Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: Support Q2_0 (#25430)

Browse Source 
* vulkan: Support Q2_0

The backend perf tests for mat-vec-mul weren't very good at first (worse than
q2_k), doubling the rows per workgroup made a big difference.

* reorder

* resolve merge conflict, adjust err threshold for f16->q2_0 set_rows
This commit is contained in:
Jeff Bolz
2026-07-17 08:42:59 +02:00
committed by GitHub
parent 0bd0ec6099
commit 788e07dc91
13 changed files with 266 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+7 -1
View File
@@ -2435,13 +2435,17 @@ struct test_set_rows : public test_case {
}
double max_nmse_err() override {
if (type_dst == GGML_TYPE_Q4_0 || type_dst == GGML_TYPE_Q4_1 || type_dst == GGML_TYPE_IQ4_NL ||
if (type_dst == GGML_TYPE_Q2_0 || type_dst == GGML_TYPE_Q4_0 || type_dst == GGML_TYPE_Q4_1 ||
type_dst == GGML_TYPE_IQ4_NL ||
type_dst == GGML_TYPE_Q5_0 || type_dst == GGML_TYPE_Q5_1 || type_dst == GGML_TYPE_Q8_0) {
// estimate what the max nmse error would be if one quantized value is
// off by one. The test values are distributed in [-1,1], so it'll be
// roughly (2.0 / 2^bits)^2, divided by the mean square value of the reference,
// which is roughly 0.25 times the number of elements.
double err_estimate = 1.0f/8.0f;
if (type_src == GGML_TYPE_F16 && type_dst == GGML_TYPE_Q2_0) {
err_estimate *= 4.0f;
}
if (type_dst == GGML_TYPE_Q5_0 || type_dst == GGML_TYPE_Q5_1) {
err_estimate /= 2.0f;
}
@@ -7964,6 +7968,7 @@ static const ggml_type all_types[] = {
GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
GGML_TYPE_Q8_0,
GGML_TYPE_Q1_0,
GGML_TYPE_Q2_0,
GGML_TYPE_MXFP4, GGML_TYPE_NVFP4,
GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
@@ -7978,6 +7983,7 @@ static const ggml_type base_types[] = {
GGML_TYPE_F32, GGML_TYPE_F16,
GGML_TYPE_Q8_0, // for I8MM tests
GGML_TYPE_Q1_0,
GGML_TYPE_Q2_0,
GGML_TYPE_Q4_0,
GGML_TYPE_Q4_1, // for I8MM tests
GGML_TYPE_Q4_K,