lm-deluge 0.0.78__tar.gz → 0.0.79__tar.gz

{lm_deluge-0.0.78/src/lm_deluge.egg-info → lm_deluge-0.0.79}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.78
+Version: 0.0.79
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10
@@ -52,7 +52,7 @@ Dynamic: license-file
 pip install lm-deluge
 ```
 
-The package relies on environment variables for API keys. Typical variables include `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `COHERE_API_KEY`, `META_API_KEY`, and `GOOGLE_API_KEY`. `LLMClient` will automatically load the `.env` file when imported; we recommend using that to set the environment variables. For Bedrock, you'll need to set `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY`.
+The package relies on environment variables for API keys. Typical variables include `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `COHERE_API_KEY`, `META_API_KEY`, and `GEMINI_API_KEY`. `LLMClient` will automatically load the `.env` file when imported; we recommend using that to set the environment variables. For Bedrock, you'll need to set `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY`.
 
 ## Quickstart
 
@@ -61,9 +61,9 @@ The package relies on environment variables for API keys. Typical variables incl
 ```python
 from lm_deluge import LLMClient
 
-client = LLMClient("gpt-4o-mini")
+client = LLMClient("gpt-4.1-mini")
 resps = client.process_prompts_sync(["Hello, world!"])
-print(resp[0].completion)
+print(resps[0].completion)
 ```
 
 ## Spraying Across Models
@@ -74,13 +74,13 @@ To distribute your requests across models, just provide a list of more than one
 from lm_deluge import LLMClient
 
 client = LLMClient(
-    ["gpt-4o-mini", "claude-3-haiku"],
+    ["gpt-4.1-mini", "claude-4.5-haiku"],
     max_requests_per_minute=10_000
 )
 resps = client.process_prompts_sync(
     ["Hello, ChatGPT!", "Hello, Claude!"]
 )
-print(resp[0].completion)
+print(resps[0].completion)
 ```
 
 ## Configuration
@@ -181,7 +181,7 @@ def get_weather(city: str) -> str:
     return f"The weather in {city} is sunny and 72°F"
 
 tool = Tool.from_function(get_weather)
-client = LLMClient("claude-3-haiku")
+client = LLMClient("claude-4.5-haiku")
 resps = client.process_prompts_sync(
     ["What's the weather in Paris?"],
     tools=[tool]
@@ -255,7 +255,7 @@ conv = (
 )
 
 # Use prompt caching to cache system message and tools
-client = LLMClient("claude-3-5-sonnet")
+client = LLMClient("claude-4.5-sonnet")
 resps = client.process_prompts_sync(
     [conv],
     cache="system_and_tools"  # Cache system message and any tools

{lm_deluge-0.0.78 → lm_deluge-0.0.79}/README.md

@@ -23,7 +23,7 @@
 pip install lm-deluge
 ```
 
-The package relies on environment variables for API keys. Typical variables include `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `COHERE_API_KEY`, `META_API_KEY`, and `GOOGLE_API_KEY`. `LLMClient` will automatically load the `.env` file when imported; we recommend using that to set the environment variables. For Bedrock, you'll need to set `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY`.
+The package relies on environment variables for API keys. Typical variables include `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `COHERE_API_KEY`, `META_API_KEY`, and `GEMINI_API_KEY`. `LLMClient` will automatically load the `.env` file when imported; we recommend using that to set the environment variables. For Bedrock, you'll need to set `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY`.
 
 ## Quickstart
 
@@ -32,9 +32,9 @@ The package relies on environment variables for API keys. Typical variables incl
 ```python
 from lm_deluge import LLMClient
 
