Skip to content

UPSTREAM PR #18849: Deepseek v3.2 dense attention support from @fairydreaming #923

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

Open
loci-dev wants to merge 1 commit into
base: main
Choose a base branch
from
Open

UPSTREAM PR #18849: Deepseek v3.2 dense attention support from @fairydreaming #923

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 17 additions & 1 deletion convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -965,13 +965,17 @@ def does_token_look_special(self, token: str | bytes) -> bool:

return seems_special

def override_tokenizer_settings(self, tokenizer):
return tokenizer

# used for GPT-2 BPE and WordPiece vocabs
def get_vocab_base(self) -> tuple[list[str], list[int], str]:
tokens: list[str] = []
toktypes: list[int] = []

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
tokenizer = self.override_tokenizer_settings(tokenizer)
vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size

Expand Down Expand Up @@ -7456,16 +7460,28 @@ def prepare_tensors(self):
@ModelBase.register(
"DeepseekV2ForCausalLM",
"DeepseekV3ForCausalLM",
"DeepseekV32ForCausalLM",
"KimiVLForConditionalGeneration",
"YoutuForCausalLM",
"YoutuVLForConditionalGeneration"
)
class DeepseekV2Model(TextModel):
model_arch = gguf.MODEL_ARCH.DEEPSEEK2

def override_tokenizer_settings(self, tokenizer):
# override add_bos_token setting to get pre-tokenizer recognized
if self.hparams.get("model_type") == "deepseek_v32":
tokenizer.add_bos_token = True
return tokenizer

def set_vocab(self):
try:
self._set_vocab_gpt2()
# in DeepSeek V3.2 adding BOS token is disabled in tokenizer configuration
# instead the BOS token is added in encode_messages() Python code
# therefore we have to override this setting
if self.hparams.get("model_type") == "deepseek_v32":
self.gguf_writer.add_add_bos_token(True)
return
except Exception:
pass
Expand Down Expand Up @@ -7576,7 +7592,7 @@ def set_gguf_parameters(self):

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# skip vision tensors and remove "language_model." for Kimi-VL
if "vision_tower" in name or "multi_modal_projector" in name:
if "vision_tower" in name or "multi_modal_projector" in name or "self_attn.indexer" in name:
return []
if name.startswith("siglip2.") or name.startswith("merger."):
return []
Expand Down