mindspore 2.4.1__cp39-none-any.whl → 2.5.0__cp39-none-any.whl

mindspore/utils/runtime_execution_order_check.py

@@ -0,0 +1,550 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""mindspore utils runtime order check."""
+import csv
+import os
+import re
+from collections import defaultdict
+from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor, as_completed
+from multiprocessing import cpu_count
+from typing import List, Dict, Union, Optional
+import sys
+import mindspore.log as logger
+
+# Set Recursion Depth Limit
+sys.setrecursionlimit(10000)
+
+
+class ExecuteOrder:
+    """Represents a single record from the execute_order.csv file."""
+
+    def __init__(self, index: str, group: str, comm_rank: str,
+                 primitive: str, src_rank: str = '', dest_rank: str = '', root_rank: str = '', input_shape: str = '',
+                 output_shape: str = ''):
+        self.index = index
+        self.group = group
+        self.comm_rank = comm_rank.split()  # Split comm_rank into a list of individual ranks
+        self.primitive = primitive
+        self.src_rank = src_rank
+        self.dest_rank = dest_rank
+        self.root_rank = root_rank
+        self.input_shape = input_shape
+        self.output_shape = output_shape
+
+    def generate_base_key(self, rank: str) -> str:
+        """
+        Generates a unique base key for counting, excluding the counter part.
+        """
+        if not self.src_rank and not self.dest_rank and not self.root_rank:
+            # Aggregate communication status, for example, allGather.
+            comm_str = ",".join(self.comm_rank)
+            return f"{self.primitive}_{self.group}_({comm_str})"
+
+        if self.primitive == "Send":
+            # Unique base key of the Send operation.
+            return f"Send_Receive_{self.group}_({rank})->({self.dest_rank})_{self.input_shape}"
+
+        if self.primitive == "Receive":
+            # Unique base key of the reception operation
+            return f"Send_Receive_{self.group}_({self.src_rank})->({rank})_{self.output_shape}"
+
+        # Other operations, such as broadCast
+        parts = [f"{self.primitive}_{self.group}", f"commRank:({','.join(self.comm_rank)})"]
+        if self.src_rank:
+            parts.append(f"srcRank:({self.src_rank})")
+        if self.dest_rank:
+            parts.append(f"destRank:({self.dest_rank})")
+        if self.root_rank:
+            parts.append(f"rootRank:({self.root_rank})")
+
+        return "_".join(parts)
+
+    def generate_key(self, rank: str, count: int) -> str:
+        """
+        Generates the final unique key, including the counterpart.
+        """
+        base_key = self.generate_base_key(rank)
+        return f"{base_key}_{count}th"
+
+
+class RankFolderParser:
+    """
+    Parses specified folder(s) and reads execute_order.csv files within rank_x folders.
+    """
+
+    REQUIRED_COLUMNS = [
+        "index", "group", "comm_rank", "primitive",
+        "src_rank", "dest_rank", "root_rank", "input_shape", "input_type", "output_shape", "output_type", "input_size",
+        "output_size"
+    ]
+
+    def __init__(self, folders: Union[str, List[str]]):
+        # Ensure folders is always a list of absolute paths
+        if isinstance(folders, str):
+            self.folders = [folders]
+        else:
+            self.folders = folders
+
+        # Validate provided folders or rank_{x} paths
+        self._validate_paths()
+
+        # Store results here
+        self.result_map: Dict[str, List[ExecuteOrder]] = {}
+
+        # Determine optimal number of workers based on CPU count
+        self.cpu_cores = cpu_count()
+        self.max_threads = min(8, self.cpu_cores * 2)  # For I/O-bound tasks
+        self.max_processes = min(8, self.cpu_cores)  # For CPU-bound tasks
+
+    def _validate_paths(self):
+        """
+        Validate all provided paths.
+        Each path must be either:
+        - A valid absolute directory containing rank_{x} folders, or
+        - A valid absolute path directly pointing to a rank_{x} folder.
+
+        Additionally, checks for duplicate rank_{x} folders and raises an error if duplicates are found.
+        """
+        seen_ranks = set()  # To track rank_x folder names and detect duplicates
+
+        for path in self.folders:
+            if not os.path.isabs(path):
+                raise ValueError(
+                    f"Invalid path: {path}. "
+                    f"Please provide an absolute path, e.g., '/absolute/path/to/folder' or '/absolute/path/to/rank_x'."
+                )
+            if not os.path.exists(path):
+                raise ValueError(
+                    f"Path does not exist: {path}. "
+                    f"Ensure the path exists and is accessible."
+                )
+
+            # Check if it is a specific rank_{x} folder
+            if os.path.isdir(path) and os.path.basename(path).startswith("rank_"):
+                rank_name = os.path.basename(path)
+                if rank_name in seen_ranks:
+                    raise ValueError(
+                        f"Duplicate rank folder detected: {rank_name}. "
+                        f"Each rank_x folder must be unique. Please remove duplicates."
+                    )
+                seen_ranks.add(rank_name)
+                continue  # Valid rank_{x} folder
+
+            # Check if it is a directory containing rank_{x} folders
+            if os.path.isdir(path):
+                rank_folders = [d for d in os.listdir(path) if
+                                d.startswith("rank_") and os.path.isdir(os.path.join(path, d))]
+                if not rank_folders:
+                    raise ValueError(
+                        f"No rank_x folders found in {path}. Ensure the directory contains at least one folder named "
+                        f"'rank_x', where x is an integer."
+                    )
+
+                # Check for duplicates within this directory
+                for rank_folder in rank_folders:
+                    if rank_folder in seen_ranks:
+                        raise ValueError(
+                            f"Duplicate rank folder detected: {rank_folder} in {path}. "
+                            f"Each rank_x folder must be unique. Please remove duplicates."
+                        )
+                    seen_ranks.add(rank_folder)
+                continue  # Valid directory containing rank_{x} folders
+
+            # Invalid case
+            raise ValueError(
+                f"Invalid path: {path}. "
+                f"Paths must either be rank_x folders or directories containing rank_x folders."
+            )
+
+    def parse(self) -> Dict[str, List[ExecuteOrder]]:
+        """
+        Main parsing function using a ProcessPoolExecutor to handle multiple paths.
+        Each rank_{x} folder is processed in parallel.
+        """
+        with ProcessPoolExecutor(max_workers=self.max_processes) as process_executor:
+            futures = [process_executor.submit(self._parse_path, path) for path in self.folders]
+            for future in as_completed(futures):
+                try:
+                    path_result = future.result()
+                    if path_result:
+                        self.result_map.update(path_result)
+                except FileNotFoundError as e:
+                    logger.error(f"File not found: {e}. Please ensure all required files are present.")
+                except ValueError as e:
+                    logger.error(f"Value error: {e}. Please check the file contents or paths.")
+
+        return self.result_map
+
+    def _parse_path(self, path: str) -> Dict[str, List]:
+        """
+        Helper function to parse a single path. Handles both:
+        - Direct rank_{x} folder paths.
+        - Parent directories containing multiple rank_{x} folders.
+        """
+        result = {}
+
+        # If the path is a rank_{x} folder, parse it directly
+        if os.path.basename(path).startswith("rank_"):
+            rank_id = os.path.basename(path).split("_")[1]
+            # Adding one more layer to access the "execute_order" folder
+            execute_order_path = os.path.join(path, "execute_order")
+            if os.path.exists(execute_order_path):
+                rank_result = self.parse_rank_folder(execute_order_path, rank_id)
+                if rank_result:
+                    result[rank_id] = rank_result[1]  # Extract execute orders
+            return result
+
+        # If the path is a directory containing rank_{x} folders, parse all
+        with ThreadPoolExecutor(max_workers=self.max_threads) as thread_executor:
+            futures = []
+            for d in os.listdir(path):
+                if d.startswith("rank_"):
+                    rank_id = d.split("_")[1]
+                    rank_folder_path = os.path.join(path, d)
+                    execute_order_path = os.path.join(rank_folder_path, "execute_order")
+
+                    if os.path.exists(execute_order_path):
+                        futures.append(thread_executor.submit(self.parse_rank_folder, execute_order_path, rank_id))
+
+            for future in as_completed(futures):
+                try:
+                    rank_id, execute_orders = future.result()
+                    if execute_orders:
+                        result[rank_id] = execute_orders
+                except FileNotFoundError as e:
+                    logger.error(f"File not found during parallel processing: {e}. "
+                                 f"Ensure the required files are present.")
+                except ValueError as e:
+                    logger.error(f"Value error during parallel processing: {e}. Check file format or contents.")
+
+        return result
+
+    def parse_rank_folder(self, rank_folder_path: str, rank_id: str) -> Optional[tuple]:
+        """
+        Parse a single rank_{x} folders execute_order.csv file with header validation.
+        """
+        execute_order_file = os.path.join(rank_folder_path, "comm_execute_order.csv")
+
+        if not os.path.exists(execute_order_file):
+            logger.error(
+                f"No execute_order.csv found in {rank_folder_path}. Skipping this folder. "
+                f"Ensure the rank_{rank_id} folder contains a valid comm_execute_order.csv file."
+            )
+            return rank_id, None
+
+        execute_orders = []
+
+        try:
+            with open(execute_order_file, mode='r', encoding='utf-8') as file:
+                csv_reader = csv.DictReader(file)
+
+                # Validate the header
+                if csv_reader.fieldnames != self.REQUIRED_COLUMNS:
+                    logger.error(
+                        f"Invalid header in {execute_order_file}. Skipping this file. "
+                        f"Expected columns: {self.REQUIRED_COLUMNS}. "
+                        f"Ensure the file contains the correct column names."
+                    )
+                    return rank_id, None
+
+                # Read and parse data rows
+                for row in csv_reader:
+                    execute_order = ExecuteOrder(
