🧹 Cleanup and fixes for TGI #96
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This is what should be used on TPUs.
Also, add timing checks in warmup test.
- TGI version updated from 2.0.3 to 0ff6ff60ada291840beed63d8bf458d6f9606f7f, that is essentially 2.3.0 + few fixes to get the v2 proto interface working again. - This update was done because otherwise debug logs were not working. This can be complicated if we need to debug something in TGI, and so far the only solution was a hack forcing to re-add the debug in the server. This is now fixed and with the Jetstream Pt generator logs are fine now. - Obviously there was a drawback 😖 Logs on threads spawned by the Pytorch/XLA generator were now all weird and always appearing even when debug was off. This has been fixed, but the workaround is not very nice (I set an env var). I think the multithread generator is going to go away soon anyway, so this should not be a big deal. - The new TGI version is built using Python 3.11, while so far with optimum-tpu we have worked on Python 3.10, because that is what they say it should be used on Pytorch/XLA front page. So the image has been updated with the python3.11 support and the required transformers installation for now, because it's easier since they run on separate processes. - TGI build process has changed a bit, so dockerfile has been changed accordingly. The text-generation-router-v2 is renamed into text-generation-router because that is what the launcher expects.
This extra step will make images leaner as it will avoid having the gpu dependencies.
mfuntowicz
approved these changes
Sep 27, 2024
| @@ -8,7 +8,7 @@ RUN tar -C /tgi -xf /tgi/sources.tar.gz --strip-components=1 | |||
|
|
|||
| # Build cargo components (adapted from TGI original Dockerfile) | |||
| # Note that the build image is aligned on the same Linux version as the base image (Debian bookworm/ Ubuntu 22.04) | |||
| FROM lukemathwalker/cargo-chef:latest-rust-1.77-bookworm AS chef | |||
| FROM lukemathwalker/cargo-chef:latest-rust-1.79-bookworm AS chef | |||
There was a problem hiding this comment.
What about using 1.81 (latest)?
There was a problem hiding this comment.
Umh I used the latest version that was on the TGI docker image.
| COPY --from=tgi /tgi/launcher launcher | ||
| RUN cargo build --release --workspace --exclude benchmark | ||
| RUN cargo build --profile release-opt | ||
|
|
||
| # Python base image | ||
| FROM ubuntu:22.04 AS base |
There was a problem hiding this comment.
ubuntu:24.04 or is it breaking stuff?
There was a problem hiding this comment.
All TPU packages and images are so far tuned for this version. I might update that if at some point I have time to, but I do not see a priority about it for now.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Many changes to prepare a proper TGI image compatible with Jetstream Pytorch.
In particular: