lm-deluge 0.0.79__tar.gz → 0.0.80__tar.gz

{lm_deluge-0.0.79/src/lm_deluge.egg-info → lm_deluge-0.0.80}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.79
+Version: 0.0.80
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/pyproject.toml

@@ -3,7 +3,7 @@ requires = ["setuptools", "wheel"]
 
 [project]
 name = "lm_deluge"
-version = "0.0.79"
+version = "0.0.80"
 authors = [{ name = "Benjamin Anderson", email = "ben@trytaylor.ai" }]
 description = "Python utility for using LLM API models."
 readme = "README.md"

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/api_requests/anthropic.py

@@ -16,6 +16,7 @@ from lm_deluge.util.schema import (
     prepare_output_schema,
     transform_schema_for_anthropic,
 )
+from lm_deluge.warnings import maybe_warn
 
 from ..models import APIModel
 from .base import APIRequestBase, APIResponse
@@ -62,20 +63,45 @@ def _build_anthropic_request(
         "max_tokens": sampling_params.max_new_tokens,
     }
 
+    if model.id == "claude-4.5-opus" and sampling_params.global_effort:
+        request_json["effort"] = sampling_params.global_effort
+        _add_beta(base_headers, "effort-2025-11-24")
+
     # handle thinking
-    if model.reasoning_model and sampling_params.reasoning_effort:
-        # translate reasoning effort of low, medium, high to budget tokens
-        budget = {"minimal": 256, "low": 1024, "medium": 4096, "high": 16384}.get(
-            sampling_params.reasoning_effort
-        )
-        request_json["thinking"] = {
-            "type": "enabled",
-            "budget_tokens": budget,
-        }
-        if "top_p" in request_json:
-            request_json["top_p"] = max(request_json["top_p"], 0.95)
-        request_json["temperature"] = 1.0
-        request_json["max_tokens"] += budget
+    if model.reasoning_model:
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.reasoning_effort is not None
+        ):
+            maybe_warn("WARN_THINKING_BUDGET_AND_REASONING_EFFORT")
+
+        if sampling_params.thinking_budget is not None:
+            budget = sampling_params.thinking_budget
+        elif sampling_params.reasoning_effort is not None:
+            # translate reasoning effort of low, medium, high to budget tokens
+            budget = {
+                "none": 0,
+                "minimal": 256,
+                "low": 1024,
+                "medium": 4096,
+                "high": 16384,
+            }.get(sampling_params.reasoning_effort)
+            assert isinstance(budget, int)
+        else:
+            budget = 0
+
+        if budget > 0:
+            request_json["thinking"] = {
+                "type": "enabled",
+                "budget_tokens": budget,
+            }
+            if "top_p" in request_json:
+                request_json["top_p"] = max(request_json["top_p"], 0.95)
+            request_json["temperature"] = 1.0
+            request_json["max_tokens"] += budget
+        else:
+            request_json["thinking"] = {"type": "disabled"}
+
     else:
         request_json["thinking"] = {"type": "disabled"}
         if sampling_params.reasoning_effort:
@@ -83,10 +109,11 @@ def _build_anthropic_request(
     if system_message is not None:
         request_json["system"] = system_message
 
-    # handle temp + top_p for opus 4.1/sonnet 4.5
+    # handle temp + top_p for opus 4.1/sonnet 4.5.
+    # TODO: make clearer / more user-friendly so there can be NotGiven
+    # and user can control which one they want to use
     if "4-1" in model.name or "4-5" in model.name:
-        if "temperature" in request_json and "top_p" in request_json:
-            request_json.pop("top_p")
+        request_json.pop("top_p")
 
     # Handle structured outputs (output_format)
     if context.output_schema:

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/api_requests/gemini.py

@@ -1,6 +1,5 @@
 import json
 import os
-from typing import Any
 
 from aiohttp import ClientResponse
 
@@ -52,47 +51,61 @@ async def _build_gemini_request(
         request_json["systemInstruction"] = {"parts": [{"text": system_message}]}
 
     # Handle reasoning models (thinking)
-    if model.reasoning_model:
-        thinking_config: dict[str, Any] | None = None
-        effort = sampling_params.reasoning_effort
-        is_gemini_3 = "gemini-3" in model.name.lower()
+    is_gemini_3 = "gemini-3" in model.name.lower()
+    if is_gemini_3:
+        # gemini3 MUST think
+        if not sampling_params.reasoning_effort:
+            maybe_warn("WARN_GEMINI3_NO_REASONING")
+            effort = "low"
+        else:
+            level_map = {
+                "none": "low",
+                "minimal": "low",
+                "low": "low",
+                "medium": "high",  # change when supported
+                "high": "high",
+            }
+            effort = level_map[sampling_params.reasoning_effort]
+        thinking_config = {"thinkingLevel": effort}
+        request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
-        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}
+    elif model.reasoning_model:
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.reasoning_effort is not None
+        ):
+            maybe_warn("WARN_THINKING_BUDGET_AND_REASONING_EFFORT")
+
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.thinking_budget > 0
+        ):
+            thinking_config = {
+                "includeThoughts": True,
+                "thinkingBudget": sampling_params.thinking_budget,
+            }
+        elif sampling_params.thinking_budget == -1:
+            # dynamic thinking
+            thinking_config = {"includeThoughts": True, "thinkingBudget": -1}
+        elif sampling_params.reasoning_effort not in [None, "none"]:
+            level_map = {
+                "minimal": 256,
+                "low": 1024,
+                "medium": 4096,
+                "high": 16384,
+            }
+            assert sampling_params.reasoning_effort in level_map
+            budget = level_map[sampling_params.reasoning_effort]
+            if "flash-lite" in model.id:
+                budget = max(budget, 512)
+            thinking_config = {"includeThoughts": True, "thinkingBudget": budget}
+        elif "2.5-pro" in model.id:
+            # 2.5 pro must think.
+            thinking_config = {"includeThoughts": True, "thinkingBudget": 128}
         else:
-            # 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
+            # no thoughts head empty
+            thinking_config = {"includeThoughts": False, "thinkingBudget": 0}
+
         request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
     else:

