lm-deluge 0.0.89__py3-none-any.whl → 0.0.91__py3-none-any.whl

lm_deluge/__init__.py

@@ -1,20 +1,6 @@
 from .client import APIResponse, LLMClient, SamplingParams
-from .file import File
-from .prompt import Conversation, Message
-from .tool import Tool
-
-try:
-    from .mock_openai import (  # noqa
-        APIError,
-        APITimeoutError,
-        BadRequestError,
-        MockAsyncOpenAI,
-        RateLimitError,
-    )
-
-    _has_openai = True
-except ImportError:
-    _has_openai = False
+from .prompt import Conversation, Message, File
+from .tool import Tool, MCPServer
 
 # dotenv.load_dotenv() - don't do this, fucks with other packages
 
@@ -25,16 +11,6 @@ __all__ = [
     "Conversation",
     "Message",
     "Tool",
+    "MCPServer",
     "File",
 ]
-
-if _has_openai:
-    __all__.extend(
-        [
-            "MockAsyncOpenAI",
-            "APIError",
-            "APITimeoutError",
-            "BadRequestError",
-            "RateLimitError",
-        ]
-    )

lm_deluge/api_requests/anthropic.py

@@ -6,10 +6,11 @@ from aiohttp import ClientResponse
 from lm_deluge.prompt import (
     Message,
     Text,
+    ThoughtSignature,
     Thinking,
     ToolCall,
 )
-from lm_deluge.request_context import RequestContext
+from lm_deluge.api_requests.context import RequestContext
 from lm_deluge.tool import MCPServer, Tool
 from lm_deluge.usage import Usage
 from lm_deluge.util.schema import (
@@ -102,7 +103,9 @@ def _build_anthropic_request(
             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
+            max_tokens = request_json["max_tokens"]
+            assert isinstance(max_tokens, int)
+            request_json["max_tokens"] = max_tokens + budget
         else:
             request_json["thinking"] = {"type": "disabled"}
             if "kimi" in model.id and "thinking" in model.id:
@@ -250,8 +253,28 @@ class AnthropicRequest(APIRequestBase):
                     if item["type"] == "text":
                         parts.append(Text(item["text"]))
                     elif item["type"] == "thinking":
-                        thinking = item["thinking"]
-                        parts.append(Thinking(item["thinking"]))
+                        thinking_content = item.get("thinking", "")
+                        thinking = thinking_content
+                        signature = item.get("signature")
+                        parts.append(
+                            Thinking(
+                                thinking_content,
+                                raw_payload=item,
+                                thought_signature=ThoughtSignature(
+                                    signature,
+                                    provider="anthropic",
+                                )
+                                if signature is not None
+                                else None,
+                            )
+                        )
+                    elif item["type"] == "redacted_thinking":
+                        parts.append(
+                            Thinking(
+                                item.get("data", ""),
+                                raw_payload=item,
+                            )
+                        )
                     elif item["type"] == "tool_use":
                         parts.append(
                             ToolCall(
@@ -265,9 +288,8 @@ class AnthropicRequest(APIRequestBase):
                 usage = Usage.from_anthropic_usage(data["usage"])
             except Exception as e:
                 is_error = True
-                error_message = (
-                    f"Error calling .json() on response w/ status {status_code}: {e}"
-                )
+                response_text = await http_response.text()
+                error_message = f"Error calling .json() on response w/ status {status_code}: {e}. Response: {response_text[:500]}"
         elif mimetype and "json" in mimetype.lower():
             is_error = True  # expected status is 200, otherwise it's an error
             data = await http_response.json()

lm_deluge/api_requests/base.py

@@ -1,4 +1,6 @@
 import asyncio
+import json
+import os
 import time
 import traceback
 from abc import ABC, abstractmethod
@@ -8,7 +10,7 @@ from aiohttp import ClientResponse
 
 from ..errors import raise_if_modal_exception
 from ..models.openai import OPENAI_MODELS
-from ..request_context import RequestContext
+from ..api_requests.context import RequestContext
 from .response import APIResponse
 
 
@@ -73,6 +75,24 @@ class APIRequestBase(ABC):
 
         # Start with base headers, then overlay filtered extra headers (extra takes precedence)
         merged = dict(base_headers)
+        if "anthropic-beta" in merged and "anthropic-beta" in filtered_extra:
+            combined = []
+            seen = set()
+            for (
+                raw
+            ) in f"{merged['anthropic-beta']},{filtered_extra['anthropic-beta']}".split(
+                ","
+            ):
+                token = raw.strip()
+                if token and token not in seen:
+                    seen.add(token)
+                    combined.append(token)
+            merged["anthropic-beta"] = ",".join(combined)
+            filtered_extra = {
+                key: value
+                for key, value in filtered_extra.items()
+                if key != "anthropic-beta"
+            }
         merged.update(filtered_extra)
 
         # Filter out None values from final merged headers
@@ -189,6 +209,23 @@ class APIRequestBase(ABC):
         await self.build_request()
         assert self.context.status_tracker
 
+        if os.getenv("DELUGE_PROXY_LOG_PROVIDER_REQUESTS", "").strip().lower() in {
+            "1",
+            "true",
+            "yes",
+            "on",
+        }:
+            print("DELUGE_PROXY_PROVIDER_REQUEST")
+            print(f"URL: {self.url}")
+            print("Headers:")
+            print(self.request_header)
+            if self.request_json is not None:
+                print("JSON:")
+                try:
+                    print(json.dumps(self.request_json, indent=2))
+                except Exception:
+                    print(self.request_json)
+
         if (
             self.context.background
             and self.context.use_responses_api

lm_deluge/api_requests/bedrock.py

@@ -16,10 +16,11 @@ except ImportError:
 from lm_deluge.prompt import (
     Message,
     Text,
+    ThoughtSignature,
     Thinking,
     ToolCall,
 )
-from lm_deluge.request_context import RequestContext
+from lm_deluge.api_requests.context import RequestContext
 from lm_deluge.tool import MCPServer, Tool
 from lm_deluge.usage import Usage
 
@@ -262,6 +263,11 @@ class BedrockRequest(APIRequestBase):
         # Create a fake requests.PreparedRequest object for AWS4Auth to sign
         import requests
 
+        assert self.url is not None, "URL must be set after build_request"
+        assert (
+            self.request_header is not None
+        ), "Headers must be set after build_request"
+
         fake_request = requests.Request(
             method="POST",
             url=self.url,
@@ -363,8 +369,28 @@ class BedrockRequest(APIRequestBase):
                         if item["type"] == "text":
                             parts.append(Text(item["text"]))
                         elif item["type"] == "thinking":
-                            thinking = item["thinking"]
-                            parts.append(Thinking(item["thinking"]))
+                            thinking_content = item.get("thinking", "")
+                            thinking = thinking_content
+                            signature = item.get("signature")
+                            parts.append(
+                                Thinking(
+                                    thinking_content,
+                                    raw_payload=item,
+                                    thought_signature=ThoughtSignature(
+                                        signature,
+                                        provider="anthropic",
+                                    )
+                                    if signature is not None
+                                    else None,
+                                )
+                            )
+                        elif item["type"] == "redacted_thinking":
+                            parts.append(
+                                Thinking(
+                                    item.get("data", ""),
+                                    raw_payload=item,
+                                )
+                            )
                         elif item["type"] == "tool_use":
                             parts.append(
                                 ToolCall(

lm_deluge/{request_context.py → api_requests/context.py}

@@ -2,9 +2,9 @@ from dataclasses import dataclass, field
 from functools import cached_property
 from typing import Any, Callable, Sequence, TYPE_CHECKING
 
-from .config import SamplingParams
-from .prompt import CachePattern, Conversation
-from .tracker import StatusTracker
+from ..config import SamplingParams
+from ..prompt import CachePattern, Conversation
+from ..tracker import StatusTracker
 
 if TYPE_CHECKING:
     from pydantic import BaseModel
@@ -83,4 +83,4 @@ class RequestContext:
         # Update with any overrides
         current_values.update(overrides)
 
-        return RequestContext(**current_values)
+        return RequestContext(**current_values)  # type: ignore[arg-type]

lm_deluge/api_requests/gemini.py

@@ -1,15 +1,16 @@
 import json
 import os
+from typing import Any
 
 from aiohttp import ClientResponse
 
-from lm_deluge.request_context import RequestContext
+from lm_deluge.api_requests.context import RequestContext
 from lm_deluge.tool import Tool
 from lm_deluge.warnings import maybe_warn
 
 from ..config import SamplingParams
 from ..models import APIModel
-from ..prompt import Conversation, Message, Text, Thinking, ToolCall
+from ..prompt import Conversation, Message, Text, ThoughtSignature, Thinking, ToolCall
 from ..usage import Usage
 from .base import APIRequestBase, APIResponse
 
@@ -37,13 +38,14 @@ async def _build_gemini_request(
                             part_type="function call",
                         )
 
-    request_json = {
+    generation_config: dict[str, Any] = {
+        "temperature": sampling_params.temperature,
+        "topP": sampling_params.top_p,
+        "maxOutputTokens": sampling_params.max_new_tokens,
+    }
+    request_json: dict[str, Any] = {
         "contents": messages,
-        "generationConfig": {
-            "temperature": sampling_params.temperature,
-            "topP": sampling_params.top_p,
-            "maxOutputTokens": sampling_params.max_new_tokens,
-        },
+        "generationConfig": generation_config,
     }
 
     # Add system instruction if present
@@ -83,7 +85,7 @@ async def _build_gemini_request(
                 }
             effort = level_map[effort_key]
         thinking_config = {"thinkingLevel": effort}
-        request_json["generationConfig"]["thinkingConfig"] = thinking_config
+        generation_config["thinkingConfig"] = thinking_config
 
     elif model.reasoning_model:
         if (
@@ -126,7 +128,7 @@ async def _build_gemini_request(
             # no thoughts head empty
             thinking_config = {"includeThoughts": False, "thinkingBudget": 0}
 
-        request_json["generationConfig"]["thinkingConfig"] = thinking_config
+        generation_config["thinkingConfig"] = thinking_config
 
     else:
         if sampling_params.reasoning_effort:
@@ -171,14 +173,14 @@ async def _build_gemini_request(
 
     # Handle JSON mode
     if sampling_params.json_mode and model.supports_json:
-        request_json["generationConfig"]["responseMimeType"] = "application/json"
+        generation_config["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"] = {
+            generation_config["mediaResolution"] = {
                 "level": sampling_params.media_resolution
             }
         else:
@@ -260,10 +262,20 @@ class GeminiRequest(APIRequestBase):
                         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")
+                                raw_sig = part.get("thoughtSignature")
+                                thought_sig = (
+                                    ThoughtSignature(raw_sig, provider="gemini")
+                                    if raw_sig is not None
+                                    else None
+                                )
 
                                 if "text" in part:
-                                    parts.append(Text(part["text"]))
+                                    parts.append(
+                                        Text(
+                                            part["text"],
+                                            thought_signature=thought_sig,
+                                        )
+                                    )
                                 elif "thought" in part:
                                     # Thought with optional signature
                                     parts.append(
@@ -286,6 +298,10 @@ class GeminiRequest(APIRequestBase):
                                             thought_signature=thought_sig,
                                         )
                                     )
+                                elif thought_sig:
+                                    parts.append(
+                                        Text("", thought_signature=thought_sig)
+                                    )
 
                     content = Message("assistant", parts)
 

lm_deluge/api_requests/mistral.py

@@ -7,7 +7,7 @@ from lm_deluge.warnings import maybe_warn
 
 from ..models import APIModel
 from ..prompt import Message
-from ..request_context import RequestContext
+from ..api_requests.context import RequestContext
 from ..usage import Usage
 from .base import APIRequestBase, APIResponse
 

lm_deluge/api_requests/openai.py

@@ -7,7 +7,7 @@ from typing import Sequence
 import aiohttp
 from aiohttp import ClientResponse
 
