Spaces:
Running
on
Zero
Running
on
Zero
Update
Browse files
app.py
CHANGED
@@ -92,7 +92,7 @@ def load_t5(device: str | torch.device = "cuda", max_length: int = 512) -> HFEmb
|
|
92 |
def load_clip(device: str | torch.device = "cuda") -> HFEmbedder:
|
93 |
return HFEmbedder("openai/clip-vit-large-patch14", max_length=77, dtype=jnp.bfloat16)
|
94 |
|
95 |
-
@spaces.GPU(duration=
|
96 |
def load_encoders():
|
97 |
is_schnell = True
|
98 |
t5 = load_t5("cuda", max_length=256 if is_schnell else 512)
|
@@ -109,7 +109,7 @@ def b64(txt,vec):
|
|
109 |
|
110 |
t5,clip=load_encoders()
|
111 |
|
112 |
-
@spaces.GPU(duration=
|
113 |
def convert(prompt):
|
114 |
if isinstance(prompt, str):
|
115 |
prompt = [prompt]
|
@@ -118,6 +118,25 @@ def convert(prompt):
|
|
118 |
vec = clip.tokenize(prompt)
|
119 |
vec = clip(vec)
|
120 |
return b64(txt,vec)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
with gr.Blocks() as demo:
|
123 |
gr.Markdown("""A workaround for flux-flax to fit into 40G VRAM""")
|
@@ -125,10 +144,14 @@ with gr.Blocks() as demo:
|
|
125 |
with gr.Column():
|
126 |
prompt = gr.Textbox(label="prompt")
|
127 |
convert_btn = gr.Button(value="Convert")
|
|
|
128 |
with gr.Column():
|
129 |
output = gr.Textbox(label="output")
|
130 |
|
131 |
convert_btn.click(convert, inputs=prompt, outputs=output, api_name="convert")
|
|
|
|
|
|
|
132 |
|
133 |
|
134 |
demo.launch()
|
|
|
92 |
def load_clip(device: str | torch.device = "cuda") -> HFEmbedder:
|
93 |
return HFEmbedder("openai/clip-vit-large-patch14", max_length=77, dtype=jnp.bfloat16)
|
94 |
|
95 |
+
@spaces.GPU(duration=60)
|
96 |
def load_encoders():
|
97 |
is_schnell = True
|
98 |
t5 = load_t5("cuda", max_length=256 if is_schnell else 512)
|
|
|
109 |
|
110 |
t5,clip=load_encoders()
|
111 |
|
112 |
+
@spaces.GPU(duration=20)
|
113 |
def convert(prompt):
|
114 |
if isinstance(prompt, str):
|
115 |
prompt = [prompt]
|
|
|
118 |
vec = clip.tokenize(prompt)
|
119 |
vec = clip(vec)
|
120 |
return b64(txt,vec)
|
121 |
+
import jax
|
122 |
+
def _to_embed(t5, clip, txt, vec):
|
123 |
+
t5=nnx.merge(*t5)
|
124 |
+
clip=nnx.merge(*clip)
|
125 |
+
return t5(txt), clip(vec)
|
126 |
+
|
127 |
+
to_embed=jax.jit(_to_embed)
|
128 |
+
|
129 |
+
t5_tuple=nnx.split(t5)
|
130 |
+
clip_tuple=nnx.split(clip)
|
131 |
+
|
132 |
+
@spaces.GPU(duration=120)
|
133 |
+
def compile(prompt):
|
134 |
+
if isinstance(prompt, str):
|
135 |
+
prompt = [prompt]
|
136 |
+
txt = t5.tokenize(prompt)
|
137 |
+
vec = clip.tokenize(prompt)
|
138 |
+
text,vec=to_embed(t5_tuple,clip_tuple,txt,vec)
|
139 |
+
return b64(txt,vec)
|
140 |
|
141 |
with gr.Blocks() as demo:
|
142 |
gr.Markdown("""A workaround for flux-flax to fit into 40G VRAM""")
|
|
|
144 |
with gr.Column():
|
145 |
prompt = gr.Textbox(label="prompt")
|
146 |
convert_btn = gr.Button(value="Convert")
|
147 |
+
compile_btn = gr.Button(value="Compile")
|
148 |
with gr.Column():
|
149 |
output = gr.Textbox(label="output")
|
150 |
|
151 |
convert_btn.click(convert, inputs=prompt, outputs=output, api_name="convert")
|
152 |
+
compile_btn.click(compile, inputs=prompt, outputs=output, api_name="compile")
|
153 |
+
|
154 |
+
|
155 |
|
156 |
|
157 |
demo.launch()
|