trueno-gpu 0.4.29

Pure Rust PTX generation for NVIDIA CUDA - no LLVM, no nvcc
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

//! Coalesced Q4_K GEMV Kernels with Optimized Memory Access
//!
//! - `CoalescedQ4KGemvKernel`: Scale loading via broadcast, vectorized qs
//! - `VectorizedQ4KGemvKernel`: Coalesced u32 loads for high bandwidth
//! - `MwvDp4aQ4KGemvKernel`: DP4A integer dot products with Q8_1-quantized activations
//! - `HalfWarpDp4aQ4KGemvKernel`: Half-warp (16 threads/SB) DP4A, 2x fewer insn/SB

mod coalesced_kernel;
mod fused_gate_up_swiglu_hw_dp4a;
mod hw_dp4a;
mod multi_warp_vectorized;
mod mwv_dp4a;
mod vectorized_kernel;
mod wide_kernel;

pub use coalesced_kernel::CoalescedQ4KGemvKernel;
pub use fused_gate_up_swiglu_hw_dp4a::FusedGateUpSwigluHwDp4aQ4KGemvKernel;
pub use hw_dp4a::HalfWarpDp4aQ4KGemvKernel;
pub use multi_warp_vectorized::MultiWarpVectorizedQ4KGemvKernel;
pub use mwv_dp4a::MwvDp4aQ4KGemvKernel;
pub use vectorized_kernel::VectorizedQ4KGemvKernel;
pub use wide_kernel::WideQ4KGemvKernel;