Update app.py

app.py

@@ -1,15 +1,32 @@
 import os
-
+import torch
 import gradio as gr
-
-os.environ["ROCR_VISIBLE_DEVICES"] = "0"
-os.environ["HIP_VISIBLE_DEVICES"] = "0"
-os.environ["HSA_OVERRIDE_GFX_VERSION"] = "9.4.2"
-
-BASE_MODEL   = "Qwen/Qwen2-1.5B"
-ADAPTER_PATH = "./outputs"
-
-
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from peft import PeftModel
+
+# ← CHANGE 1: ROCm env vars removed
+
+BASE_MODEL   = "Qwen/Qwen3-1.7B"
+ADAPTER_PATH = "HK2184/medqa-qwen3-lora"  # ← CHANGE 2: HF Hub instead of ./outputs
+
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
+tokenizer.pad_token    = tokenizer.eos_token
+tokenizer.padding_side = "left"
+
+print("Loading model...")
+DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32  # ← CHANGE 3: auto dtype
+base = AutoModelForCausalLM.from_pretrained(
+    BASE_MODEL,
+    dtype=DTYPE,
+    device_map="auto",
+    trust_remote_code=True,
+    low_cpu_mem_usage=True,
+)
+model = PeftModel.from_pretrained(base, ADAPTER_PATH)
+model = model.merge_and_unload()
+model.eval()
+print("Ready!")
 
 EXAMPLES = [
     ["Which artery is occluded in inferior MI with ST elevation in II, III, aVF?",
@@ -29,20 +46,29 @@ EXAMPLES = [
 def answer(question, opa, opb, opc, opd):
     if not question.strip():
         return "Please enter a question."
-
-    # simple mock logic (random-ish but believable)
-    import random
-    options = [opa, opb, opc, opd]
-    letters = ["A", "B", "C", "D"]
-
-    idx = random.randint(0, 3)
-
-    return f"""Answer: {letters[idx]}) {options[idx]}
-
-Explanation:
-This is a mock demo running without the full model.
-In the real system, a fine-tuned medical LLM analyzes the clinical context
-and selects the most appropriate answer based on learned patterns."""
+    if not all([opa.strip(), opb.strip(), opc.strip(), opd.strip()]):
+        return "Please fill in all four options."
+    prompt = (
+        f"### Question:\n{question}\n\n"
+        f"### Options:\n"
+        f"A) {opa}\nB) {opb}\nC) {opc}\nD) {opd}\n\n"
+        f"### Answer:\n"
+    )
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    with torch.no_grad():
+        out = model.generate(
+            **inputs,
+            max_new_tokens=200,
+            do_sample=True,
+            temperature=0.7,
+            top_p=0.9,
+            top_k=50,
+            repetition_penalty=1.3,
+            eos_token_id=tokenizer.eos_token_id,
+            pad_token_id=tokenizer.eos_token_id,
+        )
+    new = out[0][inputs["input_ids"].shape[-1]:]
+    return tokenizer.decode(new, skip_special_tokens=True)
 
 CSS = """
 @import url('https://fonts.googleapis.com/css2?family=Syne:wght@400;600;700;800&family=DM+Sans:wght@300;400;500&display=swap');
@@ -72,7 +98,6 @@ body, .gradio-container {
     padding: 0 20px 60px !important;
 }
 
-/* Header */
 #header {
     padding: 44px 0 28px;
     border-bottom: 1px solid var(--border);
@@ -105,7 +130,6 @@ h1#title {
 h1#title em { color: var(--accent); font-style: normal; }
 .subtitle { font-size: 14px; color: var(--muted); font-weight: 300; line-height: 1.6; max-width: 520px; }
 
-/* Stats */
 #stats {
     display: flex; border: 1px solid var(--border);
     border-radius: 12px; overflow: hidden;
@@ -118,7 +142,6 @@ h1#title em { color: var(--accent); font-style: normal; }
 .dot { display: inline-block; width: 6px; height: 6px; border-radius: 50%; background: var(--green); margin-right: 4px; animation: blink 2s infinite; }
 @keyframes blink { 0%,100%{opacity:1} 50%{opacity:0.3} }
 
-/* Inputs */
 label span, .label-wrap span {
     font-family: 'DM Sans', sans-serif !important;
     font-size: 11px !important; font-weight: 500 !important;
@@ -140,7 +163,6 @@ textarea:focus, input[type=text]:focus {
     outline: none !important;
 }
 
-/* Section labels */
 .section-label {
     font-size: 10px; font-weight: 600;
     letter-spacing: 0.12em; text-transform: uppercase;
@@ -152,7 +174,6 @@ textarea:focus, input[type=text]:focus {
     background: var(--accent); display: inline-block;
 }
 
-/* Button */
 button.lg.primary {
     background: linear-gradient(135deg, var(--accent2), var(--accent)) !important;
     border: none !important; border-radius: 10px !important;
@@ -165,7 +186,6 @@ button.lg.primary {
 }
 button.lg.primary:hover { opacity: 0.85 !important; transform: translateY(-1px) !important; }
 
-/* Output */
 .out-box textarea {
     background: var(--surface2) !important;
     border: 1px solid var(--border) !important;
@@ -174,7 +194,6 @@ button.lg.primary:hover { opacity: 0.85 !important; transform: translateY(-1px)
     color: var(--text) !important; min-height: 280px !important;
 }
 
-/* Examples */
 .examples-holder table {
     background: var(--surface) !important;
     border: 1px solid var(--border) !important;
@@ -187,7 +206,6 @@ button.lg.primary:hover { opacity: 0.85 !important; transform: translateY(-1px)
 }
 .examples-holder tr:hover td { background: var(--surface2) !important; cursor: pointer; }
 
-/* Footer */
 #footer {
     margin-top: 44px; padding-top: 22px;
     border-top: 1px solid var(--border);
@@ -206,7 +224,7 @@ with gr.Blocks(css=CSS, title="MedQA — AMD ROCm") as demo:
     <div id="header">
         <div class="badges">
             <span class="badge b-amd">AMD MI300X</span>
-            <span class="badge b-rocm">ROCm 6.1</span>
+            <span class="badge b-rocm">ROCm 7.2</span>
             <span class="badge b-lora">LoRA Fine-tuned</span>
             <span class="badge b-live"><span class="dot"></span>Live Inference</span>
         </div>
@@ -217,7 +235,7 @@ with gr.Blocks(css=CSS, title="MedQA — AMD ROCm") as demo:
         </p>
     </div>
     <div id="stats">
-        <div class="stat"><span class="sv">1.5B</span><span class="sl">Parameters</span></div>
+        <div class="stat"><span class="sv">1.7B</span><span class="sl">Parameters</span></div>
         <div class="stat"><span class="sv">LoRA</span><span class="sl">Fine-tuning</span></div>
         <div class="stat"><span class="sv">193k</span><span class="sl">Training QA</span></div>
         <div class="stat"><span class="sv">MI300X</span><span class="sl">AMD GPU</span></div>
@@ -256,17 +274,17 @@ with gr.Blocks(css=CSS, title="MedQA — AMD ROCm") as demo:
         examples=EXAMPLES,
         inputs=[question, opa, opb, opc, opd],
         label="",
-            )
+    )
 
     gr.HTML("""
     <div id="footer">
         <div class="fl">
             Built on <strong>AMD Developer Cloud</strong> &nbsp;·&nbsp;
-            Model: <strong>Qwen2-1.5B + LoRA</strong> &nbsp;·&nbsp;
+            Model: <strong>Qwen3-1.7B + LoRA</strong> &nbsp;·&nbsp;
             Dataset: <strong>MedMCQA</strong>
         </div>
         <div class="fr">
-            <a class="flink" href="https://github.com" target="_blank">GitHub →</a>
+            <a class="flink" href="https://github.com/HK2184/MedQA-Medical-AI-on-AMD-ROCm" target="_blank">GitHub →</a>
             <a class="flink" href="https://lablab.ai" target="_blank">lablab.ai →</a>
             <a class="flink" href="https://cloud.amd.com" target="_blank">AMD Cloud →</a>
         </div>
@@ -276,4 +294,4 @@ with gr.Blocks(css=CSS, title="MedQA — AMD ROCm") as demo:
     btn.click(fn=answer, inputs=[question, opa, opb, opc, opd], outputs=output)
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, share=True)
+    demo.launch()   # ← CHANGE 4: no server_name/port/share for HF Spaces