lm-deluge 0.0.52__py3-none-any.whl → 0.0.54__py3-none-any.whl

lm_deluge/__init__.py

@@ -1,10 +1,9 @@
-from .client import LLMClient, SamplingParams, APIResponse
+from .client import APIResponse, LLMClient, SamplingParams
+from .file import File
 from .prompt import Conversation, Message
 from .tool import Tool
-from .file import File
-import dotenv
 
-dotenv.load_dotenv()
+# dotenv.load_dotenv() - don't do this, fucks with other packages
 
 __all__ = [
     "LLMClient",

lm_deluge/api_requests/anthropic.py

@@ -60,7 +60,8 @@ def _build_anthropic_request(
             "type": "enabled",
             "budget_tokens": budget,
         }
-        request_json.pop("top_p")
+        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:
@@ -70,6 +71,11 @@ 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
+    if model.name in ["claude-sonnet-4-5-20250929", "claude-opus-4-1-20250805"]:
+        if "temperature" in request_json and "top_p" in request_json:
+            request_json.pop("top_p")
+
     if tools:
         mcp_servers = []
         tool_definitions = []
@@ -89,6 +95,9 @@ def _build_anthropic_request(
                     _add_beta(base_headers, "computer-use-2025-01-24")
                 elif tool["type"] == "code_execution_20250522":
                     _add_beta(base_headers, "code-execution-2025-05-22")
+                elif tool["type"] in ["memory_20250818", "clear_tool_uses_20250919"]:
+                    _add_beta(base_headers, "context-management-2025-06-27")
+
             elif isinstance(tool, MCPServer):
                 _add_beta(base_headers, "mcp-client-2025-04-04")
                 mcp_servers.append(tool.for_anthropic())

lm_deluge/client.py

@@ -80,6 +80,22 @@ class _LLMClient(BaseModel):
             self._tracker.log_final_status()
             self._tracker = None
 
+    def reset_tracker(self):
+        """Reset tracker by closing and reopening with fresh state.
+
+        Useful when reusing a client across multiple batches and you want
+        the progress bar to start from 0 instead of showing cumulative totals.
+        """
+        if self._tracker is None:
+            return
+
+        # Close existing tracker (including progress bar)
+        show_progress = self._tracker.use_progress_bar
+        self.close()
+
+        # Create fresh tracker
+        self.open(total=0, show_progress=show_progress)
+
     # NEW! Builder methods
     def with_model(self, model: str):
         self.model_names = [model]
@@ -353,146 +369,61 @@ class _LLMClient(BaseModel):
         cache: CachePattern | None = None,
         use_responses_api: bool = False,
     ) -> list[APIResponse | None] | list[str | None] | dict[str, int]:
-        # Convert prompts to Conversations - no upfront cache checking for dynamic caching!
-        prompts = prompts_to_conversations(prompts)
-        ids = list(range(len(prompts)))
-        results: list[APIResponse | None] = [None for _ in range(len(prompts))]
-        contexts: list[RequestContext | None] = [None for _ in range(len(prompts))]
-        inflight_tasks: set[asyncio.Task[None]] = set()
-        # Use existing tracker if client has been opened; otherwise open/close automatically
-        tracker: StatusTracker
-        tracker_preopened = self._tracker is not None
-        if tracker_preopened:
-            tracker = self._tracker  # type: ignore[assignment]
-            tracker.add_to_total(len(prompts))
-        else:
-            self.open(total=len(prompts), show_progress=show_progress)
-            tracker = self._tracker  # type: ignore[assignment]
-        assert tracker is not None
+        """Process multiple prompts asynchronously using the start_nowait/wait_for_all backend.
 
-        # Create retry queue for failed requests
-        retry_queue: asyncio.Queue[RequestContext] = asyncio.Queue()
-
-        # Calculate sleep time for rate limiting (legacy; gating happens in _wait_for_capacity)
-        seconds_to_sleep_each_loop = (60.0 * 0.9) / tracker.max_requests_per_minute
-
-        # Main dispatch loop - using original pattern but with all prompts
-        next_context = None  # Persist across iterations like original
-        next_is_retry = False  # Track whether next_context is a retry
-        prompts_not_finished = True
-        prompts_iter = iter(zip(ids, prompts))
-
-        while True:
-            # Get next context (retry or new) - only if we don't already have one waiting
-            if next_context is None:
-                if not retry_queue.empty():
-                    next_context = retry_queue.get_nowait()
-                    next_is_retry = True
-                    print(f"Retrying request {next_context.task_id}.")
-                elif prompts_not_finished:
-                    try:
-                        task_id, prompt = next(prompts_iter)
-                        model, sampling_params = self._select_model()
-                        assert isinstance(prompt, Conversation)
-                        next_context = RequestContext(
-                            task_id=task_id,
-                            model_name=model,
-                            prompt=prompt,
-                            sampling_params=sampling_params,
-                            attempts_left=self.max_attempts,
-                            request_timeout=self.request_timeout,
-                            status_tracker=tracker,
-                            tools=tools,
-                            cache=cache,
-                            use_responses_api=use_responses_api,
-                            extra_headers=self.extra_headers,
-                            force_local_mcp=self.force_local_mcp,
-                        )
-
-                        next_is_retry = False
-                    except StopIteration:
-                        prompts_not_finished = False
-
-            # Dispatch using shared capacity gate (consistent with start_nowait)
-            if next_context:
-                # Wait here until we have capacity to launch this context
-                await self._wait_for_capacity(
-                    next_context.num_tokens, tracker, retry=next_is_retry
-                )
-
-                # Launch simplified request processing
-                contexts[next_context.task_id] = next_context
-
-                async def process_and_store(ctx: RequestContext):
-                    try:
-                        response = await self.process_single_request(ctx, retry_queue)
-                        results[ctx.task_id] = response
-                    except Exception as e:
-                        # Create an error response for validation errors and other exceptions
-                        error_response = APIResponse(
-                            id=ctx.task_id,
-                            model_internal=ctx.model_name,
-                            prompt=ctx.prompt,
-                            sampling_params=ctx.sampling_params,
-                            status_code=None,
-                            is_error=True,
-                            error_message=str(e),
-                        )
-                        results[ctx.task_id] = error_response
-                        # Mark task as completed so the main loop can finish
-                        if ctx.status_tracker:
-                            ctx.status_tracker.task_failed(ctx.task_id)
-
-                task = asyncio.create_task(process_and_store(next_context))
-                inflight_tasks.add(task)
-                task.add_done_callback(inflight_tasks.discard)
-                next_context = None  # Reset after successful dispatch
-                next_is_retry = False
-
-            # Update progress - original logic
-            tracker.update_pbar()
-
-            # Check completion: consider final outcomes, not in-progress count
-            # This avoids rare hangs if in-progress is miscounted (e.g., double-increment).
-            if (tracker.num_tasks_succeeded + tracker.num_tasks_failed) >= len(
-                prompts
-            ) and retry_queue.empty():
-                break
+        This implementation creates all tasks upfront and waits for them to complete,
+        avoiding issues with tracker state accumulating across multiple calls.
+        """
+        # Convert prompts to Conversations
+        prompts = prompts_to_conversations(prompts)
 
-            # Yield briefly to allow in-flight tasks to progress
-            await asyncio.sleep(min(0.01, seconds_to_sleep_each_loop))
+        # Ensure tracker exists (start_nowait will call add_to_total for each task)
+        if self._tracker is None:
+            self.open(total=0, show_progress=show_progress)
+            tracker_preopened = False
+        else:
+            tracker_preopened = True
+
+        # Start all tasks using start_nowait - tasks will coordinate via shared capacity lock
+        task_ids = []
+        for prompt in prompts:
+            assert isinstance(prompt, Conversation)
+            task_id = self.start_nowait(
+                prompt,
+                tools=tools,
+                cache=cache,
+                use_responses_api=use_responses_api,
+            )
+            task_ids.append(task_id)
 