+                        index=row["index"],
+                        group=row["group"],
+                        comm_rank=row["comm_rank"],
+                        primitive=row["primitive"],
+                        src_rank=row["src_rank"],
+                        dest_rank=row["dest_rank"],
+                        root_rank=row["root_rank"],
+                        input_shape=row["input_shape"],
+                        output_shape=row["output_shape"],
+
+                    )
+                    execute_orders.append(execute_order)
+        except FileNotFoundError as e:
+            logger.error(f"File not found: {execute_order_file}. Ensure the file exists and is accessible. Error: {e}")
+            return rank_id, None
+
+        return rank_id, execute_orders
+
+
+def modify_execute_orders(execute_orders_map: dict) -> dict:
+    """
+     Modify and generate unique execution order keys for each rank.
+
+     This function processes a mapping of execution orders grouped by ranks. For each order,
+     it generates a unique key by combining a base key and a counter, ensuring all orders
+     are uniquely identifiable. The result is a dictionary where the keys are rank identifiers
+     and the values are lists of unique execution order keys.
+
+     Args:
+         execute_orders_map (dict): A dictionary where keys are rank identifiers (e.g., "rank_0")
+                                    and values are lists of ExecuteOrder objects. If a rank has no
+                                    orders, its value may be `None`.
+
+     Returns:
+         dict: A dictionary where keys are rank identifiers and values are lists of unique string
+               keys representing the modified execution orders for each rank.
+
+     """
+    result = {}
+
+    for rank, execute_orders in execute_orders_map.items():
+        # If execute_orders is None, use an empty list instead.
+        if execute_orders is None:
+            execute_orders = []
+
+        count_map = defaultdict(int)
+        new_orders = []
+
+        for order in execute_orders:
+            # Use generate_base_key to generate the base key for counting.
+            base_key = order.generate_base_key(rank)
+
+            count_map[base_key] += 1
+            count = count_map[base_key]
+
+            # Use generate_key to generate the final unique key.
+            new_key = order.generate_key(rank, count)
+            new_orders.append(new_key)
+
+        # Save the generated new order list to the result dictionary.
+        result[rank] = new_orders
+
+    return result
+
+
+def parse_and_validate(data: dict, all_rank: bool = True):
+    """
+     Parse and validate execution orders in a directed graph structure.
+
+     This function checks the integrity and consistency of a given dataset, ensuring all required
+     keys are present and correctly referenced. It also validates the structure of the input data
+     and parses string values to extract meaningful components.
+
+     Args:
+         data (dict): A dictionary where keys are string identifiers and values are lists of strings.
+                      Each value represents a dependency or reference to other keys.
+         all_rank (bool): If True, checks that all elements referenced in the data are present as keys
+                          in the dictionary. If False, only checks intersections.
+
+     Returns:
+         None: Log error messages to the console if validation fails, otherwise completes silently.
+
+     Raises:
+         ValueError: Raised indirectly if `parse_elements` encounters malformed input strings.
+         TypeError: Raised indirectly if data contains unexpected types.
+
+     """
+    def parse_elements(value: str, max_groups: int = 2) -> set:
+        """Extract unique elements inside the first one or two parentheses from a string."""
+        groups = re.findall(r'\((.*?)\)', value)
+        limited_groups = groups[:max_groups]  # Limit to the first `max_groups` matches
+        return {item.strip() for group in limited_groups for item in group.split(',')}
+
+    if not isinstance(data, dict):
+        logger.error("Input must be a dictionary with string keys and lists of strings as values.")
+        return
+
+    key_to_values = {key: set(values) for key, values in data.items() if
+                     isinstance(values, list) and all(isinstance(v, str) for v in values)}
+
+    for key, values in data.items():
+        if not isinstance(values, list) or not all(isinstance(v, str) for v in values):
+            logger.error(f"Values for key '{key}' must be a list of strings.")
+            continue
+
+        for value in values:
+            try:
+                elements = parse_elements(value)
+            except (ValueError, TypeError, AttributeError) as e:
+                logger.error(f"Unable to parse elements from value '{value}' in key '{key}'. Error: {e}")
+                continue
+
+            # Check for missing keys if all_rank is True
+            if all_rank:
+                missing_keys = elements - key_to_values.keys()
+                if missing_keys:
+                    logger.error(f"The following keys are missing for value '{value}': {missing_keys}")
+                    continue
+
+            # Check if the value is present in the referenced keys
+            for element in elements & key_to_values.keys() if not all_rank else elements:
+                if value not in key_to_values[element]:
+                    logger.error(f"Key '{element}' is missing the value '{value}'.")
