From: wangshuai09 <redacted>
Date: Thu, 1 Aug 2024 02:39:05 +0000 (+0800)
Subject: cann: support q8_0 for Ascend backend (#8805)
X-Git-Tag: upstream/0.0.4488~988
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=c8a0090922bad576623de4aae227717085249262;p=pkg%2Fggml%2Fsources%2Fllama.cpp

cann: support q8_0 for Ascend backend (#8805)
---

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index f2766697..90ccf3e1 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -2381,10 +2381,10 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
     size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
     size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
 
+    ggml_cann_pool_alloc input_alloctor(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        ggml_cann_pool_alloc input_alloctor(
-            ctx.pool(), ggml_nelements(src1) * input_elem_size);
+        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
         input_buffer = input_alloctor.get();
 
         int64_t* input_cast_ne = src1->ne;