add chatglm3-6b model support

huggingface model: https://hf-mirror.com/THUDM/chatglm3-6b

Signed-off-by: XingXing Qiao <qiaoxx@dingdao.com>

Author: xingxingqiao

convert-hf-to-gguf.py

@@ -56,7 +56,7 @@ def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian:
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
 
     @property
     @abstractmethod
@@ -2903,6 +2903,165 @@ def write_tensors(self):
 
             self.gguf_writer.add_tensor(new_name, data)
 
+@Model.register("ChatGLMModel")
+class ChatGLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHATGLM
+
+    def set_vocab(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+        scores: list[float] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.get_vocab().items()}
+
+        for token_id in range(vocab_size):
+            piece = tokenizer._convert_id_to_token(token_id)
+            if token_id == 0:
+                piece = "<unk>"
+            elif token_id == 1:
+                piece = "<bos>"
+            elif token_id == 2:
+                piece = "<eos>"
+
+            text = piece.encode("utf-8")
+            score = 0.0
+            if len(piece) != 0 and token_id < 64789:
+                score = tokenizer.tokenizer.sp_model.get_score(token_id)
+
+            if len(piece) == 0:
+                text = f"[PAD{token_id}]".encode("utf-8")
+
+            if token_id >= 64789:
+                toktype = SentencePieceTokenTypes.UNKNOWN
+                tokens.append(text)
+                scores.append(score)
+                toktypes.append(toktype)
+                continue
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.tokenizer.sp_model.is_unknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.tokenizer.sp_model.is_control(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.tokenizer.sp_model.is_unused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.tokenizer.sp_model.is_byte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name("ChatGLM-6b-chat")
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_head_kv = self.hparams.get("multi_query_group_num", n_head)
+        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
+        self.gguf_writer.add_embedding_length(n_embed)
+        self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed))
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_dimension_count(64)
+        self.gguf_writer.add_add_bos_token(False)
+
+    def write_tensors(self):
+        block_count = self.hparams["num_layers"]
+        tensors = dict(self.get_tensors())
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        has_lm_head = True
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        for name, data_torch in tensors.items():
+            if name.endswith(".rotary_pos_emb.inv_freq"):
+                continue
+
+            if "lm_head.weight" not in tensors.keys() and "output.weight" not in tensors.keys():
+                has_lm_head = False
+
+            name = re.sub(r'transformer\.', '', name)
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+                # Map bloom-style qkv_linear to gpt-style qkv_linear
+                # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+                # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+                qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
+                data = np.concatenate(
+                    (
+                        qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.weight")
+            elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+                qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
+                data = np.concatenate(
+                    (
+                        qkv_bias[:, 0, :].reshape((n_embed,)),
+                        qkv_bias[:, 1, :].reshape((n_embed,)),
+                        qkv_bias[:, 2, :].reshape((n_embed,)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.bias")
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+            if not has_lm_head and name == "word_embeddings.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+                print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 ###### CONVERSION LOGIC ######
 

gguf-py/gguf/constants.py

@@ -138,6 +138,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R  = auto()
     DBRX       = auto()
     OLMO       = auto()
+    CHATGLM    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -215,6 +216,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.CHATGLM:        "chatglm",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -725,6 +727,18 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.CHATGLM : [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -761,6 +775,9 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.CHATGLM: [
+        MODEL_TENSOR.ROPE_FREQS,
+    ]
 }
 
 #

gguf-py/gguf/tensor_mapping.py

@@ -24,6 +24,7 @@ class TensorNameMap:
             "backbone.embedding",                        # mamba
             "backbone.embeddings",                       # mamba-hf
             "transformer.in_out_embed",                  # Grok
+            "embedding.word_embeddings",                 # chatglm
         ),
 
         # Token type embeddings
@@ -52,6 +53,7 @@ class TensorNameMap:
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
+            "output_layer",              # chatglm
         ),
 
         # Output norm
@@ -68,11 +70,13 @@ class TensorNameMap:
             "model.norm_f",                            # mamba-qbert
             "backbone.norm_f",                         # mamba
             "transformer.rms_norm",                    # Grok
+            "encoder.final_layernorm",                 # chatglm
         ),
 
         # Rope frequencies
         MODEL_TENSOR.ROPE_FREQS: (
             "rope.freqs",  # llama-pth
+            "rotary_pos_emb.inv_freq",  # chatglm
         ),
     }
 
@@ -97,6 +101,7 @@ class TensorNameMap:
             "backbone.layers.{bid}.norm",                           # mamba
             "transformer.decoder_layer.{bid}.rms_norm",             # Grok
             "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
+            "encoder.layers.{bid}.input_layernorm",                 # chatglm
         ),
 
         # Attention norm 2
@@ -118,6 +123,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
             "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
             "model.layers.{bid}.self_attn.qkv_proj"                                # phi3
+            "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
         ),
 
         # Attention query
@@ -173,6 +179,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
             "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
             "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
+            "encoder.layers.{bid}.self_attention.dense",                    # chatglm
         ),
 
         # Attention output norm
@@ -204,6 +211,7 @@ class TensorNameMap:
             "h.{bid}.ln_2",                                                  # gpt2
             "model.layers.{bid}.ffn_norm",                                   # internlm2
             "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
+            "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -240,6 +248,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
+            "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -299,6 +308,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.w2",                     # internlm2
             "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
+            "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.9.0"
+version = "0.9.1"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [