tts : remove hardcoded constants

ggml-ci

Author: ggerganov

convert_hf_to_gguf.py

@@ -2055,7 +2055,11 @@ def set_vocab(self):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_vocab_size         (self.hparams["vocab_size"])
+        self.gguf_writer.add_features_length    (self.hparams["n_embd_features"])
+        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"])
 
 
 @Model.register("Qwen2MoeForCausalLM")

examples/tts/convert_pt_to_hf.py

@@ -147,6 +147,10 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
         "WavTokenizerDec"
     ],
     "hidden_size": 1282,
+    "n_embd_features": 512,
+    "n_embd_posnet": 768,
+    "n_embd_convnext": 768,
+    "n_ff": 2304,
     "vocab_size": 4096,
     "n_head": 1,
     "layer_norm_epsilon": 1e-6,

gguf-py/gguf/constants.py

@@ -90,6 +90,9 @@ class LLM:
         VOCAB_SIZE                        = "{arch}.vocab_size"
         CONTEXT_LENGTH                    = "{arch}.context_length"
         EMBEDDING_LENGTH                  = "{arch}.embedding_length"
+        FEATURES_LENGTH                   = "{arch}.features_length"
+        POSNET_LENGTH                     = "{arch}.posnet_length"
+        CONVNEXT_LENGTH                   = "{arch}.convnext_length"
         BLOCK_COUNT                       = "{arch}.block_count"
         LEADING_DENSE_BLOCK_COUNT         = "{arch}.leading_dense_block_count"
         FEED_FORWARD_LENGTH               = "{arch}.feed_forward_length"

gguf-py/gguf/gguf_writer.py

@@ -631,6 +631,18 @@ 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)
+
+    def add_posnet_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.POSNET_LENGTH.format(arch=self.arch), length)
+
+    def add_convnext_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.CONVNEXT_LENGTH.format(arch=self.arch), length)
+
     def add_block_count(self, length: int) -> None:
         self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)