File size: 6,250 Bytes
c5c0b58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import argparse
from safetensors import safe_open
from transformers import AutoTokenizer
from pathlib import Path
import torch
import sys
import os

if 'NO_LOCAL_GGUF' not in os.environ:
    print(str(Path(__file__).parent / 'gguf-py'))
    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))

from gguf import *


def k(raw_key: str, arch: str) -> str:
    return raw_key.format(arch=arch)


parser = argparse.ArgumentParser()
parser.add_argument("--model", type=str, default="../nanoLLaVA")
parser.add_argument("--tokenizer", type=str, default="nanoLLaVA")
args = parser.parse_args()

tensors = safe_open(f'{args.model}/model.safetensors', framework="pt", device="cpu")

### Vision encoder

ftype = 1  # fp16

fname_middle = "mmproj-"
has_text_encoder = False
has_llava_projector = True

fname_out = f"{args.model}/nanollava-mmproj-f16.gguf"
fout = GGUFWriter(fname_out, arch="clip")

fout.add_bool("clip.has_text_encoder", False)
fout.add_bool("clip.has_vision_encoder", True)
fout.add_bool("clip.has_llava_projector", True)
fout.add_file_type(ftype)  # fp16

model_name = "qnguyen3/nanoLLaVA"
fout.add_name(model_name)
fout.add_description("image encoder for " + model_name)
fout.add_string("clip.projector_type", "mlp")

# vision model hparams
VISION = "clip.vision"
fout.add_uint32("clip.vision.image_size", 378)
fout.add_uint32("clip.vision.patch_size", 14)
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), 1152)
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
fout.add_uint32("clip.vision.projection_dim", 2048)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), 27 + 1)

fout.add_array("clip.vision.image_mean", [0.5, 0.5, 0.5])
fout.add_array("clip.vision.image_std", [0.5, 0.5, 0.5])
fout.add_bool("clip.use_gelu", True)  # using regular GELU instead of quick

# vision projection
fout.add_tensor(
    "mm.0.weight",
    tensors.get_tensor("model.mm_projector.0.weight").to(
        torch.float16
    ).numpy().copy()
)
fout.add_tensor(
    "mm.0.bias",
    tensors.get_tensor("model.mm_projector.0.bias").to(torch.float32).numpy().copy(),
)
fout.add_tensor(
    "mm.2.weight",
    tensors.get_tensor("model.mm_projector.2.weight").to(
        torch.float16
    ).numpy().copy(),
)
fout.add_tensor(
    "mm.2.bias",
    tensors.get_tensor("model.mm_projector.2.bias").to(torch.float32).numpy().copy(),
)

# encoder (siglip)
fout.add_tensor(
    "v.position_embd.weight",
    tensors.get_tensor("model.vision_tower.vision_tower.vision_model.embeddings.position_embedding.weight").to(
        torch.float16
    ).numpy().copy(),
)
fout.add_tensor(
    "v.patch_embd.weight",
    tensors.get_tensor(
        "model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.weight"
    )
    .reshape(1152, 3, 14, 14)
    .to(torch.float16).numpy().copy(),
)
fout.add_tensor(
    "v.patch_embd.bias",
    tensors.get_tensor(
        "model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.bias"
    ).to(torch.float32).numpy().copy(),
)

fout.add_tensor(
    "v.post_ln.weight",
    tensors.get_tensor("model.vision_tower.vision_tower.vision_model.post_layernorm.weight").to(
        torch.float32
    ).numpy().copy(),
)
fout.add_tensor(
    "v.post_ln.bias",
    tensors.get_tensor("model.vision_tower.vision_tower.vision_model.post_layernorm.bias").to(
        torch.float32
    ).numpy().copy(),
)

def blk_tensor(i, name):
    return tensors.get_tensor(
        rf"model.vision_tower.vision_tower.vision_model.encoder.layers.{i}.{name}"
    )

def add_tensor(blk_id, gguf_id=None):
    if gguf_id is None:
        gguf_id = blk_id

    fout.add_tensor(f"v.blk.{gguf_id}.attn_q.weight", blk_tensor(blk_id, "self_attn.q_proj.weight").to(torch.float16).numpy().copy())
    fout.add_tensor(f"v.blk.{gguf_id}.attn_q.bias", blk_tensor(blk_id, "self_attn.q_proj.bias").to(torch.float32).numpy().copy())
    fout.add_tensor(f"v.blk.{gguf_id}.attn_k.weight", blk_tensor(blk_id, "self_attn.k_proj.weight").to(torch.float16).numpy().copy())
    fout.add_tensor(f"v.blk.{gguf_id}.attn_k.bias", blk_tensor(blk_id, "self_attn.k_proj.bias").to(torch.float32).numpy().copy())
    fout.add_tensor(f"v.blk.{gguf_id}.attn_v.weight", blk_tensor(blk_id, "self_attn.v_proj.weight").to(torch.float16).numpy().copy())
    fout.add_tensor(f"v.blk.{gguf_id}.attn_v.bias", blk_tensor(blk_id, "self_attn.v_proj.bias").to(torch.float32).numpy().copy())

    fout.add_tensor(
        f"v.blk.{gguf_id}.attn_out.weight",
        blk_tensor(blk_id, "self_attn.out_proj.weight").to(torch.float16).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.attn_out.bias",
        blk_tensor(blk_id, "self_attn.out_proj.bias").to(torch.float32).numpy().copy(),
    )

    fout.add_tensor(
        f"v.blk.{gguf_id}.ln1.weight",
        blk_tensor(blk_id, "layer_norm1.weight").to(torch.float32).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.ln1.bias",
        blk_tensor(blk_id, "layer_norm1.bias").to(torch.float32).numpy().copy(),
    )

    fout.add_tensor(
        f"v.blk.{gguf_id}.ffn_down.weight",
        blk_tensor(blk_id, "mlp.fc1.weight").to(torch.float16).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.ffn_down.bias",
        blk_tensor(blk_id, "mlp.fc1.bias").to(torch.float32).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.ffn_up.weight",
        blk_tensor(blk_id, "mlp.fc2.weight").to(torch.float16).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.ffn_up.bias",
        blk_tensor(blk_id, "mlp.fc2.bias").to(torch.float32).numpy().copy(),
    )

    fout.add_tensor(
        f"v.blk.{gguf_id}.ln2.weight",
        blk_tensor(blk_id, "layer_norm2.weight").to(torch.float32).numpy().copy(),
    )
    fout.add_tensor(
        f"v.blk.{gguf_id}.ln2.bias",
        blk_tensor(blk_id, "layer_norm2.bias").to(torch.float32).numpy().copy(),
    )

for i in range(27):
    add_tensor(i)

# Duplicate the last block (llava-cli skips over this)
add_tensor(26, 27)

fout.write_header_to_file()
fout.write_kv_data_to_file()
fout.write_tensors_to_file()
fout.close()
print(f'successfully exported to {fname_out}')