Skip to content

CHANGELOG.md

Latest commit

 

History

History
138 lines (101 loc) · 4.88 KB

CHANGELOG.md

File metadata and controls

138 lines (101 loc) · 4.88 KB

Changelog

All notable changes to PyGPUkit will be documented in this file.

For detailed release notes with code examples, see README.md.

[0.2.17] - 2025-12-28

Added

  • Triton Backend MVP: Optional Triton backend for rapid kernel prototyping
  • pygpukit.triton module: TritonArray wrapper, from_gpuarray, triton_available
  • Triton Kernels: RMSNorm, LayerNorm, Softmax, Rotary
  • Hybrid Execution: Mix Triton + Native CUDA in same model
  • examples/chat_cli_triton.py: Hybrid chat example demonstrating Triton + CUDA

Fixed

  • TritonArray dtype mapping: Support both PascalCase and lowercase dtype strings

[0.2.16] - 2025-12-28

Added

  • MoE (Mixture of Experts): Full Mixtral support with TopK routing, grouped GEMM
  • Thinking Model: Qwen3 <think>...</think> block parsing
  • GEMV Kernels (SM120): FP8/FP8 (W8A8), NVF4/NVF4 (W4A4), Int4
  • GEMM Kernels (SM120): W8A16, Int8 native (dp4a), Int4 via Int8, Grouped GEMM v2
  • Claude Code Skills: Build, benchmark, lint, test automation
  • Subagents: kernel-reviewer, perf-analyzer, api-designer

Changed

  • Kernel directory restructure: {gemm|gemv}/{input}/{output}/{arch}/
  • Removed redundant slow kernels (FP8 GEMV basic, Int8 via FP8)

[0.2.15] - 2025-12-26

Added

  • Whisper ASR Module: Full encoder/decoder, preprocessing, streaming transcription
  • GEMV Kernels: BF16 (vectorized BF16x2), NVF4 (pre-scaled LUT)
  • FP8 I/O GEMM (SM120): Pure FP8 E4M3 input/output with blockwise scaling
  • Pure NVF4 GEMM: 446 TFLOPS with GPU-side quantization (170x vs CPU)
  • GPUArray improvements: Scalar arithmetic, transpose, reshape, indexing
  • GPU Transpose Kernels: 2D, 3D (0,2,1), 4D (0,2,1,3), 4D (0,1,3,2)
  • Math operations: sin, cos, sqrt, rsqrt, abs, neg, clamp, where, sigmoid, tanh, argmax, min, sum_axis
  • uint8/int8 NumPy support: from_numpy for FP8 data handling

Changed

  • Linear layer uses GEMV for M=1 decode (1.3-2.4x faster than matmul)
  • SM120 compatibility via CUTLASS fork with alignment fixes

[0.2.14] - 2025-12-23

Fixed

  • Windows wheel RECORD file: Missing licenses/LICENSE entry

[0.2.13] - 2025-12-23

Fixed

  • RECORD file generation: Dynamic dist-info folder detection

Changed

  • Moved benchmark/demo files to benchmarks/ and examples/

[0.2.12] - 2025-12-22

Added

  • GPU Audio Processing (no cuFFT dependency):
    • Time-Frequency: istft, griffin_lim
    • Spectral: spectral_centroid, spectral_bandwidth, spectral_rolloff, spectral_flatness, spectral_contrast
    • Pitch: detect_pitch_yin, detect_pitch_yin_frames, autocorrelation
    • Music: cqt, chroma_stft, chroma_cqt, zero_crossing_rate
    • Source Separation: hpss, harmonic, percussive
    • Time/Pitch: time_stretch, pitch_shift

[0.2.11] - 2025-12-21

Added

  • Batch decode: Near-linear speedup (6.83x at batch=8)
  • Decode strategies: DecodeM1, DecodeM1Graph, DecodeBatch, DecodeJacobi
  • Driver API: Async memory ops, pinned malloc
  • RTX 5090 (SM120): Full support via CUDA 13.x
  • Qwen2 architecture: QWEN2_SPEC for Qwen2/2.5
  • Audio ops: STFT, Mel filterbank, MFCC, VAD, streaming

Fixed

  • CUDA Graph stream fix (RoPE/SDPA properly captured)

[0.2.10] - 2025-12-19

Added

  • Dynamic cuBLASLt loading: True driver-only deployment
  • Descriptor caching: 2.67x faster (395ms -> 148ms for 224 matmuls)

Changed

  • CUDA Graph optimizations (eliminated GPU allocations)

[0.2.9] - 2025-12-17

Added

  • Unified LLM Interface: CausalTransformerModel with ModelSpec
  • Architecture support: GPT-2, LLaMA 2/3, Qwen2/2.5, Qwen3
  • Hybrid attention: Auto CPU/GPU switching
  • LLM operations: sdpa_causal, rope_inplace, silu, rmsnorm
  • Sharded models: Auto-load split safetensors

[0.2.7] - 2025-12-14

Added

  • CUTLASS epilogue fusion: Linear + Bias + GELU
  • Multi-SM kernels: SM80/86/89/90/100/120 optimized
  • Operations: transpose, bias_add_inplace, linear_bias_gelu

Changed

  • Complete public API exports with snake_case naming

[0.2.6] - 2025-12-12

Added

  • CUTLASS backend: Default GEMM (TF32: 31 TFLOPS, FP16/BF16: 63 TFLOPS)
  • Multi-LLM scheduling: Concurrent execution with VRAM budgets
  • FP16/BF16 TensorCore: Via CUTLASS

[0.2.5] - 2025-12-10

Added

  • FP16/BF16 support: Half and brain float types
  • Type conversion: astype() method
  • Reduction ops: sum, mean, max
  • Operator overloads: +, -, *, /, @

[0.2.4] - 2025-12-08

Added

  • Driver-only mode: No cudart dependency
  • Dynamic NVRTC: JIT loaded at runtime
  • TF32 TensorCore GEMM: PTX mma.sync with cp.async pipeline

For full release notes with code examples, see the README.md "What's New" sections.