ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-12-20 05:07:52 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

llama : run all KQV ops on the CPU with no KV offload (llama/5049)

Browse Source 
ggml-ci
This commit is contained in:
slaren 2024-01-20 16:05:49 +01:00 committed by Georgi Gerganov
parent 74da3e1757
commit 078b8e23bf
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 20 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
ggml-backend.c
View File

@ -1191,6 +1191,24 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
ggml_tallocr_t src_allocr = node_allocr(src);
GGML_ASSERT(src_allocr != NULL); // all inputs should be assigned by now
if (src_allocr != node_allocr) {
// create a copy of the input in the split's backend
size_t id = hash_id(src);
if (sched->node_copies[id][cur_backend_id] == NULL) {
ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
sched->node_copies[id][cur_backend_id] = tensor_copy;
node_allocr(tensor_copy) = cur_allocr;
SET_CAUSE(tensor_copy, "4.cpy");
int n_inputs = sched->splits[cur_split].n_inputs++;
GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
sched->splits[cur_split].inputs[n_inputs] = src;
}
node->src[j] = sched->node_copies[id][cur_backend_id];
#if 0
// check if the input is already in the split
bool found = false;
for (int k = 0; k < sched->splits[cur_split].n_inputs; k++) {
@ -1206,19 +1224,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
sched->splits[cur_split].inputs[n_inputs] = src;
}
// create a copy of the input in the split's backend
size_t id = hash_id(src);
if (sched->node_copies[id][cur_backend_id] == NULL) {
ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
sched->node_copies[id][cur_backend_id] = tensor_copy;
node_allocr(tensor_copy) = cur_allocr;
SET_CAUSE(tensor_copy, "4.cpy");
}
node->src[j] = sched->node_copies[id][cur_backend_id];
#endif
}
}
}