mozempk/ggml-org-llama-cpp-mirror
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-11-11 22:32:07 +01:00
Code Issues Packages Projects Releases Wiki Activity

opencl: add basic support for q4_1 (#19534)

Browse Source 
* opencl: add q4_1 mv

* opencl: clean up

* opencl: add flattened q4_1 mv

* opencl: clean up

* opencl: add basic q4_1 mm

* opencl: fix whitespace

* opencl: add general q4_0 mm
This commit is contained in:
lhez
2026-02-12 14:52:37 -08:00
committed by GitHub
parent 338085c69e
commit 79cc0f2daf
7 changed files with 1231 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/kernels/cvt.cl
+51
View File
@@ -46,6 +46,15 @@ struct block_q4_0
uint8_t qs[QK4_0 / 2];
};
//------------------------------------------------------------------------------
// block_q4_1
//------------------------------------------------------------------------------
struct block_q4_1 {
half d; // delta
half m; // min
uchar qs[QK4_1 / 2]; // nibbles / quants
};
//------------------------------------------------------------------------------
// block_q6_K
//------------------------------------------------------------------------------
@@ -148,6 +157,48 @@ kernel void kernel_restore_block_q4_0_noshuffle(
}
}
//------------------------------------------------------------------------------
// kernel_convert_block_q4_1
// Convert the block_q4_1 format to 2 separate arrays (AOS -> SOA).
// This kernel does not deshuffle the bits.
//------------------------------------------------------------------------------
kernel void kernel_convert_block_q4_1(
global struct block_q4_1 * src0,
global uchar * dst_q,
global half * dst_d,
global half * dst_m
) {
global struct block_q4_1 * b = (global struct block_q4_1 *) src0 + get_global_id(0);
global uchar * q = (global uchar *) dst_q + QK4_1/2*get_global_id(0);
global half * d = (global half *) dst_d + get_global_id(0);
global half * m = (global half *) dst_m + get_global_id(0);
*d = b->d;
*m = b->m;
for (int i = 0; i < QK4_1/2; ++i) {
q[i] = b->qs[i];
}
}
kernel void kernel_restore_block_q4_1(
global uchar * src_q,
global half * src_d,
global half * src_m,
global struct block_q4_1 * dst
) {
global struct block_q4_1 * b = (global struct block_q4_1 *) dst + get_global_id(0);
global uchar * q = (global uchar *) src_q + QK4_1/2*get_global_id(0);
global half * d = (global half *) src_d + get_global_id(0);
global half * m = (global half *) src_m + get_global_id(0);
b->d = *d;
b->m = *m;
for (int i = 0; i < QK4_1/2; ++i) {
b->qs[i] = q[i];
}
}
//------------------------------------------------------------------------------
// block_mxfp4
//------------------------------------------------------------------------------