Skip to content

v3.4.0 – Parquet Engine, Quantized LLMs, YOLOv8-Seg, and ReAct Agents

Latest
Latest

Choose a tag to compare

View all tags
@Nodibell Nodibell released this 25 Aug 22:14
v3.4.0
f511075

🚀 SwiftSci v3.4.0

We are thrilled to announce the release of SwiftSci v3.4.0, delivering massive breakthroughs across high-throughput data processing, out-of-core memory streaming, localized quantized LLM execution on Apple Silicon GPU, token-level constrained decoding, instance segmentation, and autonomous reasoning agents.

🌟 Key Features & Highlights

📦 1. Out-of-Core Data Streaming & Apache Parquet (SwiftDataFrame)

  • ChunkedDataFrame: AsyncSequence streaming pipeline enabling memory-efficient lazy processing of datasets far exceeding RAM capacity.
  • MemoryMappedReader: Zero-copy POSIX mmap/munmap memory mapping directly bridging NVMe disk storage into unified virtual memory.
  • Pure-Swift Apache Parquet Engine: Native ParquetReader & ParquetWriter with single-pass SnappyDecompressor (LZ77) and Thrift Compact Protocol binary metadata parsing without external C dependencies.
  • SIMD Typed Hash Joins: SIMD-accelerated typed hash join algorithms for Double, Float, Int64, Int32, String, and Bool across .inner, .left, .right, and .outer joins.

🧠 2. Quantized LLMs & Constrained Decoding (SwiftLLM)

  • QuantizedLinear on Metal GPU: High-performance 4-bit (q4_0, q4_1, awq4) and 8-bit (q8_0, awq8) quantized matrix multiplications running on Apple Silicon Metal GPU via MLX.
  • JSONGrammarDecoder: Deterministic finite automaton (DFA) state machine masking token logits to $-\infty$ (-1e9) during autoregressive decoding, guaranteeing 100% compliant Codable JSON schema generation.
  • PagedKVCache: Dynamic unified memory block allocator managing fixed physical pages (pageSize: 16), completely eliminating memory fragmentation during long text generation.

👁️ 3. Advanced Vision & Instance Segmentation (SwiftVision)

  • YOLOSegHead (YOLOv8-Seg): Prototype mask generator producing 32 prototype masks (160x160) alongside mask coefficient regression branches and fast parallel GPU decodeMask reconstruction.
  • CLIPProjector: Vision-language multimodal feature projector with $L_2$-normalization, temperature-scaled cosine similarity logits, and zero-shot classification on Metal GPU.

🤖 4. Autonomous ReAct Reasoning Loop (SwiftAgent)

  • ReActAgent: Full autonomous decision cycle (Thought -> Action -> Action Input -> Observation -> Final Answer).
  • DataFrameAgentTool & CustomAgentTool: Sandboxed execution of analytical queries (filter, sample, select, head, tail, rename, dropnulls, fillnulls, groupby) over active DataFrames.

📚 Ecosystem & Documentation

  • 100.00% DocC API Coverage verified across 1,459 public symbols.
  • Fully unified interactive documentation portal generated in docs/.
  • Updated Engineering Handbook (Book/), README.md, and PERFORMANCE.md.

🧪 Verification

  • 100% test pass rate across all 15 workspace targets.
  • Strict Concurrency: Compliant with Swift 6 strict concurrency checks.

What's Changed

  • chore(deps): bump actions/deploy-pages from 4 to 5 by @dependabot[bot] in #22
  • chore(deps): bump actions/configure-pages from 5 to 6 by @dependabot[bot] in #23
  • chore(deps): bump actions/checkout from 4 to 7 by @dependabot[bot] in #24
  • Release: SwiftSci v3.4.0 — Parquet Engine, Quantized LLMs, YOLOv8-Seg, and ReAct Agents by @Nodibell in #25

Full Changelog: v3.3.0...v3.4.0

Contributors

dependabot and Nodibell
Assets 2
Loading