+
+
+def detect_cycle_in_graph(ranks_map):
+    """
+    Detects a cycle in the directed graph constructed from ranks_map.
+
+    Args:
+    - ranks_map (dict): A dictionary where keys are rank names and values are lists of nodes.
+
+    Returns:
+    - tuple: (cycle_path, cycle_ranks) where cycle_path is a list of nodes forming the cycle and cycle_ranks
+             is a list of rank transitions corresponding to the cycle path.
+    """
+    # Step 1: Build the directed graph and track edges with ranks
+    graph = defaultdict(list)
+    rank_edges = {}
+
+    for rank, nodes in ranks_map.items():
+        for i in range(len(nodes) - 1):
+            u, v = nodes[i], nodes[i + 1]
+            graph[u].append(v)
+            rank_edges[(u, v)] = rank
+
+    # Step 2: Detect cycle using DFS with path and rank tracking
+    visited = set()
+    recursion_stack = set()
+    path = []
+    cycle_path = []
+    cycle_ranks = []
+
+    def dfs(node):
+        if node in recursion_stack:  # Cycle detected
+            # Identify the cycle path from the recursion stack
+            cycle_index = path.index(node)
+            cycle_path.extend(path[cycle_index:])
+            for i in range(cycle_index, len(path) - 1):
+                u, v = path[i], path[i + 1]
+                cycle_ranks.append(f"{rank_edges[(u, v)]} {u} -> {v}")
+            # Add the closing edge for the cycle
+            cycle_ranks.append(f"{rank_edges[(path[-1], node)]} {path[-1]} -> {node}")
+            return True
+
+        if node in visited:
+            return False
+
+        visited.add(node)
+        recursion_stack.add(node)
+        path.append(node)
+
+        for neighbor in graph[node]:
+            if dfs(neighbor):
+                return True
+
+        # Backtrack
+        recursion_stack.remove(node)
+        path.pop()
+        return False
+
+    nodes = list(graph.keys())
+    for node in nodes:
+        if node not in visited:
+            if dfs(node):
+                return cycle_path, cycle_ranks
+
+    return None, None
+
+
+def determine_all_rank(folders_):
+    """
+    Determine the value of all_rank based on the input folders_.
+
+    Args:
+        folders_ (str | list): Folder path(s) to process.
+
+    Returns:
+        bool | None: Returns True/False for valid inputs, or None for invalid inputs.
+    """
+    if isinstance(folders_, str) and folders_:
+        return True
+    if isinstance(folders_, list):
+        if len(folders_) == 1:
+            return True
+        if len(folders_) > 1:
+            return False
+        return None
+    return None
+
+
+def output_cycle_results(cycle_path, cycle_ranks):
+    """
+    Helper function to output cycle detection results.
+
+    Args:
+        cycle_path (list): List of nodes forming a cycle, if any.
+        cycle_ranks (list): List of ranks involved in the cycle.
+
+    Returns:
+        None: Outputs results to the console.
+    """
+    if cycle_path:
+        logger.warning("Cycle detected:")
+        logger.warning(" -> ".join(cycle_path) + f" -> {cycle_path[0]}")  # Close the cycle
+        logger.warning("Involving ranks:")
+        for rank in cycle_ranks:
+            logger.warning(rank)
+    else:
+        logger.warning("Check success.")
+
+
+def runtime_execution_order_check(folders_, all_rank=None):
+    """
+    Verify the rank_x folder in the specified directory.
+
+    Parameter description:
+    1. folders (str or list[str]):
+    - Can be a single string or a list of strings representing the rank_x folder path or its upper-level directory path.
+    - If the input directory is the upper-level directory path, the function automatically identifies and verifies
+    all rank_x folders in the directory.
+    - If a specific rank_x folder path is passed, the function will only verify these specified folders.
+
+    2. check_all (optional, true by default):
+    - Controls whether to verify all rank_x folders.
+    - True (default): Verify all files.
+    - False: Only part of the transferred rank_x folder is verified.
+
+    Example:
+    Example 1: Verify all rank_x folders (default behavior)
+    runtime_execution_order_check("path/to/parent_folder")
+
+    Example 2: Verify only some specified rank_x folders.
+    runtime_execution_order_check(["path/to/rank_1", "path/to/rank_2"], all_rank=False)
+
+    Example 3: Verify a single rank_x folder.
+    runtime_execution_order_check("path/to/rank_x", all_rank=True)
+
+    Precautions:
+    - When folders is the upper-level directory path, the function automatically
+    searches for and verifies the rank_x folder.
+    - When check_all is set to false, only the rank_x folder specified in the input path is verified.
+    """
+    # Use the provided all_rank if available, otherwise determine it
+    if all_rank is None:
+        all_rank = determine_all_rank(folders_)
+
+    if all_rank is None:  # Input validation failed
+        logger.error("Invalid input. `folders_` must be a non-empty string or a list with at least one string element.")
+        return
+
+    # Parse folders
+    parser = RankFolderParser(folders_)
+    result_map = parser.parse()
+
+    # Modify execution orders
+    modified_orders = modify_execute_orders(result_map)
+
+    # Parse and validate execution orders
+    parse_and_validate(modified_orders, all_rank)
+
+    # Detect cycles
+    cycle_path, cycle_ranks = detect_cycle_in_graph(modified_orders)
+
+    # Output results
+    output_cycle_results(cycle_path, cycle_ranks)

