speedy-utils 1.1.45__py3-none-any.whl → 1.1.47__py3-none-any.whl

llm_utils/lm/llm.py

@@ -4,11 +4,9 @@
 Simplified LLM Task module for handling language model interactions with structured input/output.
 """
 
-import os
 import subprocess
-from typing import Any, Dict, List, Optional, Type, Union, cast
+from typing import Any, Dict, List, Optional, cast
 
-import requests
 from httpx import Timeout
 from loguru import logger
 from openai import AuthenticationError, BadRequestError, OpenAI, RateLimitError
@@ -28,16 +26,7 @@ from .mixins import (
 )
 from .utils import (
     _extract_port_from_vllm_cmd,
-    _get_port_from_client,
-    _is_lora_path,
-    _is_server_running,
-    _kill_vllm_on_port,
-    _load_lora_adapter,
-    _start_vllm_server,
-    _unload_lora_adapter,
     get_base_client,
-    kill_all_vllm_processes,
-    stop_vllm_process,
 )
 
 
@@ -216,6 +205,46 @@ class LLM(
             results.append(result_dict)
         return results
 
+    @staticmethod
+    def _strip_think_tags(text: str) -> str:
+        """Remove <think> tags if present, returning only the reasoning body."""
+        cleaned = text.strip()
+        if cleaned.startswith('<think>'):
+            cleaned = cleaned[len('<think>') :].lstrip()
+        if '</think>' in cleaned:
+            cleaned = cleaned.split('</think>', 1)[0].rstrip()
+        return cleaned
+
+    def generate_with_think_prefix(
+        self, input_data: str | BaseModel | list[dict], **runtime_kwargs
+    ) -> list[dict[str, Any]]:
+        """
+        Generate text and format output as:
+        <think>reasoning</think>
+        """
+        results = self.text_completion(input_data, **runtime_kwargs)
+
+        for result in results:
+            content = result.get('parsed') or ''
+            reasoning = result.get('reasoning_content') or ''
+
+            if not reasoning and str(content).lstrip().startswith('<think>'):
+                formatted = str(content)
+            else:
+                reasoning_body = self._strip_think_tags(str(reasoning))
+                formatted = (
+                    f'<think>\n{reasoning_body}\n</think>\n\n{str(content).lstrip()}'
+                )
+
+            result['parsed'] = formatted
+            messages = result.get('messages')
+            if isinstance(messages, list) and messages:
+                last_msg = messages[-1]
+                if isinstance(last_msg, dict) and last_msg.get('role') == 'assistant':
+                    last_msg['content'] = formatted
+
+        return results
+
     @clean_traceback
     def pydantic_parse(
         self,

speedy_utils/__init__.py

@@ -65,6 +65,8 @@ from .common.utils_error import clean_traceback, handle_exceptions_with_clean_tr
 from .multi_worker.process import multi_process
 from .multi_worker.thread import kill_all_thread, multi_thread
 from .multi_worker.dataset_ray import multi_process_dataset_ray, WorkerResources
+from .multi_worker.dataset_sharding import multi_process_dataset
+from .multi_worker.progress import report_progress
 
 
 __all__ = [
@@ -167,7 +169,9 @@ __all__ = [
     'multi_thread',
     'kill_all_thread',
     'multi_process_dataset_ray',
+    'multi_process_dataset',
     'WorkerResources',
+    'report_progress',
     # Notebook utilities
     'change_dir',
 ]

speedy_utils/multi_worker/__init__.py

@@ -1,6 +1,8 @@
 from .process import multi_process, cleanup_phantom_workers, create_progress_tracker
 from .thread import multi_thread
 from .dataset_ray import multi_process_dataset_ray, WorkerResources
+from .dataset_sharding import multi_process_dataset
+from .progress import report_progress
 
