fix: CUDA assert on T4 GPU - use eager attention, disable torch.compile

- Revert dtype param to torch_dtype for qwen_tts compatibility
- Switch from SDPA to eager attention (SDPA causes asserts on T4/Turing)
- Disable torch.compile (unstable on older GPU architectures)
- Add clickable book cards on Explore tab (navigates to Library + plays)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

app.py

@@ -140,6 +140,7 @@ with open(os.path.join(os.path.dirname(__file__), "static", "style.css"), encodi
 def generate_dashboard_html(books_list):
     featured = books_list[0]
     others = books_list[1:]
+    featured_js_title = featured['title'].replace("'", "\\'").replace('"', '\\"')
 
     html = f"""
     <div style="margin-bottom: 24px;">
@@ -148,7 +149,20 @@ def generate_dashboard_html(books_list):
     </div>
 
     <!-- Hero Spotlight Section -->
-    <div class="hero-spotlight" style="display: flex; flex-direction: row; gap: 32px; align-items: center; margin-bottom: 32px; flex-wrap: wrap;">
+    <div class="hero-spotlight" style="display: flex; flex-direction: row; gap: 32px; align-items: center; margin-bottom: 32px; flex-wrap: wrap; cursor: pointer;" onclick="
+        var tabs = document.querySelectorAll('button[role=tab]');
+        if (tabs.length > 1) {{ tabs[1].click(); }}
+        setTimeout(function() {{
+            var radios = document.querySelectorAll('.library-book-selector input[type=radio]');
+            radios.forEach(function(r) {{
+                if(r.value === '{featured_js_title}') {{
+                    r.checked = true;
+                    r.dispatchEvent(new Event('input', {{bubbles: true}}));
+                    r.dispatchEvent(new Event('change', {{bubbles: true}}));
+                }}
+            }});
+        }}, 300);
+    ">
         <div style="flex: 1; min-width: 200px; max-width: 280px; text-align: center;">
             <img src="{featured['cover_url']}" style="width: 180px; height: 260px; object-fit: cover; border-radius: 16px; box-shadow: 0 8px 24px rgba(0,0,0,0.3); transform: rotate(-2deg);" referrerPolicy="no-referrer" />
         </div>
@@ -174,8 +188,24 @@ def generate_dashboard_html(books_list):
     """
 
     for bk in books_list:
+        js_safe_title = bk['title'].replace("'", "\\'").replace('"', '\\"')
         html += f"""
-        <div class="shelf-card">
+        <div class="shelf-card" style="cursor: pointer;" onclick="
+            // Switch to My Library tab (2nd tab button)
+            var tabs = document.querySelectorAll('button[role=tab]');
+            if (tabs.length > 1) {{ tabs[1].click(); }}
+            // Select book in hidden radio
+            setTimeout(function() {{
+                var radios = document.querySelectorAll('.library-book-selector input[type=radio]');
+                radios.forEach(function(r) {{
+                    if(r.value === '{js_safe_title}') {{
+                        r.checked = true;
+                        r.dispatchEvent(new Event('input', {{bubbles: true}}));
+                        r.dispatchEvent(new Event('change', {{bubbles: true}}));
+                    }}
+                }});
+            }}, 300);
+        ">
             <div>
                 <img src="{bk['cover_url']}" class="cover-image" referrerPolicy="no-referrer" />
             </div>

voice_clone.py

@@ -99,13 +99,13 @@ def _select_dtype() -> torch.dtype:
 
 
 def _select_attn_impl() -> str:
-    """Use FlashAttention-2 if available, else default."""
+    """Use FlashAttention-2 if available, else eager (safest fallback)."""
     try:
         import flash_attn  # noqa: F401
         return "flash_attention_2"
     except ImportError:
-        logger.info("flash-attn not installed — using default attention.")
-        return "sdpa"
+        logger.info("flash-attn not installed — using eager attention.")
+        return "eager"
 
 
 def get_qwen_tts():
@@ -127,7 +127,7 @@ def get_qwen_tts():
                 _qwen_tts_model = Qwen3TTSModel.from_pretrained(
                     model_id,
                     device_map="cuda" if torch.cuda.is_available() else "cpu",
-                    dtype=dtype,
+                    torch_dtype=dtype,
                     attn_implementation=attn_impl,
                 )
 
@@ -156,7 +156,7 @@ def get_custom_voice_model():
                 _custom_voice_model = Qwen3TTSModel.from_pretrained(
                     CUSTOM_VOICE_MODEL_ID,
                     device_map="cuda" if torch.cuda.is_available() else "cpu",
-                    dtype=dtype,
+                    torch_dtype=dtype,
                     attn_implementation=attn_impl,
                 )
                 _try_torch_compile(_custom_voice_model)
@@ -165,19 +165,10 @@ def get_custom_voice_model():
 
 
 def _try_torch_compile(wrapper):
-    """Best-effort torch.compile on model submodules."""
-    if not hasattr(torch, "compile"):
-        return
-    model = wrapper.model
-    for name in ("decoder", "predictor", "speech_tokenizer"):
-        submod = getattr(model, name, None)
-        if submod is not None and isinstance(submod, torch.nn.Module):
-            try:
-                compiled = torch.compile(submod, mode="reduce-overhead")
-                setattr(model, name, compiled)
-                logger.info("torch.compile applied to model.%s", name)
-            except Exception as exc:
-                logger.debug("torch.compile skipped for %s: %s", name, exc)
+    """Best-effort torch.compile on model submodules. Disabled for stability."""
+    # torch.compile can cause CUDA asserts on some GPU architectures (T4/Turing)
+    # Disable for now in favor of stability
+    return
 
 
 def _trim_reference_audio(audio_path: str) -> str: