File size: 1,713 Bytes
c59fd93 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 | from __future__ import annotations
from typing import Callable, Iterable, TYPE_CHECKING
if TYPE_CHECKING:
from torch import Tensor
from .base import MmprojModel, ModelBase, gguf
@ModelBase.register("Llama4ForConditionalGeneration")
# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated
@ModelBase.example("unsloth/Llama-4-Scout-17B-16E-Instruct")
class Llama4VisionModel(MmprojModel):
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LLAMA4)
self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams["norm_eps"])
self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / self.hparams["pixel_shuffle_ratio"]))
assert self.hparams["hidden_act"] == "gelu"
self.gguf_writer.add_vision_use_gelu(True)
@classmethod
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
name, gen = item
if "multi_modal_projector" not in name and "vision_model" not in name:
return None
if "positional_embedding_vlm" in name and ".weight" not in name:
name += ".weight"
return super().filter_tensors((name, gen))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if "multi_modal_projector.linear_1" in name:
# despite the name with number postfix, this is a single fully connected layer
yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)
else:
yield from super().modify_tensors(data_torch, name, bid)
|