Search results

Half-warp DP4A Q4K GEMV kernel (16 threads/SB)

Status: Open.

#175 In paiml/trueno;

· noahgift opened on Mar 6, 2026

GH-174: LM head Q6K GEMV takes 35% of decode time — multi-row blocking needed

Status: Open.

#174 In paiml/trueno;

· noahgift opened on Mar 6, 2026

perf: GEMV memory bandwidth optimization for Jetson Orin (sm_87)

Status: Open.

#173 In paiml/trueno;

· noahgift opened on Mar 6, 2026

feat: FusedGradClipKernel — GPU-side norm reduction + clip scale + apply

Status: Open.

#171 In paiml/trueno;

· noahgift opened on Mar 5, 2026

fix: disable tensor cores for FP32 backward GEMMs — NaN corruption (ALB-076)

Status: Open.

#170 In paiml/trueno;

· noahgift opened on Mar 5, 2026

feat: cuBLAS GEMM benchmark infrastructure (raw C + Criterion + regression)

enhancementNew feature or requestNew feature or request

Status: Open.

#162 In paiml/trueno;

· noahgift opened on Mar 5, 2026

CUDA-Parity: Merge-based GPU SpMV kernel (PTX)

Status: Open.

#119 In paiml/trueno;

· noahgift opened on Mar 4, 2026

perf: Backward GEMM kernels use naive implementation — 2.8% compute efficiency in QLoRA training

enhancementNew feature or requestNew feature or request

Status: Open.

#109 In paiml/trueno;

· noahgift opened on Mar 2, 2026

NF4 quantized GEMM kernel for QLoRA training

enhancementNew feature or requestNew feature or request

Status: Open.

#108 In paiml/trueno;

· noahgift opened on Mar 2, 2026

gemm_tiled PTX JIT fails CUDA_ERROR_ILLEGAL_ADDRESS for Qwen3-4B dims (H=2560, I=9728)

Status: Open.

#107 In paiml/trueno;

· noahgift opened on Mar 2, 2026

feat(backend): add reducibility classification to SIMD vs GPU dispatch

Status: Open.

#99 In paiml/trueno;

· noahgift opened on Feb 1, 2026

perf: Shape-Dependent Algorithm Dispatch with Cost Models

Status: Open.

#95 In paiml/trueno;

· noahgift opened on Jan 26, 2026