refactor: apply Burn patterns for code quality (Issue #14) #15
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
…coding, GGUF TASK-062: GGUF Import Complete - Parser: types, metadata, tensor info extraction - Loader: tensor data loading with memory mapping - Dequantization: K-quants (Q4_K, Q5_K, Q6_K, Q8_0) - Converter: GGUF to Born tensor format TASK-060: Flash Attention 2 - Week 1: CPU reference with online softmax O(N) memory - Week 2: WebGPU WGSL shader with tiled computation - Supports causal masking, head dims 64-256, block sizes 64/128 - GPU vs CPU validation < 1e-4 error TASK-061: Speculative Decoding - Draft model generates K tokens speculatively - Target model verifies in parallel batch - Modified rejection sampling for token acceptance - 2-4x speedup potential for autoregressive generation Also: Fixed 226 gosec G115 lint issues across codebase
- TASK-101: Pre-slice pattern for bounds check elimination - conv2d.go, maxpool2d.go: hierarchical pre-slicing - Reduced complexity in im2col and backward functions - TASK-104: Autodiff orchestration refactor - Separate orchestration from computation (Burn pattern) - New: conv2d_backward.go, maxpool2d_backward.go in CPU backend - Autodiff conv2d.go: 409 -> 67 lines (delegation only) - Extended Backend interface with backward methods - TASK-105: Flash Attention CPU refactor - Complexity: 111 -> <30, removed //nolint:gocognit - Extracted: FlashDims, FlashConfig structs - Helpers: flashAttentionScoreBlock, flashAttentionExtractValues - flashAttentionProcessQuery for single query processing Stats: +360/-460 lines, 13 files changed Tests: all pass, golangci-lint: 0 issues
TASK-102: Stride Specialization - Added dispatch for stride==1, padding==0 in Conv2D forward/backward - Specialized functions: conv2dFloat32Stride1NoPad, conv2dInputBackwardFloat32Stride1NoPad - Enables compiler auto-vectorization for common case TASK-103: Parallel Execution Utilities - New package: internal/parallel - parallel.Config, parallel.For(), parallel.ForBatch() - Automatic fallback to sequential for small workloads - Ready for integration into CPU backend operations Stats: +634 lines, 4 files changed Tests: all pass, golangci-lint: 0 issues
Closed
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
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.
Summary
Refactoring to reduce cognitive complexity in hot path functions, following patterns from the Burn ML framework (Rust).
Issue: #14 by @marcelloh
Changes
TASK-101: Pre-Slice Bounds Elimination
TASK-102: Stride Specialization
stride=1, padding=0caseTASK-103: Parallel Execution Utilities
internal/parallelpackage withFor()andForBatch()TASK-104: Autodiff Orchestration Refactor
conv2d_backward.go,maxpool2d_backward.goTASK-105: Flash Attention CPU Refactor
//nolint:gocognitdirectiveflashAttentionScoreBlock,flashAttentionProcessQueryMetrics
flashAttentionCPUconv2dBackwardInput*im2colFloat32/64Testing