Repository navigation

feat: v0.2.16 - SM120 GEMV/GEMM optimization and comprehensive benchmarks - #121

Merged

m96-chan merged 51 commits into

mainfrom

feature/v0.2.16

Dec 28, 2025

Merged

feat: v0.2.16 - SM120 GEMV/GEMM optimization and comprehensive benchmarks#121
m96-chan merged 51 commits into
mainfrom
feature/v0.2.16

Commits on Dec 26, 2025
- docs: add Codon to Acknowledgements
  
  m96-chan
  and
  claude
  committed
  69f4871
- feat(moe): add MoE (Mixture of Experts) support for Mixtral
  
  m96-chan
  and
  claude
  committed
  d4ff0a2
- feat(examples): add chat_cli_moe.py for Mixtral inference
  
  m96-chan
  and
  claude
  committed
  97557aa
- feat(examples): add chat_cli_thinking.py for Qwen3 Thinking models
  
  m96-chan
  and
  claude
  committed
  14e7264
- feat(examples): add CUDA Graph support to chat_cli_thinking.py
  
  m96-chan
  and
  claude
  committed
  2afcea1
- feat(llm): add FP8 E4M3/E5M2 dtype support for quantized model loading
  
  m96-chan
  and
  claude
  committed
  56361d5
- fix(loader): use _native attribute for direct mmap transfer
  
  m96-chan
  and
  claude
  committed
  506e457
- feat(examples): add CUDA Graph support to chat_cli_moe.py
  
  m96-chan
  and
  claude
  committed
  0b4769a
- feat(fp8): add FP8 GEMV kernel with online dequantization
  
  m96-chan
  and
  claude
  committed
  a49cf5f
- refactor(llm): FP8 model loading without dequantization
  
  m96-chan
  and
  claude
  committed
  6acba7f
- feat(fp8): support LinearFP8 in MoE and Attention layers
  
  m96-chan
  and
  claude
  committed
  959de02
- refactor(layers): rename Linear to LinearBF16 for consistency
  
  m96-chan
  and
  claude
  committed
  cb43d99
- feat(claude): add skills and subagents for development workflow
  
  m96-chan
  and
  claude
  committed
  387e7ce
- docs: add CONTRIBUTING.md with contribution guidelines
  
  m96-chan
  and
  claude
  committed
  2f2e4a7
- docs: update README.md and CLAUDE.md with Claude Code config
  
  m96-chan
  and
  claude
  committed
  fb0d8a7
- docs: document matmul kernel directory structure
  
  m96-chan
  and
  claude
  committed
  f3b6f9a
- docs: update default GPU to RTX 5090
  
  m96-chan
  and
  claude
  committed
  bb29acd
- feat(mcp): add MCP server configuration
  
  m96-chan
  and
  claude
  committed
  22fab11
- feat(mcp): add Serena for semantic code analysis
  
  m96-chan
  and
  claude
  committed
  30791de
Commits on Dec 27, 2025
- feat(build): add automatic build log saving
  
  m96-chan
  and
  claude
  committed
  d9b7c9e
- refactor(native): reorganize matmul directory structure
  
  m96-chan
  and
  claude
  committed
  9c2128b
- docs: document LLM models directory (F:/LLM/)
  
  m96-chan
  and
  claude
  committed
  4ff5a89
- fix(moe): fix MoE layer output and add multi-template chat support
  
  m96-chan
  and
  claude
  committed
  eb4ce8a
- feat(fp8): add W8A16 GEMM kernel for SM120
  
  m96-chan
  and
  claude
  committed
  963292c
- docs: add more badges to README.md
  
  m96-chan
  and
  claude
  committed
  af4d090
- docs: add star request message to README
  
  m96-chan
  and
  claude
  committed
  58c1dbc
- feat(moe): add grouped GEMM infrastructure and uint8 concat support
  
  m96-chan
  and
  claude
  committed
  5f92ae8
- fix(moe): grouped GEMM v2 with per-row expert IDs
  
  m96-chan
  and
  claude
  committed
  2f4b3d1
- docs: sync README.md from main
  
  m96-chan
  and
  claude
  committed
  6a4ea0c
- perf(matmul): optimize GEMV FP8 and W8A16 GEMM kernels
  
  m96-chan
  and
  claude
  committed
  b01ad81
- feat(fp8): add FP8 GEMM v2 template and document SM120 constraints
  
  m96-chan
  and
  claude
  committed
  a788c41
- perf(w8a16): optimize W8A16 GEMM to 212 TFLOPS using FP8xFP8 kernel
  
  m96-chan
  and
  claude
  committed
  b7d7b13
- feat(gemm): add Int8 GEMM via FP8 TensorCore approximation for SM120
  
  m96-chan
  and
  claude
  committed
  5e6db6b
- feat(matmul): add Int4 GEMM via Int8/FP8 approximation (SM120)
  
  m96-chan
  and
  claude
  committed
  832ee47
- feat(matmul): add Int4 GEMV for M=1 decode (SM120)
  
  m96-chan
  and
  claude
  committed
  f0cecf3
- feat(gemm): add native Int8 GEMM using dp4a CUDA cores
  
  m96-chan
  and
  claude
  committed
  6d58e8a
- chore: cleanup W8A16, FP8, Int8 GEMM benchmarks and tests
  
  m96-chan
  and
  claude
  committed
  07b0005
- docs: update README with RTX 5090 benchmark results
  
  m96-chan
  and
  claude
  committed
  047619b
- docs: update README with RTX 5090 GEMV benchmarks
  
  m96-chan
  and
  claude
  committed
  3c7b31f
- feat(gemv): add pure FP8/FP8/FP8 GEMV kernel for SM120
  
  m96-chan
  and
  claude
  committed
  2920225
- perf(gemv): optimize FP8/FP8/FP8 GEMV with 128-bit loads
  
  m96-chan
  and
  claude
  committed
  dcc7dee
- feat(gemv): add pure NVF4/NVF4/NVF4 GEMV kernel for SM120
  
  m96-chan
  and
  claude
  committed
  9e2c8d2
- perf(gemv): optimize NVF4/NVF4 GEMV with row-major B layout
  
  m96-chan
  and
  claude
  committed
  3ac5c49
- perf(gemv): rewrite NVF4/NVF4 kernel with 1 thread = 1 output pattern
  
  m96-chan
  and
  claude
  committed
  8488970
- docs: add explicit GEMV quantization trade-offs section
  
  m96-chan
  and
  claude
  committed
  1e33bbd
- docs: add comprehensive GEMV benchmark results
  
  m96-chan
  and
  claude
  committed
  6f4396f
- Merge branch 'main' into feature/v0.2.16
  m96-chan
  committed
  6b8992f
Commits on Dec 28, 2025
- fix(lint): organize imports in bench_all_gemv.py
  m96-chan
  committed
  efd482b
- refactor(matmul): remove unused kernel variants
  
  m96-chan
  and
  claude
  committed
  a31bfb3
- fix(test): skip W8A16 GEMM tests when native module unavailable
  
  m96-chan
  and
  claude
  committed
  d418631
- fix(build): add missing cstdint include for uint8_t
  
  m96-chan
  and
  claude
  committed
  b9b2eb5

Commits

Commits on Dec 26, 2025

Commits on Dec 27, 2025

Commits on Dec 28, 2025