Repository navigation
fix: correct RecurrentGemma prefill state contract - #2300
Merged
Merged
Conversation
Disable chunked prefill because no-ID multi-token forwards reset model-owned RGLRU state while preserving stable-ID continuation. Add a loaded one-layer regression that compares reset prompts with a fresh model and stable-ID split execution with one-pass logits. Focused CUDA validation is running under the orchestrator. Refs #2288
…ent-gemma-chunk-contract
inureyes
added
status:review
Preserve the newly merged MLX pin and overlay baseline declarations before validating the model prefill correction against current main. The model-specific implementation and synthetic regression are unchanged.
Merge current main through the mlxcelverse migration and revalidate loaded convolution and RGLRU state on MLX 243bdf09. All seven focused CUDA tests pass, including independent-prompt reset and stable-ID split equivalence. CUDA Clippy with warnings denied, formatting, and diff checks pass. Refs #2288
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Disable cache-level chunked prefill for RecurrentGemma because multi-token no-ID forwards intentionally reset model-owned RGLRU state. Add a loaded nonzero-weight regression proving independent prompts reset convolution and recurrent state, while stable sequence-ID chunks match one-pass logits.
Integrate main
5f334629, including MLX243bdf09and mlxcelverse, without source conflicts. The rebuilt CUDA test-fast binary passes all seven RecurrentGemma tests. CUDA Clippy with warnings denied, formatting, and diff checks pass. Earlier shared integration checks passed 123 Python tests and 49 WebUI fixtures; bilingual reports distinguish those results from the fresh final-pin validation.Review found no remaining blocking correctness, security, or performance issue. The capability is global, so every honoring caller now uses whole-prompt prefill, potentially increasing transient memory for long prompts. DirectEngine's stable-ID state handling remains intact. Validation uses loaded synthetic weights, not an external checkpoint.
Closes #2288