Repository navigation
Expand file tree
/
Copy pathconvert_vla_jepa_to_gguf.py
More file actions
226 lines (204 loc) · 8.66 KB
/
Copy pathconvert_vla_jepa_to_gguf.py
File metadata and controls
226 lines (204 loc) · 8.66 KB
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
#!/usr/bin/env python3
# Copyright 2026 VinRobotics
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
from gguf_blocks import write_dit_blocks, write_qwen3_lm, write_qwen3vl_vit
from gguf_common import (
add,
arg_parser,
check_layers,
finish,
kv_f32,
kv_prefix,
kv_u32,
load_safetensors,
max_layer,
open_writer,
read_json,
resolve_out
)
ARCH = "vla_jepa"
KV = kv_prefix(ARCH)
VIT = dict(
vit_hidden=1024,
vit_layers=24,
vit_heads=16,
vit_inter=4096,
patch_size=16,
temporal_patch_size=2,
spatial_merge_size=2,
vit_num_position_embeddings=2304,
vit_ln_eps=1e-6
)
DEEPSTACK_IDXS = [5, 11, 17]
QWEN3 = dict(
lm_hidden=2048,
lm_q_heads=16,
lm_kv_heads=8,
lm_head_dim=128,
lm_inter=6144,
lm_rope_theta=5000000.0,
lm_rms_eps=1e-6
)
LM_LAYERS = 28
AH = dict(
dit_hidden=768,
dit_heads=12,
dit_head_dim=64,
dit_layers=16,
cross_dim=2048,
output_dim=1024,
action_dim=7,
state_dim=8,
action_horizon=7,
num_future_tokens=32,
time_proj_dim=256,
num_inference_timesteps=4,
num_timestep_buckets=1000,
ln_eps=1e-5,
norm_out_eps=1e-6
)
IMAGE_TOKEN_INDEX = 151655
EMBODIED_ACTION_TOKEN_ID = 151697
ACTION_TOKEN_ID_0 = 151669
VIT_ROOT = "model.qwen.model.model.visual"
LM_ROOT = "model.qwen.model.model.language_model"
AHK = "model.action_model"
AH_PROJECTORS = [
("action_encoder.layer1", "ah.act_enc.l1"),
("action_encoder.layer2", "ah.act_enc.l2"),
("action_encoder.layer3", "ah.act_enc.l3"),
("state_encoder.layer1", "ah.state_enc.l1"),
("state_encoder.layer2", "ah.state_enc.l2"),
("action_decoder.layer1", "ah.act_dec.l1"),
("action_decoder.layer2", "ah.act_dec.l2"),
("model.timestep_encoder.timestep_embedder.linear_1", "ah.time_emb.l1"),
("model.timestep_encoder.timestep_embedder.linear_2", "ah.time_emb.l2"),
("model.proj_out_1", "ah.proj_out1"),
("model.proj_out_2", "ah.proj_out2"),
]
def main() -> int:
ap = arg_parser(ARCH, "VLA-JEPA-LIBERO checkpoint dir")
args = ap.parse_args()
ckpt = args.ckpt.resolve()
out = resolve_out(args, ckpt, ARCH)
cfg_json = read_json(ckpt / "config.json")
if str(cfg_json.get("type", "")) != ARCH:
raise SystemExit(f"config.json type is {cfg_json.get('type')!r}, expected 'vla_jepa'")
AH["action_dim"] = int(cfg_json.get("action_dim", AH["action_dim"]))
AH["state_dim"] = int(cfg_json.get("state_dim", AH["state_dim"]))
AH["action_horizon"] = int(cfg_json.get("chunk_size", AH["action_horizon"]))
AH["num_future_tokens"] = int(cfg_json.get("num_embodied_action_tokens_per_instruction", AH["num_future_tokens"]))
AH["num_inference_timesteps"] = int(cfg_json.get("num_inference_timesteps", AH["num_inference_timesteps"]))
AH["num_timestep_buckets"] = int(cfg_json.get("action_num_timestep_buckets", AH["num_timestep_buckets"]))
AH["output_dim"] = int(cfg_json.get("action_hidden_size", AH["output_dim"]))
AH["dit_layers"] = int(cfg_json.get("action_num_layers", AH["dit_layers"]))
print(f"loading safetensors from {ckpt} (dropping world model) ...")
W = load_safetensors(ckpt, keep=("model.qwen.", "model.action_model."))
