Upload Moondream
Browse files- config.json +15 -0
- configuration_moondream.py +1 -1
- modeling_phi.py +2 -33
- moondream.py +6 -4
- vision_encoder.py +67 -27
config.json
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
{
|
|
|
2 |
"architectures": [
|
3 |
"Moondream"
|
4 |
],
|
@@ -10,6 +11,20 @@
|
|
10 |
"phi_config": {
|
11 |
"model_type": "phi"
|
12 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
"torch_dtype": "float16",
|
14 |
"transformers_version": "4.38.2"
|
15 |
}
|
|
|
1 |
{
|
2 |
+
"_name_or_path": "vikhyatk/moondream2",
|
3 |
"architectures": [
|
4 |
"Moondream"
|
5 |
],
|
|
|
11 |
"phi_config": {
|
12 |
"model_type": "phi"
|
13 |
},
|
14 |
+
"text_config": {
|
15 |
+
"architectures": [
|
16 |
+
"Moondream"
|
17 |
+
],
|
18 |
+
"auto_map": {
|
19 |
+
"AutoConfig": "configuration_moondream.MoondreamConfig",
|
20 |
+
"AutoModelForCausalLM": "moondream.Moondream"
|
21 |
+
},
|
22 |
+
"model_type": "phi",
|
23 |
+
"phi_config": {
|
24 |
+
"model_type": "phi"
|
25 |
+
},
|
26 |
+
"torch_dtype": "float16"
|
27 |
+
},
|
28 |
"torch_dtype": "float16",
|
29 |
"transformers_version": "4.38.2"
|
30 |
}
|
configuration_moondream.py
CHANGED
@@ -94,5 +94,5 @@ class MoondreamConfig(PretrainedConfig):
|
|
94 |
model_type = "moondream1"
|
95 |
|
96 |
def __init__(self, **kwargs):
|
97 |
-
self.
|
98 |
super().__init__(**kwargs)
|
|
|
94 |
model_type = "moondream1"
|
95 |
|
96 |
def __init__(self, **kwargs):
|
97 |
+
self.text_config = PhiConfig(**kwargs)
|
98 |
super().__init__(**kwargs)
|
modeling_phi.py
CHANGED
@@ -400,40 +400,10 @@ class PhiAttention(nn.Module):
|
|
400 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
401 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
402 |
|
403 |
-
|
404 |
-
|
405 |
-
query_states.to(torch.float32), key_states.to(torch.float32).transpose(2, 3)
|
406 |
-
) / math.sqrt(self.head_dim)
|
407 |
-
|
408 |
-
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
409 |
-
raise ValueError(
|
410 |
-
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
|
411 |
-
f" {attn_weights.size()}"
|
412 |
-
)
|
413 |
-
|
414 |
-
if attention_mask is not None:
|
415 |
-
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
416 |
-
raise ValueError(
|
417 |
-
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
418 |
-
)
|
419 |
-
attn_weights = attn_weights + attention_mask
|
420 |
-
|
421 |
-
# upcast attention to fp32
|
422 |
-
attn_weights = nn.functional.softmax(
|
423 |
-
attn_weights, dim=-1, dtype=torch.float32
|
424 |
-
).to(value_states.dtype)
|
425 |
-
attn_weights = nn.functional.dropout(
|
426 |
-
attn_weights, p=self.attention_dropout, training=self.training
|
427 |
)
|
428 |
|
429 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
430 |
-
|
431 |
-
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
432 |
-
raise ValueError(
|
433 |
-
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
434 |
-
f" {attn_output.size()}"
|
435 |
-
)
|
436 |
-
|
437 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
438 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
439 |
|
@@ -1115,7 +1085,6 @@ class PhiForCausalLM(PhiPreTrainedModel):
|
|
1115 |
|
1116 |
hidden_states = outputs[0]
|
1117 |
logits = self.lm_head(hidden_states)
|
1118 |
-
logits = logits.float()
|
1119 |
|
1120 |
loss = None
|
1121 |
if labels is not None:
|
|
|
400 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
401 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
402 |
|
403 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
404 |
+
query_states, key_states, value_states, attn_mask=attention_mask
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
405 |
)
|
406 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
408 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
409 |
|
|
|
1085 |
|
1086 |
hidden_states = outputs[0]
|
1087 |
logits = self.lm_head(hidden_states)
|
|
|
1088 |
|
1089 |
loss = None
|
1090 |
if labels is not None:
|
moondream.py
CHANGED
@@ -12,14 +12,16 @@ class Moondream(PreTrainedModel):
|
|
12 |
|
13 |
def __init__(self, config):
|
14 |
super().__init__(config)
|
15 |
-
self.vision_encoder = VisionEncoder(
|
|
|
|
|
16 |
|
17 |
-
if type(config.