-from lm_deluge.request_context import RequestContext
+from lm_deluge.api_requests.context import RequestContext
 from lm_deluge.tool import MCPServer, Tool
 from lm_deluge.util.schema import (
     prepare_output_schema,
@@ -22,6 +22,24 @@ from ..usage import Usage
 from .base import APIRequestBase, APIResponse
 
 
+def _message_contents_to_string(messages: list[dict]):
+    messages = messages.copy()
+
+    for msg in messages:
+        content = msg.get("content")
+        assert content
+        if isinstance(content, list):
+            new_content = ""
+            for part in content:
+                assert "text" in part, "Invalid text part: " + str(part)
+                new_content += part["text"]
+                new_content += "\n"
+
+            msg["content"] = new_content.strip()
+
+    return messages
+
+
 async def _build_oa_chat_request(
     model: APIModel,
     context: RequestContext,
@@ -55,6 +73,12 @@ async def _build_oa_chat_request(
                 request_json["service_tier"] = context.service_tier
         else:
             request_json["service_tier"] = context.service_tier
+    # if tinker, for now hack to mush into 1 string
+    if "tinker" in model.name:
+        messages = request_json["messages"]
+        assert isinstance(messages, list)
+        request_json["messages"] = _message_contents_to_string(messages)
+
     # set max_tokens or max_completion_tokens dep. on provider
     if "cohere" in model.api_base:
         request_json["max_tokens"] = sampling_params.max_new_tokens
@@ -217,7 +241,7 @@ class OpenAIRequest(APIRequestBase):
                         parts.append(Text(message["content"]))
 
                     # Add tool calls if present
-                    if "tool_calls" in message:
+                    if "tool_calls" in message and message["tool_calls"] is not None:
                         for tool_call in message["tool_calls"]:
                             parts.append(
                                 ToolCall(
@@ -238,9 +262,9 @@ class OpenAIRequest(APIRequestBase):
                         and "logprobs" in data["choices"][0]
                     ):
                         logprobs = data["choices"][0]["logprobs"]["content"]
-                except Exception:
+                except Exception as e:
                     is_error = True
-                    error_message = f"Error getting 'choices' and 'usage' from {self.model.name} response."
+                    error_message = f"Error getting 'choices' and 'usage' from {self.model.name} response: {data}. Error: {e}"
         elif mimetype and "json" in mimetype.lower():
             is_error = True  # expected status is 200, otherwise it's an error
             data = await http_response.json()
@@ -655,7 +679,12 @@ async def stream_chat(
         request_header.update(filtered_extra)
 
     context = SimpleNamespace(
-        prompt=prompt, tools=tools, sampling_params=sampling_params
+        prompt=prompt,
+        tools=tools,
+        sampling_params=sampling_params,
+        service_tier=None,
+        output_schema=None,
+        model_name=model_name,
     )
 
     request_json = await _build_oa_chat_request(model, context)  # type: ignore

lm_deluge/batches.py

@@ -3,7 +3,7 @@ import json
 import os
 import tempfile
 import time
-from typing import Literal, Sequence, cast
+from typing import Any, Literal, Sequence, cast
 
 import aiohttp
 from rich.console import Console
@@ -22,7 +22,7 @@ from lm_deluge.prompt import (
     Prompt,
     prompts_to_conversations,
 )
-from lm_deluge.request_context import RequestContext
+from lm_deluge.api_requests.context import RequestContext
 
 
 def _create_batch_status_display(
@@ -141,31 +141,22 @@ async def submit_batch_oa(file_path: str):
         return batch_id
 
 
-async def _submit_anthropic_batch(file_path: str, headers: dict, model: str):
-    """Upload a JSONL file and create one Anthropic batch."""
+async def _submit_anthropic_batch(requests: list[dict], headers: dict, model: str):
+    """Submit batch requests to Anthropic's Message Batches API."""
 
     async with aiohttp.ClientSession() as session:
         url = f"{registry[model].api_base}/messages/batches"
-        data = aiohttp.FormData()
-        with open(file_path, "rb") as f:
-            data.add_field(
-                "file",
-                f,
-                filename=os.path.basename(file_path),
-                content_type="application/json",
-            )
-
-            async with session.post(url, data=data, headers=headers) as response:
-                if response.status != 200:
-                    text = await response.text()
-                    raise ValueError(f"Error creating batch: {text}")
+        payload = {"requests": requests}
 
-                batch_data = await response.json()
-                batch_id = batch_data["id"]
-                print(f"Anthropic batch job started successfully: id = {batch_id}")
+        async with session.post(url, json=payload, headers=headers) as response:
+            if response.status != 200:
+                text = await response.text()
+                raise ValueError(f"Error creating batch: {text}")
 
-        os.remove(file_path)
-        return batch_id
+            batch_data = await response.json()
+            batch_id = batch_data["id"]
+            print(f"Anthropic batch job started successfully: id = {batch_id}")
+            return batch_id
 
 
 async def create_batch_files_oa(
@@ -409,20 +400,10 @@ async def submit_batches_anthropic(
 
         if current_batch and (would_exceed_size or would_exceed_items):
             # Submit current batch
-            def write_batch_file():
-                with tempfile.NamedTemporaryFile(
-                    mode="w+", suffix=".jsonl", delete=False
-                ) as f:
-                    for batch_request in current_batch:
-                        json.dump(batch_request, f)
-                        f.write("\n")
-                    print("wrote", len(current_batch), "items")
-                    return f.name
-
-            file_path = await asyncio.to_thread(write_batch_file)
+            print("wrote", len(current_batch), "items")
             batch_tasks.append(
                 asyncio.create_task(
-                    _submit_anthropic_batch(file_path, request_headers, model)  # type: ignore
+                    _submit_anthropic_batch(current_batch, request_headers, model)  # type: ignore
                 )
             )
 
@@ -436,21 +417,10 @@ async def submit_batches_anthropic(
 
     # Submit final batch if it has items
     if current_batch:
-
-        def write_final_batch_file():
-            with tempfile.NamedTemporaryFile(
-                mode="w+", suffix=".jsonl", delete=False
-            ) as f:
-                for batch_request in current_batch:
-                    json.dump(batch_request, f)
-                    f.write("\n")
-                print("wrote", len(current_batch), "items")
-                return f.name
-
-        file_path = await asyncio.to_thread(write_final_batch_file)
+        print("wrote", len(current_batch), "items")
         batch_tasks.append(
             asyncio.create_task(
-                _submit_anthropic_batch(file_path, request_headers, model)  # type: ignore
+                _submit_anthropic_batch(current_batch, request_headers, model)  # type: ignore
             )
         )
 
@@ -510,7 +480,7 @@ async def _wait_for_anthropic_batch_completion_async(
 
     # Event to signal when to stop the display updater
     stop_display_event = asyncio.Event()
-    current_status = {"status": "processing", "counts": None}
+    current_status: dict[str, Any] = {"status": "processing", "counts": None}
 
     async def display_updater():
         """Update display independently of polling."""
@@ -662,7 +632,7 @@ async def _wait_for_openai_batch_completion_async(
 
     # Event to signal when to stop the display updater
     stop_display_event = asyncio.Event()
-    current_status = {"status": "pending", "counts": None}
+    current_status: dict[str, Any] = {"status": "pending", "counts": None}
 
     async def display_updater():
         """Update display independently of polling."""

lm_deluge/cache.py

@@ -8,7 +8,7 @@ from .api_requests.base import APIResponse
 try:
     import plyvel  # type: ignore
 except ImportError:
-    plyvel = None
+    plyvel: Any = None
     print("Warning: plyvel not installed, cannot use LevelDB.")