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/client.py

@@ -79,7 +79,7 @@ class _LLMClient(BaseModel):
     background: bool = False
     # sampling params - if provided, and sampling_params is not,
     # these override the defaults
-    temperature: float = 0.75
+    temperature: float = 1.0
     top_p: float = 1.0
     json_mode: bool = False
     max_new_tokens: int = 512
@@ -337,7 +337,7 @@ class _LLMClient(BaseModel):
         if "sampling_params" not in data or len(data.get("sampling_params", [])) == 0:
             data["sampling_params"] = [
                 SamplingParams(
-                    temperature=data.get("temperature", 0.75),
+                    temperature=data.get("temperature", 1.0),
                     top_p=data.get("top_p", 1.0),
                     json_mode=data.get("json_mode", False),
                     max_new_tokens=data.get("max_new_tokens", 512),
@@ -1067,7 +1067,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,
@@ -1096,7 +1096,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,
@@ -1124,7 +1124,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/config.py

@@ -4,11 +4,13 @@ from pydantic import BaseModel
 
 
 class SamplingParams(BaseModel):
-    temperature: float = 0.0
+    temperature: float = 1.0  # more typical for new models
     top_p: float = 1.0
     json_mode: bool = False
     max_new_tokens: int = 2_048
+    global_effort: Literal["low", "medium", "high"] = "high"  # for opus-4.5
     reasoning_effort: Literal["low", "medium", "high", "minimal", "none", None] = None
+    thinking_budget: int | None = None
     logprobs: bool = False
     top_logprobs: int | None = None
     strict_tools: bool = True

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/models/anthropic.py

@@ -10,6 +10,19 @@ ANTHROPIC_MODELS = {
     #                                                                   ░███
     #                                                                   █████
     #
+    "claude-4.5-opus": {
+        "id": "claude-4.5-opus",
+        "name": "claude-opus-4-5-20251101",
+        "api_base": "https://api.anthropic.com/v1",
+        "api_key_env_var": "ANTHROPIC_API_KEY",
+        "supports_json": False,
+        "api_spec": "anthropic",
+        "input_cost": 5.0,
+        "cached_input_cost": 0.50,
+        "cache_write_cost": 6.25,
+        "output_cost": 25.0,
+        "reasoning_model": True,
+    },
     "claude-4.5-haiku": {
         "id": "claude-4.5-haiku",
         "name": "claude-haiku-4-5-20251001",
@@ -21,6 +34,7 @@ ANTHROPIC_MODELS = {
         "cached_input_cost": 0.10,
         "cache_write_cost": 1.25,
         "output_cost": 3.0,
+        "reasoning_model": True,
     },
     "claude-4.5-sonnet": {
         "id": "claude-4.5-sonnet",
@@ -33,6 +47,7 @@ ANTHROPIC_MODELS = {
         "cached_input_cost": 0.30,
         "cache_write_cost": 3.75,
         "output_cost": 15.0,
+        "reasoning_model": True,
     },
     "claude-4.1-opus": {
         "id": "claude-4.1-opus",

{lm_deluge-0.0.79 → lm_deluge-0.0.80}/src/lm_deluge/warnings.py

@@ -11,6 +11,8 @@ WARNINGS: dict[str, str] = {
     "WARN_MINIMAL_TO_NONE": "GPT-5.1 models don't support 'minimal' reasoning effort. Converting to 'none' for {model_name}.",
     "WARN_MEDIA_RESOLUTION_UNSUPPORTED": "media_resolution parameter is only supported for Gemini 3 models, ignoring for {model_name}.",
     "WARN_GEMINI3_MISSING_SIGNATURE": "Gemini 3 thought signature missing in {part_type}, injecting dummy signature 'context_engineering_is_the_way_to_go' to avoid API error.",
+    "WARN_GEMINI3_NO_REASONING": "Gemini 3 requires reasoning (thinkingConfig). Setting thinkingConfig to low.",
+    "WARN_THINKING_BUDGET_AND_REASONING_EFFORT": "`reasoning_effort` and `thinking_budget` both provided. `thinking_budget` will take priority.",
 }
 
 

{lm_deluge-0.0.79 → lm_deluge-0.0.80/src/lm_deluge.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.79
+Version: 0.0.80
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10