CUDA: mul_mat_id always on GPU for batches >= 32

ggerganov · Dec 20, 2023 · 54a9a77 · 54a9a77
1 parent 799fc22
commit 54a9a77
Showing 1 changed file with 28 additions and 7 deletions.
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
@@ -8782,8 +8782,6 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     // TODO: mmq/mmv support
 #endif
 
-    GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
-
     const int64_t nb11 = src1->nb[1];
     const int64_t nb1  =  dst->nb[1];
 
@@ -8812,13 +8810,24 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     ggml_tensor src1_row = *src1;
     ggml_tensor dst_row = *dst;
 
+    ggml_backend_type src1_original_backend = src1_row.backend;
+    ggml_backend_type  dst_original_backend =  dst_row.backend;
+
+    src1_row.backend = GGML_BACKEND_GPU;
+    dst_row.backend  = GGML_BACKEND_GPU;
+
     src1_row.extra = &src1_row_extra;
     dst_row.extra = &dst_row_extra;
 
-    char * src1_original = (char *) src1_extra->data_device[g_main_device];
-    char * dst_original  = (char *)  dst_extra->data_device[g_main_device];
+    char * src1_original = src1_original_backend == GGML_BACKEND_CPU ?
+        (char *) src1->data : (char *) src1_extra->data_device[g_main_device];
+    char * dst_original  =  dst_original_backend == GGML_BACKEND_CPU ?
+        (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device];
 
     if (src1->ne[1] == 1) {
+        GGML_ASSERT(src1_original_backend == GGML_BACKEND_GPU);
+        GGML_ASSERT(dst_original_backend  == GGML_BACKEND_GPU);
+
         for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
             //int32_t row_id;
             //CUDA_CHECK(cudaMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
@@ -8846,6 +8855,11 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
         src1_row_extra.data_device[g_main_device] = src1_contiguous;
         dst_row_extra.data_device[g_main_device]  =  dst_contiguous;
 
+        const cudaMemcpyKind src1_kind = src1_original_backend == GGML_BACKEND_CPU ?
+            cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
+        const cudaMemcpyKind dst_kind  = src1_original_backend == GGML_BACKEND_CPU ?
+            cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
+
         for (int32_t row_id = 0; row_id < n_as; ++row_id) {
             const struct ggml_tensor * src0_row = dst->src[row_id + 2];
 
@@ -8860,7 +8874,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
                 GGML_ASSERT(row_id >= 0 && row_id < n_as);
 
                 CUDA_CHECK(cudaMemcpyAsync(src1_contiguous + num_src1_rows*nb11, src1_original + i01*nb11,
-                                        nb11, cudaMemcpyDeviceToDevice, stream));
+                                        nb11, src1_kind, stream));
                 num_src1_rows++;
             }
 
@@ -8892,14 +8906,21 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
                 GGML_ASSERT(row_id >= 0 && row_id < n_as);
 
                 CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous + num_src1_rows*nb1,
-                                        nb1, cudaMemcpyDeviceToDevice, stream));
+                                        nb1, dst_kind, stream));
                 num_src1_rows++;
             }
         }
 
         ggml_cuda_pool_free(src1_contiguous, as_src1);
         ggml_cuda_pool_free(dst_contiguous,  as_dst);
     }
+
+    src1_row.backend = src1_original_backend;
+
+    if (dst_original_backend == GGML_BACKEND_CPU) {
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+    }
+    dst_row.backend  =  dst_original_backend;
 }
 
 static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -9298,7 +9319,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
         || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
 
-    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT) {
+    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
         return false;
     }