Repository navigation
chore: bump mlx-swift-lm to b620ae8 - #243
Merged
Merged
Conversation
…-lm#85) Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The server integration job for Qwen2.5-0.5B dies with SIGSEGV during prefill on the mlx-swift-lm b620ae8 bump, intermittently (4 of 5 attempts so far, passing on the rest). The CI log has no stack, so this turns on the Swift runtime backtracer for the E2E step and dumps the macOS crash report on failure. Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The integration job retries the server test up to three times, so a crash in attempt 1 that passes on attempt 2 shows as a green job. The crash report dump only ran when the whole job failed, which hid every such crash. Collect the report after each failed attempt instead, limited to reports newer than that attempt's start. Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The onPrefillDone save ran on the first streamed token, while the producer had already started the next decode step on the same KVCacheSimple. PromptCache.save read the live key/value arrays on the HTTP thread during that update. With mlx-swift f93decf freeing the old context before replacing it, the race surfaces as a use-after-free (SIGBUS in CI). Move the save to the .info case, which is the producer's last yield, so no cache writes happen after it. Early exits (stop sequence, client disconnect, error before .info) no longer save the prompt cache; multi-turn clients re-prefill those prompts. Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Points the submodule at the merge of SharpAI/mlx-swift-lm#85 (upstream ml-explore/mlx-swift-lm bb0d5b0 sync: LoRA training fix). SwiftLM does not reference the LoRA training code, so this is a pin update only.
🤖 Generated with Claude Code