2025 winter project [trudging] 03_nf4_dequant - #49

Open

trudging wants to merge 36 commits into

InfiniTensor:2025-winter-projectfrom

trudging:2025-winter-project

Commits on Jan 19, 2026

feat.: add 2025_winter assignment
Ziminli
committed
fcdc822

Commits on Jan 20, 2026

docs: add more info in README.md
Ziminli
committed
b52d284
docs: enrich README with detailed domestic GPU platform support info.
Ziminli
committed
6ab326b
fix: rebuild tester_nv.o with CUDA 11.8 for compatibility
Ziminli
committed
9e7376e

Commits on Feb 3, 2026

refactor: relax the tolerance settings for flashAttention
Ziminli
committed
72ff7d2

Commits on Feb 4, 2026

Implement trace and flashAttention CUDA kernels
trudging
committed
2e3eb4f
Optimize kernels: warp shuffle reduction, tiled attention, improved code quality
trudging
committed
d4808b6
Performance tuning: grid-stride trace, larger attention tiles, __ldg cached loads
trudging
committed
b2ce4d6
Fix attention kernel: correct warp reduction, stable online softmax
trudging
committed
e30445f
Debug: force fallback kernel
trudging
committed
dd16a3a
Restrict optimized kernel to known-good dimensions
trudging
committed
4715318
Disable optimized kernel for correctness
trudging
committed
0f85a1c
Add MetaX (沐曦) platform adaptation
trudging
committed
b047f61
Add Moore Threads (摩尔线程) platform adaptation
trudging
committed
a19d735
Add Iluvatar (天数) platform compatibility with LDG macro
trudging
committed
6179d29

Commits on Feb 5, 2026

Commits on Feb 6, 2026

chore: update test standards (tester object files) from commit 65b2898
trudging
committed
309f99a
fix: remove extraneous closing brace in kernels.cu
trudging
committed
2a468e9
revert: restore float softmax for attention
trudging
committed
b3d44b4

Commits on Mar 14, 2026

feat(nf4): add NF4 double-dequantization CUDA kernel and test scripts
trudging
committed
8ad6610
Merge remote README
trudging
committed
e43ef1e
refactor(nf4): move trudging user files into trudging subdirectory
trudging
committed
5cb34c0
fix(nf4): correctly apply templates, params reading, and output generation
trudging
committed
09d90c5
Refactor nf4_dequant to support templates and speedup calculation
trudging
committed
7c2364a
Support MUSA, MACA, and ILUVATAR platforms seamlessly
trudging
committed
9c1f288

Commits on Mar 15, 2026