 __all__ = [
     'multi_process',
@@ -9,4 +11,6 @@ __all__ = [
     'create_progress_tracker',
     'multi_process_dataset_ray',
     'WorkerResources',
+    'multi_process_dataset',
+    'report_progress',
 ]

speedy_utils/multi_worker/_multi_process.py

@@ -0,0 +1,425 @@
+"""
+Multi-process implementation with sequential and threadpool backends.
+
+Provides the public `multi_process` dispatcher and implementations for:
+- 'seq': Sequential execution
+- 'mp': ThreadPoolExecutor-based parallelism
+- 'safe': Same as 'mp' but without process tracking (for tests)
+"""
+from __future__ import annotations
+
+import concurrent.futures
+import inspect
+import os
+import sys
+import warnings
+from typing import Any, Callable, Iterable, Literal
+
+from tqdm import tqdm
+
+from .common import (
+    ErrorHandlerType,
+    ErrorStats,
+    _build_cache_dir,
+    _call_with_log_control,
+    _cleanup_log_gate,
+    _exit_on_worker_error,
+    _prune_dead_processes,
+    _track_multiprocessing_processes,
+    create_log_gate_path,
+    wrap_dump,
+)
+
+# Import thread tracking functions if available
+try:
+    from .thread import _prune_dead_threads, _track_executor_threads
+except ImportError:
+    _prune_dead_threads = None  # type: ignore[assignment]
+    _track_executor_threads = None  # type: ignore[assignment]
+
+
+def multi_process(
+    func: Callable[[Any], Any],
+    items: Iterable[Any] | None = None,
+    *,
+    inputs: Iterable[Any] | None = None,
+    workers: int | None = None,
+    lazy_output: bool = False,
+    progress: bool = True,
+    backend: Literal['seq', 'ray', 'mp', 'safe'] = 'mp',
+    desc: str | None = None,
+    shared_kwargs: list[str] | None = None,
+    dump_in_thread: bool = True,
+    ray_metrics_port: int | None = None,
+    log_worker: Literal['zero', 'first', 'all'] = 'first',
+    total_items: int | None = None,
+    poll_interval: float = 0.3,
+    error_handler: ErrorHandlerType = 'log',
+    max_error_files: int = 100,
+    process_update_interval: int | None = None,
+    batch: int | None = None,
+    ordered: bool = True,
+    stop_on_error: bool = True,
+    **func_kwargs: Any,
+) -> list[Any]:
+    """
+    Multi-process map with selectable backend.
+
+    backend:
+        - "seq": run sequentially
+        - "ray": run in parallel with Ray
+        - "mp": run in parallel with thread pool (uses ThreadPoolExecutor)
+        - "safe": run in parallel with thread pool (explicitly safe for tests)
+
+    shared_kwargs:
+        - Optional list of kwarg names that should be shared via Ray's
+          zero-copy object store
+        - Only works with Ray backend
+        - Useful for large objects (e.g., models, datasets)
+        - Example: shared_kwargs=['model', 'tokenizer']
+
+    dump_in_thread:
+        - Whether to dump results to disk in a separate thread (default: True)
+        - If False, dumping is done synchronously
+
+    ray_metrics_port:
+        - Optional port for Ray metrics export (Ray backend only)
+
+    log_worker:
+        - Control worker stdout/stderr noise
+        - 'first': only first worker emits logs (default)
+        - 'all': allow worker prints
+        - 'zero': silence all worker output
+
+    total_items:
+        - Optional item-level total for progress tracking (Ray backend only)
+
+    poll_interval:
+        - Poll interval in seconds for progress actor updates (Ray only)
+
+    error_handler:
+        - 'raise': raise exception on first error
+        - 'ignore': continue processing, return None for failed items
+        - 'log': same as ignore, but logs errors to files (default)
+        - Note: for 'mp' and 'ray' backends, 'raise' prints a formatted
+          traceback and exits the process.
+
+    max_error_files:
+        - Maximum number of error log files to write (default: 100)
+        - Error logs are written to .cache/speedy_utils/error_logs/{idx}.log
+        - First error is always printed to screen with the log file path
+
+    process_update_interval:
+        - Legacy parameter, accepted for backward compatibility but not implemented
+
+    batch:
+        - Legacy parameter, accepted for backward compatibility but not implemented
+
+    ordered:
+        - Whether to maintain order of results (default: True)
+        - Legacy parameter, accepted for backward compatibility but not implemented
+
+    stop_on_error:
+        - Whether to stop on first error (default: True)
+        - Legacy parameter, accepted for backward compatibility
+        - Use error_handler parameter instead for error handling control
+
+    If lazy_output=True, every result is saved to .pkl and
+    the returned list contains file paths.
