Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
import gradio as gr
|
2 |
-
from transformers import AutoModel, AutoTokenizer, pipeline, AutoConfig
|
3 |
-
from huggingface_hub import cached_download, hf_hub_url, list_models
|
4 |
from transformers.modeling_utils import PreTrainedModel
|
5 |
import requests
|
6 |
import json
|
@@ -10,7 +10,6 @@ from io import BytesIO
|
|
10 |
import base64
|
11 |
import torch
|
12 |
from torch.nn.utils import prune
|
13 |
-
from transformers.models.auto import AutoModelForCausalLM # Import for CausalLM
|
14 |
|
15 |
# Function to fetch open-weight LLM models
|
16 |
def fetch_open_weight_models():
|
@@ -18,20 +17,15 @@ def fetch_open_weight_models():
|
|
18 |
return models
|
19 |
|
20 |
# Function to prune a model using the "merge-kit" approach
|
21 |
-
def prune_model(llm_model_name, target_size,
|
22 |
try:
|
23 |
# Load the LLM model and tokenizer
|
24 |
llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
|
25 |
# Handle cases where the model is split into multiple safetensors
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
torch_dtype=torch.float16, # Adjust dtype as needed
|
31 |
-
use_auth_token=None,
|
32 |
-
)
|
33 |
-
else:
|
34 |
-
llm_model = AutoModel.from_pretrained(llm_model_name)
|
35 |
|
36 |
# Get the model config
|
37 |
config = AutoConfig.from_pretrained(llm_model_name)
|
@@ -41,8 +35,12 @@ def prune_model(llm_model_name, target_size, output_dir):
|
|
41 |
# Use merge-kit to prune the model
|
42 |
pruned_model = merge_kit_prune(llm_model, target_num_parameters)
|
43 |
|
44 |
-
# Save the pruned model
|
45 |
-
|
|
|
|
|
|
|
|
|
46 |
|
47 |
# Create a visualization
|
48 |
fig, ax = plt.subplots(figsize=(10, 5))
|
@@ -53,7 +51,7 @@ def prune_model(llm_model_name, target_size, output_dir):
|
|
53 |
fig.savefig(buf, format="png")
|
54 |
buf.seek(0)
|
55 |
image_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
56 |
-
return f"Pruned model saved to {
|
57 |
|
58 |
except Exception as e:
|
59 |
return f"Error: {e}", None
|
@@ -61,23 +59,19 @@ def prune_model(llm_model_name, target_size, output_dir):
|
|
61 |
# Merge-kit Pruning Function (adjust as needed)
|
62 |
def merge_kit_prune(model: PreTrainedModel, target_num_parameters: int) -> PreTrainedModel:
|
63 |
"""Prunes a model using a merge-kit approach.
|
64 |
-
|
65 |
Args:
|
66 |
model (PreTrainedModel): The model to be pruned.
|
67 |
target_num_parameters (int): The target number of parameters after pruning.
|
68 |
-
|
69 |
Returns:
|
70 |
PreTrainedModel: The pruned model.
|
71 |
"""
|
72 |
-
|
73 |
# Define the pruning method
|
74 |
pruning_method = "unstructured"
|
75 |
|
76 |
# Calculate the pruning amount
|
77 |
-
amount = 1 - (target_num_parameters / model.
|
78 |
|
79 |
-
# Prune the model using the selected method
|
80 |
-
# Example: If Llama uses specific layers, adjust the pruning logic here
|
81 |
for name, module in model.named_modules():
|
82 |
if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
|
83 |
prune.random_unstructured(module, name="weight", amount=amount)
|
@@ -107,22 +101,25 @@ def create_interface():
|
|
107 |
interactive=True,
|
108 |
)
|
109 |
|
110 |
-
#
|
111 |
-
|
|
|
|
|
|
|
112 |
|
113 |
-
# Output for
|
114 |
-
|
115 |
|
116 |
# Button to start pruning
|
117 |
prune_button = gr.Button("Prune Model")
|
118 |
|
119 |
# Output for visualization
|
120 |
-
visualization = gr.Image(label="Model Size Comparison")
|
121 |
|
122 |
# Connect components
|
123 |
prune_button.click(
|
124 |
fn=prune_model,
|
125 |
-
inputs=[llm_model_name, target_size,
|
126 |
outputs=[pruning_status, visualization],
|
127 |
)
|
128 |
|
@@ -133,11 +130,11 @@ def create_interface():
|
|
133 |
# Generate text button
|
134 |
generate_button = gr.Button("Generate Text")
|
135 |
|
136 |
-
def generate_text(text,
|
137 |
try:
|
138 |
# Load the pruned model and tokenizer
|
139 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
140 |
-
model = AutoModelForCausalLM.from_pretrained(
|
141 |
|
142 |
# Use the pipeline for text generation
|
143 |
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
@@ -146,7 +143,7 @@ def create_interface():
|
|
146 |
except Exception as e:
|
147 |
return f"Error: {e}"
|
148 |
|
149 |
-
generate_button.click(fn=generate_text, inputs=[text_input,
|
150 |
|
151 |
return demo
|
152 |
|
|
|
1 |
import gradio as gr
|
2 |
+
from transformers import AutoModel, AutoTokenizer, pipeline, AutoConfig, AutoModelForCausalLM
|
3 |
+
from huggingface_hub import cached_download, hf_hub_url, list_models, create_repo, HfApi
|
4 |
from transformers.modeling_utils import PreTrainedModel
|
5 |
import requests
|
6 |
import json
|
|
|
10 |
import base64
|
11 |
import torch
|
12 |
from torch.nn.utils import prune
|
|
|
13 |
|
14 |
# Function to fetch open-weight LLM models
|
15 |
def fetch_open_weight_models():
|
|
|
17 |
return models
|
18 |
|
19 |
# Function to prune a model using the "merge-kit" approach
|
20 |
+
def prune_model(llm_model_name, target_size, hf_write_token, repo_name):
|
21 |
try:
|
22 |
# Load the LLM model and tokenizer
|
23 |
llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
|
24 |
# Handle cases where the model is split into multiple safetensors
|
25 |
+
llm_model = AutoModelForCausalLM.from_pretrained(
|
26 |
+
llm_model_name,
|
27 |
+
torch_dtype=torch.float16, # Adjust dtype as needed
|
28 |
+
)
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
# Get the model config
|
31 |
config = AutoConfig.from_pretrained(llm_model_name)
|
|
|
35 |
# Use merge-kit to prune the model
|
36 |
pruned_model = merge_kit_prune(llm_model, target_num_parameters)
|
37 |
|
38 |
+
# Save the pruned model to Hugging Face repository
|
39 |
+
api = HfApi()
|
40 |
+
repo_id = f"{hf_write_token}/{repo_name}"
|
41 |
+
create_repo(repo_id, token=hf_write_token, private=False, exist_ok=True)
|
42 |
+
pruned_model.push_to_hub(repo_id, use_auth_token=hf_write_token)
|
43 |
+
llm_tokenizer.push_to_hub(repo_id, use_auth_token=hf_write_token)
|
44 |
|
45 |
# Create a visualization
|
46 |
fig, ax = plt.subplots(figsize=(10, 5))
|
|
|
51 |
fig.savefig(buf, format="png")
|
52 |
buf.seek(0)
|
53 |
image_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
54 |
+
return f"Pruned model saved to Hugging Face Hub in repository {repo_id}", f"data:image/png;base64,{image_base64}"
|
55 |
|
56 |
except Exception as e:
|
57 |
return f"Error: {e}", None
|
|
|
59 |
# Merge-kit Pruning Function (adjust as needed)
|
60 |
def merge_kit_prune(model: PreTrainedModel, target_num_parameters: int) -> PreTrainedModel:
|
61 |
"""Prunes a model using a merge-kit approach.
|
|
|
62 |
Args:
|
63 |
model (PreTrainedModel): The model to be pruned.
|
64 |
target_num_parameters (int): The target number of parameters after pruning.
|
|
|
65 |
Returns:
|
66 |
PreTrainedModel: The pruned model.
|
67 |
"""
|
|
|
68 |
# Define the pruning method
|
69 |
pruning_method = "unstructured"
|
70 |
|
71 |
# Calculate the pruning amount
|
72 |
+
amount = 1 - (target_num_parameters / sum(p.numel() for p in model.parameters()))
|
73 |
|
74 |
+
# Prune the model using the selected method
|
|
|
75 |
for name, module in model.named_modules():
|
76 |
if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
|
77 |
prune.random_unstructured(module, name="weight", amount=amount)
|
|
|
101 |
interactive=True,
|
102 |
)
|
103 |
|
104 |
+
# Input for Hugging Face write token
|
105 |
+
hf_write_token = gr.Textbox(label="Hugging Face Write Token", placeholder="Enter your HF write token", interactive=True, type="password")
|
106 |
+
|
107 |
+
# Input for repository name
|
108 |
+
repo_name = gr.Textbox(label="Repository Name", placeholder="Enter the name of the repository", interactive=True)
|
109 |
|
110 |
+
# Output for pruning status
|
111 |
+
pruning_status = gr.Textbox(label="Pruning Status", interactive=False)
|
112 |
|
113 |
# Button to start pruning
|
114 |
prune_button = gr.Button("Prune Model")
|
115 |
|
116 |
# Output for visualization
|
117 |
+
visualization = gr.Image(label="Model Size Comparison", interactive=False)
|
118 |
|
119 |
# Connect components
|
120 |
prune_button.click(
|
121 |
fn=prune_model,
|
122 |
+
inputs=[llm_model_name, target_size, hf_write_token, repo_name],
|
123 |
outputs=[pruning_status, visualization],
|
124 |
)
|
125 |
|
|
|
130 |
# Generate text button
|
131 |
generate_button = gr.Button("Generate Text")
|
132 |
|
133 |
+
def generate_text(text, repo_name):
|
134 |
try:
|
135 |
# Load the pruned model and tokenizer
|
136 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_name, use_auth_token=hf_write_token)
|
137 |
+
model = AutoModelForCausalLM.from_pretrained(repo_name, use_auth_token=hf_write_token)
|
138 |
|
139 |
# Use the pipeline for text generation
|
140 |
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
|
|
143 |
except Exception as e:
|
144 |
return f"Error: {e}"
|
145 |
|
146 |
+
generate_button.click(fn=generate_text, inputs=[text_input, repo_name], outputs=text_output)
|
147 |
|
148 |
return demo
|
149 |
|