llm-codegen-research 2.16__tar.gz → 2.17__tar.gz

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llm-codegen-research
-Version: 2.16
+Version: 2.17
 Summary: Useful classes and methods for researching code-generation by LLMs.
 Author-email: Lukas Twist <itsluketwist@gmail.com>
 Project-URL: Homepage, https://github.com/itsluketwist/llm-codegen-research
@@ -13,6 +13,7 @@ License-File: LICENSE
 Requires-Dist: esprima
 Provides-Extra: api
 Requires-Dist: anthropic; extra == "api"
+Requires-Dist: google-genai; extra == "api"
 Requires-Dist: mistralai; extra == "api"
 Requires-Dist: openai; extra == "api"
 Requires-Dist: together; extra == "api"

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/pyproject.toml

@@ -28,6 +28,7 @@ dependencies = [
 [project.optional-dependencies]
 api = [
     "anthropic",
+    "google-genai",
     "mistralai",
     "openai",
     "together",

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_cgr/__init__.py

@@ -43,9 +43,11 @@ try:
         Base_LLM,
         DeepSeek_LLM,
         GenerationProtocol,
+        Google_LLM,
         Mistral_LLM,
         OpenAI_LLM,
         OpenAI_Tool_LLM,
+        Qwen_LLM,
         TogetherAI_LLM,
         Tool,
         generate,
@@ -64,9 +66,11 @@ try:
         "Base_LLM",
         "DeepSeek_LLM",
         "GenerationProtocol",
+        "Google_LLM",
         "Mistral_LLM",
         "OpenAI_LLM",
         "OpenAI_Tool_LLM",
+        "Qwen_LLM",
         "Tool",
         "TogetherAI_LLM",
         "generate",

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_cgr/llm/__init__.py

@@ -3,9 +3,11 @@ from llm_cgr.llm.clients import (
     Base_LLM,
     DeepSeek_LLM,
     GenerationProtocol,
+    Google_LLM,
     Mistral_LLM,
     OpenAI_LLM,
     OpenAI_Tool_LLM,
+    Qwen_LLM,
     TogetherAI_LLM,
     Tool,
     get_llm,
@@ -24,9 +26,11 @@ __all__ = [
     "Base_LLM",
     "DeepSeek_LLM",
     "GenerationProtocol",
+    "Google_LLM",
     "Mistral_LLM",
     "OpenAI_LLM",
     "OpenAI_Tool_LLM",
+    "Qwen_LLM",
     "Tool",
     "TogetherAI_LLM",
     "get_llm",

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_cgr/llm/clients/__init__.py

@@ -3,6 +3,7 @@
 from llm_cgr.llm.clients.anthropic import Anthropic_LLM
 from llm_cgr.llm.clients.base import Base_LLM
 from llm_cgr.llm.clients.deepseek import DeepSeek_LLM
+from llm_cgr.llm.clients.google import Google_LLM
 from llm_cgr.llm.clients.mistral import Mistral_LLM
 from llm_cgr.llm.clients.nscale import Nscale_LLM
 from llm_cgr.llm.clients.openai import OpenAI_LLM
@@ -13,14 +14,17 @@ from llm_cgr.llm.clients.openai_tool import (
     Tool,
 )
 from llm_cgr.llm.clients.protocol import GenerationProtocol
+from llm_cgr.llm.clients.qwen import Qwen_LLM
 from llm_cgr.llm.clients.together import TogetherAI_LLM
 
 
 PROVIDER_MAP: dict[str, type[Base_LLM]] = {
     "anthropic": Anthropic_LLM,
     "deepseek": DeepSeek_LLM,
+    "google": Google_LLM,
     "mistral": Mistral_LLM,
     "openai": OpenAI_LLM,
+    "qwen": Qwen_LLM,
     "together": TogetherAI_LLM,
     "nscale": Nscale_LLM,
 }
@@ -43,7 +47,7 @@ def get_llm(
 
     If tools are provided, returns an OpenAI_Tool_LLM instance. Tool calls
     are currently only supported for OpenAI models. enable_reasoning is only
-    supported by Anthropic, DeepSeek, Mistral, and TogetherAI models.
+    supported by Anthropic, DeepSeek, Google, Mistral, Qwen, and TogetherAI models.
     """
     llm_class: type[Base_LLM]
     if provider is not None:
@@ -56,6 +60,10 @@ def get_llm(
         llm_class = Mistral_LLM
     elif "deepseek" in model:
         llm_class = DeepSeek_LLM
+    elif "qwen" in model.lower():
+        llm_class = Qwen_LLM
+    elif "gemini" in model.lower():
+        llm_class = Google_LLM
     else:
         llm_class = TogetherAI_LLM
 
