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

Fix projector linear substitution

Browse source 
Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
This commit is contained in:
Alex-Brooks 2025-01-26 01:40:07 -07:00
parent 61d4ae4699
commit 7905f9dd40
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/llava/convert_image_encoder_to_gguf.py
View file

@ -43,9 +43,9 @@ def get_tensor_name(name: str) -> str:
return "model.image_newline"
if name.startswith("multi_modal_projector"):
name = name.replace("multi_modal_projector", "mm")
if name.endswith("linear_1"):
if "linear_1" in name:
name = name.replace("linear_1", "0")
if name.endswith("linear_2"):
if "linear_2" in name:
name = name.replace("linear_2", "1")
return name
@ -251,7 +251,7 @@ if has_vision_encoder:
fout.add_uint32("clip.vision.projection_dim", visual_projection_dim)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"] # Why is this decremented? Should be 27...
block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
# /**
# "image_grid_pinpoints": [