cuda : fix disabling device with --tensor-split 1,0

cebtenzzre · cebtenzzre · commit 699ea6d3b778 · 2023-11-05T00:35:07.000-04:00
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
@@ -7117,6 +7117,9 @@ static void ggml_cuda_op_mul_mat(
 
         CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         for (int64_t id = 0; id < g_device_count; ++id) {
+            if (g_tensor_split[id] >= (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
+                continue;
+            }
             for (int64_t is = 0; is < is_max; ++is) {
                 CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
             }

Original file line number	Diff line number	Diff line change
`@@ -7117,6 +7117,9 @@ static void ggml_cuda_op_mul_mat(`
`7117`	`7117`
`7118`	`7118`	`CUDA_CHECK(ggml_cuda_set_device(g_main_device));`
`7119`	`7119`	`for (int64_t id = 0; id < g_device_count; ++id) {`
	`7120`	`+ if (g_tensor_split[id] >= (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {`
	`7121`	`+ continue;`
	`7122`	`+ }`
`7120`	`7123`	`for (int64_t is = 0; is < is_max; ++is) {`
`7121`	`7124`	`CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));`
`7122`	`7125`	`}`