Repository navigation
Free the playground's GPU models before training; show why a run failed - #17
Merged
Merged
Conversation
A big base trained next to a cached inference copy of itself ran the GPU out of memory (Qwen3-8B MoRE on a 48 GB L40S: OOM at step 0, or accelerate offloading to meta and the load failing). The local runner now drops the inference cache and releases the allocator before it loads the base, and says so in the run's console.
The error only showed under the Artifact tab, so a failed run looked like one with no data yet.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
Two Qwen3-8B MoRE runs on the studio box failed at step 0 — one with
CUDA out of memory(44.37 of 44.39 GiB in use), one withCannot copy out of meta tensor. The playground had Qwen3-8B cached on the GPU, and training loaded a second copy next to it. The same config trains fine (200 steps) on a free card.What
serve.py: before the local runner loads a base, it drops the inference cache and releases the allocator (shared withclear_vram), and logsunloaded N playground models to free the GPU for trainingin the run's console.Runs.tsx: a failed run shows Why it failed above the tabs; it was only under Artifact._static.Testing
make test: 224 passed (2 new intests/test_serve_vram.py).