Add all popular open-source models to ALL_MODELS for Spaces

app.py

@@ -10,41 +10,44 @@ image = modal.Image.debian_slim().pip_install(
     "gradio"
 )
 
-app = modal.App("agentic-demo")
+app = modal.App("agentic-demo", image=image)
 
 ALL_MODELS = [
-    "meta-llama/Llama-2-70b-hf",
-    "mistralai/Mixtral-8x7B-v0.1",
     "Qwen/Qwen-72B",
-    "mistralai/Mistral-7B-v0.1"
+    "deepseek-ai/deepseek-llm-67b-base",
+    "openchat/openchat-3.5-1210",
+    "microsoft/phi-2",
+    "google/gemma-7b",
+    "01-ai/Yi-34B",
+    "upstage/SOLAR-10.7B-v1.0",
+    "microsoft/Orca-2-13b",
+    "lmsys/vicuna-13b-v1.5"
 ]
 
 def debate_agent(topic, agent_a_model, agent_b_model, judge_model):
     if len({agent_a_model, agent_b_model, judge_model}) < 3:
         return {"error": "Please select three different models."}
     # Agent A
-    tokenizer_a = AutoTokenizer.from_pretrained(agent_a_model, token=os.environ["HUGGINGFACE_API_KEY"])
+    tokenizer_a = AutoTokenizer.from_pretrained(agent_a_model)
     model_a = AutoModelForCausalLM.from_pretrained(
         agent_a_model,
-        token=os.environ["HUGGINGFACE_API_KEY"],
         load_in_4bit=True,
         device_map="auto"
     )
     prompt_a = f"Debate as Agent A: {topic}"
     inputs_a = tokenizer_a(prompt_a, return_tensors="pt").to(model_a.device)
-    outputs_a = model_a.generate(**inputs_a, max_new_tokens=200)
+    outputs_a = model_a.generate(**inputs_a, max_new_tokens=10000)
     arg_a = tokenizer_a.decode(outputs_a[0], skip_special_tokens=True)
     # Agent B
-    tokenizer_b = AutoTokenizer.from_pretrained(agent_b_model, token=os.environ["HUGGINGFACE_API_KEY"])
+    tokenizer_b = AutoTokenizer.from_pretrained(agent_b_model)
     model_b = AutoModelForCausalLM.from_pretrained(
         agent_b_model,
-        token=os.environ["HUGGINGFACE_API_KEY"],
         load_in_4bit=True,
         device_map="auto"
     )
     prompt_b = f"Debate as Agent B: {topic}"
     inputs_b = tokenizer_b(prompt_b, return_tensors="pt").to(model_b.device)
-    outputs_b = model_b.generate(**inputs_b, max_new_tokens=200)
+    outputs_b = model_b.generate(**inputs_b, max_new_tokens=10000)
     arg_b = tokenizer_b.decode(outputs_b[0], skip_special_tokens=True)
     # Judge
     judge_prompt = (
@@ -54,15 +57,14 @@ def debate_agent(topic, agent_a_model, agent_b_model, judge_model):
         f"Agent B says: {arg_b}\n"
         f"Summarize both arguments and pick a winner (A or B) with a short justification."
     )
-    tokenizer_j = AutoTokenizer.from_pretrained(judge_model, token=os.environ["HUGGINGFACE_API_KEY"])
+    tokenizer_j = AutoTokenizer.from_pretrained(judge_model)
     model_j = AutoModelForCausalLM.from_pretrained(
         judge_model,
-        token=os.environ["HUGGINGFACE_API_KEY"],
         load_in_4bit=True,
         device_map="auto"
     )
     inputs_j = tokenizer_j(judge_prompt, return_tensors="pt").to(model_j.device)
-    outputs_j = model_j.generate(**inputs_j, max_new_tokens=200)
+    outputs_j = model_j.generate(**inputs_j, max_new_tokens=10000)
     judge_summary = tokenizer_j.decode(outputs_j[0], skip_special_tokens=True)
     return {
         "Agent A": arg_a,