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

llama : add Command R Plus support (#6491)

Browse source 
* Add Command R Plus GGUF

* Add Command R Plus GGUF

* Loading works up to LayerNorm2D

* Export new tensors in 1D so they are not quantized.

* Fix embedding layer based on Noeda's example

* Whitespace

* Add line

* Fix unexpected tokens on MPS. Re-add F16 fix. ((Noeda)

* dranger003: Fix block index overflow in CUDA dequantizing.

* Reverted blocked multiplication code as it still has issues and could affect other Llama arches

* export norms as f32

* fix overflow issues during quant and other cleanup

* Type convention

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* dranger003: Fix more int overflow during quant.

---------

Co-authored-by: S <seast@Ss-Mac-Studio.local>
Co-authored-by: S <s@example.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Carolinabanana 2024-04-09 09:16:13 +01:00 • committed by GitHub
parent e11a8999b5
commit 5dc9dd7152
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 358 additions and 318 deletions
Download patch file Download diff file Expand all files Collapse all files

16
ggml-cuda/quantize.cu
View file

@ -1,20 +1,20 @@
#include "quantize.cuh"
static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) {
const int ix = blockDim.x*blockIdx.x + threadIdx.x;
static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx_padded) {
const int64_t ix = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
if (ix >= kx_padded) {
return;
}
const int iy = blockDim.y*blockIdx.y + threadIdx.y;
const int64_t iy = (int64_t)blockDim.y*blockIdx.y + threadIdx.y;
const int i_padded = iy*kx_padded + ix;
const int64_t i_padded = (int64_t)iy*kx_padded + ix;
block_q8_1 * y = (block_q8_1 *) vy;
const int ib = i_padded / QK8_1; // block index
const int iqs = i_padded % QK8_1; // quant index
const int64_t ib = i_padded / QK8_1; // block index
const int64_t iqs = i_padded % QK8_1; // quant index
const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
float amax = fabsf(xi);
@ -36,8 +36,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
reinterpret_cast<half&>(y[ib].ds.y) = sum;
}
void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) {
const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream) {
const int64_t block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
const dim3 num_blocks(block_num_x, ky, 1);
const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);