speedy-utils 1.0.4__py3-none-any.whl → 1.0.5__py3-none-any.whl

speedy_utils/multi_worker/process.py

@@ -0,0 +1,198 @@
+import inspect
+import multiprocessing
+import os
+import time
+import traceback
+from collections.abc import Callable, Iterable, Iterator, Sequence
+from concurrent.futures import ProcessPoolExecutor, as_completed
+from itertools import islice
+from typing import Any, List, TypeVar, cast
+
+T = TypeVar("T")
+
+if hasattr(multiprocessing, "set_start_method"):
+    try:
+        multiprocessing.set_start_method("spawn", force=True)
+    except RuntimeError:
+        pass
+
+try:
+    from tqdm import tqdm
+except ImportError:  # pragma: no cover
+    tqdm = None  # type: ignore[assignment]
+
+
+# ──── internal helpers ────────────────────────────────────────────────────
+
+
+def _group_iter(src: Iterable[Any], size: int) -> Iterable[list[Any]]:
+    "Yield *size*-sized chunks from *src*."
+    it = iter(src)
+    while chunk := list(islice(it, size)):
+        yield chunk
+
+
+def _short_tb() -> str:
+    tb = "".join(traceback.format_exc())
+    return "\n".join(ln for ln in tb.splitlines() if "multi_process" not in ln)
+
+
+def _safe_call(func: Callable, obj, fixed):
+    try:
+        return func(obj, **fixed)
+    except Exception as exc:
+        func_name = getattr(func, "__name__", str(func))
+        raise RuntimeError(
+            f"{func_name}({obj!r}) failed: {exc}\n{_short_tb()}"
+        ) from exc
+
+
+def _worker_process(
+    func: Callable, item_batch: Any, fixed_kwargs: dict, batch_size: int
+):
+    """Worker function executed in each process."""
+    if batch_size > 1:
+        results = []
+        for itm in item_batch:
+            try:
+                results.append(_safe_call(func, itm, fixed_kwargs))
+            except Exception:
+                results.append(None)
+        return results
+    return _safe_call(func, item_batch, fixed_kwargs)
+
+
+# ──── public API ──────────────────────────────────────────────────────────
+def multi_process(
+    func: Callable[[Any], Any],
+    inputs: Iterable[Any],
+    *,
+    workers: int | None = None,
+    batch: int = 1,
+    ordered: bool = True,
+    progress: bool = False,
+    inflight: int | None = None,
+    timeout: float | None = None,
+    stop_on_error: bool = True,
+    process_update_interval=10,
+    **fixed_kwargs,
+) -> list[Any]:
+    """
+    Simple multi‑processing parallel map that returns a *list*.
+
+    Parameters
+    ----------
+    func          – target callable executed in separate processes, must be of the form f(obj, ...).
+    inputs        – iterable with the objects.
+    workers       – process pool size (defaults to :pyfunc:`os.cpu_count()`).
+    batch         – package *batch* inputs into one call to reduce IPC overhead.
+    ordered       – keep original order; if ``False`` results stream as finished.
+    progress      – show a tqdm bar (requires *tqdm*).
+    inflight      – max logical items concurrently submitted
+                    *(default: ``workers × 4``)*.
+    timeout       – overall timeout for the mapping (seconds).
+    stop_on_error – raise immediately on first exception (default) or
+                    substitute failing result with ``None``.
+    **fixed_kwargs – static keyword args forwarded to every ``func()`` call.