mindspore/version.py

@@ -1 +1 @@
-__version__ = '2.4.1'
+__version__ = '2.5.0'

{mindspore-2.4.1.dist-info → mindspore-2.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mindspore
-Version: 2.4.1
+Version: 2.5.0
 Summary: MindSpore is a new open source deep learning training/inference framework that could be used for mobile, edge and cloud scenarios.
 Home-page: https://www.mindspore.cn
 Download-URL: https://github.com/mindspore-ai/mindspore/tags
@@ -38,6 +38,7 @@ Requires-Dist: packaging (>=20.0)
 Requires-Dist: psutil (>=5.6.1)
 Requires-Dist: astunparse (>=1.6.3)
 Requires-Dist: safetensors (>=0.4.0)
+Requires-Dist: dill (>=0.3.7)
 
 ![MindSpore Logo](https://gitee.com/mindspore/mindspore/raw/master/docs/MindSpore-logo.png "MindSpore logo")
 
@@ -46,7 +47,6 @@ Requires-Dist: safetensors (>=0.4.0)
 [![Downloads](https://static.pepy.tech/badge/mindspore)](https://pepy.tech/project/mindspore)
 [![DockerHub](https://img.shields.io/docker/pulls/mindspore/mindspore-cpu.svg)](https://hub.docker.com/r/mindspore/mindspore-cpu)
 [![LICENSE](https://img.shields.io/github/license/mindspore-ai/mindspore.svg?style=flat-square)](https://github.com/mindspore-ai/mindspore/blob/master/LICENSE)
-[![Slack](https://img.shields.io/badge/slack-chat-green.svg?logo=slack)](https://join.slack.com/t/mindspore/shared_invite/zt-dgk65rli-3ex4xvS4wHX7UDmsQmfu8w)
 [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg?style=flat-square)](https://gitee.com/mindspore/mindspore/pulls)
 
 [查看中文](./README_CN.md)
@@ -118,7 +118,7 @@ MindSpore offers build options across multiple backends:
 
 | Hardware Platform | Operating System | Status |
 | :---------------- | :--------------- | :----- |
-| Ascend910 | Ubuntu-x86 | ✔️ |
+| Ascend | Ubuntu-x86 | ✔️ |
 |  | Ubuntu-aarch64 | ✔️ |
 |  | EulerOS-aarch64 | ✔️ |
 |  | CentOS-x86 | ✔️ |
@@ -296,7 +296,6 @@ Check out how MindSpore Open Governance [works](https://gitee.com/mindspore/comm
 
 ### Communication
 
-- [MindSpore Slack](https://join.slack.com/t/mindspore/shared_invite/zt-dgk65rli-3ex4xvS4wHX7UDmsQmfu8w) - Communication platform for developers.
 - IRC channel at `#mindspore` (only for meeting minutes logging purpose)
 - Video Conferencing: TBD
 - Mailing-list: <https://mailweb.mindspore.cn/postorius/lists>