DeepLink-org — 15 mods

probing AGENTS.md

01

Instructions file CodexOpenCode

Instructions for DeepLink-org/probing, covering agent instructions — probing, module boundaries, coding principles, minimal diff (pragmatic) and high cohesion, low diffusion.

11 2d ago A 3,615 tokens original Apache-2.0

acme_health

02

DeepLink-org/probing

Skill Claude CodeCodex

Demo skill from probing-acme vendor package. Use to verify pip-installed extension discovery (skills + magics).

11 2d ago A 26 tokens original Apache-2.0

comm_bottleneck

03

DeepLink-org/probing

Skill Claude CodeCodex

Rank collective ops by total and p99 latency.

11 2d ago A 15 tokens original Apache-2.0

crash_triage

04

DeepLink-org/probing

Skill Claude CodeCodex

Crash triage: grep PROBING CRASH in rank logs; read spill JSON under {PROBINGDATADIR}/crash/ /latest.json.

11 2d ago A 38 tokens original Apache-2.0

gpu_pressure

05

DeepLink-org/probing

Skill Claude CodeCodex

GPU memory and utilization headroom.

11 2d ago A 10 tokens original Apache-2.0

health_overview

06

DeepLink-org/probing

Skill Claude CodeCodex

One-shot health check: CPU, GPU, tables, torch step progress, TorchProbe overhead, cluster nodes, NCCL profiler health.

11 2d ago A 31 tokens original Apache-2.0

job_health

07

DeepLink-org/probing

Skill Claude CodeCodex

Job-level health: cluster alive, step lag, slowdown proxy, comm trend.

11 2d ago A 19 tokens original Apache-2.0

memory_leak

08

DeepLink-org/probing

Skill Claude CodeCodex

Detect monotonic GPU memory growth across training steps.

11 2d ago A 14 tokens original Apache-2.0

module_bottleneck

09

DeepLink-org/probing

Skill Claude CodeCodex

Find slowest PyTorch modules in recent steps.

11 2d ago A 15 tokens original Apache-2.0

nccl_culprit_victim

10

DeepLink-org/probing

Skill Claude CodeCodex

NCCL proxy wait decomposition — culprit (sendgpuwait) vs victim (recvwait).

11 2d ago A 27 tokens original Apache-2.0

persistent_straggler

11

DeepLink-org/probing

Skill Claude CodeCodex

Find ranks that are slow on many steps (worstfraction), not just once.

11 2d ago A 22 tokens original Apache-2.0

slow_rank

12

DeepLink-org/probing

Skill Claude CodeCodex

Find straggler ranks via collective latency imbalance.

11 2d ago A 13 tokens original Apache-2.0

sre_triage

13

DeepLink-org/probing

Skill Claude CodeCodex

SRE first-response triage for distributed training incidents. Automates the manual checks from PyTorch/NCCL debugging runbooks.

11 2d ago A 31 tokens original Apache-2.0

training_hang

14

DeepLink-org/probing

Skill Claude CodeCodex

Training stalled — capture stacks and thread/blocking hints.

11 2d ago A 15 tokens original Apache-2.0

watchdog_timeout

15

DeepLink-org/probing

Skill Claude CodeCodex

Diagnose PyTorch NCCL watchdog timeouts with Flight Recorder collective sequence alignment.

11 2d ago A 20 tokens original Apache-2.0