|
18 |
phi_config = PhiConfig(
|
19 |
-
**config.
|
20 |
)
|
21 |
else:
|
22 |
-
phi_config = config.
|
23 |
self.text_model = PhiForCausalLM(phi_config)
|
24 |
|
25 |
@property
|
|
|
12 |
|
13 |
def __init__(self, config):
|
14 |
super().__init__(config)
|
15 |
+
self.vision_encoder = VisionEncoder(
|
16 |
+
use_flash_attn=config._attn_implementation == "flash_attention_2"
|
17 |
+
)
|
18 |
|
19 |
+
if type(config.text_config) == dict:
|
20 |
phi_config = PhiConfig(
|
21 |
+
**config.text_config, attn_implementation=config._attn_implementation
|
22 |
)
|
23 |
else:
|
24 |
+
phi_config = config.text_config
|
25 |
self.text_model = PhiForCausalLM(phi_config)
|
26 |
|
27 |
@property
|
vision_encoder.py
CHANGED
@@ -10,10 +10,20 @@ from torchvision.transforms.v2 import (
|
|
10 |
ToDtype,
|
11 |
Normalize,
|
12 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
|
15 |
class Attention(nn.Module):
|
16 |
-
|
|
|
17 |
super().__init__()
|
18 |
assert dim % num_heads == 0, "dim should be divisible by num_heads"
|
19 |
|
@@ -23,6 +33,11 @@ class Attention(nn.Module):
|
|
23 |
self.qkv = nn.Linear(dim, dim * 3)
|
24 |
self.proj = nn.Linear(dim, dim)
|
25 |
|
|
|
|
|
|
|
|
|
|
|
26 |
torch.nn.init.kaiming_normal_(
|
27 |
self.qkv.weight, mode="fan_in", nonlinearity="relu"
|
28 |
)
|
@@ -31,25 +46,36 @@ class Attention(nn.Module):
|
|
31 |
)
|
32 |
|
33 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
|
49 |
class VitBlock(nn.Module):
|
50 |
-
|
|
|
51 |
super().__init__()
|
52 |
-
self.attn = Attention(embed_dim)
|
53 |
self.mlp = MLP(embed_dim, 4304)
|
54 |
self.norm1 = nn.LayerNorm(embed_dim)
|
55 |
self.norm2 = nn.LayerNorm(embed_dim)
|
@@ -62,7 +88,7 @@ class VitBlock(nn.Module):
|
|
62 |
|
63 |
class VisionTransformer(nn.Module):
|
64 |
|
65 |
-
def __init__(self):
|
66 |
super().__init__()
|
67 |
|
68 |
embed_len = 729
|
@@ -70,7 +96,9 @@ class VisionTransformer(nn.Module):
|
|
70 |
|
71 |
self.patch_embed = LinearPatchEmbedding()
|
72 |
self.pos_embed = nn.Parameter(torch.randn(1, embed_len, embed_dim) * 0.02)
|
73 |
-
self.blocks = nn.Sequential(
|
|
|
|
|
74 |
self.norm = nn.LayerNorm(embed_dim)
|
75 |
|
76 |
def forward(self, x):
|
@@ -83,9 +111,9 @@ class VisionTransformer(nn.Module):
|
|
83 |
|
84 |
class EncoderWrapper(nn.Module):
|
85 |
|
86 |
-
def __init__(self):
|
87 |
super().__init__()
|
88 |
-
self.model = nn.ModuleDict({"visual": VisionTransformer()})
|
89 |
|
90 |
def forward(self, x):
|
91 |
return self.model["visual"](x)
|
@@ -98,6 +126,13 @@ class LinearPatchEmbedding(nn.Module):
|
|
98 |
self.linear = nn.Linear(588, 1152)
