mlcommons / mlcommons/storage

KV-Cache does not support Object Storage

Open
#285 3 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Future KVCache TF
Dominant language
Python
Stars
203
Forks
67
Avg merge
20m
Merged PRs (30d)
8

Description

Problem

The KV-Cache workload does currently not support the use of any object storage

Background

All other workloads within MLCommons storage support the use of both file and object storage types, specifically:

  • Training
  • Checkpointing
  • VectorDB (as the backend)

The notable exception is KV-Cache. This support should be enabled IDEALLY in the following way:

  • As an optional, 4th tier (Tier-1 = GPU memory, Tier-2 = System memory, Tier-3 = (local) File storage), New: Tier-4 Object Storage

In a future version of KV-Cache, object storage could be used optionally as Tier-3, but this will likely require support for RDMA - GDS access, necessitating the use of GPUs, which is beyond the scope of this effort. Hence, Object Storage for this round should be as an optional 4th tier only.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing the KV-Cache workload and comparing its storage handling with the Training, Checkpointing, and VectorDB workloads mentioned in the issue. The work is done when KV-Cache can optionally use object storage as a fourth tier, while local file storage remains Tier-3.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
data, machine-learning
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.