Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -7,7 +7,7 @@ import torch
|
|
7 |
import edge_tts
|
8 |
import asyncio
|
9 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
10 |
-
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
11 |
from transformers.image_utils import load_image
|
12 |
import time
|
13 |
|
@@ -35,6 +35,7 @@ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
|
|
35 |
|
36 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
37 |
|
|
|
38 |
model_id = "prithivMLmods/FastThink-0.5B-Tiny"
|
39 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
40 |
model = AutoModelForCausalLM.from_pretrained(
|
@@ -53,6 +54,7 @@ TTS_VOICES = [
|
|
53 |
"en-US-JasonNeural", # @tts6
|
54 |
]
|
55 |
|
|
|
56 |
MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
57 |
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
58 |
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
@@ -77,29 +79,39 @@ def generate(
|
|
77 |
top_k: int = 50,
|
78 |
repetition_penalty: float = 1.2,
|
79 |
):
|
80 |
-
"""
|
|
|
|
|
|
|
|
|
81 |
text = input_dict["text"]
|
82 |
files = input_dict.get("files", [])
|
83 |
|
84 |
# Check if input includes image(s)
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
-
# Check if message is for TTS
|
88 |
tts_prefix = "@tts"
|
89 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 7))
|
90 |
voice_index = next((i for i in range(1, 7) if text.strip().lower().startswith(f"{tts_prefix}{i}")), None)
|
91 |
-
|
92 |
if is_tts and voice_index:
|
93 |
voice = TTS_VOICES[voice_index - 1]
|
94 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
|
|
|
|
95 |
else:
|
96 |
voice = None
|
97 |
text = text.replace(tts_prefix, "").strip()
|
|
|
98 |
|
99 |
-
|
100 |
-
|
101 |
if images:
|
102 |
-
# Process multimodal input
|
103 |
messages = [
|
104 |
{"role": "user", "content": [
|
105 |
*[{"type": "image", "image": image} for image in images],
|
@@ -109,9 +121,9 @@ def generate(
|
|
109 |
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
110 |
inputs = processor(text=[prompt], images=images, return_tensors="pt", padding=True).to("cuda")
|
111 |
|
|
|
112 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
113 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=max_new_tokens)
|
114 |
-
|
115 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
116 |
thread.start()
|
117 |
|
@@ -124,7 +136,7 @@ def generate(
|
|
124 |
yield buffer
|
125 |
|
126 |
else:
|
127 |
-
# Process text-only input
|
128 |
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
129 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
130 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
@@ -147,21 +159,18 @@ def generate(
|
|
147 |
t.start()
|
148 |
|
149 |
outputs = []
|
150 |
-
for
|
151 |
-
outputs.append(
|
152 |
yield "".join(outputs)
|
153 |
|
154 |
final_response = "".join(outputs)
|
155 |
|
156 |
-
# Yield text response first
|
157 |
-
yield final_response
|
158 |
|
|
|
159 |
if is_tts and voice:
|
160 |
-
|
161 |
-
asyncio.set_event_loop(loop)
|
162 |
-
output_file = loop.run_until_complete(text_to_speech(final_response, voice))
|
163 |
-
|
164 |
-
# Separate yield for audio output
|
165 |
yield gr.Audio(output_file, autoplay=True)
|
166 |
|
167 |
demo = gr.ChatInterface(
|
|
|
7 |
import edge_tts
|
8 |
import asyncio
|
9 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
10 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, TextIteratorStreamer
|
11 |
from transformers.image_utils import load_image
|
12 |
import time
|
13 |
|
|
|
35 |
|
36 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
37 |
|
38 |
+
# Load the text-only model and tokenizer
|
39 |
model_id = "prithivMLmods/FastThink-0.5B-Tiny"
|
40 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
41 |
model = AutoModelForCausalLM.from_pretrained(
|
|
|
54 |
"en-US-JasonNeural", # @tts6
|
55 |
]
|
56 |
|
57 |
+
# Load the multimodal (OCR) model and processor
|
58 |
MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
59 |
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
60 |
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
|
|
79 |
top_k: int = 50,
|
80 |
repetition_penalty: float = 1.2,
|
81 |
):
|
82 |
+
"""
|
83 |
+
Generates chatbot response and handles TTS requests with multimodal input support.
|
84 |
+
If the query starts with a TTS command (e.g. '@tts1'), the chat history is cleared
|
85 |
+
to avoid non-text responses (like Audio) interfering with template rendering.
|
86 |
+
"""
|
87 |
text = input_dict["text"]
|
88 |
files = input_dict.get("files", [])
|
89 |
|
90 |
# Check if input includes image(s)
|
91 |
+
if len(files) > 1:
|
92 |
+
images = [load_image(image) for image in files]
|
93 |
+
elif len(files) == 1:
|
94 |
+
images = [load_image(files[0])]
|
95 |
+
else:
|
96 |
+
images = []
|
97 |
|
98 |
+
# Check if the message is for TTS
|
99 |
tts_prefix = "@tts"
|
100 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 7))
|
101 |
voice_index = next((i for i in range(1, 7) if text.strip().lower().startswith(f"{tts_prefix}{i}")), None)
|
102 |
+
|
103 |
if is_tts and voice_index:
|
104 |
voice = TTS_VOICES[voice_index - 1]
|
105 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
106 |
+
# Clear conversation history to avoid issues with non-text outputs.
|
107 |
+
conversation = [{"role": "user", "content": text}]
|
108 |
else:
|
109 |
voice = None
|
110 |
text = text.replace(tts_prefix, "").strip()
|
111 |
+
conversation = [*chat_history, {"role": "user", "content": text}]
|
112 |
|
113 |
+
# If there are images, process multimodal input
|
|
|
114 |
if images:
|
|
|
115 |
messages = [
|
116 |
{"role": "user", "content": [
|
117 |
*[{"type": "image", "image": image} for image in images],
|
|
|
121 |
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
122 |
inputs = processor(text=[prompt], images=images, return_tensors="pt", padding=True).to("cuda")
|
123 |
|
124 |
+
# Handle generation for multimodal input using model_m
|
125 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
126 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=max_new_tokens)
|
|
|
127 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
128 |
thread.start()
|
129 |
|
|
|
136 |
yield buffer
|
137 |
|
138 |
else:
|
139 |
+
# Process text-only input using model
|
140 |
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
141 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
142 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
|
|
159 |
t.start()
|
160 |
|
161 |
outputs = []
|
162 |
+
for new_text in streamer:
|
163 |
+
outputs.append(new_text)
|
164 |
yield "".join(outputs)
|
165 |
|
166 |
final_response = "".join(outputs)
|
167 |
|
168 |
+
# Yield text response first.
|
169 |
+
yield final_response
|
170 |
|
171 |
+
# If TTS was requested, yield audio output separately.
|
172 |
if is_tts and voice:
|
173 |
+
output_file = asyncio.run(text_to_speech(final_response, voice))
|
|
|
|
|
|
|
|
|
174 |
yield gr.Audio(output_file, autoplay=True)
|
175 |
|
176 |
demo = gr.ChatInterface(
|