lm-deluge 0.0.25__tar.gz → 0.0.27__tar.gz

{lm_deluge-0.0.25/src/lm_deluge.egg-info → lm_deluge-0.0.27}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.25
+Version: 0.0.27
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10
@@ -274,7 +274,7 @@ We support all models in `src/lm_deluge/models.py`. Vertex support is not planne
 
 ## Feature Support
 
-We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
+We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Passing `None` (or the string `"none"`) disables Gemini thoughts entirely. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
 
 ## Built‑in tools
 

{lm_deluge-0.0.25 → lm_deluge-0.0.27}/README.md

@@ -247,7 +247,7 @@ We support all models in `src/lm_deluge/models.py`. Vertex support is not planne
 
 ## Feature Support
 
-We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
+We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Passing `None` (or the string `"none"`) disables Gemini thoughts entirely. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
 
 ## Built‑in tools
 

{lm_deluge-0.0.25 → lm_deluge-0.0.27}/pyproject.toml

@@ -3,7 +3,7 @@ requires = ["setuptools", "wheel"]
 
 [project]
 name = "lm_deluge"
-version = "0.0.25"
+version = "0.0.27"
 authors = [{ name = "Benjamin Anderson", email = "ben@trytaylor.ai" }]
 description = "Python utility for using LLM API models."
 readme = "README.md"

{lm_deluge-0.0.25 → lm_deluge-0.0.27}/src/lm_deluge/api_requests/gemini.py

@@ -37,14 +37,17 @@ async def _build_gemini_request(
 
     # Handle reasoning models (thinking)
     if model.reasoning_model:
-        request_json["generationConfig"]["thinkingConfig"] = {"includeThoughts": True}
-        if sampling_params.reasoning_effort and "flash" in model.id:
-            budget = {"low": 1024, "medium": 4096, "high": 16384}.get(
-                sampling_params.reasoning_effort
-            )
-            request_json["generationConfig"]["thinkingConfig"]["thinkingBudget"] = (
-                budget
-            )
+        thinking_config = None
+        effort = sampling_params.reasoning_effort
+        if effort is None or effort == "none":
+            # Explicitly disable thoughts when no effort is requested
+            thinking_config = {"includeThoughts": False, "thinkingBudget": 0}
+        else:
+            thinking_config = {"includeThoughts": True}
+            if effort in {"low", "medium", "high"} and "flash" in model.id:
+                budget = {"low": 1024, "medium": 4096, "high": 16384}[effort]
+                thinking_config["thinkingBudget"] = budget
+        request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
     else:
         if sampling_params.reasoning_effort:

{lm_deluge-0.0.25 → lm_deluge-0.0.27}/src/lm_deluge/file.py

@@ -1,3 +1,4 @@
+from functools import cached_property
 import os
 import io
 import requests
@@ -68,13 +69,13 @@ class File:
             return encoded
         return f"data:{self._mime()};base64,{encoded}"
 
-    @property
+    @cached_property
     def fingerprint(self) -> str:
         # Hash the file contents for fingerprinting
         file_bytes = self._bytes()
         return xxhash.xxh64(file_bytes).hexdigest()
 
-    @property
+    @cached_property
     def size(self) -> int:
         """Return file size in bytes."""
         return len(self._bytes())

{lm_deluge-0.0.25 → lm_deluge-0.0.27}/src/lm_deluge/request_context.py

@@ -1,4 +1,5 @@
 from dataclasses import dataclass, field
+from functools import cached_property
 from typing import Any, Callable
 
 from .config import SamplingParams
@@ -39,14 +40,18 @@ class RequestContext:
 
     # Computed properties
     cache_key: str = field(init=False)
-    num_tokens: int = field(init=False)
+    # num_tokens: int = field(init=False)
 
-    def __post_init__(self):
-        # Compute cache key from prompt fingerprint
-        self.cache_key = self.prompt.fingerprint
+    # def __post_init__(self):
+    #     # Compute cache key from prompt fingerprint
+    #     # self.cache_key = self.prompt.fingerprint
 
-        # Compute token count
-        self.num_tokens = self.prompt.count_tokens(self.sampling_params.max_new_tokens)
+    #     # Compute token count
+    #     self.num_tokens =
+
+    @cached_property
+    def num_tokens(self):
+        return self.prompt.count_tokens(self.sampling_params.max_new_tokens)
 
     def maybe_callback(self, response, tracker):
         if not self.callback:

{lm_deluge-0.0.25 → lm_deluge-0.0.27/src/lm_deluge.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.25
+Version: 0.0.27
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10
@@ -274,7 +274,7 @@ We support all models in `src/lm_deluge/models.py`. Vertex support is not planne
 
 ## Feature Support
 
-We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
+We support structured outputs via `json_mode` parameter provided to `SamplingParams`. Structured outputs with a schema are planned. Reasoning models are supported via the `reasoning_effort` parameter, which is translated to a thinking budget for Claude/Gemini. Passing `None` (or the string `"none"`) disables Gemini thoughts entirely. Image models are supported. We support tool use as documented above. We support logprobs for OpenAI models that return them.
 
 ## Built‑in tools