feat: multi-gpu inference, trajectory analyzer #2
+6,880
−1,175
Conversation
This file contains hidden or 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
Collaborator
BjarniHaukur
left a comment
BjarniHaukur
left a comment
There was a problem hiding this comment.
Looks good.
Had to do some ad-hoc change to get the infer script running. (and to run vllm serve myself)
Probably some artifact I can fix on my end. Possibly something with the pyproject.
I verified whether the performance degradation I observed with increased parallelism was still in play. And excitingly it does not look like it is! (8 concurrent and 32 concurrent got the same score)
🏆 Current Leaderboard
Performance on SWE-bench Lite subset, ranked by code similarity
| # | Ver | Model | Code Sim | Test Sim | Tokens | Tools |
|---|---|---|---|---|---|---|
| 11 | v5.0.1 | qwen3-32b | 0.276 | 0.000 | 4,409 / 16,384 | 23.2 / 100 |
| 12 | v5.0.1 | qwen3-32b | 0.273 | 0.005 | 5,514 / 16,384 | 32.1 / 100 |
| 13 | v3.2.0 | qwen-2.5-72b-instruct | 0.272 | 0.000 | 5,873 / 16,384 | 35.1 / 100 |
| 14 | v3.2.0 | qwen3-32b | 0.255 | 0.000 | 5,281 / 16,384 | 28.3 / 100 |
| 15 | v3.2.0 | llama-4-maverick | 0.255 | 0.000 | 4,647 / 16,384 | 10.4 / 100 |
| #SBATCH --array=0 | ||
|
|
||
| set -euo pipefail | ||
|
|
Collaborator
There was a problem hiding this comment.
Suggested change
| # === NSC Cluster Setup === | |
| # Load GCC build environment for Triton JIT compilation | |
| module load buildenv-gcccuda/12.1.1-gcc12.3.0 | |
| # Add Python 3.11 headers (required for Triton to compile cuda_utils) | |
| # These headers were extracted from Python source since python3.11-devel is not installed | |
| export CPATH="$HOME/.local/include/python3.11:${CPATH:-}" |
Collaborator
There was a problem hiding this comment.
had to add something like this to get it working. Obviously a hack.
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.
Most of the PR is the trajectory analyzer, but this also includes the changes to the dependencies that made it possible to run inference outisde of apptainer and on the full 8 GPU node with no issues from vllm coming up.
@BjarniHaukur you should be able to just
uv syncand run it. See benchmarks/swe_bench/run_harness_eval.sh for the batchscript.I created a pyproject from scratch since I was facing some version conflicts, but haven't added training dependencies back to it. Let me know if adding them to the new setup is not enough, we can debug if that's not the case. After that we can make a run to see if the parallelism is also working for training.