convert : add phi-3 support

convert-hf-to-gguf.py

@@ -62,6 +62,7 @@ def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian:
     def model_arch(self) -> gguf.MODEL_ARCH:
         pass
 
+    # TODO: add "default" argument
     def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
         key = next((k for k in keys if k in self.hparams), None)
         if key is not None:
@@ -89,7 +90,12 @@ def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                     yield name, data
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_name(self.dir_model.name)
+        if (mtype := self.find_hparam(["model_type"], optional=True)) is not None:
+            self.gguf_writer.add_name(mtype)
+            print(f"gguf: model type = {mtype}")
+        else:
+            self.gguf_writer.add_name(self.dir_model.name)
+
         self.gguf_writer.add_block_count(self.block_count)
 
         if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
@@ -332,7 +338,12 @@ def _set_vocab_sentencepiece(self):
         tokenizer = SentencePieceProcessor(str(tokenizer_path))
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
         for token_id in range(tokenizer.vocab_size()):
+
             piece = tokenizer.id_to_piece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.get_score(token_id)
@@ -347,23 +358,24 @@ def _set_vocab_sentencepiece(self):
             elif tokenizer.is_byte(token_id):
                 toktype = SentencePieceTokenTypes.BYTE
 
-            tokens.append(text)
-            scores.append(score)
-            toktypes.append(toktype)
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
 
         added_tokens_file = self.dir_model / 'added_tokens.json'
         if added_tokens_file.is_file():
             with open(added_tokens_file, "r", encoding="utf-8") as f:
                 added_tokens_json = json.load(f)
 
                 for key in added_tokens_json:
-                    key = key.encode("utf-8")
-                    if key not in tokens:
-                        tokens.append(key)
-                        scores.append(-1000.0)
-                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    token_id = added_tokens_json[key]
+                    if (token_id >= vocab_size):
+                        print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
 
-        assert len(tokens) == vocab_size
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_token_list(tokens)
@@ -1293,7 +1305,7 @@ def _stack_qk_norm(self, block_count, name, tensor_map, n_head, norms, n_dims, l
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM", "Phi3ForCausalLM")
 class LlamaModel(Model):
     model_arch = gguf.MODEL_ARCH.LLAMA
 
@@ -1322,18 +1334,39 @@ def set_vocab(self):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
+
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
         self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
 
     # Same as super class, but permuting q_proj, k_proj
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
         tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        n_embd = self.hparams.get("hidden_size")
         n_head = self.hparams.get("num_attention_heads")
         n_kv_head = self.hparams.get("num_key_value_heads")
         n_experts = self.hparams.get("num_local_experts")
         experts = dict()
-        for name, data_torch in self.get_tensors():
+
+        head_dim = n_embd // n_head
+
+        tensors = dict(self.get_tensors())
+        for i in range(block_count):
+            # Phi-3 transformations
+            # ref: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/8b29aca7bb785d6336fc19819b045bc7bc584b06/modeling_phi3.py#L379-L384
+            if (w := tensors.get(f"model.layers.{i}.self_attn.qkv_proj.weight")) is not None:
+                qpos = n_head * head_dim
+                tensors[f"model.layers.{i}.self_attn.q_proj.weight"] = w[:qpos]
+                tensors[f"model.layers.{i}.self_attn.k_proj.weight"] = w[qpos:qpos + n_kv_head * head_dim]
+                tensors[f"model.layers.{i}.self_attn.v_proj.weight"] = w[qpos + n_kv_head * head_dim:]
+                del tensors[f"model.layers.{i}.self_attn.qkv_proj.weight"]
+            if (w := tensors.get(f"model.layers.{i}.mlp.gate_up_proj.weight")) is not None:
+                ff_dim = w.shape[0] // 2
+                tensors[f"model.layers.{i}.mlp.gate_proj.weight"] = w[:ff_dim]
+                tensors[f"model.layers.{i}.mlp.up_proj.weight"] = w[ff_dim:]
+                del tensors[f"model.layers.{i}.mlp.gate_up_proj.weight"]
+
+        for name, data_torch in tensors.items():
             # we don't need these
             if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
                 continue

llama.cpp

@@ -4352,6 +4352,7 @@ static void llm_load_vocab(
                         //vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL &&
                         (t.first == "<|eot_id|>" ||
                          t.first == "<|im_end|>" ||
+                         t.first == "<|end|>" ||
                          t.first == "<end_of_turn>"
                         )
                    ) {