Skip to content

support interns1-mini #15412

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Aug 25, 2025
+65 −68
Merged

support interns1-mini #15412

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
23c7da7
support interns1-mini
RunningLeon Aug 19, 2025
37888ba
fix comment
RunningLeon Aug 22, 2025
dbbe1f7
update
RunningLeon Aug 25, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
132 changes: 64 additions & 68 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1210,6 +1210,54 @@ def _try_set_pooling_type(self) -> None:
raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
self.gguf_writer.add_pooling_type(pooling_type)

def _set_vocab_interns1(self):
tokens: list[str] = []
toktypes: list[int] = []

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
vocab_size = self.hparams.get("vocab_size", len(vocab))
assert max(vocab.values()) < vocab_size

tokpre = self.get_vocab_base_pre(tokenizer)

reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
added_vocab = tokenizer.get_added_vocab()

added_tokens_decoder = tokenizer.added_tokens_decoder

for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.UNUSED)
else:
token: str = reverse_vocab[i]
if token in added_vocab:
# The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
# To avoid unexpected issues - we make sure to normalize non-normalized tokens
if not added_tokens_decoder[i].normalized:
previous_token = token
token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
if previous_token != token:
logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

if added_tokens_decoder[i].special or self.does_token_look_special(token):
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
toktypes.append(gguf.TokenType.NORMAL)
tokens.append(token)

self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_tokenizer_pre(tokpre)
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)

special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
special_vocab.add_to_gguf(self.gguf_writer)


class MmprojModel(ModelBase):
model_type = ModelType.MMPROJ
Expand Down Expand Up @@ -2917,7 +2965,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
if "language_model." in name:
name = name.replace("language_model.", "") # for InternVL
if name.startswith("mlp") or name.startswith("multi_modal_projector") \
or name.startswith("vision_model") or name.startswith("audio_tower"):
or name.startswith("vision_model") or name.startswith("audio_tower") \
or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
# skip vision and audio tensors
return []
yield from super().modify_tensors(data_torch, name, bid)
Expand Down Expand Up @@ -3589,6 +3638,19 @@ def prepare_tensors(self):
class Qwen3Model(Qwen2Model):
model_arch = gguf.MODEL_ARCH.QWEN3

def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
self.origin_hf_arch = hparams.get('architectures', [None])[0]

def set_vocab(self):
# deal with intern-s1-mini
if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
self._set_vocab_interns1()
return

super().set_vocab()


@ModelBase.register("Qwen3MoeForCausalLM")
class Qwen3MoeModel(Qwen2MoeModel):
Expand All @@ -3605,73 +3667,7 @@ def set_vocab(self):
self._set_vocab_interns1()
return

try:
self._set_vocab_sentencepiece()
except FileNotFoundError:
self._set_vocab_gpt2()

def _set_vocab_interns1(self):
tokens: list[str] = []
toktypes: list[int] = []

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
vocab_size = self.hparams.get("vocab_size", len(vocab))
assert max(vocab.values()) < vocab_size

tokpre = self.get_vocab_base_pre(tokenizer)

reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
added_vocab = tokenizer.get_added_vocab()

added_tokens_decoder = tokenizer.added_tokens_decoder

for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.UNUSED)
else:
token: str = reverse_vocab[i]
if token in added_vocab:
# The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
# To avoid unexpected issues - we make sure to normalize non-normalized tokens
if not added_tokens_decoder[i].normalized:
previous_token = token
token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
if previous_token != token:
logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

if added_tokens_decoder[i].special or self.does_token_look_special(token):
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
toktypes.append(gguf.TokenType.NORMAL)
tokens.append(token)

self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_tokenizer_pre(tokpre)
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)

special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
special_tokens_map_file = self.dir_model / 'special_tokens_map.json'
additional_special_tokens = []
if special_tokens_map_file.is_file():
with open(special_tokens_map_file, encoding = 'utf-8') as f:
additional_special_tokens = json.load(f).get('additional_special_tokens', [])
tokenizer_cfg_file = self.dir_model / 'special_tokens_map.json'
if tokenizer_cfg_file.is_file():
with open(tokenizer_cfg_file, encoding = 'utf-8') as f:
added_tokens_decoder = json.load(f).get('added_tokens_decoder', {})
token2ids_map = {data['content'] : int(token) for token, data in added_tokens_decoder.items() if data['special']}
for token in additional_special_tokens:
if token in token2ids_map:
special_vocab._set_special_token(token, token2ids_map[token])
special_vocab._set_special_token('eos', 151645)
special_vocab._set_special_token("bos", 151643)
special_vocab.add_to_gguf(self.gguf_writer)
super().set_vocab()


@ModelBase.register("GPT2LMHeadModel")
Expand Down