root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-05 16:24:34 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

Try find other single threaded operator to run

Browse Source
This commit is contained in:
Howard Su 2023-04-08 00:34:05 +08:00
parent 43dde039b0
commit 455f6f79bc
1 changed files with 26 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
ggml.c
View File

@ -9557,8 +9557,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} }
else else
{ {
int start = i;
int end = i;
if (i + 1 < cgraph->n_nodes) if (i + 1 < cgraph->n_nodes)
{ {
struct ggml_tensor * next = cgraph->nodes[i + 1]; struct ggml_tensor * next = cgraph->nodes[i + 1];
@ -9568,14 +9566,37 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
.type = GGML_TASK_COMPUTE | GGML_TASK_INIT, .type = GGML_TASK_COMPUTE | GGML_TASK_INIT,
.ith = 0, .ith = 0,
.nth = 1, .nth = 1,
.wsize = cgraph->work ? ggml_nbytes(cgraph->work) : 0, .wsize = 0,
.wdata = cgraph->work ? cgraph->work->data : NULL, .wdata = NULL,
}; };
workers[next_task].node = next;
thpool_add_work(ctx->tpool, ggml_graph_compute_thread, &workers[next_task]); thpool_add_work(ctx->tpool, ggml_graph_compute_thread, &workers[next_task]);
next_task++; next_task++;
if (i + 2 < cgraph->n_nodes)
{
struct ggml_tensor * prev = cgraph->nodes[i + 1];
struct ggml_tensor * next = cgraph->nodes[i + 2];
if (next->src0 != node && next->src1 != node && next->n_tasks == 1 &&
next->src0 != prev && next->src1 != prev
)
{
workers[next_task].params = (struct ggml_compute_params) {
.type = GGML_TASK_COMPUTE | GGML_TASK_INIT,
.ith = 0,
.nth = 1,
.wsize = 0,
.wdata = NULL,
};
workers[next_task].node = next;
thpool_add_work(ctx->tpool, ggml_graph_compute_thread, &workers[next_task]);
next_task++;
}
}
} }
} }
} }
params.type = GGML_TASK_COMPUTE; params.type = GGML_TASK_COMPUTE;