+    """
+
+    # default backend selection
+    if backend is None:
+        try:
+            import ray as _ray_module
+            backend = 'ray'
+        except ImportError:
+            backend = 'mp'
+
+    # Validate shared_kwargs
+    if shared_kwargs:
+        sig = inspect.signature(func)
+        valid_params = set(sig.parameters.keys())
+
+        for kw in shared_kwargs:
+            if kw not in func_kwargs:
+                raise ValueError(
+                    f"shared_kwargs key '{kw}' not found in "
+                    f"provided func_kwargs"
+                )
+            has_var_keyword = any(
+                p.kind == inspect.Parameter.VAR_KEYWORD
+                for p in sig.parameters.values()
+            )
+            if kw not in valid_params and not has_var_keyword:
+                raise ValueError(
+                    f"shared_kwargs key '{kw}' is not a valid parameter "
+                    f"for function '{func.__name__}'. "
+                    f"Valid parameters: {valid_params}"
+                )
+
+    # Prefer Ray backend when shared kwargs are requested
+    if shared_kwargs and backend != 'ray':
+        warnings.warn(
+            "shared_kwargs only supported with 'ray' backend, "
+            "switching backend to 'ray'",
+            UserWarning,
+        )
+        backend = 'ray'
+
+    # unify items and coerce to concrete list
+    if items is None and inputs is not None:
+        items = list(inputs)
+    if items is not None and not isinstance(items, list):
+        items = list(items)
+    if items is None:
+        raise ValueError("'items' or 'inputs' must be provided")
+
+    if workers is None and backend != 'ray':
+        workers = os.cpu_count() or 1
+
+    # build cache dir + wrap func
+    cache_dir = _build_cache_dir(func, items) if lazy_output else None
+    f_wrapped = wrap_dump(func, cache_dir, dump_in_thread)
+
+    log_gate_path = create_log_gate_path(log_worker)
+
+    total = len(items)
+    if desc:
+        desc = desc.strip() + f'[{backend}]'
+    else:
+        desc = f'Multi-process [{backend}]'
+
+    # Initialize error stats for error handling
+    func_name = getattr(func, '__name__', repr(func))
+    error_stats = ErrorStats(
+        func_name=func_name,
+        max_error_files=max_error_files,
+        write_logs=error_handler == 'log'
+    )
+
+    def _update_pbar_postfix(pbar: tqdm) -> None:
+        """Update pbar with success/error counts."""
+        postfix = error_stats.get_postfix_dict()
+        pbar.set_postfix(postfix)
+
+    # ---- sequential backend ----
+    if backend == 'seq':
+        return _run_seq_backend(
+            f_wrapped=f_wrapped,
+            items=items,
+            total=total,
+            desc=desc,
+            progress=progress,
+            func_kwargs=func_kwargs,
+            log_worker=log_worker,
+            log_gate_path=log_gate_path,
+            error_handler=error_handler,
+            error_stats=error_stats,
+            func_name=func_name,
+            update_pbar_postfix=_update_pbar_postfix,
+        )
+
+    # ---- ray backend ----
+    if backend == 'ray':
+        from ._multi_process_ray import run_ray_backend
+        return run_ray_backend(
+            f_wrapped=f_wrapped,
+            items=items,
+            total=total,
+            workers=workers,
+            progress=progress,
+            desc=desc,
+            func_kwargs=func_kwargs,
+            shared_kwargs=shared_kwargs,
+            log_worker=log_worker,
+            log_gate_path=log_gate_path,
+            total_items=total_items,
+            poll_interval=poll_interval,
+            ray_metrics_port=ray_metrics_port,
+            error_handler=error_handler,
+            error_stats=error_stats,
+            func_name=func_name,
+        )
+
+    # ---- threadpool backends (mp / safe) ----
+    if backend == 'mp':
+        return _run_threadpool_backend(
+            backend_label='mp',
+            track_processes=True,
+            f_wrapped=f_wrapped,
+            items=items,
+            total=total,
+            workers=workers,
+            desc=desc,
+            progress=progress,
+            func_kwargs=func_kwargs,
+            log_worker=log_worker,
+            log_gate_path=log_gate_path,
+            error_handler=error_handler,
+            error_stats=error_stats,
+            func_name=func_name,
+            update_pbar_postfix=_update_pbar_postfix,
+        )
+
+    if backend == 'safe':
+        return _run_threadpool_backend(
+            backend_label='safe',
+            track_processes=False,
+            f_wrapped=f_wrapped,
+            items=items,
