fix

Author: orca-zhang

ggml/src/ggml-cuda/pad.cu

@@ -74,8 +74,6 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->type == src0->type);
     GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-    GGML_LOG_INFO("ggml_cuda_op_pad: type=%ld, ne0=%d, ne1=%d, ne2=%d, ne3=%d, ne0=%d, ne1=%d, ne2=%d, ne3=%d\n",
-        src0->type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
 
     if (src0->type == GGML_TYPE_F32) {
         const float * src0_d = (const float *)src0->data;

src/llama.cpp

@@ -1338,9 +1338,8 @@ struct llm_build_context {
     }
 
     struct ggml_tensor * build_inp_KQ_mask(bool causal = true) {
-        const auto i = std::max(n_embd_head_k, n_embd_head_v);
         lctx.inp_KQ_mask = causal
-            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, cparams.flash_attn ? i * i : n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
+            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
             : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         cb(lctx.inp_KQ_mask, "KQ_mask", -1);
         ggml_set_input(lctx.inp_KQ_mask);