+    """
+    if workers is None:
+        workers = os.cpu_count() or 1
+    if inflight is None:
+        inflight = workers * 4
+    if batch < 1:
+        raise ValueError("batch must be ≥ 1")
+
+    try:
+        n_inputs = len(inputs)  # type: ignore[arg-type]
+    except Exception:
+        n_inputs = None
+
+    src_iter: Iterator[Any] = iter(inputs)
+    if batch > 1:
+        src_iter = cast(Iterator[Any], _group_iter(src_iter, batch))
+
+    logical_total = n_inputs
+    bar = None
+    last_bar = 0
+    if progress and tqdm is not None and logical_total is not None:
+        bar = tqdm(
+            total=logical_total,
+            ncols=80,
+            colour="green",
+            bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt}"
+            " [{elapsed}<{remaining}, {rate_fmt}{postfix}]",
+        )
+
+    if ordered and logical_total is not None:
+        results: list[Any] = [None] * logical_total
+    else:
+        results = []
+
+    completed = 0
+    next_idx = 0
+
+    with ProcessPoolExecutor(max_workers=workers) as pool:
+        futures = set()
+
+        for _ in range(min(inflight, workers)):
+            try:
+                arg = next(src_iter)
+            except StopIteration:
+                break
+            fut = pool.submit(_worker_process, func, arg, fixed_kwargs, batch)
+            fut.idx = next_idx  # type: ignore[attr-defined]
+            futures.add(fut)
+            next_idx += len(arg) if batch > 1 else 1
+
+        while futures:
+            for fut in as_completed(futures, timeout=timeout):
+                futures.remove(fut)
+                idx = fut.idx  # type: ignore[attr-defined]
+                try:
+                    res = fut.result()
+                except Exception:
+                    if stop_on_error:
+                        raise
+                    num_items = batch if batch > 1 else 1
+                    res = [None] * num_items if batch > 1 else None
+
+                out_items = res if batch > 1 else [res]
+                if out_items is None:
+                    out_items = []
+
+                if ordered and logical_total is not None:
+                    if isinstance(out_items, list) and len(out_items) > 0:
+                        for i, item in enumerate(out_items):
+                            if idx + i < len(results):
+                                results[idx + i] = item
+                else:
+                    if isinstance(out_items, list):
+                        results.extend(out_items)
+
+                completed += len(out_items)
+
+                if bar and completed - last_bar >= process_update_interval:
+                    bar.update(completed - last_bar)
+                    last_bar = completed
+
+                try:
+                    while next_idx - completed < inflight:
+                        arg = next(src_iter)
+                        fut2 = pool.submit(
+                            _worker_process, func, arg, fixed_kwargs, batch
+                        )
+                        fut2.idx = next_idx  # type: ignore[attr-defined]
+                        futures.add(fut2)
+                        next_idx += len(arg) if batch > 1 else 1
+                except StopIteration:
+                    pass
+                break
+
+    if bar:
+        bar.update(completed - last_bar)
+        bar.close()
+
+    return results
+
+
+__all__ = ["multi_process"]

speedy_utils/multi_worker/thread.py

@@ -0,0 +1,327 @@
+"""Provides thread-based parallel execution utilities."""
+
+import os
+import time
+import traceback
+from collections.abc import Callable, Iterable
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from itertools import islice
+from typing import Any, TypeVar
+
+from loguru import logger
+
+try:
+    from tqdm import tqdm
+except ImportError:  # pragma: no cover
+    tqdm = None  # type: ignore[assignment]
+
+# Sensible defaults
+DEFAULT_WORKERS = (os.cpu_count() or 4) * 2
+
+T = TypeVar("T")
+R = TypeVar("R")
+
+
+def _group_iter(src: Iterable[T], size: int) -> Iterable[list[T]]:
+    """Yield successive chunks from iterable of specified size."""
+    it = iter(src)
+    while chunk := list(islice(it, size)):
+        yield chunk
+
+
+def _short_tb() -> str:
+    """Return a shortened traceback, excluding internal frames."""
+    tb = "".join(traceback.format_exc())
+    # hide frames inside this helper to keep logs short
+    return "\n".join(ln for ln in tb.splitlines() if "multi_thread.py" not in ln)
+
+
+def _worker(item: T, func: Callable[[T], R], fixed_kwargs: dict[str, Any]) -> R:
+    """Execute the function with an item and fixed kwargs."""
+    return func(item, **fixed_kwargs)
+
+
+# ────────────────────────────────────────────────────────────
+# main API
+# ────────────────────────────────────────────────────────────
+def multi_thread(
+    func: Callable,
+    inputs: Iterable[Any],
+    *,
+    workers: int | None = DEFAULT_WORKERS,
+    batch: int = 1,
+    ordered: bool = True,
+    progress: bool = True,
+    progress_update: int = 10,
+    prefetch_factor: int = 4,
+    timeout: float | None = None,
+    stop_on_error: bool = True,
+    n_proc=0,
+    store_output_pkl_file: str | None = None,
+    **fixed_kwargs,
+) -> list[Any]:
+    """
+    ThreadPool **map** that returns a *list*.
+
+    Parameters
+    ----------
+    func            – target callable.
+    inputs          – iterable with the arguments.
+    workers         – defaults to ``os.cpu_count()*2``.
+    batch           – package *batch* inputs into one call for low‑overhead.
+    ordered         – keep original order (costs memory); if ``False`` results
+                      are yielded as soon as they finish.
+    progress        – show a tqdm bar (requires *tqdm* installed).
+    progress_update – bar redraw frequency (logical items, *not* batches).
+    prefetch_factor – in‑flight tasks ≈ ``workers * prefetch_factor``.
+    timeout         – overall timeout (seconds) for the mapping.
+    stop_on_error   – raise immediately on first exception (default).  If
+                      ``False`` the failing task’s result becomes ``None``.
+    **fixed_kwargs  – static keyword args forwarded to every ``func()`` call.
+    """
+    from speedy_utils import dump_json_or_pickle, load_by_ext
+
+    if n_proc > 1:
+        import tempfile
+
+        from fastcore.all import threaded
+
+        # split the inputs by nproc
+        inputs = list(inputs)
+        n_per_proc = max(len(inputs) // n_proc, 1)
+        proc_inputs_list = []
+        for i in range(0, len(inputs), n_per_proc):
+            proc_inputs_list.append(inputs[i : i + n_per_proc])
+        procs = []
+        in_process_multi_thread = threaded(process=True)(multi_thread)
+
+        for proc_id, proc_inputs in enumerate(proc_inputs_list):
+            with tempfile.NamedTemporaryFile(
+                delete=False, suffix="multi_thread.pkl"
+            ) as tmp_file:
+                file_pkl = tmp_file.name
+            assert isinstance(in_process_multi_thread, Callable)
+            proc = in_process_multi_thread(
+                func,
+                proc_inputs,
+                workers=workers,
+                batch=batch,
+                ordered=ordered,
+                progress=proc_id == 0,
+                progress_update=progress_update,
+                prefetch_factor=prefetch_factor,
+                timeout=timeout,
+                stop_on_error=stop_on_error,
+                n_proc=0,  # prevent recursion
+                store_output_pkl_file=file_pkl,
+                **fixed_kwargs,
+            )
+            procs.append([proc, file_pkl])
+        # join
+        results = []
+
+        for proc, file_pkl in procs:
+            proc.join()
+            logger.info(f"Done proc {proc=}")
+            results.extend(load_by_ext(file_pkl))
+        return results
+
+    try:
+        import pandas as pd
+
+        if isinstance(inputs, pd.DataFrame):
+            inputs = inputs.to_dict(orient="records")
+    except ImportError:
+        pass
+
+    try:
+        n_inputs = len(inputs)  # type: ignore[arg-type]
+    except Exception:
+        n_inputs = None
+    workers_val = workers if workers is not None else DEFAULT_WORKERS
+
+    if batch == 1 and n_inputs and n_inputs / max(workers_val, 1) > 20_000:
+        batch = 32  # empirically good for sub‑ms tasks
+
+    # ── build (maybe‑batched) source iterator ────────────────────────────
+    src_iter: Iterable[Any] = iter(inputs)
+    if batch > 1:
+        src_iter = _group_iter(src_iter, batch)
+    # Ensure src_iter is always an iterator
+    src_iter = iter(src_iter)
+
+    # total logical items (for bar & ordered pre‑allocation)
+    logical_total = n_inputs
+    if logical_total is not None and batch > 1:
+        logical_total = n_inputs  # still number of *items*, not batches
+
+    # ── progress bar ─────────────────────────────────────────────────────
+    bar = None
+    last_bar_update = 0
+    if progress and tqdm is not None and logical_total is not None:
+        bar = tqdm(
+            total=logical_total,
+            ncols=128,
+            colour="green",
+            bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt}"
+            " [{elapsed}<{remaining}, {rate_fmt}{postfix}]",
+        )
+
+    # ── prepare result container ─────────────────────────────────────────
+    if ordered and logical_total is not None:
+        results: list[Any] = [None] * logical_total
+    else:
+        results = []
+
+    # ── main execution loop ──────────────────────────────────────────────────
+    workers_val = workers if workers is not None else DEFAULT_WORKERS
+    max_inflight = workers_val * max(prefetch_factor, 1)
+    completed_items = 0
+    next_logical_idx = 0  # index assigned to the next submission
+
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        inflight = set()
+
+        # prime the pool
+        for _ in range(max_inflight):
+            try:
+                arg = next(src_iter)
+            except StopIteration:
+                break
+            if batch > 1:
+                fut = pool.submit(
+                    lambda items: [_worker(item, func, fixed_kwargs) for item in items],
+                    arg,
+                )
+                fut.idx = next_logical_idx  # type: ignore[attr-defined]
+                inflight.add(fut)
+                next_logical_idx += len(arg)
+            else:
+                fut = pool.submit(_worker, arg, func, fixed_kwargs)
+                fut.idx = next_logical_idx  # type: ignore[attr-defined]
+                inflight.add(fut)
+                next_logical_idx += 1
