CUDA: use min compute capability of GPUs actually used (#2506)

cebtenzzre · web-flow · commit 4329d1acb01c · 2023-08-04T17:35:22.000+02:00
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
@@ -5347,7 +5347,8 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_
         } else {
             int min_compute_capability = INT_MAX;
             for (int id = 0; id < g_device_count; ++id) {
-                if (min_compute_capability > g_compute_capabilities[id]) {
+                if (min_compute_capability > g_compute_capabilities[id]
+                        && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
                     min_compute_capability = g_compute_capabilities[id];
                 }
             }

Original file line number	Diff line number	Diff line change
`@@ -5347,7 +5347,8 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_`
`5347`	`5347`	`} else {`
`5348`	`5348`	`int min_compute_capability = INT_MAX;`
`5349`	`5349`	`for (int id = 0; id < g_device_count; ++id) {`
`5350`		`- if (min_compute_capability > g_compute_capabilities[id]) {`
	`5350`	`+ if (min_compute_capability > g_compute_capabilities[id]`
	`5351`	`+ && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {`
`5351`	`5352`	`min_compute_capability = g_compute_capabilities[id];`
`5352`	`5353`	`}`
`5353`	`5354`	`}`