From: Cebtenzzre <redacted>
Date: Fri, 4 Aug 2023 15:35:22 +0000 (-0400)
Subject: CUDA: use min compute capability of GPUs actually used (#2506)
X-Git-Tag: gguf-v0.4.0~356
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=4329d1acb01c353803a54733b8eef9d93d0b84b2;p=pkg%2Fggml%2Fsources%2Fllama.cpp

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

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4321e467..d64d7045 100644
--- 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];
                 }
             }