+
+        try:
+            # Process futures as they complete and add new ones to keep the pool busy
+            while inflight:  # Continue until all in-flight tasks are processed
+                for fut in as_completed(inflight, timeout=timeout):
+                    inflight.remove(fut)
+                    idx = fut.idx  # type: ignore[attr-defined]
+                    try:
+                        res = fut.result()
+                    except Exception:
+                        if stop_on_error:
+                            raise
+                        res = None
+
+                    # flatten res to list of logical outputs
+                    out_items = res if batch > 1 else [res]
+
+                    # Ensure out_items is a list (and thus Sized)
+                    if out_items is None:
+                        out_items = [None]
+                    elif not isinstance(out_items, list):
+                        out_items = (
+                            list(out_items)
+                            if isinstance(out_items, Iterable)
+                            else [out_items]
+                        )
+
+                    # store outputs
+                    if ordered and logical_total is not None:
+                        results[idx : idx + len(out_items)] = out_items
+                    else:
+                        results.extend(out_items)
+
+                    completed_items += len(out_items)
+
+                    # progress bar update
+                    if bar and completed_items - last_bar_update >= progress_update:
+                        bar.update(completed_items - last_bar_update)
+                        last_bar_update = completed_items
+                        # Show pending, submitted, processing in the bar postfix
+                        submitted = next_logical_idx
+                        processing = min(len(inflight), workers_val)
+                        pending = (
+                            (logical_total - submitted)
+                            if logical_total is not None
+                            else None
+                        )
+                        postfix = {
+                            "pending": pending if pending is not None else "-",
+                            # 'submitted': submitted,
+                            "processing": processing,
+                        }
+                        bar.set_postfix(postfix)
+
+                    # keep queue full
+                    try:
+                        while next_logical_idx - completed_items < max_inflight:
+                            arg = next(src_iter)
+                            if batch > 1:
+                                fut2 = pool.submit(
+                                    lambda items: [
+                                        _worker(item, func, fixed_kwargs)
+                                        for item in items
+                                    ],
+                                    arg,
+                                )
+                                fut2.idx = next_logical_idx  # type: ignore[attr-defined]
+                                inflight.add(fut2)
+                                next_logical_idx += len(arg)
+                            else:
+                                fut2 = pool.submit(_worker, arg, func, fixed_kwargs)
+                                fut2.idx = next_logical_idx  # type: ignore[attr-defined]
+                                inflight.add(fut2)
+                                next_logical_idx += 1
+                    except StopIteration:
+                        pass
+
+                    # Break the inner loop as we've processed one future
+                    break
+
+                # If we've exhausted the inner loop without processing anything,
+                # and there are still in-flight tasks, we need to wait for them
+                if inflight and timeout is not None:
+                    # Use a small timeout to avoid hanging indefinitely
+                    time.sleep(min(0.01, timeout / 10))
+
+        finally:
+            if bar:
+                bar.update(completed_items - last_bar_update)
+                bar.close()
+    if store_output_pkl_file:
+        dump_json_or_pickle(results, store_output_pkl_file)
+    return results
+
+
+def multi_thread_standard(
+    fn: Callable[[Any], Any], items: Iterable[Any], workers: int = 4
+) -> list[Any]:
+    """Execute a function using standard ThreadPoolExecutor.
+
+    A standard implementation of multi-threading using ThreadPoolExecutor.
+    Ensures the order of results matches the input order.
+
+    Parameters
+    ----------
+    fn : Callable
+        The function to execute for each item.
+    items : Iterable
+        The items to process.
+    workers : int, optional
+        Number of worker threads, by default 4.
+
+    Returns
+    -------
+    list
+        Results in same order as input items.