@@ -93,10 +101,12 @@ __all__ = [
     "Base_LLM",
     "DeepSeek_LLM",
     "GenerationProtocol",
+    "Google_LLM",
     "OpenAI_LLM",
     "OpenAI_Tool_LLM",
     "TogetherAI_LLM",
     "Mistral_LLM",
+    "Qwen_LLM",
     "Tool",
     "get_llm",
 ]

llm_codegen_research-2.17/src/llm_cgr/llm/clients/google.py

@@ -0,0 +1,106 @@
+"""Class to access LLMs via the Google Gemini API."""
+
+import os
+from typing import Any, cast
+
+from google import genai
+from google.genai import types
+
+from llm_cgr.llm.clients.base import Base_LLM
+
+
+class Google_LLM(Base_LLM):
+    """Class to access LLMs via the Google Gemini API."""
+
+    def __init__(
+        self,
+        model: str | None = None,
+        system: str | None = None,
+        temperature: float | None = None,
+        top_p: float | None = None,
+        max_tokens: int | None = None,
+        enable_reasoning: bool = False,
+    ) -> None:
+        """
+        Initialise the Google client.
+
+        Requires the GOOGLE_API_KEY environment variable to be set.
+        Set enable_reasoning=True to include the model's thinking summary on
+        supported models (e.g. gemini-2.5-pro); thinking output is hidden by
+        default.
+        """
+        super().__init__(
+            model=model,
+            system=system,
+            temperature=temperature,
+            top_p=top_p,
+            max_tokens=max_tokens,
+            enable_reasoning=enable_reasoning,
+        )
+        self._client = genai.Client(
+            api_key=os.environ["GOOGLE_API_KEY"],
+        )
+
+    def _build_message(
+        self,
+        role: str,
+        content: str,
+    ) -> dict[str, Any]:
+        """Build a Gemini model message."""
+        # gemini uses "model" rather than "assistant" for past responses
+        return {
+            "role": "model" if role == "assistant" else "user",
+            "parts": [{"text": content}],
+        }
+
+    def _build_input(
+        self,
+        user: str,
+        system: str | None = None,
+    ) -> list[dict[str, Any]]:
+        """Build the full Gemini model input."""
+        # the system prompt is passed separately as part of the request config
+        return [self._build_message(role="user", content=user)]
+
+    def _get_response(
+        self,
+        model: str,
+        input: list[dict[str, Any]],
+        system: str | None = None,
+        temperature: float | None = None,
+        top_p: float | None = None,
+        max_tokens: int | None = None,
+    ) -> tuple[str, str | None]:
+        """Generate a model response from the Google Gemini API."""
+        # built-in tools (google search, code execution, url context, etc.) are
+        # deliberately left unset, so they remain off by default
+        config = types.GenerateContentConfig(
+            system_instruction=system or self._system,
+            temperature=temperature,
+            top_p=top_p,
+            max_output_tokens=max_tokens,
+            thinking_config=(
+                types.ThinkingConfig(include_thoughts=True)
+                if self._enable_reasoning
+                else None
+            ),
+        )
+        response = self._client.models.generate_content(
+            model=model,
+            contents=cast(types.ContentListUnionDict, input),
+            config=config,
+        )
+        candidate = response.candidates[0] if response.candidates else None
+        content = candidate.content if candidate else None
+        parts = (content.parts if content else None) or []
+
+        # chain-of-thought from thinking-enabled models; None otherwise
+        reasoning_parts = [
+            part.text for part in parts if part.thought and part.text is not None
+        ]
+        reasoning = "\n".join(reasoning_parts) if reasoning_parts else None
+
+        text_parts = [
+            part.text for part in parts if not part.thought and part.text is not None
+        ]
+        return "\n".join(text_parts), reasoning