+            total=total,
+            workers=workers,
+            desc=desc,
+            progress=progress,
+            func_kwargs=func_kwargs,
+            log_worker=log_worker,
+            log_gate_path=log_gate_path,
+            error_handler=error_handler,
+            error_stats=error_stats,
+            func_name=func_name,
+            update_pbar_postfix=_update_pbar_postfix,
+        )
+
+    raise ValueError(f'Unsupported backend: {backend!r}')
+
+
+def _run_seq_backend(
+    *,
+    f_wrapped,
+    items: list,
+    total: int,
+    desc: str,
+    progress: bool,
+    func_kwargs: dict,
+    log_worker,
+    log_gate_path,
+    error_handler,
+    error_stats: ErrorStats,
+    func_name: str,
+    update_pbar_postfix,
+) -> list[Any]:
+    """Run sequential (single-threaded) backend."""
+    results: list[Any] = []
+    with tqdm(
+        total=total,
+        desc=desc,
+        disable=not progress,
+        file=sys.stdout,
+    ) as pbar:
+        for idx, x in enumerate(items):
+            try:
+                result = _call_with_log_control(
+                    f_wrapped,
+                    x,
+                    func_kwargs,
+                    log_worker,
+                    log_gate_path,
+                )
+                error_stats.record_success()
+                results.append(result)
+            except Exception as e:
+                if error_handler == 'raise':
+                    raise
+                error_stats.record_error(idx, e, x, func_name)
+                results.append(None)
+            pbar.update(1)
+            update_pbar_postfix(pbar)
+    _cleanup_log_gate(log_gate_path)
+    return results
+
+
+def _run_threadpool_backend(
+    *,
+    backend_label: str,
+    track_processes: bool,
+    f_wrapped,
+    items: list,
+    total: int,
+    workers: int | None,
+    desc: str,
+    progress: bool,
+    func_kwargs: dict,
+    log_worker,
+    log_gate_path,
+    error_handler,
+    error_stats: ErrorStats,
+    func_name: str,
+    update_pbar_postfix,
+) -> list[Any]:
+    """Run ThreadPoolExecutor backend for 'mp' and 'safe' modes."""
+    # Capture caller frame for better error reporting
+    caller_frame = inspect.currentframe()
+    caller_info = None
+    if caller_frame and caller_frame.f_back and caller_frame.f_back.f_back:
+        # Go back two frames: _run_threadpool_backend -> multi_process -> user
+        outer = caller_frame.f_back.f_back
+        caller_info = {
+            'filename': outer.f_code.co_filename,
+            'lineno': outer.f_lineno,
+            'function': outer.f_code.co_name,
+        }
+
+    def worker_func(x):
+        return _call_with_log_control(
+            f_wrapped,
+            x,
+            func_kwargs,
+            log_worker,
+            log_gate_path,
+        )
+
+    results: list[Any] = [None] * total
+    with tqdm(
+        total=total,
+        desc=desc,
+        disable=not progress,
+        file=sys.stdout,
+    ) as pbar:
+        with concurrent.futures.ThreadPoolExecutor(
+            max_workers=workers
+        ) as executor:
+            if _track_executor_threads is not None:
+                _track_executor_threads(executor)
+
+            # Submit all tasks
+            future_to_idx = {
+                executor.submit(worker_func, x): idx
+                for idx, x in enumerate(items)
+            }
+
+            # Process results as they complete
+            for future in concurrent.futures.as_completed(future_to_idx):
+                idx = future_to_idx[future]
+                try:
+                    result = future.result()
+                    error_stats.record_success()
+                    results[idx] = result
+                except Exception as e:
+                    if error_handler == 'raise':
+                        # Cancel remaining futures
+                        for f in future_to_idx:
+                            f.cancel()
+                        _exit_on_worker_error(
+                            e,
+                            pbar,
+                            caller_info,
+                            backend=backend_label,
+                        )
+                    error_stats.record_error(idx, e, items[idx], func_name)
+                    results[idx] = None
+                pbar.update(1)
+                update_pbar_postfix(pbar)
+
+        if _prune_dead_threads is not None:
+            _prune_dead_threads()
+
+    if track_processes:
+        _track_multiprocessing_processes()
+        _prune_dead_processes()
+
+    _cleanup_log_gate(log_gate_path)
+    return results
+
+
+__all__ = ['multi_process']