fr-analysis

bởi nvidia

Phân tích các bản ghi flight-recorder (FR) NCCL của PyTorch để xác định các lần treo hoạt động tập thể và cô lập các rank chịu trách nhiệm bằng CollectiveAnalyzer. Sử dụng khi một…

npx skills add https://github.com/nvidia/nvidia-resiliency-ext --skill fr-analysis

Tải ZIP GitHub

324

Dùng thử kỹ năng nàyĐược tài trợ

Skill: fr_analysis

Analyze PyTorch NCCL flight-recorder (FR) dumps to identify the collective operation hang and isolate the ranks responsible, using CollectiveAnalyzer.

Script: scripts/fr_attribution.py → attribution/trace_analyzer/fr_attribution.py

What it does

Loads all FR dump files matching a glob pattern under --fr-path.
Parses each dump into Collective records (op type, ranks, process group, timing, state).
Groups collectives by process group and sequence ID across ranks to detect mismatches.
Identifies the wavefront — the process group boundary where collectives diverge — and returns the missing ranks at that boundary as the root-cause suspects.
Optionally runs an LLM pass (--llm-analyze) over the structured findings for a human-readable summary.

CLI

python scripts/fr_attribution.py \
    --fr-path /path/to/fr_dumps/ \
    [-p "_dump_*"] \
    [--verbose] \
    [--health-check] \
    [--llm-analyze] \
    [--model MODEL] \
    [--debug]

Flag	Default	Description
`--fr-path`	required	Path to a directory (or single file) containing FR dump files
`--pattern`, `-p`	`_dump_*`	Glob pattern for dump files within `--fr-path`
`--verbose`, `-v`	off	Print detailed per-rank collective tables
`--health-check`, `-c`	off	Include node health check results in output
`--llm-analyze`, `-l`	off	Pass structured findings to the LLM for a narrative summary
`--model`, `-m`	`nvidia/nemotron-3-super-120b-a12b`	LLM model (only used with `--llm-analyze`)
`--debug`	off	Convert binary trace files to JSON for inspection

Programmatic API

from nvidia_resiliency_ext.attribution.trace_analyzer.fr_attribution import CollectiveAnalyzer

analyzer = CollectiveAnalyzer({
    "fr_path": "/path/to/fr_dumps/",
    "pattern": "_dump_*",
    "verbose": False,
    "health_check": False,
    "llm_analyze": False,
    "model": "nvidia/nemotron-3-super-120b-a12b",
})
results = analyzer.run_sync({
    "fr_path": "/path/to/fr_dumps/",
})
# results: tuple[FRAnalysisResult | str, AttributionState]

Output

Returns (result, AttributionState) where result is the FR analysis table and describes:

The selected wavefront/front process group
Missing ranks at that process group (root-cause suspects)
Per-rank collective status tables (when --verbose)
Node health summary (when --health-check)
LLM narrative (when --llm-analyze)

AttributionState.STOP indicates the hang is unrecoverable; CONTINUE indicates the job may be restartable after isolating the identified ranks.

Dump file formats

Format	Notes
`_dump_*` files	PyTorch FR dump prefix pattern used by the feedback loop
Binary pickle / JSON payloads	Detected automatically; use `--debug` to convert binary traces to JSON

FR dumps are typically written to the directory specified by TORCH_NCCL_DEBUG_INFO_TEMP_FILE or triggered automatically on NCCL timeout.