model : Plamo3 support (#17304)
* plamo3 * fix plamo3 * clean code * clean up the code * fix diff * clean up the code * clean up the code * clean up the code * clean up the code * clean up the code * clean up the code * add chat_template if exist * clean up the code * fix cpu-backend * chore: whitespace trim fix + typo fix * Fix: address review feedback * restore `FREQ_BASE_SWA` constant * Fix: address review feedback2 * Fix:typecheck * Fix: address review feedback3 * final cleanup --------- Co-authored-by: mmngays <146910567+mmngays@users.noreply.github.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
momonga
co-authored by
mmngays
Sigbjørn Skjæret
mmngays
Sigbjørn Skjæret
parent
07a0c4ba92
commit
9c675c7140
+129
-81
@@ -1696,6 +1696,84 @@ class TextModel(ModelBase):
|
||||
if template is not None:
|
||||
self.gguf_writer.add_chat_template(template)
|
||||
|
||||
def _set_vocab_plamo(self):
|
||||
# PLaMo models use a custom tokenizer with a .jsonl file
|
||||
tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
|
||||
tokenizer_config_path = self.dir_model / "tokenizer_config.json"
|
||||
|
||||
if not tokenizer_jsonl_path.is_file():
|
||||
raise FileNotFoundError(f"PLaMo tokenizer file not found: {tokenizer_jsonl_path}")
|
||||
|
||||
# Load tokenizer config
|
||||
with open(tokenizer_config_path, "r", encoding="utf-8") as f:
|
||||
tokenizer_config = json.load(f)
|
||||
|
||||
# Load tokens from JSONL file (actually a list format)
|
||||
tokens = []
|
||||
scores = []
|
||||
toktypes = []
|
||||
|
||||
with open(tokenizer_jsonl_path, "r", encoding="utf-8") as f:
|
||||
for line_num, line in enumerate(f):
|
||||
if line.strip():
|
||||
token_data = json.loads(line)
|
||||
# Format: [token, score, type, ?, ?, ?, ?]
|
||||
token = token_data[0].encode("utf-8")
|
||||
score = float(token_data[1])
|
||||
token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
|
||||
|
||||
tokens.append(token)
|
||||
scores.append(score)
|
||||
|
||||
if token_type_str == "UNKNOWN":
|
||||
toktypes.append(gguf.TokenType.UNKNOWN)
|
||||
elif token_type_str == "CONTROL":
|
||||
toktypes.append(gguf.TokenType.CONTROL)
|
||||
elif token_type_str == "BYTE":
|
||||
toktypes.append(gguf.TokenType.BYTE)
|
||||
else:
|
||||
token_str = token_data[0]
|
||||
if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
|
||||
toktypes.append(gguf.TokenType.CONTROL)
|
||||
else:
|
||||
toktypes.append(gguf.TokenType.NORMAL)
|
||||
|
||||
vocab_size = self.hparams["vocab_size"]
|
||||
if vocab_size > len(tokens):
|
||||
pad_count = vocab_size - len(tokens)
|
||||
logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
|
||||
for i in range(1, pad_count + 1):
|
||||
tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
|
||||
scores.append(-1000.0)
|
||||
toktypes.append(gguf.TokenType.UNUSED)
|
||||
|
||||
self.gguf_writer.add_tokenizer_model("plamo2")
|
||||
self.gguf_writer.add_tokenizer_pre("default")
|
||||
self.gguf_writer.add_token_list(tokens)
|
||||
self.gguf_writer.add_token_scores(scores)
|
||||
self.gguf_writer.add_token_types(toktypes)
|
||||
|
||||
if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_bos_token_id(token_id)
|
||||
if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_eos_token_id(token_id)
|
||||
if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_pad_token_id(token_id)
|
||||
if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_sep_token_id(token_id)
|
||||
if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_unk_token_id(token_id)
|
||||
|
||||
# Add <|plamo:op|> as EOT to ensure appropriate end of generation
|
||||
self.gguf_writer.add_eot_token_id(4)
|
||||
|
||||
self.gguf_writer.add_add_space_prefix(False)
|
||||
|
||||
|
||||
class MmprojModel(ModelBase):
|
||||
model_type = ModelType.MMPROJ
|
||||
@@ -4798,87 +4876,7 @@ class Plamo2Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.PLAMO2
|
||||
|
||||
def set_vocab(self):
|
||||
# PLaMo 2 uses a custom tokenizer with a .jsonl file
|
||||
# We need to handle this specially
|
||||
tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
|
||||
tokenizer_config_path = self.dir_model / "tokenizer_config.json"
|
||||
|
||||
if not tokenizer_jsonl_path.is_file():
|
||||
raise FileNotFoundError(f"PLaMo 2 tokenizer file not found: {tokenizer_jsonl_path}")
|
||||
|
||||
# Load tokenizer config
|
||||
with open(tokenizer_config_path, 'r', encoding='utf-8') as f:
|
||||
tokenizer_config = json.load(f)
|
||||
|
||||
# Load tokens from JSONL file (actually a list format)
|
||||
tokens = []
|
||||
scores = []
|
||||
toktypes = []
|
||||
|
||||
with open(tokenizer_jsonl_path, 'r', encoding='utf-8') as f:
|
||||
for line_num, line in enumerate(f):
|
||||
if line.strip():
|
||||
token_data = json.loads(line)
