CUDA: fix partial offloading for ne0 % 256 != 0 (#8572)

Author: JohannesGaessler

ggml/include/ggml-backend.h

@@ -29,21 +29,23 @@ extern "C" {
     enum ggml_backend_buffer_usage {
         GGML_BACKEND_BUFFER_USAGE_ANY = 0,
         GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+        GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
     };
 
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+    GGML_API           const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API           void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API GGML_CALL void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API           bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API           enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
+    GGML_API           ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     //
     // Backend

ggml/src/ggml-alloc.c

@@ -776,6 +776,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
                 fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
                 return false;
             }
+            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
         }
     }
 

ggml/src/ggml-backend.c

@@ -134,6 +134,10 @@ void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backe
     }
 }
 
+enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
+    return buffer->usage;
+}
+
 ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
     return buffer->buft;
 }

ggml/src/ggml-cuda.cu

@@ -464,12 +464,12 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
         return;
     }
 
-    if (ggml_is_quantized(tensor->type)) {
+    if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
         // initialize padding to 0 to avoid possible NaN values
         size_t original_size = ggml_nbytes(tensor);
         size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
-        if (padded_size > original_size && tensor->view_src == nullptr) {
+        if (padded_size > original_size) {
             ggml_cuda_set_device(ctx->device);
             CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
         }
@@ -1485,6 +1485,13 @@ static void ggml_cuda_op_mul_mat(
             dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
         }
 
+        // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+            const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
+            const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
+            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
+        }
+
         if (src1_on_device && src1_is_contiguous) {
             dev[id].src1_ddf = (float *) src1->data;
         } else {