vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into xsn/fix_lora

Browse source
This commit is contained in:
ngxson 2024-07-10 19:52:39 +02:00
commit e68344cb06
79 changed files with 3369 additions and 411 deletions
Download patch file Download diff file Expand all files Collapse all files

90
convert_hf_to_gguf.py
View file

@ -265,7 +265,7 @@ class Model:
break
for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
data: np.ndarray = data # type hint
data: np.ndarray # type hint
n_dims = len(data.shape)
data_dtype = data.dtype
data_qtype: gguf.GGMLQuantizationType | None = None
@ -602,10 +602,6 @@ class Model:
tokenizer_path = self.dir_model / 'tokenizer.model'
tokens: list[bytes] = []
scores: list[float] = []
toktypes: list[int] = []
if not tokenizer_path.is_file():
raise FileNotFoundError(f"File not found: {tokenizer_path}")
@ -1363,7 +1359,7 @@ class LlamaModel(Model):
def set_vocab(self):
try:
self. _set_vocab_sentencepiece()
self._set_vocab_sentencepiece()
except FileNotFoundError:
try:
self._set_vocab_llama_hf()
@ -2127,7 +2123,7 @@ class InternLM2Model(Model):
logger.error(f'Error: Missing {tokenizer_path}')
sys.exit(1)
sentencepiece_model = model.ModelProto()
sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue]
sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
@ -2155,6 +2151,9 @@ class InternLM2Model(Model):
toktype = SentencePieceTokenTypes.UNUSED
elif tokenizer.IsByte(token_id):
toktype = SentencePieceTokenTypes.BYTE
# take care of ununsed raw token
if piece.startswith('[UNUSED'):
toktype = SentencePieceTokenTypes.UNKNOWN
tokens.append(text)
scores.append(score)
@ -2170,6 +2169,47 @@ class InternLM2Model(Model):
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
chat_eos_token = '<|im_end|>'
chat_eos_token_id = None
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
for token_id, foken_data in added_tokens_decoder.items():
token_id = int(token_id)
token = foken_data["content"]
if token == chat_eos_token:
chat_eos_token_id = token_id
token = token.encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert(tokens[token_id] == token)
tokens[token_id] = token
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
if foken_data.get("special"):
toktypes[token_id] = SentencePieceTokenTypes.CONTROL
tokenizer_file = self.dir_model / 'tokenizer.json'
if tokenizer_file.is_file():
with open(tokenizer_file, "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
added_tokens = tokenizer_json.get("added_tokens", [])
for foken_data in added_tokens:
token_id = int(foken_data["id"])
token = foken_data["content"]
if token == chat_eos_token:
chat_eos_token_id = token_id
token = token.encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert(tokens[token_id] == token)
tokens[token_id] = token
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
if foken_data.get("special"):
toktypes[token_id] = SentencePieceTokenTypes.CONTROL
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_tokenizer_pre("default")
self.gguf_writer.add_token_list(tokens)
@ -2179,28 +2219,16 @@ class InternLM2Model(Model):
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
old_eos = special_vocab.special_token_ids["eos"]
if "chat" in os.path.basename(self.dir_model.absolute()):
if chat_eos_token_id is not None:
# For the chat model, we replace the eos with '<|im_end|>'.
# TODO: this is a hack, should be fixed
# https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer)
logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
in chat mode so that the conversation can end normally.")
special_vocab.special_token_ids["eos"] = chat_eos_token_id
logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}"
" in chat mode so that the conversation can end normally.")
special_vocab.add_to_gguf(self.gguf_writer)
def _try_get_sft_eos(self, tokenizer):
unused_145_list = tokenizer.Encode('[UNUSED_TOKEN_145]')
im_end_list = tokenizer.Encode('<|im_end|>')
eos_token = None
assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1)
if len(unused_145_list) == 1:
eos_token = unused_145_list[0]
if len(im_end_list) == 1:
eos_token = im_end_list[0]
assert eos_token
return eos_token
def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int):
if n_head_kv is not None and n_head != n_head_kv:
n_head = n_head_kv
@ -2219,6 +2247,10 @@ in chat mode so that the conversation can end normally.")
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
self.gguf_writer.add_file_type(self.ftype)
if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
if self.hparams["rope_scaling"].get("type") == "linear":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
num_heads = self.hparams["num_attention_heads"]
@ -2979,16 +3011,16 @@ class T5Model(Model):
if not tokenizer_path.is_file():
raise FileNotFoundError(f"File not found: {tokenizer_path}")
sentencepiece_model = model.ModelProto()
sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue]
sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
# some models like Pile-T5 family use BPE tokenizer instead of Unigram
if sentencepiece_model.trainer_spec.model_type == 2: # BPE
if sentencepiece_model.trainer_spec.model_type == 2: # BPE
# assure the tokenizer model file name is correct
assert tokenizer_path.name == 'tokenizer.model'
return self._set_vocab_sentencepiece()
else:
assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM
assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM
add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
@ -3159,7 +3191,7 @@ class JaisModel(Model):
# but Jais's PyTorch model simply precalculates the slope values and places them
# in relative_pes.slopes
n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"]))
first_val = float(data_torch._data[0])
first_val = float(data_torch[0].item())
self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2)
return tensors
@ -3193,7 +3225,7 @@ class ChatGLMModel(Model):
def set_vocab_chatglm3(self):
dir_model = self.dir_model
hparams = self.hparams
tokens: list[bytearray] = []
tokens: list[bytes] = []
toktypes: list[int] = []
scores: list[float] = []
@ -3342,7 +3374,7 @@ class ChatGLMModel(Model):
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
self.gguf_writer.add_name(self.hparams.get("_name_or_path").split("/")[1]) # THUDM/glm4-9b-chat or THUDM/chatglm3-6b
self.gguf_writer.add_name(self.hparams["_name_or_path"].split("/")[1]) # THUDM/glm4-9b-chat or THUDM/chatglm3-6b
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)