freQuensy23 commited on
Commit
c0be431
1 Parent(s): 3bf54fe
Files changed (4) hide show
  1. app.py +44 -142
  2. generators.py +106 -0
  3. requirements.txt +0 -0
  4. utils.py +25 -0
app.py CHANGED
@@ -1,146 +1,48 @@
 
 
1
  import gradio as gr
2
- import numpy as np
3
- import random
4
- from diffusers import DiffusionPipeline
5
- import torch
6
 
7
- device = "cuda" if torch.cuda.is_available() else "cpu"
8
-
9
- if torch.cuda.is_available():
10
- torch.cuda.max_memory_allocated(device=device)
11
- pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
12
- pipe.enable_xformers_memory_efficient_attention()
13
- pipe = pipe.to(device)
14
- else:
15
- pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True)
16
- pipe = pipe.to(device)
17
-
18
- MAX_SEED = np.iinfo(np.int32).max
19
- MAX_IMAGE_SIZE = 1024
20
-
21
- def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
22
-
23
- if randomize_seed:
24
- seed = random.randint(0, MAX_SEED)
25
-
26
- generator = torch.Generator().manual_seed(seed)
27
-
28
- image = pipe(
29
- prompt = prompt,
30
- negative_prompt = negative_prompt,
31
- guidance_scale = guidance_scale,
32
- num_inference_steps = num_inference_steps,
33
- width = width,
34
- height = height,
35
- generator = generator
36
- ).images[0]
37
-
38
- return image
39
-
40
- examples = [
41
- "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
42
- "An astronaut riding a green horse",
43
- "A delicious ceviche cheesecake slice",
44
- ]
45
-
46
- css="""
47
- #col-container {
48
- margin: 0 auto;
49
- max-width: 520px;
50
- }
51
- """
52
-
53
- if torch.cuda.is_available():
54
- power_device = "GPU"
55
- else:
56
- power_device = "CPU"
57
-
58
- with gr.Blocks(css=css) as demo:
59
-
60
- with gr.Column(elem_id="col-container"):
61
- gr.Markdown(f"""
62
- # Text-to-Image Gradio Template
63
- Currently running on {power_device}.
64
- """)
65
-
66
- with gr.Row():
67
-
68
- prompt = gr.Text(
69
- label="Prompt",
70
- show_label=False,
71
- max_lines=1,
72
- placeholder="Enter your prompt",
73
- container=False,
74
- )
75
-
76
- run_button = gr.Button("Run", scale=0)
77
-
78
- result = gr.Image(label="Result", show_label=False)
79
-
80
- with gr.Accordion("Advanced Settings", open=False):
81
-
82
- negative_prompt = gr.Text(
83
- label="Negative prompt",
84
- max_lines=1,
85
- placeholder="Enter a negative prompt",
86
- visible=False,
87
- )
88
-
89
- seed = gr.Slider(
90
- label="Seed",
91
- minimum=0,
92
- maximum=MAX_SEED,
93
- step=1,
94
- value=0,
95
- )
96
-
97
- randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
98
-
99
- with gr.Row():
100
-
101
- width = gr.Slider(
102
- label="Width",
103
- minimum=256,
104
- maximum=MAX_IMAGE_SIZE,
105
- step=32,
106
- value=512,
107
- )
108
-
109
- height = gr.Slider(
110
- label="Height",
111
- minimum=256,
112
- maximum=MAX_IMAGE_SIZE,
113
- step=32,
114
- value=512,
115
- )
116
-
117
- with gr.Row():
118
-
119
- guidance_scale = gr.Slider(
120
- label="Guidance scale",
121
- minimum=0.0,
122
- maximum=10.0,
123
- step=0.1,
124
- value=0.0,
125
- )
126
-
127
- num_inference_steps = gr.Slider(
128
- label="Number of inference steps",
129
- minimum=1,
130
- maximum=12,
131
- step=1,
132
- value=2,
133
- )
134
-
135
- gr.Examples(
136
- examples = examples,
137
- inputs = [prompt]
138
- )
139
-
140
- run_button.click(
141
- fn = infer,
142
- inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
143
- outputs = [result]
144
  )
145
 
146
- demo.queue().launch()
 
1
+ from dotenv import load_dotenv
2
+ from generators import *
3
  import gradio as gr
 
 
 
 
4
 
5
+ from utils import async_zip_stream
6
+
7
+ load_dotenv()
8
+
9
+
10
+ async def handle(system_input: str, user_input: str):
11
+ print(system_input, user_input)
12
+ buffers = ["", "", "", ""]
13
+ async for outputs in async_zip_stream(
14
+ generate_gpt2(system_input, user_input),
15
+ generate_mistral_7bvo1(system_input, user_input),
16
+ generate_llama2(system_input, user_input),
17
+ generate_llama3(system_input, user_input),
18
+ ):
19
+ # gpt_output, mistral_output, llama_output, llama2_output, llama3_output, llama4_output = outputs
20
+ for i, b in enumerate(buffers):
21
+ buffers[i] += str(outputs[i])
22
+
23
+ yield list(buffers) + ["", ""]
24
+ yield list(buffers) + [await generate_openllama(system_input, user_input),
25
+ await generate_bloom(system_input, user_input)]
26
+
27
+
28
+ with gr.Blocks() as demo:
29
+ system_input = gr.Textbox(label='System Input', value='You are AI assistant', lines=2)
30
+ with gr.Row():
31
+ gpt = gr.Textbox(label='gpt-2', lines=4, interactive=False)
32
+ mistral = gr.Textbox(label='mistral', lines=4, interactive=False)
33
+ llama = gr.Textbox(label='openllama', lines=4, interactive=False)
34
+ with gr.Row():
35
+ llama2 = gr.Textbox(label='llama-2', lines=4, interactive=False)
36
+ llama3 = gr.Textbox(label='llama-3', lines=4, interactive=False)
37
+ bloom = gr.Textbox(label='bloom', lines=4, interactive=False)
38
+
39
+ user_input = gr.Textbox(label='User Input', lines=2)
40
+ gen_button = gr.Button('Generate')
41
+
42
+ gen_button.click(
43
+ fn=handle,
44
+ inputs=[system_input, user_input],
45
+ outputs=[gpt, mistral, llama2, llama3, llama, bloom],
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46
  )
47
 