keys = set(W.keys())
print(f" {len(W)} kept tensors")
check_layers(max_layer(keys, f"{VIT_ROOT}.blocks."), VIT["vit_layers"], "ViT layers")
check_layers(max_layer(keys, f"{LM_ROOT}.layers."), LM_LAYERS, "LM layers")
check_layers(max_layer(keys, f"{AHK}.model.transformer_blocks."), AH["dit_layers"], "DiT blocks")
check_layers(max_layer(keys, f"{VIT_ROOT}.deepstack_merger_list."), len(DEEPSTACK_IDXS), "deepstack mergers")
vocab = int(W[f"{LM_ROOT}.embed_tokens.weight"].shape[0])
pe_w = W[f"{VIT_ROOT}.patch_embed.proj.weight"]
assert pe_w.shape == (VIT["vit_hidden"], 3, VIT["temporal_patch_size"], VIT["patch_size"], VIT["patch_size"]), pe_w.shape
patch_flat = 3 * VIT["temporal_patch_size"] * VIT["patch_size"] ** 2
c_merged = VIT["vit_hidden"] * VIT["spatial_merge_size"] ** 2
assert W[f"{VIT_ROOT}.merger.norm.weight"].shape == (VIT["vit_hidden"],), W[f"{VIT_ROOT}.merger.norm.weight"].shape
assert W[f"{VIT_ROOT}.deepstack_merger_list.0.norm.weight"].shape == (c_merged,)
assert W[f"{AHK}.model.transformer_blocks.0.attn1.to_k.weight"].shape == (AH["dit_hidden"], AH["cross_dim"])
assert W[f"{AHK}.model.transformer_blocks.1.attn1.to_k.weight"].shape == (AH["dit_hidden"], AH["dit_hidden"])
print(f"resolved: vit=Qwen3-VL {VIT['vit_hidden']}d×{VIT['vit_layers']}L (deepstack@{DEEPSTACK_IDXS}, merge÷{VIT['spatial_merge_size']}) "
f"lm=Qwen3-VL {QWEN3['lm_hidden']}d×{LM_LAYERS}L ({QWEN3['lm_q_heads']}q/{QWEN3['lm_kv_heads']}kv×{QWEN3['lm_head_dim']}, θ={QWEN3['lm_rope_theta']:g}) vocab={vocab} "
f"dit-B {AH['dit_layers']}L×{AH['dit_heads']}h×{AH['dit_head_dim']}(inner {AH['dit_hidden']}, cross {AH['cross_dim']}, out {AH['output_dim']}) "
f"horizon={AH['action_horizon']} action_dim={AH['action_dim']} state_dim={AH['state_dim']} future={AH['num_future_tokens']} N_steps={AH['num_inference_timesteps']} "
f"img_tok={IMAGE_TOKEN_INDEX} emb_tok={EMBODIED_ACTION_TOKEN_ID}")
writer = open_writer(out, ARCH)
kv_u32(
writer,
KV,
dict(
vit_hidden=VIT["vit_hidden"],
vit_layers=VIT["vit_layers"],
vit_heads=VIT["vit_heads"],
vit_inter=VIT["vit_inter"],
patch_size=VIT["patch_size"],
temporal_patch_size=VIT["temporal_patch_size"],
spatial_merge_size=VIT["spatial_merge_size"],
vit_num_position_embeddings=VIT["vit_num_position_embeddings"],
vit_patch_flat=patch_flat,
vit_merged_dim=c_merged,
image_target_size=256,
deepstack_idx_0=DEEPSTACK_IDXS[0],
deepstack_idx_1=DEEPSTACK_IDXS[1],
deepstack_idx_2=DEEPSTACK_IDXS[2],
lm_hidden=QWEN3["lm_hidden"],
lm_layers=LM_LAYERS,
lm_q_heads=QWEN3["lm_q_heads"],
lm_kv_heads=QWEN3["lm_kv_heads"],
lm_head_dim=QWEN3["lm_head_dim"],
lm_inter=QWEN3["lm_inter"],
vocab_size=vocab,
image_token_index=IMAGE_TOKEN_INDEX,
embodied_action_token_id=EMBODIED_ACTION_TOKEN_ID,
action_token_id_0=ACTION_TOKEN_ID_0,
dit_hidden=AH["dit_hidden"],
dit_heads=AH["dit_heads"],
dit_head_dim=AH["dit_head_dim"],
dit_layers=AH["dit_layers"],
cross_dim=AH["cross_dim"],
output_dim=AH["output_dim"],
time_proj_dim=AH["time_proj_dim"],
action_dim=AH["action_dim"],
state_dim=AH["state_dim"],
action_horizon=AH["action_horizon"],
num_future_tokens=AH["num_future_tokens"],
num_inference_timesteps=AH["num_inference_timesteps"],
num_timestep_buckets=AH["num_timestep_buckets"],
)
)
writer.add_float64(KV("lm_rope_theta"), float(QWEN3["lm_rope_theta"]))
kv_f32(
writer,
KV,
dict(
lm_rms_eps=QWEN3["lm_rms_eps"],
vit_ln_eps=VIT["vit_ln_eps"],
vit_rope_theta=10000.0,
connector_ln_eps=1e-6,
dit_ln_eps=AH["ln_eps"],
dit_norm_out_eps=AH["norm_out_eps"]
)
)
g = W.__getitem__
write_qwen3vl_vit(
writer,
g,
VIT_ROOT,
VIT["vit_layers"],
len(DEEPSTACK_IDXS),
VIT["vit_hidden"],
patch_flat
)
write_qwen3_lm(writer, g, LM_ROOT, LM_LAYERS)
for src, dst in AH_PROJECTORS:
add(writer, f"{dst}.weight", g(f"{AHK}.{src}.weight")); add(writer, f"{dst}.bias", g(f"{AHK}.{src}.bias"))
add(writer, "ah.future_tokens", g(f"{AHK}.future_tokens.weight"))
add(writer, "ah.pos_embd", g(f"{AHK}.position_embedding.weight"))
write_dit_blocks(writer, g, f"{AHK}.model.transformer_blocks", "ah.dit", AH["dit_layers"])
return finish(writer, out, " - combined GGUF (Qwen3-VL backbone + deepstack + DiT-B action head + cfg)")
if __name__ == "__main__":
raise SystemExit(main())