Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix for docs on deepspeed inference #482

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion docs/source/Inference.md
Original file line number Diff line number Diff line change
Expand Up @@ -147,7 +147,7 @@ Some commonly used command line flags are here. A full list of flags can be view

The **DeepSpeed DS4Sci_EvoformerAttention kernel** is a memory-efficient attention kernel developed as part of a collaboration between OpenFold and the DeepSpeed4Science initiative.

If your system supports deepseed, using deepspeed generally leads an inference speedup of 2 - 3x without significant additional memory use. You may specify this option by selecting the `--use_deepspeed_inference` argument.
If your system supports deepspeed, using deepspeed generally leads an inference speedup of 2 - 3x without significant additional memory use. You may specify this option by selecting the `--use_deepspeed_evoformer_attention` argument. An additional requirement for this option is the [CUTLASS repository](https://github.com/NVIDIA/cutlass). You will need to clone it and set environment variable `CUTLASS_PATH` to point to it, see [instructions](https://www.deepspeed.ai/tutorials/ds4sci_evoformerattention/).

If DeepSpeed is unavailable for your system, you may also try using [FlashAttention](https://github.com/HazyResearch/flash-attention) by adding `globals.use_flash = True` to the `--experiment_config_json`. Note that FlashAttention appears to work best for sequences with < 1000 residues.

Expand Down
Loading