48
+ demo.launch()
generators.py ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import asyncio
2
+ import json
3
+ import os
4
+
5
+ import aiohttp
6
+ import gradio as gr
7
+ import numpy as np
8
+ import spaces
9
+ from huggingface_hub import InferenceClient
10
+
11
+ import random
12
+ import torch
13
+ from huggingface_hub import AsyncInferenceClient
14
+ from transformers import LlamaTokenizer, LlamaForCausalLM, AutoTokenizer
15
+
16
+
17
+ async def query_llm(payload, model_name):
18
+ headers = {"Authorization": f"Bearer {os.getenv('HF_TOKEN')}"}
19
+ async with aiohttp.ClientSession() as session:
20
+ async with session.post(f"https://api-inference.huggingface.co/models/{model_name}", headers=headers,
21
+ json=payload) as response:
22
+ return await response.json()
23
+
24
+
25
+ async def generate_mistral_7bvo1(system_input, user_input):
26
+ client = AsyncInferenceClient(
27
+ "mistralai/Mistral-7B-Instruct-v0.1",
28
+ token=os.getenv('HF_TOKEN'),
29
+ )
30
+
31
+ async for message in await client.chat_completion(
32
+ messages=[
33
+ {"role": "system", "content": system_input},
34
+ {"role": "user", "content": user_input}, ],
35
+ max_tokens=256,
36
+ stream=True,
37
+ ):
38
+ yield message.choices[0].delta.content
39
+
40
+
41
+ async def generate_gpt2(system_input, user_input):
42
+ output = await query_llm({
43
+ "inputs": (inputs:=f"{system_input}\n{user_input}"),
44
+ }, "openai-community/gpt2")
45
+ yield output[0]["generated_text"].replace(inputs, '')
46
+
47
+
48
+ async def generate_llama2(system_input, user_input):
49
+ client = AsyncInferenceClient(
50
+ "meta-llama/Llama-2-7b-chat-hf",
51
+ token=os.getenv('HF_TOKEN')
52
+ )
53
+ async for message in await client.chat_completion(
54
+ messages=[
55
+ {"role": "system", "content": system_input},
56
+ {"role": "user", "content": user_input}, ],
57
+ max_tokens=256,
58
+ stream=True,
59
+ ):
60
+ yield message.choices[0].delta.content
61
+
62
+
63
+ @spaces.GPU
64
+ async def generate_openllama(system_input, user_input):
65
+ model_path = 'openlm-research/open_llama_3b_v2'
66
+ tokenizer = LlamaTokenizer.from_pretrained(model_path)
67
+ model = LlamaForCausalLM.from_pretrained(
68
+ model_path, torch_dtype=torch.float16, device_map='cuda',
69
+ )
70
+ # model = model.to("cuda")
71
+ input_text = f"{system_input}\n{user_input}"
72
+ input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
73
+ output = model.generate(input_ids, max_length=128)
74
+ return tokenizer.decode(output[0], skip_special_tokens=True)
75
+
76
+
77
+ @spaces.GPU
78
+ async def generate_bloom(system_input, user_input):
79
+ model_path = 'bigscience/bloom-7b1'
80
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
81
+ model = LlamaForCausalLM.from_pretrained(
82
+ model_path, torch_dtype=torch.float16, device_map='cuda',
83
+ )
84
+ input_text = f"{system_input}\n{user_input}"
85
+ input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
86
+ output = model.generate(input_ids, max_length=128)
87
+ return tokenizer.decode(output[0], skip_special_tokens=True)
88
+
89
+
90
+
91
+ async def generate_llama3(system_input, user_input):
92
+ client = AsyncInferenceClient(
93
+ "meta-llama/Meta-Llama-3.1-8B-Instruct",
94
+ token=os.getenv('HF_TOKEN')
95
+ )
96
+ try:
97
+ async for message in await client.chat_completion(
98
+ messages=[
99
+ {"role": "system", "content": system_input},
100
+ {"role": "user", "content": user_input}, ],
101
+ max_tokens=256,
102
+ stream=True,
103
+ ):
104
+ yield message.choices[0].delta.content
105
+ except json.JSONDecodeError:
106
+ pass
requirements.txt CHANGED
Binary files a/requirements.txt and b/requirements.txt differ
 
utils.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import asyncio
2
+
3
+
4
+ async def async_zip_stream(*iterators, default_value=''):
5
+ tasks = [asyncio.create_task(iterator.__anext__()) for iterator in iterators]
6
+ done = [False] * len(iterators)
7
+
8
+ while not all(done):
9
+ results = []
10
+
11
+ for i, task in enumerate(tasks):
12
+ if done[i]:
13
+ results.append(default_value)
14
+ elif task.done():
15
+ try:
16
+ results.append(task.result())
17
+ tasks[i] = asyncio.create_task(iterators[i].__anext__())
18
+ except StopAsyncIteration:
19
+ done[i] = True
20
+ results.append(default_value)
21
+ else:
22
+ results.append(default_value)
23
+
24
+ yield tuple(results)
25
+ await asyncio.sleep(0.01) # Slight delay to allow other tasks to progress