-        if inflight_tasks:
-            await asyncio.gather(*inflight_tasks, return_exceptions=True)
+        # Wait for all tasks to complete
+        results = await self.wait_for_all(task_ids)
 
+        # Close tracker if we opened it
         if not tracker_preopened:
             self.close()
 
+        # Defensive check: This should rarely happen, but provides a safety net
         for idx, response in enumerate(results):
             if response is None:
-                ctx = contexts[idx]
-                prompt = ctx.prompt if ctx else prompts[idx]
-                sampling_params = (
-                    ctx.sampling_params
-                    if ctx
-                    else self.sampling_params[0]
-                    if self.sampling_params
-                    else SamplingParams()
+                # This should only happen if there's a bug in _run_context
+                print(
+                    f"WARNING: result[{idx}] is None! Creating defensive error response. "
+                    f"Please report this bug."
                 )
-                model_name = ctx.model_name if ctx else self.model_names[0]
-                assert isinstance(
-                    prompt, Conversation
-                ), "expected prompt to be a conversation"
                 results[idx] = APIResponse(
                     id=idx,
-                    model_internal=model_name,
-                    prompt=prompt,
-                    sampling_params=sampling_params,
+                    model_internal=self.model_names[0],
+                    prompt=prompts[idx],  # type: ignore
+                    sampling_params=self.sampling_params[0]
+                    if self.sampling_params
+                    else SamplingParams(),
                     status_code=None,
                     is_error=True,
                     error_message="Internal error: no response produced.",
                 )
 
+        # Handle return format
         if return_completions_only:
             return [r.completion if r is not None else None for r in results]
 

lm_deluge/llm_tools/extract.py

@@ -1,11 +1,12 @@
 import asyncio
 import io
 import json
+import os
 from typing import Any
 
+from lm_deluge.client import _LLMClient
 from lm_deluge.file import File
 
-from ..client import LLMClient
 from ..prompt import Conversation
 from ..util.json import load_json
 
@@ -18,7 +19,7 @@ except ImportError:
 async def extract_async(
     inputs: list[str | Any],
     schema: Any,
-    client: LLMClient,
+    client: _LLMClient,
     document_name: str | None = None,
     object_name: str | None = None,
     show_progress: bool = True,
@@ -32,12 +33,13 @@ async def extract_async(
         raise ValueError("schema must be a pydantic model or a dict.")
 
     # warn if json_mode is not True
+    has_warned = os.environ.get("LM_DELUGE_WARN_JSON_MODE", False)
     for sp in client.sampling_params:
-        if sp.json_mode is False:
+        if sp.json_mode is False and not has_warned:
             print(
                 "Warning: json_mode is False for one or more sampling params. You may get invalid output."
             )
-            break
+            os.environ["LM_DELUGE_WARN_JSON_MODE"] = "True"
     # check_schema(schema_dict) -- figure out later
     if document_name is None:
         document_name = "text"
@@ -111,7 +113,7 @@ async def extract_async(
 def extract(
     inputs: list[str | Any],
     schema: Any,
-    client: LLMClient,
+    client: _LLMClient,
     document_name: str | None = None,
     object_name: str | None = None,
     show_progress: bool = True,

lm_deluge/models/anthropic.py

@@ -10,6 +10,18 @@ ANTHROPIC_MODELS = {
     #                                                                   ░███
     #                                                                   █████
     #
+    "claude-4.5-sonnet": {
+        "id": "claude-4.5-sonnet",
+        "name": "claude-sonnet-4-5-20250929",
+        "api_base": "https://api.anthropic.com/v1",
+        "api_key_env_var": "ANTHROPIC_API_KEY",
+        "supports_json": False,
+        "api_spec": "anthropic",
+        "input_cost": 3.0,
+        "output_cost": 15.0,
+        "requests_per_minute": 4_000,
+        "tokens_per_minute": 400_000,
+    },
     "claude-4.1-opus": {
         "id": "claude-4.1-opus",
         "name": "claude-opus-4-1-20250805",

{lm_deluge-0.0.52.dist-info → lm_deluge-0.0.54.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lm_deluge
-Version: 0.0.52
+Version: 0.0.54
 Summary: Python utility for using LLM API models.
 Author-email: Benjamin Anderson <ben@trytaylor.ai>
 Requires-Python: >=3.10

{lm_deluge-0.0.52.dist-info → lm_deluge-0.0.54.dist-info}/RECORD

@@ -1,9 +1,9 @@
-lm_deluge/__init__.py,sha256=mAztMuxINmh7dGbYnT8tsmw1eryQAvd0jpY8yHzd0EE,315
+lm_deluge/__init__.py,sha256=D01sxqvAuW1QPYQfdSOlBFVhf7QUr78fGgCNPvKXbAc,346
 lm_deluge/agent.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lm_deluge/batches.py,sha256=rQocJLyIs3Ko_nRdAE9jT__5cKWYxiIRAH_Lw3L0E1k,24653
 lm_deluge/cache.py,sha256=xO2AIYvP3tUpTMKQjwQQYfGRJSRi6e7sMlRhLjsS-u4,4873
 lm_deluge/cli.py,sha256=Ilww5gOw3J5v0NReq_Ra4hhxU4BCIJBl1oTGxJZKedc,12065
-lm_deluge/client.py,sha256=WIz1M6PLZi08Y3SFhGS3Pxs1fP8P0nKSLMYzFUdNnOQ,35814
+lm_deluge/client.py,sha256=iEyqSgYt_BZ-46nHbAAnqDFcTpZNDmrJV_oLUqixwNo,32018
 lm_deluge/config.py,sha256=H1tQyJDNHGFuwxqQNL5Z-CjWAC0luHSBA3iY_pxmACM,932
 lm_deluge/embed.py,sha256=CO-TOlC5kOTAM8lcnicoG4u4K664vCBwHF1vHa-nAGg,13382
 lm_deluge/errors.py,sha256=oHjt7YnxWbh-eXMScIzov4NvpJMo0-2r5J6Wh5DQ1tk,209
@@ -17,7 +17,7 @@ lm_deluge/tool.py,sha256=3weKo09E_srEKwHlz2WMVhk2BuDr5pJpi1UP0-qlcmo,16210
 lm_deluge/tracker.py,sha256=EHFPsS94NmsON2u97rSE70q1t6pwCsixUmGV-kIphMs,11531
 lm_deluge/usage.py,sha256=VMEKghePFIID5JFBObqYxFpgYxnbYm_dnHy7V1-_T6M,4866
 lm_deluge/api_requests/__init__.py,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
-lm_deluge/api_requests/anthropic.py,sha256=8MledxnN0S-H_fZRq8DGUokcjZPQ154mr8tPWAussJ8,7992
+lm_deluge/api_requests/anthropic.py,sha256=Iy-AMo1o7xliwWhamFIUc601PZ0YWLhwAgszgfws42I,8467
 lm_deluge/api_requests/base.py,sha256=EVHNFtlttKbN7Tt1MnLaO-NjvKHPSV5CqlRv-OnpVAE,5593
 lm_deluge/api_requests/bedrock.py,sha256=GmVxXz3ERAeQ7e52Nlztt81O4H9eJOQeOnS6b65vjm4,15453
 lm_deluge/api_requests/common.py,sha256=BZ3vRO5TB669_UsNKugkkuFSzoLHOYJIKt4nV4sf4vc,422
@@ -38,13 +38,13 @@ lm_deluge/built_in_tools/anthropic/computer_use.py,sha256=47DEQpj8HBSa-_TImW-5JC
 lm_deluge/built_in_tools/anthropic/editor.py,sha256=DyC_DrHVTm1khU9QDL39vBuhu4tO5mS5H7xMRIT0Ng4,23327
 lm_deluge/llm_tools/__init__.py,sha256=TbZTETq9i_9yYskFWQKOG4pGh5ZiyE_D-h3RArfhGp4,231
