CUDA: fix tensor core logic for Pascal and HIP

Author: JohannesGaessler

ggml-cuda.cu

@@ -134,7 +134,7 @@
 // TODO: improve this to be correct for more hardware
 //       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
 //       probably other such cases, and not sure what happens on AMD hardware
-#if !defined(GGML_CUDA_FORCE_MMQ)
+#if !defined(GGML_CUDA_FORCE_MMQ) && !defined(GGML_USE_HIPBLAS)
 #define CUDA_USE_TENSOR_CORES
 #endif
 
@@ -8663,7 +8663,7 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
     }
 
 #ifdef CUDA_USE_TENSOR_CORES
-    const bool use_tensor_cores = true;
+    const bool use_tensor_cores = min_compute_capability >= CC_VOLTA;
 #else
     const bool use_tensor_cores = false;
 #endif
@@ -8706,7 +8706,7 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
 
             // when tensor cores are available, use them for large batch size
             // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-            if (use_tensor_cores && min_compute_capability >= CC_VOLTA && src1->ne[1] > MMQ_MAX_BATCH_SIZE) {
+            if (use_tensor_cores && src1->ne[1] > MMQ_MAX_BATCH_SIZE) {
                 use_mul_mat_q = false;
             }