deepspeedai / deepspeedai/DeepSpeedExamples

[inference] when will deepspeed support inference with int8 gemm?

Open
#611 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
6.8k
Forks
1.1k
Avg merge
2d 16h
Merged PRs (30d)
1

Description

For now, deepspeed inference with int8 model is still use fp16 cublas gemm function and the quantized weights(int8) are dequantized back to fp16 as shown here:

https://github.com/microsoft/DeepSpeed/blob/master/csrc/transformer/inference/csrc/pt_binding.cpp#L863

When will it support real int8 inference which uses the int8 data to do gemm ? In this case, the activation or hidden states are also quantized to int8 data type which can reduce the memory requirements dramatically

Contributor guide

No contributing guide indexed for this repository

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 at csrc/transformer/inference/csrc/pt_binding.cpp around line 863, where the issue identifies the current inference behavior. Determine the scope needed for real int8 GEMM using quantized weights and activations without dequantizing weights to fp16; the issue does not name tests or a specific implementation plan.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp, python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.