update app

app.py

@@ -29,7 +29,16 @@ import numpy as np
 import random
 
 import threading
-generation_lock = threading.Lock()
+# generation_lock = threading.Lock()
+
+# from transformers import StoppingCriteria, StoppingCriteriaList
+# class StopGenerationCriteria(StoppingCriteria):
+# 	def __init__(self, stop_event):
+# 		self.stop_event = stop_event
+
+# 	def __call__(self, input_ids, scores, **kwargs):
+# 		return self.stop_event.is_set()
+
 
 device = 'cuda' if torch.cuda.is_available() else 'cpu'
 
@@ -131,6 +140,11 @@ openers = [
 	"Hmmm I think that this message",
 	"Reflecting on the message here",
 	"Considering what this poster is trying to say",
+	"Seeing this message makes me think",
+	"Thinking about what this poster is communicating",
+	"After reading what's on here",
+	"Based on what’s written here",
+	"After I look at this whole thing",
 ]
 openers_generic = [
 	"Hmmm when thinking about",
@@ -151,27 +165,33 @@ openers_poster_summary = [
 	"This poster seems to",
 	"My interpretation of the poster is",
 	"From what this poster shows, it seems to",
-    "Looking at the poster as a whole, it appears to",
-    "Based on the imagery and tone, the poster seems to",
-    "Visually, the poster comes across as trying to",
-    "To me, this poster is trying to",
-    "When I look at this poster, it feels like it aims to",
+	"Looking at the poster as a whole, it appears to",
+	"Based on the imagery and tone, the poster seems to",
+	"Visually, the poster comes across as trying to",
+	"To me, this poster is trying to",
+	"When I look at this poster, it feels like it aims to",
 	"The poster gives me the impression that it intends to",
 ]
 openers_explain = [
-    "The reason why I think that is because",
-    "To explain why I",
-    "Well, to explain my thoughts",
-    "To put it simply, I feel this way because",
-    "My reasoning behind that is",
-    "What leads me to that view is",
-    "A big part of why I think that is",
-    "To give some context for my view,",
-    "Here’s why I lean that way:",
-    "I see it that way mainly because",
-    "Let me explain why I think so",
-    "Thinking through it, I realize it's because",
-    "To unpack my thinking a bit,"
+	"The reason why I think that is because",
+	"To explain why I",
+	"Well, to explain my thoughts",
+	"To put it simply, I feel this way because",
+	"My reasoning behind that is",
+	"What leads me to that view is",
+	"A big part of why I think that is",
+	"To give some context for my view,",
+	"Here’s why I lean that way:",
+	"I see it that way mainly because",
+	"Let me explain why I think so",
+	"Thinking through it, I realize it's because",
+	"To unpack my thinking a bit,",
+	"I guess it’s because",
+	"The thing that really shapes my view is",
+	"It’s pretty much because",
+	"A lot of it comes down to",
+	"I feel that way mostly because",
+	"My thinking comes from the idea that",
 ]
 
 
@@ -184,89 +204,99 @@ def vlm_response(user_input, history, health_topic,
 		political, education, income, family_status,
 		# extraversion, agreeableness, conscientiousness, neuroticism, openness,
 	):
-	with generation_lock:
-		"""Clear any lingering state"""
-		torch.cuda.empty_cache()  # Clear GPU memory
-
-		""" [NOTE] we have not use `history` for this generation """
-		# get uploaded image
-		image = Image.open(user_input['files'][0]) if user_input['files'] else None
-		image_uploaded = True
-		if image is None:
-			image = Image.new('RGB', (24,24))
-			image_uploaded = False
-		# image_b64 = convert_to_base64(image)
-		print(health_topic)
-		# print("Image uploaded:", image_uploaded)
+	# # 1. Initialize Stop Event for this session
+	# stop_event = threading.Event()
+	# # Create the stopping criteria to pass to the model
+	# stopping_criteria = StoppingCriteriaList([StopGenerationCriteria(stop_event)])
 
-		
-		
-		#################################################
-		# 1. Construct traits from user inputs
-		#################################################
-		demo_dict = {
-			"Gender": gender,
-			"Age": age,
-			"Current Profession": profession,
-			"Race/Ethnicity": race,
-			"Religious/Cultural Group": religion,
-			"Political Affiliation": political,
-			"Highest Education": education,
-			"Annual Household Income": income,
-			"Family Status": family_status,
-		}
-		# big5_dict = {
-		# 	"Extraversion": extraversion,
-		# 	"Agreeableness": agreeableness,
-		# 	"Conscientiousness": conscientiousness,
-		# 	"Neuroticism": neuroticism,
-		# 	"Open-Mindedness": openness,
-		# }
-		
-		demo_info = ""
-		for trait, value in demo_dict.items():
-			if value != "Leave Blank": # only add non-blank values
-				demo_info += f"{trait}: {value}\n"
-			else:
-				demo_info += f"{trait}: [Not specified]\n"
-		persona_score = ""
-		persona_score += "Big-Five Trait Scores:\n"
-		# for trait, value in big5_dict.items():
-		# 	persona_score += f"{trait}: {value}\n"
-		# no locus of control trait score
-		locus = None
-
-		######################################################################################
-		# 1*. modify trait info based on trait selection setings
-		#       demo_full: wheter include full demographic traits or only selected ones
-		#       include_big5, include_facet, include_locus: include big5 / facet / locus of control traits or not
-		#     format: <trait>: <value> if available; else <trait>: [Not specified]
-		######################################################################################
-		demo_info, persona_score, locus = process_trait_info(
-			demo_info, persona_score, locus,
-			demo_full=False, include_big5=True,
-			include_facet=False, include_locus=False,
-			train_mode=False,
-		)
-		# print(demo_info)
-		# print(persona_score)
-
-		###############################################
-		### 		Add style variability           ###
-		###############################################
-		style_hint = random.choice(style_variants) # increase style variant
-		lexical_hint = random.choice(lexical_flavors) # increase lexical variant
-		opening_phrase = random.choice(openers) # increase opening variant
-		opening_generic = random.choice(openers_generic) # increase opening variant
-		opening_poster = random.choice(openers_poster_summary) # poster summary variation
-		opening_explain = random.choice(openers_explain) # thought explanation
-		print('Style:', style_hint)
-		print('Lexical:', lexical_hint)
-		print('Opening:', opening_phrase)
-		print('Generic opening:', opening_generic)
+	# 1. Clear any lingering state
+	torch.cuda.empty_cache()  # Clear GPU memory
+	# 2. Initialize Streamers LOCALLY (Fresh for every request)
+	#    Note: We need to re-initialize these for every single generation call 
+	#    or just once per function call if we share them.
+	streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+	# streamer_aux = TextIteratorStreamer(tokenizer_aux, skip_prompt=True, skip_special_tokens=True)
+
+	""" [NOTE] we have not use `history` for this generation """
+	# get uploaded image
+	image = Image.open(user_input['files'][0]) if user_input['files'] else None
+	image_uploaded = True
+	if image is None:
+		image = Image.new('RGB', (24,24))
+		image_uploaded = False
+	# image_b64 = convert_to_base64(image)
+	print(health_topic)
+	# print("Image uploaded:", image_uploaded)
+
+	
+	
+	#################################################
+	# 1. Construct traits from user inputs
+	#################################################
+	demo_dict = {
+		"Gender": gender,
+		"Age": age,
+		"Current Profession": profession,
+		"Race/Ethnicity": race,
+		"Religious/Cultural Group": religion,
+		"Political Affiliation": political,
+		"Highest Education": education,
+		"Annual Household Income": income,
+		"Family Status": family_status,
+	}
+	# big5_dict = {
+	# 	"Extraversion": extraversion,
+	# 	"Agreeableness": agreeableness,
+	# 	"Conscientiousness": conscientiousness,
+	# 	"Neuroticism": neuroticism,
+	# 	"Open-Mindedness": openness,
+	# }
+	
+	demo_info = ""
+	for trait, value in demo_dict.items():
+		if value != "Leave Blank": # only add non-blank values
+			demo_info += f"{trait}: {value}\n"
+		else:
+			demo_info += f"{trait}: [Not specified]\n"
+	persona_score = ""
+	persona_score += "Big-Five Trait Scores:\n"
+	# for trait, value in big5_dict.items():
+	# 	persona_score += f"{trait}: {value}\n"
+	# no locus of control trait score
+	locus = None
+
+	######################################################################################
+	# 1*. modify trait info based on trait selection setings
+	#       demo_full: wheter include full demographic traits or only selected ones
+	#       include_big5, include_facet, include_locus: include big5 / facet / locus of control traits or not
+	#     format: <trait>: <value> if available; else <trait>: [Not specified]
+	######################################################################################
+	demo_info, persona_score, locus = process_trait_info(
+		demo_info, persona_score, locus,
+		demo_full=False, include_big5=True,
+		include_facet=False, include_locus=False,
+		train_mode=False,
+	)
+	# print(demo_info)
+	# print(persona_score)
+
+	###############################################
+	### 		Add style variability           ###
+	###############################################
+	style_hint = random.choice(style_variants) # increase style variant
+	lexical_hint = random.choice(lexical_flavors) # increase lexical variant
+	opening_phrase = random.choice(openers) # increase opening variant
+	opening_generic = random.choice(openers_generic) # increase opening variant
+	opening_poster = random.choice(openers_poster_summary) # poster summary variation
+	opening_explain = random.choice(openers_explain) # thought explanation
+	print('Style:', style_hint)
+	print('Lexical:', lexical_hint)
+	print('Opening:', opening_phrase)
+	print('Generic opening:', opening_generic)
 
 		
-		
+	# Wrap the GENERATION logic in try/finally to handle cleanup
+	try:	
 		if image_uploaded:
 			"""###############################################################
 				Case 1: a health poster is uploaded
@@ -275,6 +305,8 @@ def vlm_response(user_input, history, health_topic,
 			################################################
 			# * IMAGE UNDERSTANDING
 			################################################
+			yield "Analyzing image content..." # UI Feedback
+			
 			PROMPT = (
 				f"Describe the content and main message in given heatlh campaign poster and how it's related to {health_topic}. ",
 				"Note that the message could be non-direct or subtle (e.g. irony, fear-driven evoke without explicit texts, etc). Only provide the answer (in 2-4 sentences). ",
@@ -461,8 +493,15 @@ def vlm_response(user_input, history, health_topic,
 					{"type": "text", "text": SYSTEM_PROMPT + USER_PROMPT}
 				]}
 			]
-			input_text = tokenizer_aux.apply_chat_template(messages, add_generation_prompt = True)
-			inputs = tokenizer_aux(
+			# input_text = tokenizer_aux.apply_chat_template(messages, add_generation_prompt = True)
+			# inputs = tokenizer_aux(
+			# 	# image.convert("RGB"),
+			# 	input_text,
+			# 	add_special_tokens = False,
+			# 	return_tensors = "pt",
+			# ).to(device)
+			input_text = tokenizer.apply_chat_template(messages, add_generation_prompt = True)
+			inputs = tokenizer(
 				# image.convert("RGB"),
 				input_text,
 				add_special_tokens = False,
@@ -475,7 +514,7 @@ def vlm_response(user_input, history, health_topic,
 			# generation with streamer
 			generate_kwargs = dict(
 				**inputs,
-				streamer=streamer_aux,
+				streamer=streamer, # streamer_aux,
 				max_new_tokens=512,
 				use_cache=True,
 				# min_p=0.3,
@@ -485,7 +524,7 @@ def vlm_response(user_input, history, health_topic,
 			)
 			# separate thread to run generation
 			thread = threading.Thread(
-				target=model_aux.generate,
+				target=model.generate, # model_aux.generate,
 				kwargs=generate_kwargs
 			)
 			thread.start()
@@ -494,11 +533,14 @@ def vlm_response(user_input, history, health_topic,
 				f"Emulated traits:\n {demo_info}\n" + '='*20 + "\n\n", 
 				image_desc + "\n\n"
 			]
-			for new_token in streamer_aux:
+			for new_token in streamer: # streamer_aux:
 				outputs.append(new_token)
 				final_output = ''.join(outputs)
 				yield final_output
 
+			# Ensure thread finishes
+			thread.join()
+
 			# text representation of final response
 			response = "".join(outputs[2:]) # ignore trait summary & image description
 			print(colored('Traits', 'green'), demo_info)
@@ -534,8 +576,15 @@ def vlm_response(user_input, history, health_topic,
 					{"type": "text", "text": SYSTEM_PROMPT + USER_PROMPT}
 				]}
 			]
-			input_text = tokenizer_aux.apply_chat_template(messages, add_generation_prompt = True)
-			inputs = tokenizer_aux(
+			# input_text = tokenizer_aux.apply_chat_template(messages, add_generation_prompt = True)
+			# inputs = tokenizer_aux(
+			# 	image.convert("RGB"),
+			# 	input_text,
+			# 	add_special_tokens = False,
+			# 	return_tensors = "pt",
+			# ).to(device)
+			input_text = tokenizer.apply_chat_template(messages, add_generation_prompt = True)
+			inputs = tokenizer(
 				image.convert("RGB"),
 				input_text,
 				add_special_tokens = False,
@@ -548,7 +597,7 @@ def vlm_response(user_input, history, health_topic,
 			# generation with streamer
 			generate_kwargs = dict(
 				**inputs,
-				streamer=streamer_aux,
+				streamer=streamer, # streamer_aux,
 				max_new_tokens=512,
 				use_cache=True,
 				min_p=0.85,
@@ -557,17 +606,19 @@ def vlm_response(user_input, history, health_topic,
 			)
 			# separate thread to run generation
 			thread = threading.Thread(
-				target=model_aux.generate,
+				target=model.generate, # model_aux.generate,
 				kwargs=generate_kwargs
 			)
 			thread.start()
 			# stream out generation
 			# outputs = [image_desc + "\n\n"]
 			outputs += ["\n"]
-			for new_token in streamer_aux:
+			for new_token in streamer: # streamer_aux:
 				outputs.append(new_token)
 				final_output = ''.join(outputs)
 				yield final_output
+
+			thread.join()
 			
 
 			return answer
@@ -658,6 +709,15 @@ def vlm_response(user_input, history, health_topic,
 				outputs.append(new_token)
 				final_output = ''.join(outputs)
 				yield final_output
+			thread.join()
+	
+	except GeneratorExit:
+		print("User disconnected. Waiting for generation to complete...")
+	finally:
+		# Ensure cleanup happens even on normal finish or errors
+		if thread is not None and thread.is_alive():
+			thread.join()
+		torch.cuda.empty_cache()
 
 """###########################################################################
 	Evaluate a given model (specified in model_cfgs)
@@ -723,25 +783,25 @@ if __name__ == '__main__':
 		### => summarization model
 		###	=> larger (12b) for better summarization
 		#################################################
-		model_aux, tokenizer_aux = FastVisionModel.from_pretrained(
-			model_name=cfgs["model_summarize"],
-			load_in_4bit=True,
-		)
-		FastVisionModel.for_inference(model)
-		if "gemma" in cfgs["model"]:
-			# gemma-specific tokenizer chat template
-			tokenizer_aux = get_chat_template(
-				tokenizer_aux,
-				chat_template = "gemma-3",
-			)
+		# model_aux, tokenizer_aux = FastVisionModel.from_pretrained(
+		# 	model_name=cfgs["model_summarize"],
+		# 	load_in_4bit=True,
+		# )
+		# FastVisionModel.for_inference(model)
+		# if "gemma" in cfgs["model"]:
+		# 	# gemma-specific tokenizer chat template
+		# 	tokenizer_aux = get_chat_template(
+		# 		tokenizer_aux,
+		# 		chat_template = "gemma-3",
+		# 	)
 	
-	# initialize streamer tokens
-	streamer = TextIteratorStreamer(
-		tokenizer, skip_prompt=True, skip_special_tokens=True
-	)
-	streamer_aux = TextIteratorStreamer(
-		tokenizer_aux, skip_prompt=True, skip_special_tokens=True
-	)
+	# # initialize streamer tokens
+	# streamer = TextIteratorStreamer(
+	# 	tokenizer, skip_prompt=True, skip_special_tokens=True
+	# )
+	# streamer_aux = TextIteratorStreamer(
+	# 	tokenizer_aux, skip_prompt=True, skip_special_tokens=True
+	# )
 	
 	"""=============================================
 		4. User-input Dropdown Traits
@@ -907,13 +967,6 @@ if __name__ == '__main__':
 		###    Health Topic	   ###
 		##########################
 		gr.Markdown("## 2. Please specify the main Health Topic of the poster here:")
-		gr.Markdown("""
-		#### Notes:
-		* Select the main Health Topic of the poster you will upload next.
-		* Please make sure the selected Health Topic matches the uploaded poster for best results.
-		* If you don’t upload a poster, the model will produce a general response emulating how a person with given traits would say about the selected Health Topic.
-		""",
-		)
 		# ---- dropdown at ~50% page width and centered ----
 		with gr.Row():
 			with gr.Column(scale=1):
@@ -928,15 +981,22 @@ if __name__ == '__main__':
 		##########################
 		gr.Markdown("## 3. Upload Public Health Poster here (if no poster is uploaded, the model emulates General Response to the topic):")
 		gr.Markdown("""
-		#### Notes:
-		##### Use case 1: Trait-conditioned Response Emulation to specific Health Poster 
-		* Upload **one poster image at a time** — only the first image will be processed. The model does **not retain memory**, so please re-upload an image for each new submission. 
-		* Ensure the uploaded poster **matches the selected Health Topic** for best results.  You do **not** need to type any text prompt — simply upload a poster and click **Submit**. 
-		* Please allow our model 5-10 seconds to summarize its poster understanding before emulating the response.
-		##### Use case 2: Trait-conditioned Generic Response Emulation to a Health Topic
-		* Simply select a Health Topic and click "Send" (do not upload any poster). 
+		#### ▶️ Use Case 1: Poster-Based Response
+		+ Upload **only one** poster image — the first file is the one processed.
+		+ The model has **no memory**, so re-upload the image for each new request.
+		+ Must choose a **Health Topic** that matches the poster content for best results.
+		+ No text prompt is needed: upload the poster and click **Submit**.
+		#### ▶️ Use Case 2: General Response (No Poster)
+		+ Simply select a Health Topic and click **Send**.  
 		"""
 		)
+		gr.Markdown("""
+		### 📘 Important Notes
+		- ⚠️ **Do not interrupt the generation process.** Stopping midway can cause backend issues. Please allow the response to complete.
+		- 🏷️ Before uploading a poster, select its **corresponding health topic**.
+		- 🎯 For the best experience, ensure the **topic accurately matches the poster content**.
+		- 🧩 If you choose not to upload a poster, the model will produce a **general, trait-conditioned response** for the selected topic.
+		""")
 		chat = gr.ChatInterface(
 			fn=vlm_response,
 			multimodal=True,  # text + image
@@ -947,18 +1007,20 @@ if __name__ == '__main__':
 				political, education, income, family_status,
 				# extraversion, agreeableness, conscientiousness, neuroticism, openness,
 			],
-			chatbot=gr.Chatbot(height=600), # height=330
+			chatbot=gr.Chatbot(height=500), # height=330
 			autofocus=False,
 		)
 
 	"""=============================================
 		5. Chat Interface Launch
 	============================================="""
-	# interface.queue(
-	# 	max_size=20, 
-	# 	default_concurrency_limit=1,
-	# ).launch(
-	# 	share=True, 
-	# 	max_threads=1,
-	# )
-	interface.queue().launch()
+	interface.queue(
+		max_size=20, 
+		default_concurrency_limit=1,
+	).launch(
+		share=True, 
+		max_threads=1,
+		# show_error=True,
+		# prevent_thread_lock=False,
+		# debug=True,
+	)

app.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+#SBATCH -c 16                   # 16 CPUs
+#SBATCH --mem=32g               # 32 GB RAM
+#SBATCH --gres=gpu:rtxa5000:1   # 1 GPU (A6000)
+#SBATCH --time=3-00:00:00       # 8 days
+#SBATCH --account=gamma
+#SBATCH --partition=gamma
+#SBATCH --qos=gamma-huge-long
+#SBATCH --output=/fs/nexus-projects/health_sim_ai/src_hf_deploy/app_logs/app_%j.out
+
+export HOME=/fs/nexus-projects/health_sim_ai
+cd /fs/nexus-projects/health_sim_ai
+source venvs/llm/bin/activate
+cd src_hf_deploy
+python -u app.py
+# python inference_pred_llm.py
+# python inference_rec_llm.py