Skip to content

CHANGELOG.md

Latest commit

 

History

History
119 lines (96 loc) · 6.36 KB

CHANGELOG.md

File metadata and controls

119 lines (96 loc) · 6.36 KB
Raw

Changelog

[Unreleased] - 2026-04-22

Added

  • WebGPU Conv2d kernel (webgpu-engine.js): Direct convolution compute shader (molt_conv2d) with 16x16 workgroup, fma()-optimized inner loop, zero-padding support. Conv2d is 60% of PaddleOCR compute -- now GPU-accelerated via conv2d() and conv2dGPU() methods.
  • Node.js WASM inference harness (tests/e2e/run_paddleocr_wasm.js): Loads PaddleOCR WASM, discovers ONNX weights, measures instantiation time (11.0 ms), and enumerates exported OCR functions.
  • WASM SHA256 checksums: wasm/molt_runtime.wasm.sha256 and wasm/molt_runtime_reloc.wasm.sha256 for deploy integrity verification.

Performance

  • WASM instantiation: 11.0 ms (Node.js, Apple Silicon) for 10.8 MB PaddleOCR binary
  • Conv+Activation fusion: 62 nodes fused in ONNX graph optimization
  • Chinese OCR parity verified (你好世界 round-trip correct)

Deployed

  • Falcon-OCR Worker redeployed (version ee5fadc6) -- all endpoints verified (/test, /test/paddle, /dashboard, /health)
  • freeinvoicemaker.app verified live (HTTP 200)

[Unreleased] - 2026-04-14

Added

  • GPU inference proxy (gpu-proxy.js): External GPU service forwarding for bfloat16-quality Falcon-OCR inference. Supports HuggingFace Inference Endpoints, Replicate, RunPod, Modal, and Fly.io. Wired into Worker as X-Use-Backend: gpu route.
  • ScanButton UX phases: Extended InitProgressPhase with inferring, decoding, and done phases. Added OcrProgress interface for unified lifecycle progress reporting. MoltOcrBackend.recognize() now emits progress callbacks during inference.
  • Load test harness (tests/e2e/test_load.sh): Concurrent load testing for /health, /invoice/fill, and /ocr GPU proxy endpoints with latency statistics.
  • GPU inference status included in /health endpoint response

Changed

  • WASM binary analysis documented: 10.1 MB total (8.2 MB code / 9,934 functions, 1.8 MB data / 1,358 segments)
  • Production status docs updated with GPU proxy status, WASM analysis, and load test results

[Unreleased] - 2026-04-21

Performance

  • matmul_f32_tiled: K-loop unrolled by 4 with precomputed row pointers
  • Rust SIMD release profile restored to opt-level = "z" for the core parity artifact.

Changed

  • Rust-only matmul_f32_fast removed from the core SIMD artifact until it has cross-backend parity coverage and a separate size/performance lane.
  • Production status docs updated with SIMD benchmark numbers, CF services inventory, differential test results

[Unreleased] - 2026-04-20

Added

  • Falcon-OCR WASM compilation and execution (13 MB, 3.9 MB gzipped)
  • 26 tinygrad-conformant GPU primitives (molt-gpu crate)
  • 7 shader renderers (MSL, WGSL, GLSL, CUDA, HIP, OpenCL, MIL)
  • Workers AI OCR with retry logic and model fallback
  • x402 payment integration ($0.001/request USDC)
  • Template-from-scan feature
  • NL invoice filling
  • Tiered KV cache with H2O scoring
  • Browser WASM loader with IndexedDB caching
  • INT4 quantized model (129 MB, 5 shards)

Fixed

  • SCCP dead block elimination (SSA dominance violation)
  • WASM import alias resolution (frontend)
  • WASM linker table ref preservation (null function traps)
  • Turnstile iPad Safari login (explicit render mode)

Performance

  • Fused matmul: 10.9x speedup
  • SIMD everywhere (WASM, Rust, all shaders)
  • WASM binary: 44 MB -> 13 MB (strip + optimize)
  • 7,359 dead exports eliminated

[Unreleased] - 2026-03-20

WASM Optimization (Complete Sections 1-5 of WASM Optimization Plan)

Codegen Quality

  • br_table dispatch: O(1) state machine dispatch for generators/coroutines with 5+ states (2-5x faster resume)
  • Box/unbox elimination: Skip NaN-box tag checks for statically-typed integer ops (10-20% faster)
  • Dead local elimination: Shared __dead_sink variable reduces WASM local count (2-5% smaller)
  • Local variable coalescing: Greedy linear-scan reuse of __tmp/__v temporaries (5-15% smaller)
  • Constant folding: Compile-time evaluation of fast_int arithmetic on known constants (3-5% smaller)
  • Instruction combining: Const propagation through box/unbox, 5->2 insns for known-const unbox (3-8% faster)
  • local.tee: 37 eliminated redundant LocalGet instructions across arithmetic hot paths
  • Constant caching: Pre-materialized INT_SHIFT/INT_MIN/INT_MAX in function locals

WASM Proposals

  • Multi-value returns: Internal functions returning 2+ values use WASM multi-value ABI (eliminates tuple allocation)
  • Tail calls: return_call for tail-position calls in non-stateful functions
  • Native exception handling: try_table/catch/throw (gated by MOLT_WASM_NATIVE_EH=1)
  • Bulk memory: memory.fill for generator control blocks, memory_copy intrinsic
  • SIMD support: Full SIMD instruction handling in WASI stub rewriter

Binary Size & Startup

  • Full LTO: wasm-release profile with lto=true, codegen-units=1
  • wasm-opt integration: Oz/O3 pass pipelines (DCE, code folding, local CSE, inlining)
  • --precompile: Generate .cwasm via wasmtime compile (10-50x faster startup)
  • --wasm-profile pure: Compile-time IO/async/time import stripping (30-50% smaller for pure compute)

WASM Freestanding Target

  • --target wasm-freestanding: Post-link WASI import stubbing for zero-WASI binaries
  • Strict allowlist: --allow-undefined-file with 26 WASI + 14 trampoline symbols
  • wasm-validate integration: Optional post-build validation

Virtual Filesystem (VFS)

  • Mount-oriented VFS: /bundle (read-only), /tmp (ephemeral r/w), /dev (stdio), /state (stub)
  • BundleFs: In-memory from tar archive, symlink rejection, deterministic iteration
  • TmpFs: RwLock-based, quota enforcement, rename support, between-request clear()
  • DevFs: stdin/stdout/stderr pseudo-devices with 16 MB buffer cap
  • Capability system: fs.bundle.read, fs.tmp.read/write, fs.state.read/write
  • open() routing: VFS dispatch for WASM targets with fallback to std::fs
  • Module import: VFS-aware is_file/read helpers, /bundle in sys.path
  • Snapshot: molt.snapshot.json with SHA-256 integrity hash

Packaging & Deployment

  • --bundle: Tar archive packaging with manifest
  • --profile: cloudflare/browser/wasi/fastly presets
  • Worker template: Cloudflare Worker entry point generator

Testing

  • 34 Rust VFS unit tests
  • 32+ Python tests (freestanding, stubbing, bundle, worker, snapshot, size tracking)
  • 10 Rust WASM backend tests (br_table, dead local)