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

vulkan/cpu: Support f16 as SET_ROWS src. (#25432)

Browse Source 
* vulkan/cpu: Support f16 as SET_ROWS src.

This adds full support for f16 SET_ROWS (equivalent to f32) to vulkan and CPU
backends, and adds more backend tests.

* Set DenormPreserve 16 when supported, to try to fix failures on Intel

* tune error threshold

* update metal supports_op
This commit is contained in:
Jeff Bolz
2026-07-14 08:26:55 -05:00
committed by GitHub
parent 8ff8c4299d
commit 7cbd61002d
7 changed files with 150 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu.c
+6
View File
@@ -2863,6 +2863,12 @@ struct ggml_cplan ggml_graph_plan(
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SET_ROWS:
{
if (node->src[0]->type == GGML_TYPE_F16 && node->type != GGML_TYPE_F16) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SOFT_MAX:
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK:
ggml/src/ggml-cpu/ops.cpp
+16 -11
View File
@@ -5041,7 +5041,7 @@ static void ggml_compute_forward_set_rows_impl(
assert(ne0 == nc);
assert(ne2 == ne02);
assert(ne3 == ne03);
GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
assert(ne02 % ne11 == 0);
assert(ne03 % ne12 == 0);
@@ -5075,10 +5075,19 @@ static void ggml_compute_forward_set_rows_impl(
(const float *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc);
} else if constexpr (std::is_same_v<src_t, ggml_fp16_t>) {
memcpy(
if (dst->type == GGML_TYPE_F16) {
memcpy(
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3),
((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
rs);
} else {
float * wdata = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
ggml_fp16_to_fp32_row(
(const ggml_fp16_t *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
wdata, nc);
from_float(wdata,
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc);
}
} else {
GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
}
@@ -5107,16 +5116,12 @@ void ggml_compute_forward_set_rows(
} break;
case GGML_TYPE_F16:
{
if (dst->type == GGML_TYPE_F16) {
if (src1->type == GGML_TYPE_I64) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
} else if (src1->type == GGML_TYPE_I32) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
} else {
GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
}
if (src1->type == GGML_TYPE_I64) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
} else if (src1->type == GGML_TYPE_I32) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
} else {
GGML_ABORT("dst->type = %d (%s) not supported with src0->type = %d (%s)", dst->type, ggml_type_name(dst->type), src0->type, ggml_type_name(src0->type));
GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
}
} break;
default: