vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : group all experts in a single ggml_mul_mat_id (#6505)

Browse source 
* ggml : group all experts in a single ggml_mul_mat_id
cuda : improve mmid row copy

* cuda : fix bin bcast with non-cont src0

* test-backend-ops : only run all mul mat tests for base types

* llama : disable moe offloading with SYCL

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
slaren 2024-04-18 15:18:48 +02:00 • committed by GitHub
parent 03c0946d73
commit 0d56246f4b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 971 additions and 821 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-cuda/convert.cu
View file

@ -45,6 +45,8 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
vals[ix] = x0[ix];
}
__syncthreads();
#pragma unroll
for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
if (need_check && i0 + iy + 2*threadIdx.x >= k) {