Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-11 04:56:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

Remove ggml_cpu_iqp_scratch_offset and ggml_cpu_iqp_src1_conv_size

Browse Source
This commit is contained in:
Colin Kealty
2026-08-30 08:10:13 -04:00
parent ee71e14866
commit 1d7caa0180
3 changed files with 5 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu.c
+1 -2
View File
@@ -2885,8 +2885,7 @@ struct ggml_cplan ggml_graph_plan(
// the IQ panel path needs one scratch panel per thread past the q8_K rows
if (ggml_cpu_iqp_supported_mul_mat(node)) {
cur =
MAX(cur, ggml_cpu_iqp_scratch_offset(node) + n_tasks * ggml_cpu_iqp_scratch_size(node));
cur = GGML_PAD(cur, 64) + n_tasks * ggml_cpu_iqp_scratch_size(node);
}
} break;
case GGML_OP_MUL_MAT_ID:
ggml/src/ggml-cpu/iqp.cpp
+4 -14
View File
@@ -1111,12 +1111,6 @@ bool ggml_cpu_iqp_supported_mul_mat_id(const struct ggml_tensor * dst) {
return true;
}
size_t ggml_cpu_iqp_src1_conv_size(const struct ggml_tensor * dst) {
const enum ggml_type vec_dot_type = ggml_get_type_traits_cpu(dst->src[0]->type)->vec_dot_type;
return ggml_row_size(vec_dot_type, ggml_nelements(dst->src[1]));
}
void ggml_compute_forward_mul_mat_id_iqp(const struct ggml_compute_params * params,
struct ggml_tensor * dst,
int64_t cur_a,
@@ -1176,11 +1170,6 @@ void ggml_compute_forward_mul_mat_id_iqp(const struct ggml_compute_params * para
}
}
// ggml_graph_plan sizes the work buffer with these two helpers and the compute pass addresses it with them, so the two cannot drift
size_t ggml_cpu_iqp_scratch_offset(const struct ggml_tensor * dst) {
return GGML_PAD(ggml_cpu_iqp_src1_conv_size(dst), 64);
}
size_t ggml_cpu_iqp_scratch_size(const struct ggml_tensor * dst) {
return GGML_PAD((dst->src[0]->ne[0] / QK_K) * sizeof(block_iqp_x8), 64);
}
@@ -1201,10 +1190,11 @@ void ggml_compute_forward_mul_mat_iqp(const struct ggml_compute_params * params,
const size_t scratch_size = ggml_cpu_iqp_scratch_size(dst);
GGML_ASSERT(ggml_cpu_iqp_scratch_offset(dst) + (size_t) nth * scratch_size <= params->wsize);
const size_t scratch_offset = GGML_PAD(nbw2 * ne12, 64);
block_iqp_x8 * panel =
(block_iqp_x8 *) ((char *) params->wdata + ggml_cpu_iqp_scratch_offset(dst) + (size_t) ith * scratch_size);
GGML_ASSERT(scratch_offset + (size_t) nth * scratch_size <= params->wsize);
block_iqp_x8 * panel = (block_iqp_x8 *) ((char *) params->wdata + scratch_offset + (size_t) ith * scratch_size);
const int64_t nrows = ne11;
ggml/src/ggml-cpu/iqp.h
-5
View File
@@ -31,11 +31,6 @@ bool ggml_cpu_iqp_supported_mul_mat(const struct ggml_tensor * dst);
// node level test only - per expert eligibility is decided with ggml_cpu_iqp_expert_eligible
bool ggml_cpu_iqp_supported_mul_mat_id(const struct ggml_tensor * dst);
size_t ggml_cpu_iqp_src1_conv_size(const struct ggml_tensor * dst);
// offset of the panel scratch area inside the work buffer (past the q8_K conversion of src1)
size_t ggml_cpu_iqp_scratch_offset(const struct ggml_tensor * dst);
// per thread panel scratch bytes, padded
size_t ggml_cpu_iqp_scratch_size(const struct ggml_tensor * dst);