Skip to main content

PD disaggregation accuracy diagnostic tool for xPyD

Project description

xPyD-acc

PD disaggregation accuracy diagnostic tool.

When PD disaggregated inference produces wrong output, xPyD-acc pinpoints where the problem is: Prefill, KV transfer, or Decode.

Key Features

  • Logprobs comparison — token-by-token divergence detection between endpoints
  • KV cache analysis — numerical accuracy checks (max abs diff, cosine similarity)
  • Automated diagnostics — full pipeline: baseline → isolate → compare → report
  • Interactive REPL — exploratory comparison with live parameter tuning
  • Batch & offline modes — multi-dataset runs and file-based comparison

Install

pip install xpyd-acc

Quick Start

# Run full diagnostic
xpyd-acc diagnose \
  --baseline-url http://aggregated:8000 \
  --prefill-url http://prefill:8001 \
  --decode-url http://decode:8002 \
  --prompt "The quick brown fox"

# Compare logprobs between two endpoints
xpyd-acc compare-logprobs \
  --endpoint-a http://aggregated:8000 \
  --endpoint-b http://prefill:8001 \
  --prompt "Hello world"

Part of xPyD

Component Description
xPyD-proxy PD-aware reverse proxy
xPyD-sim PD disaggregation simulator
xPyD-bench Benchmarking tool
xPyD-acc Accuracy diagnostics (this repo)
xPyD-plan Project planning

Resources

License

Apache 2.0

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

xpyd_acc-0.1.0.tar.gz (261.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

xpyd_acc-0.1.0-py3-none-any.whl (196.3 kB view details)

Uploaded Python 3

File details

Details for the file xpyd_acc-0.1.0.tar.gz.

File metadata

  • Download URL: xpyd_acc-0.1.0.tar.gz
  • Upload date:
  • Size: 261.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.13

File hashes

Hashes for xpyd_acc-0.1.0.tar.gz
Algorithm Hash digest
SHA256 7820d31c3b451919cc77e7a3c64a68d0ed0f45aac3d9325c7645e634c20435ae
MD5 a2c1afc5c3b152a9fb901ba21ff338af
BLAKE2b-256 0801b16f855cb74ea4c48567f1bd6789fa36ebcedeff022c34544bb12b88147a

See more details on using hashes here.

File details

Details for the file xpyd_acc-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: xpyd_acc-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 196.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.13

File hashes

Hashes for xpyd_acc-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 08d5c719efc6b800940102b33c2bf755918a60eb1704c15b8c56af4cd34016d4
MD5 4ca8b1a5ea4e8326c9e83992b821ef2b
BLAKE2b-256 6cac27971f39e7df6a69c242b4f92704bb0d3fff0cefd9b3e48ac47457e446ff

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page