llm_codegen_research-2.17/src/llm_cgr/llm/clients/qwen.py

@@ -0,0 +1,90 @@
+"""Class to access LLMs via the Qwen API."""
+
+import os
+from typing import Any, cast
+
+import openai
+from openai.types.chat import ChatCompletionMessageParam
+
+from llm_cgr.llm.clients.base import Base_LLM
+
+
+class Qwen_LLM(Base_LLM):
+    """Class to access LLMs via the Qwen API, using the OpenAI interfaces."""
+
+    def __init__(
+        self,
+        model: str | None = None,
+        system: str | None = None,
+        temperature: float | None = None,
+        top_p: float | None = None,
+        max_tokens: int | None = None,
+        enable_reasoning: bool = False,
+    ) -> None:
+        """
+        Initialise the Qwen client.
+
+        Requires the QWEN_API_KEY environment variable to be set.
+        Set enable_reasoning=True to enable thinking mode on supported models
+        (e.g. qwen3-235b-a22b); thinking is off by default.
+        """
+        super().__init__(
+            model=model,
+            system=system,
+            temperature=temperature,
+            top_p=top_p,
+            max_tokens=max_tokens,
+            enable_reasoning=enable_reasoning,
+        )
+        self._client = openai.OpenAI(
+            api_key=os.environ["QWEN_API_KEY"],
+            base_url=os.environ["QWEN_API_URL"],
+        )
+
+    def _build_message(
+        self,
+        role: str,
+        content: str,
+    ) -> dict[str, str]:
+        """Build an OpenAI model message."""
+        return {"role": role, "content": content}
+
+    def _build_input(
+        self,
+        user: str,
+        system: str | None = None,
+    ) -> list[dict[str, Any]]:
+        """Build the full OpenAI model input."""
+        input = []
+        if system:
+            input.append(self._build_message(role="system", content=system))
+        input.append(self._build_message(role="user", content=user))
+        return input
+
+    def _get_response(
+        self,
+        model: str,
+        input: list[dict[str, Any]],
+        system: str | None = None,
+        temperature: float | None = None,
+        top_p: float | None = None,
+        max_tokens: int | None = None,
+    ) -> tuple[str, str | None]:
+        """Generate a model response from the Qwen API."""
+        response = self._client.chat.completions.create(
+            messages=cast(list[ChatCompletionMessageParam], input),
+            model=model,
+            temperature=temperature if temperature is not None else openai.omit,
+            top_p=top_p if top_p is not None else openai.omit,
+            max_completion_tokens=max_tokens if max_tokens is not None else openai.omit,
+            # thinking mode and other dashscope extras (e.g. web search) are left
+            # off unless reasoning is explicitly requested
+            extra_body={"enable_thinking": True} if self._enable_reasoning else None,
+        )
+        message = response.choices[0].message
+
+        # chain-of-thought from thinking-enabled models; None otherwise
+        reasoning = getattr(message, "reasoning_content", None)
+
+        # cast to str as text completions always return string content
+        return cast(str, message.content), reasoning

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_codegen_research.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llm-codegen-research
-Version: 2.16
+Version: 2.17
 Summary: Useful classes and methods for researching code-generation by LLMs.
 Author-email: Lukas Twist <itsluketwist@gmail.com>
 Project-URL: Homepage, https://github.com/itsluketwist/llm-codegen-research
@@ -13,6 +13,7 @@ License-File: LICENSE
 Requires-Dist: esprima
 Provides-Extra: api
 Requires-Dist: anthropic; extra == "api"
+Requires-Dist: google-genai; extra == "api"
 Requires-Dist: mistralai; extra == "api"
 Requires-Dist: openai; extra == "api"
 Requires-Dist: together; extra == "api"

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_codegen_research.egg-info/SOURCES.txt

@@ -23,11 +23,13 @@ src/llm_cgr/llm/clients/__init__.py
 src/llm_cgr/llm/clients/anthropic.py
 src/llm_cgr/llm/clients/base.py
 src/llm_cgr/llm/clients/deepseek.py
+src/llm_cgr/llm/clients/google.py
 src/llm_cgr/llm/clients/mistral.py
 src/llm_cgr/llm/clients/nscale.py
 src/llm_cgr/llm/clients/openai.py
 src/llm_cgr/llm/clients/openai_tool.py
 src/llm_cgr/llm/clients/protocol.py
+src/llm_cgr/llm/clients/qwen.py
 src/llm_cgr/llm/clients/together.py
 src/llm_cgr/scripts/test_cuda.py
 src/llm_codegen_research.egg-info/PKG-INFO

{llm_codegen_research-2.16 → llm_codegen_research-2.17}/src/llm_codegen_research.egg-info/requires.txt

@@ -2,6 +2,7 @@ esprima
 
 [api]
 anthropic
+google-genai
 mistralai
 openai
 together