tts : fix tensor shapes

Author: ggerganov

convert_hf_to_gguf.py

@@ -326,6 +326,8 @@ def prepare_tensors(self):
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
+                            gguf.MODEL_TENSOR.POS_NET_NORM1,
+                            gguf.MODEL_TENSOR.POS_NET_NORM2,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -2060,6 +2062,8 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_posnet_length      (self.hparams["n_embd_posnet"])
         self.gguf_writer.add_convnext_length    (self.hparams["n_embd_convnext"])
         self.gguf_writer.add_feed_forward_length(self.hparams["n_ff"])
+        self.gguf_writer.add_group_norm_eps     (self.hparams["group_norm_epsilon"])
+        self.gguf_writer.add_group_norm_groups  (self.hparams["group_norm_groups"])
 
 
 @Model.register("Qwen2MoeForCausalLM")

examples/tts/convert_pt_to_hf.py

@@ -98,6 +98,13 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
         if new_key.endswith("gamma"):
             new_key = new_key.replace("gamma", "gamma.weight")
 
+        # convert from 1D [768] to 2D [768, 1] so that ggml_add can broadcast the bias
+        if (new_key.endswith("norm.weight") or new_key.endswith("norm1.weight") or new_key.endswith("norm2.weight") or new_key.endswith(".bias")) and (new_key.startswith("backbone.pos_net") or new_key.startswith("backbone.embed.bias")):
+            value = value.unsqueeze(1)
+
+        if new_key.endswith("dwconv.bias"):
+            value = value.unsqueeze(1)
+
         size_mb = value.element_size() * value.nelement() / (1024 * 1024)
         print(f"{size_mb:8.2f} MB - {new_key}: {value.shape}")
 
@@ -154,6 +161,8 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
     "vocab_size": 4096,
     "n_head": 1,
     "layer_norm_epsilon": 1e-6,
+    "group_norm_epsilon": 1e-6,
+    "group_norm_groups": 32,
     "max_position_embeddings": 8192, # ?
     "num_hidden_layers": 12
 }

gguf-py/gguf/constants.py

@@ -125,6 +125,8 @@ class Attention:
         VALUE_LENGTH      = "{arch}.attention.value_length"
         LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
         LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
+        GROUPNORM_EPS     = "{arch}.attention.group_norm_epsilon"
+        GROUPNORM_GROUPS  = "{arch}.attention.group_norm_groups"
         CAUSAL            = "{arch}.attention.causal"
         Q_LORA_RANK       = "{arch}.attention.q_lora_rank"
         KV_LORA_RANK      = "{arch}.attention.kv_lora_rank"

gguf-py/gguf/gguf_writer.py

@@ -631,9 +631,6 @@ def add_context_length(self, length: int) -> None:
     def add_embedding_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
 
-    def add_embedding_length(self, length: int) -> None:
-        self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
-
     def add_features_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.FEATURES_LENGTH.format(arch=self.arch), length)
 
@@ -739,6 +736,12 @@ def add_layer_norm_eps(self, value: float) -> None:
     def add_layer_norm_rms_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
 
+    def add_group_norm_eps(self, value: float) -> None:
+        self.add_float32(Keys.Attention.GROUPNORM_EPS.format(arch=self.arch), value)
+
+    def add_group_norm_groups(self, value: int) -> None:
+        self.add_uint32(Keys.Attention.GROUPNORM_GROUPS.format(arch=self.arch), value)
+
     def add_causal_attention(self, value: bool) -> None:
         self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
 

gguf-py/tests/test_quants.py

@@ -136,7 +136,7 @@ def compare_tensors(t1: np.ndarray, t2: np.ndarray, qtype: GGMLQuantizationType)
         logger.debug(f"Sample bad block ({diff_bits[bad_block_id]} differing bits):\n{t1[bad_block_id]}\nReference:\n{t2[bad_block_id]}")
 
         sum_diff_bits = np.sum(diff_bits)
-        logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits/(x.size * 8):.6f}%)")
+        logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits / (x.size * 8):.6f}%)")
         return False