|
99 |
|
100 |
def forward(self, x):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
return self.linear(x)
|
102 |
|
103 |
|
@@ -148,10 +183,11 @@ class VisionProjection(nn.Module):
|
|
148 |
|
149 |
|
150 |
class VisionEncoder(nn.Module):
|
151 |
-
|
|
|
152 |
super().__init__()
|
153 |
|
154 |
-
self.encoder = EncoderWrapper()
|
155 |
self.projection = VisionProjection()
|
156 |
|
157 |
self.preprocess = Compose(
|
@@ -172,16 +208,20 @@ class VisionEncoder(nn.Module):
|
|
172 |
return self.projection.mlp.fc1.weight.dtype
|
173 |
|
174 |
def __call__(self, images) -> torch.Tensor:
|
175 |
-
if not isinstance(images, list):
|
176 |
images = [images]
|
177 |
|
178 |
with torch.no_grad():
|
179 |
-
|
180 |
-
|
181 |
-
|
|
|
|
|
182 |
|
183 |
-
|
|
|
184 |
|
|
|
185 |
x = self.encoder(x)
|
186 |
x = self.projection(x)
|
187 |
|
|
|
10 |
ToDtype,
|
11 |
Normalize,
|
12 |
)
|
13 |
+
from transformers.utils import is_flash_attn_2_available
|
14 |
+
|
15 |
+
try:
|
16 |
+
if is_flash_attn_2_available():
|
17 |
+
from flash_attn.modules.mha import FlashSelfAttention
|
18 |
+
else:
|
19 |
+
FlashSelfAttention = None
|
20 |
+
except ImportError:
|
21 |
+
FlashSelfAttention = None
|
22 |
|
23 |
|
24 |
class Attention(nn.Module):
|
25 |
+
|
26 |
+
def __init__(self, dim, num_heads=16, use_flash_attn=False):
|
27 |
super().__init__()
|
28 |
assert dim % num_heads == 0, "dim should be divisible by num_heads"
|
29 |
|
|
|
33 |
self.qkv = nn.Linear(dim, dim * 3)
|
34 |
self.proj = nn.Linear(dim, dim)
|
35 |
|
36 |
+
if use_flash_attn and FlashSelfAttention is not None:
|
37 |
+
self.flash_attn = FlashSelfAttention()
|
38 |
+
else:
|
39 |
+
self.flash_attn = None
|
40 |
+
|
41 |
torch.nn.init.kaiming_normal_(
|
42 |
self.qkv.weight, mode="fan_in", nonlinearity="relu"
|
43 |
)
|
|
|
46 |
)
|
47 |
|
48 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
49 |
+
if self.flash_attn is not None:
|
50 |
+
qkv = self.qkv(x)
|
51 |
+
qkv = rearrange(
|
52 |
+
qkv, "... (three h d) -> ... three h d", three=3, h=self.num_heads
|
53 |
+
)
|
54 |
+
attn_output = self.flash_attn(qkv)
|
55 |
+
output = rearrange(attn_output, "... h d -> ... (h d)")
|
56 |
+
output = self.proj(output)
|
57 |
+
return output
|
58 |
+
else:
|
59 |
+
B, N, C = x.shape
|
60 |
+
qkv = (
|
61 |
+
self.qkv(x)
|
62 |
+
.reshape(B, N, 3, self.num_heads, self.head_dim)
|
63 |
+
.permute(2, 0, 3, 1, 4)
|
64 |
+
)
|
65 |
+
q, k, v = qkv.unbind(0)
|
66 |
+
|
67 |
+
x = F.scaled_dot_product_attention(q, k, v)
|
68 |
+
|
69 |
+
x = x.transpose(1, 2).reshape(B, N, C)
|
70 |
+
x = self.proj(x)
|
71 |
