feat: upgrade holographic resolution to 2048 bands on NVIDIA A100

app.py

@@ -60,9 +60,9 @@ def kalpana_generate(prompt: str, max_tokens: int = 256, temperature: float = 0.
     inputs = _LOCAL_TOKENIZER(fmt, return_tensors="pt").to(device)
 
     num_layers = getattr(_LOCAL_QWEN_MODEL.config, "num_hidden_layers", 24)
-    # Use bands=512 to keep VRAM footprint small on ZeroGPU shared GPU
-    cache = KalpanaDynamicCache(num_layers=num_layers, bands=512)
-    print(f"[API] Cache: {num_layers} layers, is_sliding[:3]={cache.is_sliding[:3]}")
+    # 2048 bands for high-fidelity holographic resolution (4x quality)
+    cache = KalpanaDynamicCache(num_layers=num_layers, bands=2048)
+    print(f"[API] Cache: {num_layers} layers, bands=2048, is_sliding[:3]={cache.is_sliding[:3]}")
 
     t0 = time.perf_counter()
     with torch.inference_mode():

kalpana_embed_to_kv/core.py

@@ -24,7 +24,7 @@ class KalpanaRIFTensor(nn.Module):
         self,
         batch_size: int = 1,
         num_heads: int = 8,
-        bands: int = 512,
+        bands: int = 2048,
         dim: int = 128,
         kappa: float = 1.0,
         min_freq: float = 0.1,