Merge pull request #179 from MollySophia/fix-linux-cuda

Fix build with CUDA under linux

Author: LaylBongers

CMakeLists.txt

@@ -172,7 +172,6 @@ elseif()
     set(GGML_STANDALONE ON)
 endif()
 
-set(BUILD_SHARED_LIBS OFF)
 if (NOT TARGET ggml)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt

rwkv_graph.inc

@@ -548,7 +548,7 @@ static struct ggml_tensor * rwkv_att_v6(
         ggml_reshape_1d(ctx, layer.att_time_decay, n_embed)
     );
 
-    w = rwkv_exp(ctx, ggml_neg_inplace(ctx, rwkv_exp(ctx, w)));
+    w = rwkv_exp(ctx, ggml_neg(ctx, rwkv_exp(ctx, w)));
     w = ggml_reshape_4d(ctx, w, 1, head_size, head_count, sequence_length);
 
     // dup is not strictly required; doing it just in case.
@@ -576,9 +576,9 @@ static struct ggml_tensor * rwkv_att_v6(
     x = rwkv_group_norm_eps_64e_minus5(ctx, x, head_count);
     // Convert back to a regular vector.
     x = ggml_reshape_2d(ctx, x, n_embed, sequence_length);
-    x = ggml_add_inplace(
+    x = ggml_add(
         ctx,
-        ggml_mul_inplace(
+        ggml_mul(
             ctx,
             x,
             layer.att_ln_x_weight

rwkv_operators.inc

@@ -158,7 +158,7 @@ static void rwkv_groupnorm_impl(
 
 // Element-wise exp(x)
 struct ggml_tensor * rwkv_exp(struct ggml_context * ctx, struct ggml_tensor * x) {
-    return ggml_map_custom1_inplace(ctx, x, rwkv_exp_impl, 1, NULL);
+    return ggml_map_custom1(ctx, x, rwkv_exp_impl, 1, NULL);
 }
 
 // Element-wise 1 - x