Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -49,11 +49,12 @@ demo_script_path = Path("demo/gradio_demo.py")
|
|
49 |
print(f"Reading {demo_script_path} to apply environment-specific modifications...")
|
50 |
|
51 |
try:
|
52 |
-
|
53 |
|
54 |
-
#
|
55 |
-
|
56 |
-
|
|
|
57 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
58 |
' self.model_path,',
|
59 |
' torch_dtype=torch.bfloat16,',
|
@@ -61,61 +62,77 @@ try:
|
|
61 |
' attn_implementation="flash_attention_2",',
|
62 |
' )'
|
63 |
]
|
64 |
-
|
65 |
|
66 |
-
#
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
73 |
if USE_ZEROGPU:
|
74 |
print("Optimizing for ZeroGPU execution...")
|
75 |
|
76 |
-
#
|
77 |
-
|
|
|
|
|
|
|
|
|
78 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
79 |
' self.model_path,',
|
80 |
' torch_dtype=torch.bfloat16,',
|
81 |
" device_map='cuda',",
|
82 |
' )'
|
83 |
]
|
84 |
-
|
85 |
|
86 |
-
# Add
|
87 |
-
|
88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
else:
|
90 |
-
|
91 |
-
|
92 |
-
#
|
93 |
-
if
|
94 |
-
modified_content = modified_content.replace(
|
95 |
-
|
96 |
-
|
97 |
-
)
|
98 |
-
|
99 |
-
# Replace the model loading block
|
100 |
-
modified_content = modified_content.replace(original_block, replacement_block_gpu)
|
101 |
-
print("Script modified for ZeroGPU successfully.")
|
102 |
|
103 |
else: # Pure CPU execution
|
104 |
print("Modifying for pure CPU execution...")
|
105 |
|
106 |
-
# New block for CPU: Use float32 and map
|
107 |
-
|
108 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
109 |
' self.model_path,',
|
110 |
' torch_dtype=torch.float32, # Use float32 for CPU',
|
111 |
' device_map="cpu",',
|
112 |
' )'
|
113 |
]
|
114 |
-
|
115 |
|
116 |
-
#
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
119 |
|
120 |
# Write the dynamically modified content back to the demo file
|
121 |
demo_script_path.write_text(modified_content)
|
@@ -125,7 +142,7 @@ except Exception as e:
|
|
125 |
sys.exit(1)
|
126 |
|
127 |
# --- 4. Launch the Gradio Demo ---
|
128 |
-
model_id = "microsoft/
|
129 |
|
130 |
# Construct the command to run the modified demo script
|
131 |
command = [
|
|
|
49 |
print(f"Reading {demo_script_path} to apply environment-specific modifications...")
|
50 |
|
51 |
try:
|
52 |
+
modified_content = demo_script_path.read_text()
|
53 |
|
54 |
+
# --- Patch Definitions ---
|
55 |
+
|
56 |
+
# Define the original model loading block to be replaced.
|
57 |
+
original_model_lines = [
|
58 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
59 |
' self.model_path,',
|
60 |
' torch_dtype=torch.bfloat16,',
|
|
|
62 |
' attn_implementation="flash_attention_2",',
|
63 |
' )'
|
64 |
]
|
65 |
+
original_model_block = "\n".join(original_model_lines)
|
66 |
|
67 |
+
# Define the generation method signature to add the GPU decorator to.
|
68 |
+
original_method_lines = [
|
69 |
+
' def generate_podcast_streaming(self, ',
|
70 |
+
' num_speakers: int,',
|
71 |
+
' script: str,',
|
72 |
+
' speaker_1: str = None,',
|
73 |
+
' speaker_2: str = None,',
|
74 |
+
' speaker_3: str = None,',
|
75 |
+
' speaker_4: str = None,',
|
76 |
+
' cfg_scale: float = 1.3) -> Iterator[tuple]:'
|
77 |
+
]
|
78 |
+
original_method_signature = "\n".join(original_method_lines)
|
79 |
|
80 |
if USE_ZEROGPU:
|
81 |
print("Optimizing for ZeroGPU execution...")
|
82 |
|
83 |
+
# Add 'import spaces' if it's not already there.
|
84 |
+
if "import spaces" not in modified_content:
|
85 |
+
modified_content = "import spaces\n" + modified_content
|
86 |
+
|
87 |
+
# New block for ZeroGPU model loading: remove `attn_implementation`.
|
88 |
+
replacement_model_lines_gpu = [
|
89 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
90 |
' self.model_path,',
|
91 |
' torch_dtype=torch.bfloat16,',
|
92 |
" device_map='cuda',",
|
93 |
' )'
|
94 |
]
|
95 |
+
replacement_model_block_gpu = "\n".join(replacement_model_lines_gpu)
|
96 |
|
97 |
+
# Add the @spaces.GPU decorator to the generation method instead of the class.
|
98 |
+
replacement_method_signature_gpu = "@spaces.GPU(duration=120)\n" + original_method_signature
|
99 |
+
|
100 |
+
# --- Apply Patches for GPU ---
|
101 |
+
|
102 |
+
# Patch 1: Decorate the generation method
|
103 |
+
if original_method_signature in modified_content:
|
104 |
+
modified_content = modified_content.replace(original_method_signature, replacement_method_signature_gpu)
|
105 |
+
print("Successfully applied GPU decorator to the generation method.")
|
106 |
else:
|
107 |
+
print("\033[91mWarning: Could not find the generation method signature to apply the GPU decorator.\033[0m")
|
108 |
+
|
109 |
+
# Patch 2: Modify the model loading
|
110 |
+
if original_model_block in modified_content:
|
111 |
+
modified_content = modified_content.replace(original_model_block, replacement_model_block_gpu)
|
112 |
+
print("Successfully patched the model loading block for ZeroGPU.")
|
113 |
+
else:
|
114 |
+
print("\033[91mWarning: The original model loading block was not found. Patching may have failed.\033[0m")
|
|
|
|
|
|
|
|
|
115 |
|
116 |
else: # Pure CPU execution
|
117 |
print("Modifying for pure CPU execution...")
|
118 |
|
119 |
+
# New block for CPU: Use float32 and map to CPU.
|
120 |
+
replacement_model_lines_cpu = [
|
121 |
' self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
|
122 |
' self.model_path,',
|
123 |
' torch_dtype=torch.float32, # Use float32 for CPU',
|
124 |
' device_map="cpu",',
|
125 |
' )'
|
126 |
]
|
127 |
+
replacement_model_block_cpu = "\n".join(replacement_model_lines_cpu)
|
128 |
|
129 |
+
# Apply patch for CPU
|
130 |
+
if original_model_block in modified_content:
|
131 |
+
modified_content = modified_content.replace(original_model_block, replacement_model_block_cpu)
|
132 |
+
print("Script modified for CPU successfully.")
|
133 |
+
else:
|
134 |
+
print("\033[91mWarning: The original model loading block was not found. Patching may have failed.\033[0m")
|
135 |
+
|
136 |
|
137 |
# Write the dynamically modified content back to the demo file
|
138 |
demo_script_path.write_text(modified_content)
|
|
|
142 |
sys.exit(1)
|
143 |
|
144 |
# --- 4. Launch the Gradio Demo ---
|
145 |
+
model_id = "microsoft/V_VibeVoice-1.5B"
|
146 |
|
147 |
# Construct the command to run the modified demo script
|
148 |
command = [
|