+
return x
|
72 |
|
73 |
|
74 |
class VitBlock(nn.Module):
|
75 |
+
|
76 |
+
def __init__(self, embed_dim, use_flash_attn=False):
|
77 |
super().__init__()
|
78 |
+
self.attn = Attention(embed_dim, use_flash_attn=use_flash_attn)
|
79 |
self.mlp = MLP(embed_dim, 4304)
|
80 |
self.norm1 = nn.LayerNorm(embed_dim)
|
81 |
self.norm2 = nn.LayerNorm(embed_dim)
|
|
|
88 |
|
89 |
class VisionTransformer(nn.Module):
|
90 |
|
91 |
+
def __init__(self, use_flash_attn=False):
|
92 |
super().__init__()
|
93 |
|
94 |
embed_len = 729
|
|
|
96 |
|
97 |
self.patch_embed = LinearPatchEmbedding()
|
98 |
self.pos_embed = nn.Parameter(torch.randn(1, embed_len, embed_dim) * 0.02)
|
99 |
+
self.blocks = nn.Sequential(
|
100 |
+
*[VitBlock(embed_dim, use_flash_attn=use_flash_attn) for _ in range(27)]
|
101 |
+
)
|
102 |
self.norm = nn.LayerNorm(embed_dim)
|
103 |
|
104 |
def forward(self, x):
|
|
|
111 |
|
112 |
class EncoderWrapper(nn.Module):
|
113 |
|
114 |
+
def __init__(self, use_flash_attn=False):
|
115 |
super().__init__()
|
116 |
+
self.model = nn.ModuleDict({"visual": VisionTransformer(use_flash_attn)})
|
117 |
|
118 |
def forward(self, x):
|
119 |
return self.model["visual"](x)
|
|
|
126 |
self.linear = nn.Linear(588, 1152)
|
127 |
|
128 |
def forward(self, x):
|
129 |
+
b, c, hp1, wp2 = x.shape
|
130 |
+
p1, p2 = 14, 14
|
131 |
+
h, w = hp1 // p1, wp2 // p2
|
132 |
+
x = x.reshape(b, c, h, p1, w, p2)
|
133 |
+
x = x.permute(0, 2, 4, 1, 3, 5)
|
134 |
+
x = x.reshape(b, h * w, c * p1 * p2)
|
135 |
+
|
136 |
return self.linear(x)
|
137 |
|
138 |
|
|
|
183 |
|
184 |
|
185 |
class VisionEncoder(nn.Module):
|
186 |
+
|
187 |
+
def __init__(self, use_flash_attn=False):
|
188 |
super().__init__()
|
189 |
|
190 |
+
self.encoder = EncoderWrapper(use_flash_attn)
|
191 |
self.projection = VisionProjection()
|
192 |
|
193 |
self.preprocess = Compose(
|
|
|
208 |
return self.projection.mlp.fc1.weight.dtype
|
209 |
|
210 |
def __call__(self, images) -> torch.Tensor:
|
211 |
+
if not isinstance(images, list) and not isinstance(images, torch.Tensor):
|
212 |
images = [images]
|
213 |
|
214 |
with torch.no_grad():
|
215 |
+
# Skip preprocess if images are already tensors
|
216 |
+
if not isinstance(images, torch.Tensor) and not isinstance(
|
217 |
+
images[0], torch.Tensor
|
218 |
+
):
|
219 |
+
images = [self.preprocess(image.convert("RGB")) for image in images]
|
220 |
|
221 |
+
if isinstance(images, list):
|
222 |
+
images = torch.stack(images)
|
223 |
|
224 |
+
x = images.to(self.device, dtype=self.dtype)
|
225 |
x = self.encoder(x)
|
226 |
x = self.projection(x)
|
227 |
|