File size: 9,711 Bytes
c3e525b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
import comfy.utils

LORA_CLIP_MAP = {
    "mlp.fc1": "mlp_fc1",
    "mlp.fc2": "mlp_fc2",
    "self_attn.k_proj": "self_attn_k_proj",
    "self_attn.q_proj": "self_attn_q_proj",
    "self_attn.v_proj": "self_attn_v_proj",
    "self_attn.out_proj": "self_attn_out_proj",
}


def load_lora(lora, to_load):
    patch_dict = {}
    loaded_keys = set()
    for x in to_load:
        alpha_name = "{}.alpha".format(x)
        alpha = None
        if alpha_name in lora.keys():
            alpha = lora[alpha_name].item()
            loaded_keys.add(alpha_name)

        regular_lora = "{}.lora_up.weight".format(x)
        diffusers_lora = "{}_lora.up.weight".format(x)
        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
        A_name = None

        if regular_lora in lora.keys():
            A_name = regular_lora
            B_name = "{}.lora_down.weight".format(x)
            mid_name = "{}.lora_mid.weight".format(x)
        elif diffusers_lora in lora.keys():
            A_name = diffusers_lora
            B_name = "{}_lora.down.weight".format(x)
            mid_name = None
        elif transformers_lora in lora.keys():
            A_name = transformers_lora
            B_name ="{}.lora_linear_layer.down.weight".format(x)
            mid_name = None

        if A_name is not None:
            mid = None
            if mid_name is not None and mid_name in lora.keys():
                mid = lora[mid_name]
                loaded_keys.add(mid_name)
            patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid))
            loaded_keys.add(A_name)
            loaded_keys.add(B_name)


        ######## loha
        hada_w1_a_name = "{}.hada_w1_a".format(x)
        hada_w1_b_name = "{}.hada_w1_b".format(x)
        hada_w2_a_name = "{}.hada_w2_a".format(x)
        hada_w2_b_name = "{}.hada_w2_b".format(x)
        hada_t1_name = "{}.hada_t1".format(x)
        hada_t2_name = "{}.hada_t2".format(x)
        if hada_w1_a_name in lora.keys():
            hada_t1 = None
            hada_t2 = None
            if hada_t1_name in lora.keys():
                hada_t1 = lora[hada_t1_name]
                hada_t2 = lora[hada_t2_name]
                loaded_keys.add(hada_t1_name)
                loaded_keys.add(hada_t2_name)

            patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2))
            loaded_keys.add(hada_w1_a_name)
            loaded_keys.add(hada_w1_b_name)
            loaded_keys.add(hada_w2_a_name)
            loaded_keys.add(hada_w2_b_name)


        ######## lokr
        lokr_w1_name = "{}.lokr_w1".format(x)
        lokr_w2_name = "{}.lokr_w2".format(x)
        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
        lokr_t2_name = "{}.lokr_t2".format(x)
        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
        lokr_w2_b_name = "{}.lokr_w2_b".format(x)

        lokr_w1 = None
        if lokr_w1_name in lora.keys():
            lokr_w1 = lora[lokr_w1_name]
            loaded_keys.add(lokr_w1_name)

        lokr_w2 = None
        if lokr_w2_name in lora.keys():
            lokr_w2 = lora[lokr_w2_name]
            loaded_keys.add(lokr_w2_name)

        lokr_w1_a = None
        if lokr_w1_a_name in lora.keys():
            lokr_w1_a = lora[lokr_w1_a_name]
            loaded_keys.add(lokr_w1_a_name)

        lokr_w1_b = None
        if lokr_w1_b_name in lora.keys():
            lokr_w1_b = lora[lokr_w1_b_name]
            loaded_keys.add(lokr_w1_b_name)

        lokr_w2_a = None
        if lokr_w2_a_name in lora.keys():
            lokr_w2_a = lora[lokr_w2_a_name]
            loaded_keys.add(lokr_w2_a_name)

        lokr_w2_b = None
        if lokr_w2_b_name in lora.keys():
            lokr_w2_b = lora[lokr_w2_b_name]
            loaded_keys.add(lokr_w2_b_name)

        lokr_t2 = None
        if lokr_t2_name in lora.keys():
            lokr_t2 = lora[lokr_t2_name]
            loaded_keys.add(lokr_t2_name)

        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
            patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2))

        #glora
        a1_name = "{}.a1.weight".format(x)
        a2_name = "{}.a2.weight".format(x)
        b1_name = "{}.b1.weight".format(x)
        b2_name = "{}.b2.weight".format(x)
        if a1_name in lora:
            patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha))
            loaded_keys.add(a1_name)
            loaded_keys.add(a2_name)
            loaded_keys.add(b1_name)
            loaded_keys.add(b2_name)

        w_norm_name = "{}.w_norm".format(x)
        b_norm_name = "{}.b_norm".format(x)
        w_norm = lora.get(w_norm_name, None)
        b_norm = lora.get(b_norm_name, None)

        if w_norm is not None:
            loaded_keys.add(w_norm_name)
            patch_dict[to_load[x]] = ("diff", (w_norm,))
            if b_norm is not None:
                loaded_keys.add(b_norm_name)
                patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (b_norm,))

        diff_name = "{}.diff".format(x)
        diff_weight = lora.get(diff_name, None)
        if diff_weight is not None:
            patch_dict[to_load[x]] = ("diff", (diff_weight,))
            loaded_keys.add(diff_name)

        diff_bias_name = "{}.diff_b".format(x)
        diff_bias = lora.get(diff_bias_name, None)
        if diff_bias is not None:
            patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (diff_bias,))
            loaded_keys.add(diff_bias_name)

    for x in lora.keys():
        if x not in loaded_keys:
            print("lora key not loaded", x)
    return patch_dict

def model_lora_keys_clip(model, key_map={}):
    sdk = model.state_dict().keys()

    text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
    clip_l_present = False
    for b in range(32): #TODO: clean up
        for c in LORA_CLIP_MAP:
            k = "clip_h.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
            if k in sdk:
                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
                key_map[lora_key] = k
                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
                key_map[lora_key] = k
                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
                key_map[lora_key] = k

            k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
            if k in sdk:
                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
                key_map[lora_key] = k
                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
                key_map[lora_key] = k
                clip_l_present = True
                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
                key_map[lora_key] = k

            k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
            if k in sdk:
                if clip_l_present:
                    lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
                    key_map[lora_key] = k
                    lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
                    key_map[lora_key] = k
                else:
                    lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
                    key_map[lora_key] = k
                    lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
                    key_map[lora_key] = k
                    lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
                    key_map[lora_key] = k


    k = "clip_g.transformer.text_projection.weight"
    if k in sdk:
        key_map["lora_prior_te_text_projection"] = k #cascade lora?
        # key_map["text_encoder.text_projection"] = k #TODO: check if other lora have the text_projection too
        # key_map["lora_te_text_projection"] = k

    return key_map

def model_lora_keys_unet(model, key_map={}):
    sdk = model.state_dict().keys()

    for k in sdk:
        if k.startswith("diffusion_model.") and k.endswith(".weight"):
            key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
            key_map["lora_unet_{}".format(key_lora)] = k
            key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config

    diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config)
    for k in diffusers_keys:
        if k.endswith(".weight"):
            unet_key = "diffusion_model.{}".format(diffusers_keys[k])
            key_lora = k[:-len(".weight")].replace(".", "_")
            key_map["lora_unet_{}".format(key_lora)] = unet_key

            diffusers_lora_prefix = ["", "unet."]
            for p in diffusers_lora_prefix:
                diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
                if diffusers_lora_key.endswith(".to_out.0"):
                    diffusers_lora_key = diffusers_lora_key[:-2]
                key_map[diffusers_lora_key] = unet_key
    return key_map