `Metadata` - Platform metadata
+- `DefaultParameters` - Default parameter values
+- `CostProperties` - Pricing information
+- `ProviderProperties` - Provider-specific config (dynamically generated)
+
+## Migration Notes
+
+### Previous Structure (Deprecated)
+The old configuration used `class_properties` which mixed support flags and metadata:
+
+```yaml
+# OLD - Do not use
+class_properties:
+  supports_images: true
+  supports_batch_requests: true
+  deprecated: false
+  available_for_everyone: true
+properties:
+  max_token_output: 32_000
+```

{model_library-0.1.2 → model_library-0.1.4}/model_library/config/ai21labs_models.yaml

@@ -1,22 +1,23 @@
 base-config:
   company: AI21 Labs
   documentation_url: https://www.ai21.com/jamba
-  class_properties:
-    supports_images: false
+  supports:
+    images: false
+    files: false
+  metadata:
     available_as_evaluator: false
-    supports_files: false
     available_for_everyone: false
     ignored_for_cost: false
   properties:
     context_window: null
-    max_token_output: 4096
+    max_tokens: 4096
     training_cutoff: null
 
 ai21labs-models:
   base-config:
     open_source: true
-    class_properties:
-      supports_temperature: true
+    supports:
+      temperature: true
     default_parameters:
       temperature: 0.4
 
@@ -46,7 +47,7 @@ ai21labs-models:
     label: Jamba 1.6 Large
     description: The most powerful and efficient long context model
     release_date: 2025-03-06
-    class_properties:
+    metadata:
       deprecated: true
     properties:
       context_window: 256_000
@@ -60,7 +61,7 @@ ai21labs-models:
     label: Jamba 1.6 Mini
     description: The most powerful and efficient long context model
     release_date: 2025-03-06
-    class_properties:
+    metadata:
       deprecated: true
     properties:
       context_window: 256_000
@@ -76,7 +77,7 @@ ai21labs-models:
     release_date: 2024-08-22
     properties:
       context_window: 256_000
-    class_properties:
+    metadata:
       deprecated: true
     costs_per_million_token:
       input: 2.00
@@ -90,7 +91,7 @@ ai21labs-models:
     release_date: 2024-08-22
     properties:
       context_window: 256_000
-    class_properties:
+    metadata:
       deprecated: true
     costs_per_million_token:
       input: 0.2

{model_library-0.1.2 → model_library-0.1.4}/model_library/config/alibaba_models.yaml

@@ -3,13 +3,11 @@ qwen-models:
     company: Alibaba
     open_source: false
 
-    class_properties:
+    supports:
+      temperature: true
+    metadata:
       available_for_everyone: false
       available_as_evaluator: false
-      supports_metadata: true
-      supports_files: false
-      ignored_for_cost: false
-      supports_temperature: true
     default_parameters:
       temperature: 0.7
     properties:
@@ -21,15 +19,16 @@ qwen-models:
     release_date: 2025-09-05
     properties:
       context_window: 262_144
-      max_token_output: 65_536
+      max_tokens: 65_536
       training_cutoff: ""
     costs_per_million_token:
       input: 1.2
       output: 6
-    class_properties:
+    supports:
+      images: false
+      tools: true
+    metadata:
       available_for_everyone: false
-      supports_images: false
-      supports_tools: true
 
   alibaba/qwen3-max-2025-09-23:
     label: Qwen 3 Max 2025-09-23
@@ -37,16 +36,17 @@ qwen-models:
     release_date: 2025-09-23
     properties:
       context_window: 262_144
-      max_token_output: 65_536
+      max_tokens: 65_536
       training_cutoff: ""
       reasoning_model: true
     costs_per_million_token:
       input: 1.2
       output: 6
-    class_properties:
+    supports:
+      images: false
+      tools: true
+    metadata:
       available_for_everyone: false
-      supports_images: false
-      supports_tools: true
 
   alibaba/qwen3-max:
     label: Qwen 3 Max
@@ -54,7 +54,7 @@ qwen-models:
     release_date: 2025-09-23
     properties:
       context_window: 262_144
-      max_token_output: 65_536
+      max_tokens: 65_536
       training_cutoff: ""
       reasoning_model: false
     costs_per_million_token:
@@ -67,12 +67,11 @@ qwen-models:
         threshold: 32_000
         input: 2.4
         output: 12
-
-
-    class_properties:
+    supports:
+      images: false
+      tools: true
+    metadata:
       available_for_everyone: false
-      supports_images: false
-      supports_tools: true
 
   alibaba/qwen3-vl-plus-2025-09-23:
     label: Qwen 3 VL Plus
@@ -81,11 +80,11 @@ qwen-models:
     release_date: 2025-09-23
     properties:
       context_window: 262_144
-      max_token_output: 32_768
+      max_tokens: 32_768
       training_cutoff: ""
       reasoning_model: false
     costs_per_million_token:
       input: 0.2
       output: 1.6
-    class_properties:
-      supports_images: true
+    supports:
+      images: true