+    """
+    with ThreadPoolExecutor(max_workers=workers) as executor:
+        futures = [executor.submit(fn, item) for item in items]
+        results = [fut.result() for fut in futures]
+    return results
+
+
+__all__ = ["multi_thread", "multi_thread_standard"]

speedy_utils/scripts/mpython.py

@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+import argparse
+import itertools
+import multiprocessing  # Import multiprocessing module
+import os
+import shlex  # To properly escape command line arguments
+import shutil
+import subprocess
+
+taskset_path = shutil.which("taskset")
+
+
+def assert_script(python_path):
+    code_str = open(python_path).read()
+    if "MP_ID" not in code_str or "MP_TOTAL" not in code_str:
+        example_code = (
+            'import os; MP_TOTAL = int(os.environ.get("MP_TOTAL"));MP_ID = int(os.environ.get("MP_ID"))\n'
+            "inputs = list(inputs[MP_ID::MP_TOTAL])"
+        )
+        # ANSI escape codes for coloring
+        YELLOW = "\033[93m"
+        RESET = "\033[0m"
+        raise_msg = (
+            f"MP_ID and MP_TOTAL not found in {python_path}, please add them.\n\n"
+            f"Example:\n{YELLOW}{example_code}{RESET}"
+        )
+        raise Exception(raise_msg)
+
+
+def run_in_tmux(commands_to_run, tmux_name, num_windows):
+    with open("/tmp/start_multirun_tmux.sh", "w") as script_file:
+        # first cmd is to kill the session if it exists
+
+        script_file.write("#!/bin/bash\n\n")
+        script_file.write(f"tmux kill-session -t {tmux_name}\nsleep .1\n")
+        script_file.write(f"tmux new-session -d -s {tmux_name}\n")
+        for i, cmd in enumerate(itertools.cycle(commands_to_run)):
+            if i >= num_windows:
+                break
+            window_name = f"{tmux_name}:{i}"
+            if i == 0:
+                script_file.write(f"tmux send-keys -t {window_name} '{cmd}' C-m\n")
+            else:
+                script_file.write(f"tmux new-window -t {tmux_name}\n")
+                script_file.write(f"tmux send-keys -t {window_name} '{cmd}' C-m\n")
+
+        # Make the script executable
+        script_file.write("chmod +x /tmp/start_multirun_tmux.sh\n")
+        print("Run /tmp/start_multirun_tmux.sh")
+
+
+def main():
+    # Assert that MP_ID and MP_TOTAL are not already set
+
+    parser = argparse.ArgumentParser(description="Process fold arguments")
+    parser.add_argument(
+        "--total_fold", "-t", default=16, type=int, help="total number of folds"
+    )
+    parser.add_argument("--gpus", type=str, default="0,1,2,3,4,5,6,7")
+    parser.add_argument("--ignore_gpus", "-ig", type=str, default="")
+    parser.add_argument(
+        "--total_cpu",
+        type=int,
+        default=multiprocessing.cpu_count(),
+        help="total number of cpu cores available",
+    )
+    parser.add_argument(
+        "cmd", nargs=argparse.REMAINDER
+    )  # This will gather the remaining unparsed arguments
+
+    args = parser.parse_args()
+    if not args.cmd or (args.cmd[0] == "--" and len(args.cmd) == 1):
+        parser.error("Invalid command provided")
+    assert_script(args.cmd[0])
+
+    cmd_str = None
+    if args.cmd[0] == "--":
+        cmd_str = shlex.join(args.cmd[1:])
+    else:
+        cmd_str = shlex.join(args.cmd)
+
+    gpus = args.gpus.split(",")
+    gpus = [gpu for gpu in gpus if not gpu in args.ignore_gpus.split(",")]
+    num_gpus = len(gpus)
+
+    cpu_per_process = max(args.total_cpu // args.total_fold, 1)
+    cmds = []
+    for i in range(args.total_fold):
+        gpu = gpus[i % num_gpus]
+        cpu_start = (i * cpu_per_process) % args.total_cpu
+        cpu_end = ((i + 1) * cpu_per_process - 1) % args.total_cpu
+        ENV = f"CUDA_VISIBLE_DEVICES={gpu} MP_ID={i} MP_TOTAL={args.total_fold}"
+        if taskset_path:
+            fold_cmd = (
+                f"{ENV} {taskset_path} -c {cpu_start}-{cpu_end}  python {cmd_str}"
+            )
+        else:
+            fold_cmd = f"{ENV} python {cmd_str}"
+
+        cmds.append(fold_cmd)
+
+    run_in_tmux(cmds, "mpython", args.total_fold)
+    os.chmod("/tmp/start_multirun_tmux.sh", 0o755)  # Make the script executable
+    os.system("/tmp/start_multirun_tmux.sh")
+
+
+if __name__ == "__main__":
+    main()