cont [no ci]

Author: ggerganov

examples/tts/convert_pt_to_hf.py

@@ -172,10 +172,6 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
         "n_embd": 768,
         "n_layer": 12
     },
-    #"n_embd_posnet": 768,
-    #"n_embd_convnext": 768,
-    #"n_layer_posnet": 6,
-    #"n_layer_convnext": 12
 }
 
 with open(path_dst + '/config.json', 'w') as f:

src/llama.cpp

@@ -2747,37 +2747,37 @@ struct llama_cparams {
 
 struct llama_layer_posnet {
     // resnet
-    struct ggml_tensor * norm1 = nullptr;
-    struct ggml_tensor * norm1_b = nullptr;
+    struct ggml_tensor * norm1;
+    struct ggml_tensor * norm1_b;
 
-    struct ggml_tensor * conv1 = nullptr;
-    struct ggml_tensor * conv1_b = nullptr;
+    struct ggml_tensor * conv1;
+    struct ggml_tensor * conv1_b;
 
-    struct ggml_tensor * norm2 = nullptr;
-    struct ggml_tensor * norm2_b = nullptr;
+    struct ggml_tensor * norm2;
+    struct ggml_tensor * norm2_b;
 
-    struct ggml_tensor * conv2 = nullptr;
-    struct ggml_tensor * conv2_b = nullptr;
+    struct ggml_tensor * conv2;
+    struct ggml_tensor * conv2_b;
 
     // attention
-    struct ggml_tensor * attn_norm = nullptr;
-    struct ggml_tensor * attn_norm_b = nullptr;
+    struct ggml_tensor * attn_norm;
+    struct ggml_tensor * attn_norm_b;
 
-    struct ggml_tensor * attn_q = nullptr;
-    struct ggml_tensor * attn_q_b = nullptr;
+    struct ggml_tensor * attn_q;
+    struct ggml_tensor * attn_q_b;
 
-    struct ggml_tensor * attn_k = nullptr;
-    struct ggml_tensor * attn_k_b = nullptr;
+    struct ggml_tensor * attn_k;
+    struct ggml_tensor * attn_k_b;
 
-    struct ggml_tensor * attn_v = nullptr;
-    struct ggml_tensor * attn_v_b = nullptr;
+    struct ggml_tensor * attn_v;
+    struct ggml_tensor * attn_v_b;
 
-    struct ggml_tensor * attn_o = nullptr;
-    struct ggml_tensor * attn_o_b = nullptr;
+    struct ggml_tensor * attn_o;
+    struct ggml_tensor * attn_o_b;
 
     // normalize
-    struct ggml_tensor * norm = nullptr;
-    struct ggml_tensor * norm_b = nullptr;
+    struct ggml_tensor * norm;
+    struct ggml_tensor * norm_b;
 };
 
 struct llama_layer_convnext {
@@ -2796,7 +2796,6 @@ struct llama_layer_convnext {
     struct ggml_tensor * gamma;
 };
 
-// TODO: separate into "llama_layer_enc" and "llama_layer_dec"
 struct llama_layer {
     llama_layer() {
         // initialize all pointers to NULL
@@ -3167,6 +3166,7 @@ struct llama_sbatch {
     // batch indices of the output
     std::vector<size_t> out_ids;
     std::vector<llama_sbatch_seq> seq;
+
     const llama_batch * batch = nullptr;
 
     // buffers for the ubatch