update readme and demo

demo.py

@@ -23,6 +23,11 @@ import glob
 import os
 import time
 
+# Must be set before `import torch` / any CUDA init. Reduces the reserved-vs-allocated
+# memory gap by letting the caching allocator grow segments on demand instead of
+# pre-reserving fixed-size blocks.
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
 import cv2
 import numpy as np
 import torch
@@ -113,7 +118,7 @@ def load_model(args, device):
         enable_3d_rope=args.enable_3d_rope,
         max_frame_num=args.max_frame_num,
         kv_cache_sliding_window=args.kv_cache_sliding_window,
-        kv_cache_scale_frames=args.kv_cache_scale_frames,
+        kv_cache_scale_frames=args.num_scale_frames,
         kv_cache_cross_frame_special=True,
         kv_cache_include_scale_frames=True,
         use_sdpa=args.use_sdpa,
@@ -247,7 +252,7 @@ def main():
     # Streaming options
     parser.add_argument("--enable_3d_rope", action="store_true", default=True)
     parser.add_argument("--max_frame_num", type=int, default=1024)
-    parser.add_argument("--num_scale_frames", type=int, default=4)
+    parser.add_argument("--num_scale_frames", type=int, default=8)
     parser.add_argument(
         "--keyframe_interval",
         type=int,
@@ -255,7 +260,6 @@ def main():
         help="Streaming only. Every N-th frame after scale frames is kept as a keyframe. 1 = every frame.",
     )
     parser.add_argument("--kv_cache_sliding_window", type=int, default=64)
-    parser.add_argument("--kv_cache_scale_frames", type=int, default=8)
     parser.add_argument("--use_sdpa", action="store_true", default=False,
                         help="Use SDPA backend (no flashinfer needed). Default: FlashInfer")
     parser.add_argument(
@@ -265,7 +269,6 @@ def main():
         help="Offload per-frame predictions to CPU during inference to cut GPU peak memory. "
              "Use --no-offload_to_cpu to keep outputs on GPU.",
     )
-
     # Windowed options
     parser.add_argument("--window_size", type=int, default=64, help="Frames per window (windowed mode)")
     parser.add_argument("--overlap_size", type=int, default=16, help="Overlap between windows")
@@ -313,13 +316,21 @@ def main():
     model = load_model(args, device)
     print(f"Total load time: {time.time() - t0:.1f}s")
 
-    # Keep model in its loaded dtype — autocast handles bf16/fp16 for the ops
-    # that benefit from it and keeps LayerNorm / reductions in fp32.
+    # Pick inference dtype; autocast still runs for the ops that need fp32 (e.g. LayerNorm).
     if torch.cuda.is_available():
         dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] >= 8 else torch.float16
     else:
         dtype = torch.float32
 
+    # Cast the aggregator (DINOv2-style trunk) to the inference dtype to remove the
+    # redundant fp32 master weight copy + autocast bf16 weight cache (~2-3 GB saved,
+    # no measurable quality change). gct_base._predict_* upcasts inputs to fp32 and
+    # runs each head under `autocast(enabled=False)`, so camera/depth/point heads
+    # keep fp32 weights automatically.
+    if dtype != torch.float32 and getattr(model, "aggregator", None) is not None:
+        print(f"Casting aggregator to {dtype} (heads kept in fp32)")
+        model.aggregator = model.aggregator.to(dtype=dtype)
+
     images = images.to(device)
     num_frames = images.shape[0]
     print(f"Input: {num_frames} frames, shape {tuple(images.shape)}")