Repository navigation
fix(compile-hlo): read block-quant weights through the packed view - #1479
Merged
Merged
Conversation
BlockQuantMatmulConverter read Q8_0 / Q4_0 blocks from heap packedData only. Weights from the GGUF loader are off-heap BufferPackedTensorData, whose packedData throws, so exporting a real checkpoint failed. The converter now decodes the canonical ggml blocks (Q8_0: f16 d + 32 i8; Q4_0: f16 d + 16 nibble bytes) by copying rows out of packedView.storage, which works for heap and loader-backed weights alike. Found exporting the full FunctionGemma decode step (127 folded projections). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Nov 6, 2026
Merged
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Follow-up to #1478.
BlockQuantMatmulConverterread Q8_0 / Q4_0 blocks only from heappackedData. Weights loaded from a GGUF are off-heapBufferPackedTensorData, whosepackedDatathrows, so exporting a real checkpoint failed with "has no block accessors".The converter now decodes the canonical ggml blocks (Q8_0: f16
d+ 32i8; Q4_0: f16d+ 16 nibble bytes) by copying rows out ofpackedView.storage, which works for heap and loader-backed weights alike.Found while exporting the full FunctionGemma 270M decode step: all 127 projections fold, and the module runs on an SL2610 NPU with Torq 2.3.0 in one dispatch, picking the same top-1 token as the f32 reference.
Tests: new
offHeapStyleStorageReadsTheSameBlocks(a storage whosepackedDatathrows gives byte-identical codes, scales and zero points for Q8_0 and Q4_0). TheBlockQuantMatmulExportTestsuite andapiCheckpass.🤖 Generated with Claude Code