Ascend/msagent / skills/profiler/ascend-cluster-fast-slow-rank-detector
ascend-cluster-fast-slow-rank-detector
Verified专门用于 Ascend 集群 Profiling 性能数据的“快慢卡”诊断专家技能。当用户提供【集群性能数据目录/路径】并要求分析【快慢卡】、【慢节点】、【负载不均衡】或【集群瓶颈】时,必须触发此技能。该技能会自动接收集群路径,调度相关工具输出快慢卡的宏观定性与微观根因(如 Host 下发瓶颈、算子计算劣化)。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.0 KB |
| scripts/compare_api_stats.py | script | 8.0 KB |
| scripts/compare_op_stats.py | script | 8.2 KB |
| scripts/rank_data_finder.py | script | 5.3 KB |
scripts/rank_data_finder.py
View on GitHub"""
Discover slow/fast rank profiler data paths under a cluster root.
Supports text (op_statistic.csv / api_statistic.csv) and db (ascend_pytorch_profiler_{rank}.db).
"""
from __future__ import annotations
import json
import os
from typing import Tuple, Optional, Literal
DataType = Literal["text", "db"]
# 限制从 cluster_path 起向下遍历的最大深度,避免误入深层无关目录
DEFAULT_MAX_WALK_DEPTH = 6
def find_profiler_info_dirs(
cluster_path: str,
max_depth: int = DEFAULT_MAX_WALK_DEPTH,
) -> dict[int, str]:
"""
Find all profiler_info_{rank_id}.json under cluster_path and return
rank_id -> directory (parent of the json file).
Traversal is limited to max_depth levels below cluster_path.
"""
result = {}
cluster_path = os.path.abspath(cluster_path)
if not os.path.isdir(cluster_path):
return result
cluster_path = os.path.normpath(cluster_path)
for root, dirs, files in os.walk(cluster_path):
try:
rel = os.path.relpath(root, cluster_path)
except ValueError:
continue
depth = len(rel.split(os.sep)) if rel != "." else 0
if depth > max_depth:
dirs.clear()
continue
for f in files:
if f.startswith("profiler_info_") and f.endswith(".json"):
try:
rank_str = f.replace("profiler_info_", "").replace(".json", "")
rank_id = int(rank_str)
result[rank_id] = root
except ValueError:
continue
return result
def find_rank_csv_path(
cluster_path: str, rank_id: int, csv_name: str
) -> Optional[str]:
"""
Find path to csv_name (op_statistic.csv or api_statistic.csv) for the given rank.
csv_name must be 'op_statistic.csv' or 'api_statistic.csv'.
"""
dirs = find_profiler_info_dirs(cluster_path)
rank_dir = dirs.get(rank_id)
if not rank_dir:
return None
candidate = os.path.join(rank_dir, "ASCEND_PROFILER_OUTPUT", csv_name)
return candidate if os.path.isfile(candidate) else None
def find_rank_db_path(cluster_path: str, rank_id: int) -> Optional[str]:
"""Find path to ascend_pytorch_profiler_{rank_id}.db for the given rank."""
dirs = find_profiler_info_dirs(cluster_path)
rank_dir = dirs.get(rank_id)
if not rank_dir:
return None
candidate = os.path.join(
rank_dir, "ASCEND_PROFILER_OUTPUT", f"ascend_pytorch_profiler_{rank_id}.db"
)
return candidate if os.path.isfile(candidate) else None
def _data_type_from_path(path: str) -> DataType:
return "db" if path.lower().endswith(".db") else "text"
def resolve_rank_paths(
cluster_path: str,
slow_rank: int,
fast_rank: int,
mode: Literal["op", "api"],
slow_path: Optional[str] = None,
fast_path: Optional[str] = None,
) -> Tuple[Optional[str], Optional[str], DataType, str]:
"""
Resolve paths for slow and fast rank data.
mode: 'op' -> op_statistic.csv or db for op; 'api' -> api_statistic.csv or db for api.
Returns:
(slow_path, fast_path, data_type, error_message)
If error_message is non-empty, path resolution failed.
"""
csv_name = "op_statistic.csv" if mode == "op" else "api_statistic.csv"
if slow_path and fast_path:
if not os.path.isfile(slow_path):
return None, None, "text", f"慢卡路径不存在或不是文件: {slow_path}"
if not os.path.isfile(fast_path):
return None, None, "text", f"快卡路径不存在或不是文件: {fast_path}"
dt = _data_type_from_path(slow_path)
return slow_path, fast_path, dt, ""
cluster_path = os.path.abspath(cluster_path)
if not os.path.isdir(cluster_path):
return None, None, "text", f"集群数据路径不存在或不是目录: {cluster_path}"
# Resolve slow/fast paths (explicit or discovered)
def resolve_one(rank_id: int, explicit: Optional[str], get_csv=None, get_db=None):
if explicit and os.path.isfile(explicit):
if explicit.lower().endswith(".csv"):
return explicit, None
if explicit.lower().endswith(".db"):
return None, explicit
return (get_csv() if get_csv else None), (get_db() if get_db else None)
slow_csv, slow_db = resolve_one(
slow_rank, slow_path,
get_csv=lambda: find_rank_csv_path(cluster_path, slow_rank, csv_name),
get_db=lambda: find_rank_db_path(cluster_path, slow_rank),
)
fast_csv, fast_db = resolve_one(
fast_rank, fast_path,
get_csv=lambda: find_rank_csv_path(cluster_path, fast_rank, csv_name),
get_db=lambda: find_rank_db_path(cluster_path, fast_rank),
)
if slow_csv and fast_csv:
return slow_csv, fast_csv, "text", ""
if slow_db and fast_db:
return slow_db, fast_db, "db", ""
if slow_db and fast_db:
return slow_db, fast_db, "db", ""
missing = []
if not slow_csv and not slow_db:
missing.append(f"Rank {slow_rank}")
if not fast_csv and not fast_db:
missing.append(f"Rank {fast_rank}")
return (
None,
None,
"text",
f"未找到以下卡的数据(未找到 {csv_name} 或 ascend_pytorch_profiler_{{rank}}.db):{', '.join(missing)}。请确认集群路径或手动指定 --slow-path / --fast-path。",
)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |