skills.rest
Rockyeast
RockyeastCOMMUNITY
Explore all skills by this author →
GitHub

collect-workloads

Collect and sanitize SGLang ML workloads into flashinfer-trace JSONL and safetensors.

Updated Aug 27, 2026
Search Tags:#sanitization#benchmarking#flashinfer#sglang#workloads#safetensors
One-click install
npx skills add https://github.com/Rockyeast/flashinfer-trace-dev --skill collect-workloads
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: collect-workloads
Source: https://github.com/Rockyeast/flashinfer-trace-dev/tree/main/.claude/skills/collect-workloads
Command: npx skills add https://github.com/Rockyeast/flashinfer-trace-dev --skill collect-workloads

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve?

Auto-collect real-world workloads produced by SGLang inferences using FlashInfer logging, then sanitize results for inclusion in the flashinfer-trace dataset and preparation for PR submissions.

Core Features & Use Cases

  • Auto-detect and collect workloads from definitions across op_types using SGLang + FlashInfer Level 10 logging.
  • Sanitize per-call dumps into flashinfer-trace JSONL and safetensors for compact, reproducible workloads.
  • Default to sglang mode for end-to-end collection, with a direct mode fallback when FlashInfer APIs are not wired into SGLang.
  • Map raw dumps to the flashinfer-trace schema, align plan/run dumps, and prepare metadata for PRs to HuggingFace.
  • Output is ready for per-definition HuggingFace PRs, including workload JSONL, safetensors, and definition metadata.

Quick Start

Run the collect_workloads.py script in sglang mode with your target definitions and a flashinfer-trace directory to start automatic collection, sanitization, and optional PR submission.

Frequently Asked Questions about collect-workloads

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I collect real-world ML workloads from SGLang inferences for benchmarking?▼

You can collect SGLang inference workloads automatically by running the collection script in sglang mode, which uses FlashInfer Level 10 logging to auto-detect and gather per-call dumps for benchmarking.

What is FlashInfer logging used for when gathering inference traces?▼

FlashInfer logging captures raw per-call dumps during SGLang inferences, filtering logs by fi_api tags and mapping plan/run dumps to a workload schema for reproducible dataset preparation.

How do I sanitize raw inference dumps into JSONL and safetensors for HuggingFace PRs?▼

The collection script sanitizes raw dumps into flashinfer-trace JSONL and safetensors formats, aligning plan and run dumps while preparing definition metadata for direct HuggingFace PR submission.

Can I collect workloads if FlashInfer APIs are not wired into my SGLang environment?▼

Yes, you can use the direct API mode as a fallback when FlashInfer APIs are not integrated into SGLang, allowing you to gather workloads without the default end-to-end sglang mode.

How do I configure environment variables to ensure safe and targeted workload data collection?▼

You can set environment variables like FLASHINFER_LOGLEVEL and FLASHINFER_DUMP_INCLUDE to control logging behavior, ensuring targeted and safe data collection exported to a flashinfer-trace directory.

What are the limitations of using direct mode versus sglang mode for workload collection?▼

Sglang mode provides end-to-end workload collection across multiple op_types definitions, while direct mode serves as a fallback when FlashInfer APIs are absent, lacking the full automated pipeline integration.

Related Skills

Explore More Data & Analytics Skills→
hey-konghey-kong

debug-cuda-crash

Diagnoses CUDA crashes in SGLang via level-based kernel API logging and crash dumps.

Community
Advanced
trueratruera
3.5k

trulens-evaluation-workflow

Orchestrates end-to-end TruLens evaluation workflows for LLM apps.

Official
Advanced
QingHuan-6QingHuan-6

debug-cuda-crash

Diagnose CUDA crashes in SGLang by exposing kernel API call traces and tensor metadata.

Community
Intermediate
annealing-inversionannealing-inversion

debug-distributed-hang

Diagnose distributed-inference hangs in SGLang across parallel setups using per-rank traces.

Community
Advanced