 lm_deluge/llm_tools/classify.py,sha256=OdMwV5u4XoPlVhjOHX0sng5KPBIKFJmQeOE2fmnPgLU,21
-lm_deluge/llm_tools/extract.py,sha256=C3drVAMaoFx5jNE38Xi5cXxrqboyoZ9cE7nX5ylWbXw,4482
+lm_deluge/llm_tools/extract.py,sha256=p61JW8yv5gQxPp4P8Hkm90ERgfD_Ek5IABzjIIlX-M0,4631
 lm_deluge/llm_tools/locate.py,sha256=lYNbKTmy9dTvj0lEQkOQ7yrxyqsgYzjD0C_byJKI_4w,6271
 lm_deluge/llm_tools/ocr.py,sha256=7fDlvs6uUOvbxMasvGGNJx5Fj6biM6z3lijKZaGN26k,23
 lm_deluge/llm_tools/score.py,sha256=9oGA3-k2U5buHQXkXaEI9M4Wb5yysNhTLsPbGeghAlQ,2580
 lm_deluge/llm_tools/translate.py,sha256=iXyYvQZ8bC44FWhBk4qpdqjKM1WFF7Shq-H2PxhPgg4,1452
 lm_deluge/models/__init__.py,sha256=Dh2CuTZeCAddIIXwWJXOjM10B0CpKqjTdMXWYuBP0s8,4289
-lm_deluge/models/anthropic.py,sha256=3pW7fyBY9Xh1m1RtfncU9amWTtKnjGZD0STjpu8iUSQ,5700
+lm_deluge/models/anthropic.py,sha256=zTOjwNh00OYPZMFiaTniZzTtwL1Vmnlm750GhVaczqk,6112
 lm_deluge/models/bedrock.py,sha256=PIaXvho2agCm1hSSAEy8zHCITjApXT2eUOGDKW425tE,5424
 lm_deluge/models/cerebras.py,sha256=u2FMXJF6xMr0euDRKLKMo_NVTOcvSrrEpehbHr8sSeE,2050
 lm_deluge/models/cohere.py,sha256=M_7cVA9QD4qe1X4sZXCpKEkKrKz2jibaspiTnzsZ1GU,3998
@@ -66,8 +66,8 @@ lm_deluge/util/logprobs.py,sha256=UkBZakOxWluaLqHrjARu7xnJ0uCHVfLGHJdnYlEcutk,11
 lm_deluge/util/spatial.py,sha256=BsF_UKhE-x0xBirc-bV1xSKZRTUhsOBdGqsMKme20C8,4099
 lm_deluge/util/validation.py,sha256=hz5dDb3ebvZrZhnaWxOxbNSVMI6nmaOODBkk0htAUhs,1575
 lm_deluge/util/xml.py,sha256=Ft4zajoYBJR3HHCt2oHwGfymGLdvp_gegVmJ-Wqk4Ck,10547
-lm_deluge-0.0.52.dist-info/licenses/LICENSE,sha256=uNNXGXPCw2TC7CUs7SEBkA-Mz6QBQFWUUEWDMgEs1dU,1058
-lm_deluge-0.0.52.dist-info/METADATA,sha256=X1JJBjExVA0NNXSaoB2NkOpT9f660AFe9u58BmKdb2w,13443
-lm_deluge-0.0.52.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lm_deluge-0.0.52.dist-info/top_level.txt,sha256=hqU-TJX93yBwpgkDtYcXyLr3t7TLSCCZ_reytJjwBaE,10
-lm_deluge-0.0.52.dist-info/RECORD,,
+lm_deluge-0.0.54.dist-info/licenses/LICENSE,sha256=uNNXGXPCw2TC7CUs7SEBkA-Mz6QBQFWUUEWDMgEs1dU,1058
+lm_deluge-0.0.54.dist-info/METADATA,sha256=5s1jfrJkLOXWWU3P4S_9Eketc8o3lJv137x4ZnyV1CA,13443
+lm_deluge-0.0.54.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lm_deluge-0.0.54.dist-info/top_level.txt,sha256=hqU-TJX93yBwpgkDtYcXyLr3t7TLSCCZ_reytJjwBaE,10
+lm_deluge-0.0.54.dist-info/RECORD,,