|
||||
# Format: [token, score, type, ?, ?, ?, ?]
|
||||
token = token_data[0].encode("utf-8")
|
||||
score = float(token_data[1])
|
||||
token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
|
||||
|
||||
tokens.append(token)
|
||||
scores.append(score)
|
||||
|
||||
# Map token type strings to GGUF token types
|
||||
if token_type_str == "UNKNOWN":
|
||||
toktypes.append(gguf.TokenType.UNKNOWN)
|
||||
elif token_type_str == "CONTROL":
|
||||
toktypes.append(gguf.TokenType.CONTROL)
|
||||
elif token_type_str == "BYTE":
|
||||
toktypes.append(gguf.TokenType.BYTE)
|
||||
else:
|
||||
# Check for PLaMo-2 special tokens
|
||||
token_str = token_data[0]
|
||||
if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
|
||||
toktypes.append(gguf.TokenType.CONTROL)
|
||||
else:
|
||||
toktypes.append(gguf.TokenType.NORMAL)
|
||||
|
||||
vocab_size = self.hparams["vocab_size"]
|
||||
if vocab_size > len(tokens):
|
||||
pad_count = vocab_size - len(tokens)
|
||||
logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
|
||||
for i in range(1, pad_count + 1):
|
||||
tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
|
||||
scores.append(-1000.0)
|
||||
toktypes.append(gguf.TokenType.UNUSED)
|
||||
|
||||
# Use "plamo2" tokenizer type for PLaMo-2's custom Aho-Corasick tokenizer
|
||||
self.gguf_writer.add_tokenizer_model("plamo2")
|
||||
self.gguf_writer.add_tokenizer_pre("default")
|
||||
self.gguf_writer.add_token_list(tokens)
|
||||
self.gguf_writer.add_token_scores(scores)
|
||||
self.gguf_writer.add_token_types(toktypes)
|
||||
|
||||
# Add special tokens from config
|
||||
if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_bos_token_id(token_id)
|
||||
if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_eos_token_id(token_id)
|
||||
if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_pad_token_id(token_id)
|
||||
if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_sep_token_id(token_id)
|
||||
if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
|
||||
token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
|
||||
self.gguf_writer.add_unk_token_id(token_id)
|
||||
|
||||
# Add <|plamo:op|> as EOT to ensure appropriate end of generation
|
||||
self.gguf_writer.add_eot_token_id(4)
|
||||
|
||||
self.gguf_writer.add_add_space_prefix(False)
|
||||
self._set_vocab_plamo()
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
hparams = self.hparams
|
||||
@@ -4966,6 +4964,56 @@ class Plamo2Model(TextModel):
|
||||
return [(new_name, data_torch)]
|
||||
|
||||
|
||||
@ModelBase.register("Plamo3ForCausalLM", "PLaMo3ForCausalLM")
|
||||
class Plamo3Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.PLAMO3
|
||||
|
||||
def set_vocab(self):
|
||||
self._set_vocab_plamo()
|
||||
|
||||
tokenizer_config_path = self.dir_model / "tokenizer_config.json"
|
||||
tokenizer_config = {}
|
||||
|
||||
if tokenizer_config_path.is_file():
|
||||
with open(tokenizer_config_path, encoding="utf-8") as f:
|
||||
tokenizer_config = json.load(f)
|
||||
|
||||
chat_template = tokenizer_config.get("chat_template")
|
||||
chat_template_jinja = self.dir_model / "chat_template.jinja"
|
||||
|
||||
if chat_template_jinja.is_file():
|
||||
with open(chat_template_jinja, encoding="utf-8") as f:
|
||||
chat_template = f.read()
|
||||
|
||||
if chat_template:
|
||||
self.gguf_writer.add_chat_template(chat_template)
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
super().set_gguf_parameters()
|
||||
self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
|
||||
if (sliding_window := self.find_hparam(["window_size", "sliding_window"], optional=True)) is not None:
|
||||
self.gguf_writer.add_sliding_window(sliding_window)
|
||||
self.gguf_writer.add_sliding_window_pattern(self.hparams["sliding_window_pattern"])
|
||||
self.gguf_writer.add_rope_freq_base_swa(self.rope_parameters.get("sliding_attention", {"rope_theta": self.hparams.get("rope_local_theta")})["rope_theta"])
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
|
||||
if name.endswith(".pre_mixer_norm.weight"):
|
||||
data_torch = data_torch + 1.0
|
||||
elif name.endswith(".post_mixer_norm.weight"):
|
||||
data_torch = data_torch + 1.0 / 5
|
||||
elif name.endswith(".pre_mlp_norm.weight"):
|
||||
data_torch = data_torch + 1.0
|
||||
elif name.endswith(".post_mlp_norm.weight"):
|
||||
data_torch = data_torch + 1.0 / (5**1.5)
|
||||
elif name.endswith((".mixer.q_norm.weight", ".mixer.k_norm.weight")):
|
||||
data_torch = data_torch + 1.0
|
||||
elif name.endswith(".norm.weight"):
|
||||
data_torch = data_torch + 1.0
|
||||
|
||||
return [(self.map_tensor_name(name), data_torch)]
|
||||
|
||||
|
||||
@ModelBase.register("CodeShellForCausalLM")
|
||||
class CodeShellModel(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.CODESHELL
|
||||
|
||||
Reference in New Issue
Block a user