-client = LLMClient("gpt-4o-mini")
+client = LLMClient("gpt-4.1-mini")
 resps = client.process_prompts_sync(["Hello, world!"])
-print(resp[0].completion)
+print(resps[0].completion)
 ```
 
 ## Spraying Across Models
@@ -45,13 +45,13 @@ To distribute your requests across models, just provide a list of more than one
 from lm_deluge import LLMClient
 
 client = LLMClient(
-    ["gpt-4o-mini", "claude-3-haiku"],
+    ["gpt-4.1-mini", "claude-4.5-haiku"],
     max_requests_per_minute=10_000
 )
 resps = client.process_prompts_sync(
     ["Hello, ChatGPT!", "Hello, Claude!"]
 )
-print(resp[0].completion)
+print(resps[0].completion)
 ```
 
 ## Configuration
@@ -152,7 +152,7 @@ def get_weather(city: str) -> str:
     return f"The weather in {city} is sunny and 72°F"
 
 tool = Tool.from_function(get_weather)
-client = LLMClient("claude-3-haiku")
+client = LLMClient("claude-4.5-haiku")
 resps = client.process_prompts_sync(
     ["What's the weather in Paris?"],
     tools=[tool]
@@ -226,7 +226,7 @@ conv = (
 )
 
 # Use prompt caching to cache system message and tools
-client = LLMClient("claude-3-5-sonnet")
+client = LLMClient("claude-4.5-sonnet")
 resps = client.process_prompts_sync(
     [conv],
     cache="system_and_tools"  # Cache system message and any tools

{lm_deluge-0.0.78 → lm_deluge-0.0.79}/pyproject.toml

@@ -3,7 +3,7 @@ requires = ["setuptools", "wheel"]
 
 [project]
 name = "lm_deluge"
-version = "0.0.78"
+version = "0.0.79"
 authors = [{ name = "Benjamin Anderson", email = "ben@trytaylor.ai" }]
 description = "Python utility for using LLM API models."
 readme = "README.md"

{lm_deluge-0.0.78 → lm_deluge-0.0.79}/src/lm_deluge/api_requests/gemini.py

@@ -23,6 +23,21 @@ async def _build_gemini_request(
 ) -> dict:
     system_message, messages = prompt.to_gemini()
 
+    # For Gemini 3, inject dummy signatures when missing for function calls
+    is_gemini_3 = "gemini-3" in model.name.lower()
+    if is_gemini_3:
+        dummy_sig = "context_engineering_is_the_way_to_go"
+        for msg in messages:
+            if "parts" in msg:
+                for part in msg["parts"]:
+                    # For function calls, inject dummy signature if missing
+                    if "functionCall" in part and "thoughtSignature" not in part:
+                        part["thoughtSignature"] = dummy_sig
+                        maybe_warn(
+                            "WARN_GEMINI3_MISSING_SIGNATURE",
+                            part_type="function call",
+                        )
+
     request_json = {
         "contents": messages,
         "generationConfig": {
@@ -40,17 +55,44 @@ async def _build_gemini_request(
     if model.reasoning_model:
         thinking_config: dict[str, Any] | None = None
         effort = sampling_params.reasoning_effort
-        if effort is None or effort == "none":
-            budget = 128 if "2.5-pro" in model.id else 0
-            # Explicitly disable thoughts when no effort is requested
-            thinking_config = {"includeThoughts": False, "thinkingBudget": budget}
+        is_gemini_3 = "gemini-3" in model.name.lower()
+
+        if is_gemini_3:
+            # Gemini 3 uses thinkingLevel instead of thinkingBudget
+            if effort in {"none", "minimal"}:
+                thinking_config = {"thinkingLevel": "low"}
+            elif effort is None:
+                # Default to high when reasoning is enabled but no preference was provided
+                thinking_config = {"thinkingLevel": "high"}
+            else:
+                # Map reasoning_effort to thinkingLevel
+                level_map = {
+                    "minimal": "low",
+                    "low": "low",
+                    "medium": "medium",  # Will work when supported
+                    "high": "high",
+                }
+                thinking_level = level_map.get(effort, "high")
+                thinking_config = {"thinkingLevel": thinking_level}
         else:
-            thinking_config = {"includeThoughts": True}
-            if effort in {"minimal", "low", "medium", "high"} and "flash" in model.id:
-                budget = {"minimal": 256, "low": 1024, "medium": 4096, "high": 16384}[
-                    effort
-                ]
-                thinking_config["thinkingBudget"] = budget
+            # Gemini 2.5 uses thinkingBudget (legacy)
+            if effort is None or effort == "none":
+                budget = 128 if "2.5-pro" in model.id else 0
+                # Explicitly disable thoughts when no effort is requested
+                thinking_config = {"includeThoughts": False, "thinkingBudget": budget}
+            else:
+                thinking_config = {"includeThoughts": True}
+                if (
+                    effort in {"minimal", "low", "medium", "high"}
+                    and "flash" in model.id
+                ):
+                    budget = {
+                        "minimal": 256,
+                        "low": 1024,
+                        "medium": 4096,
+                        "high": 16384,
+                    }[effort]
+                    thinking_config["thinkingBudget"] = budget
         request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
     else:
@@ -66,6 +108,21 @@ async def _build_gemini_request(
     if sampling_params.json_mode and model.supports_json:
         request_json["generationConfig"]["responseMimeType"] = "application/json"
 
+    # Handle media_resolution for Gemini 3 (requires v1alpha)
+    if sampling_params.media_resolution is not None:
+        is_gemini_3 = "gemini-3" in model.name.lower()
+        if is_gemini_3:
+            # Add global media resolution to generationConfig
+            request_json["generationConfig"]["mediaResolution"] = {
+                "level": sampling_params.media_resolution
+            }
+        else:
+            # Warn if trying to use media_resolution on non-Gemini-3 models
+            maybe_warn(
+                "WARN_MEDIA_RESOLUTION_UNSUPPORTED",
+                model_name=model.name,
+            )
+
     return request_json
 
 
@@ -137,10 +194,19 @@ class GeminiRequest(APIRequestBase):
                         candidate = data["candidates"][0]
                         if "content" in candidate and "parts" in candidate["content"]:
                             for part in candidate["content"]["parts"]:
+                                # Extract thought signature if present
+                                thought_sig = part.get("thoughtSignature")
+
                                 if "text" in part:
                                     parts.append(Text(part["text"]))
                                 elif "thought" in part:
-                                    parts.append(Thinking(part["thought"]))
+                                    # Thought with optional signature
+                                    parts.append(
+                                        Thinking(
+                                            content=part["thought"],
+                                            thought_signature=thought_sig,
+                                        )
+                                    )
                                 elif "functionCall" in part:
                                     func_call = part["functionCall"]
                                     # Generate a unique ID since Gemini doesn't provide one
@@ -152,6 +218,7 @@ class GeminiRequest(APIRequestBase):
                                             id=tool_id,
                                             name=func_call["name"],
                                             arguments=func_call.get("args", {}),
+                                            thought_signature=thought_sig,
                                         )
                                     )
 

{lm_deluge-0.0.78 → lm_deluge-0.0.79}/src/lm_deluge/client.py

@@ -262,6 +262,7 @@ class _LLMClient(BaseModel):
             self.max_tokens_per_minute = max_tokens_per_minute
         if max_concurrent_requests:
             self.max_concurrent_requests = max_concurrent_requests
+        return self
 
     def _get_tracker(self) -> StatusTracker:
         if self._tracker is None:

{lm_deluge-0.0.78 → lm_deluge-0.0.79}/src/lm_deluge/config.py

@@ -12,6 +12,13 @@ class SamplingParams(BaseModel):
     logprobs: bool = False
     top_logprobs: int | None = None
     strict_tools: bool = True
+    # Gemini 3 only - controls multimodal vision processing fidelity
+    media_resolution: (
+        Literal[
+            "media_resolution_low", "media_resolution_medium", "media_resolution_high"
+        ]
+        | None
+    ) = None
 
     def to_vllm(self):
         try: