add minimal functionality using tensorrt decoder/encoder engines

nodes.py

@@ -1,20 +1,25 @@
+#nodes.py
+
 import torch
 import os
 import requests
-import numpy as np
 import logging
-import json
 import ast
-
 import sys
 
+import numpy as np #ugh
+import cv2 #double ugh
+
+import pycuda.driver as cuda
+import pycuda.autoinit  # Automatically initializes a CUDA context
+
 # Add the directory containing 'sam2_realtime' to sys.path
 current_directory = os.path.dirname(os.path.abspath(__file__))
 sam2_realtime_path = os.path.join(current_directory)  # Adjust the relative path
 sys.path.append(sam2_realtime_path)
 
 from sam2_realtime.sam2_tensor_predictor import SAM2TensorPredictor
-from comfy.utils import load_torch_file
+from sam2_realtime.sam2_tensorrt_predictor import Sam2TensorrtPredictor
 
 from omegaconf import OmegaConf
 from hydra.utils import instantiate
@@ -141,13 +146,13 @@ class Sam2RealtimeSegmentation:
         return {
             "required": {
                 "images": ("IMAGE",),
-                "sam2_model": ("SAM2MODEL",),
+                # "sam2_model": ("SAM2MODEL",),
                 # "keep_model_loaded": ("BOOLEAN", {"default": True}),
             },
             "optional": {
-                "coordinates_positive": ("STRING", ),
+                # "coordinates_positive": ("STRING", ),
-                "coordinates_negative": ("STRING", ),
+                # "coordinates_negative": ("STRING", ),
-                "reset_tracking": ("BOOLEAN", {"default": False}),
+                # "reset_tracking": ("BOOLEAN", {"default": False}),
                 # "bboxes": ("BBOX", ),
                 # "individual_objects": ("BOOLEAN", {"default": False}),
                 # "mask": ("MASK", ),
@@ -160,7 +165,26 @@ class Sam2RealtimeSegmentation:
     CATEGORY = "SAM2-Realtime"
 
     def __init__(self):
-        self.predictor = None
+        ##############################################################################
+        # SETUP: Load Engines, Create Contexts, Allocate Buffers
+        ##############################################################################
+        self.device = torch.device("cuda")
+        
+        cuda.init()
+        device = cuda.Device(0)
+        context = device.make_context()
+        print(f"Using device: {device.name()}")
+
+        # Allocate memory
+        mem = cuda.mem_alloc(1024)
+        print(f"Memory allocated at: {mem}")
+        # context.pop()
+
+        tensor_models_path = os.path.join(folder_paths.models_dir, "tensorrt")
+        self.predictor = Sam2TensorrtPredictor(
+            encoder_engine_path=os.path.join(tensor_models_path, "sam2_hiera_tiny.encoder.engine"),
+            decoder_engine_path=os.path.join(tensor_models_path, "sam2_hiera_tiny.decoder.engine")
+        )
         self.if_init = False
 
     def _process_coordinate_input(self, coordinates, label):
@@ -192,71 +216,60 @@ class Sam2RealtimeSegmentation:
     def segment_images(
         self,
         images,
-        sam2_model,
+        # sam2_model,
-        # keep_model_loaded,
+        # coordinates_positive=None,
-        coordinates_positive=None,
+        # coordinates_negative=None,
-        coordinates_negative=None,
+        # reset_tracking=False,
-        reset_tracking=False,
-        #point_labels=None,
-        # bboxes=None,
-        # individual_objects=False,
-        # mask=None,
     ):
-        model = sam2_model["model"]
+        
-        device = torch.device("cuda")
-        model.to(device)
 
         processed_frames = []
         mask_list = []
-        # The `model` variable is now ready and equivalent to `predictor` returned by sam2.build_sam.build_sam2_camera_predictor
         
-        if reset_tracking:
+        # if reset_tracking:
-            self.if_init = False
+        #     self.if_init = False
-            self.predictor = None
-
-        if self.predictor is None:
-            self.predictor = model
-
-        # Process coordinates once, outside the frame loop
-        pos_points, pos_labels = self._process_coordinate_input(coordinates_positive, 1)
-        neg_points, neg_labels = self._process_coordinate_input(coordinates_negative, 0)
-        all_points = pos_points + neg_points
-        all_labels = pos_labels + neg_labels
-
-        if all_points:
-            points_tensor = torch.tensor([all_points], device=device)
-            labels_tensor = torch.tensor([all_labels], device=device)
 
         with torch.inference_mode(), torch.autocast("cuda", dtype=torch.float16):
             for frame_idx, frame in enumerate(images):
-                frame = frame.to(device).float()
+                
+                # Tensor -> numpy array
+                frame_numpy = frame.numpy()  # shape: already (H,W,C)!!
+
+                # Convert from RGB -> BGR
+                frame_bgr = cv2.cvtColor(frame_numpy, cv2.COLOR_RGB2BGR)  # shape: (H,W,3), float32 in [0..1]
+
+                # Scale to [0..255] and cast to uint8
+                frame_bgr = (frame_bgr * 255).clip(0,255).astype(np.uint8)
 
                 if not self.if_init:
-                    self.predictor.load_first_frame(frame)
                     self.if_init = True
 
-                    if all_points:
+                    out_mask_logits = self.predictor.load_first_frame_and_prompt(
-                        _, _, out_mask_logits = self.predictor.add_new_prompt(
+                        frame_bgr,
-                            frame_idx=0,
+                        point_coord=(512, 512)
-                            obj_id=1,
+                    )
-                            points=points_tensor,
-                            labels=labels_tensor,
-                        )
-                    else:
-                        out_mask_logits = torch.zeros((0,), device=device)
                 else:
-                    out_obj_ids, out_mask_logits = self.predictor.track(frame)
+                    out_mask_logits = self.predictor.track(frame_bgr)
 
                 # Process mask logits
-                mask = self._process_mask_logits(out_mask_logits, frame.shape, device)
+                # mask = self._process_mask_logits(out_mask_logits, frame.shape, self.device)
+                mask = out_mask_logits
 
-                # Create colored overlay for processed frames
+                # # Create colored overlay for processed frames
-                mask_colored = torch.stack([mask] * 3, dim=2)
+                # mask_colored = torch.stack([mask] * 3, dim=2)
 
-                overlayed_frame = torch.add(frame * 0.7, mask_colored * 0.3)
+                # overlayed_frame = torch.add(frame * 0.7, mask_colored * 0.3)
                 
-                processed_frames.append(overlayed_frame)
+                # Create colored overlay for processed frames
-                mask_list.append(mask)
+                mask_colored = np.stack([mask] * 3, axis=2)  # Stack along the last dimension (HxWxC format)
+
+                mask_colored_resized = cv2.resize(mask_colored, (frame.shape[1], frame.shape[0]))
+
+                # Overlay the frame and the colored mask
+                overlaid_frame = frame * 0.7 + mask_colored_resized * 0.3
+
+                processed_frames.append(torch.tensor(overlaid_frame))
+                mask_list.append(torch.tensor(mask))
 
         # Stack masks and frames
         stacked_masks = torch.stack(mask_list, dim=0)
@@ -264,12 +277,63 @@ class Sam2RealtimeSegmentation:
         
         return (stacked_frames, stacked_masks)
 
+
+
+
+class Sam2RealtimeSegmentationTest:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                # "sam2_model": ("SAM2MODEL",),
+                # "keep_model_loaded": ("BOOLEAN", {"default": True}),
+            },
+            "optional": {
+                # "coordinates_positive": ("STRING", ),
+                # "coordinates_negative": ("STRING", ),
+                # "reset_tracking": ("BOOLEAN", {"default": False}),
+                # "bboxes": ("BBOX", ),
+                # "individual_objects": ("BOOLEAN", {"default": False}),
+                # "mask": ("MASK", ),
+            },
+        }
+
+    RETURN_NAMES = ("PROCESSED_IMAGES", "MASK",)
+    RETURN_TYPES = ("IMAGE", "MASK",)
+    FUNCTION = "segment_images"
+    CATEGORY = "SAM2-Realtime"
+
+    def segment_images(
+        self,
+        images,
+        # sam2_model,
+        # coordinates_positive=None,
+        # coordinates_negative=None,
+        # reset_tracking=False,
+    ):
+        processed_frames = []
+        mask_list = []
+        # processed_frames.append(images)
+        # images_temp = images
+        # coordinates_positive_temp = coordinates_positive.__str__
+        # reset_tracking_temp = reset_tracking
+
+        # Stack masks and frames
+        stacked_masks = torch.stack(mask_list, dim=0)
+        stacked_frames = torch.stack(processed_frames, dim=0)
+        
+        return (stacked_frames, stacked_masks)
+
+
 NODE_CLASS_MAPPINGS = {
     "DownloadAndLoadSAM2RealtimeModel": DownloadAndLoadSAM2RealtimeModel,
-    "Sam2RealtimeSegmentation": Sam2RealtimeSegmentation
+    "Sam2RealtimeSegmentation": Sam2RealtimeSegmentation,
+    "Sam2RealtimeSegmentationTest": Sam2RealtimeSegmentationTest
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "DownloadAndLoadSAM2RealtimeModel": "(Down)Load sam2_realtime Model",
-    "Sam2RealtimeSegmentation": "Sam2RealtimeSegmentation"
+    "Sam2RealtimeSegmentation": "Sam2RealtimeSegmentation",
+    "Sam2RealtimeSegmentationTest": "Sam2RealtimeSegmentationTest"
 }

requirements.txt

@@ -1,7 +1,16 @@
 pyyaml
+ninja
 numpy>=1.24.4
 tqdm>=4.66.1
 hydra-core>=1.3.2
 iopath>=0.1.10
 pillow>=9.4.0
-git+https://github.com/pschroedl/ComfyUI-SAM2-Realtime.git@main#egg=sam2_realtime
+opencv-python>=4.5.0
+pycuda>=2022.1
+tensorrt-cu12==10.7.0
+tensorrt-cu12-bindings==10.7.0
+tensorrt-cu12-libs==10.7.0
+tensorrt==10.7.0
+# git+https://github.com/pschroedl/ComfyUI-SAM2-Realtime.git@main#egg=sam2_realtime
+pip install -e .
+

sam2_realtime/sam2_tensorrt.py

@@ -0,0 +1,241 @@
+import sys
+import numpy as np
+import cv2
+
+import pycuda.driver as cuda
+import pycuda.autoinit  # Automatically initializes a CUDA context
+import tensorrt as trt
+
+TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
+
+ENCODER_ENGINE_PATH = "sam2_hiera_tiny.encoder.engine"
+DECODER_ENGINE_PATH = "sam2_hiera_tiny.decoder.engine"
+
+def load_engine(engine_path: str) -> trt.ICudaEngine:
+    """
+    Load a TensorRT engine from a file.
+    """
+    with open(engine_path, "rb") as f:
+        engine_data = f.read()
+    runtime = trt.Runtime(TRT_LOGGER)
+    engine = runtime.deserialize_cuda_engine(engine_data)
+    return engine
+
+def allocate_io_tensors(engine: trt.ICudaEngine):
+    """
+    Allocates device+host buffers for each I/O tensor in name-based I/O.
+    Returns (device_buffers, host_buffers, input_names, output_names).
+    """
+    device_buffers = {}
+    host_buffers = {}
+    input_names = []
+    output_names = []
+
+    n_tensors = engine.num_io_tensors
+    for i in range(n_tensors):
+        tname = engine.get_tensor_name(i)
+        mode = engine.get_tensor_mode(tname)  # trt.TensorIOMode.INPUT or OUTPUT
+        shape = engine.get_tensor_shape(tname)
+        dtype = engine.get_tensor_dtype(tname)
+
+        # Convert TRT dtype -> numpy dtype
+        if dtype == trt.float32:
+            np_dtype = np.float32
+        elif dtype == trt.float16:
+            np_dtype = np.float16
+        else:
+            raise ValueError(f"Unsupported TRT dtype: {dtype}")
+
+        volume = np.prod(shape)
+        host_mem = np.zeros(volume, dtype=np_dtype)
+        device_mem = cuda.mem_alloc(host_mem.nbytes)
+
+        device_buffers[tname] = device_mem
+        host_buffers[tname] = host_mem
+
+        if mode == trt.TensorIOMode.INPUT:
+            input_names.append(tname)
+        else:
+            output_names.append(tname)
+
+    return device_buffers, host_buffers, input_names, output_names
+
+def do_inference_async(
+    context: trt.IExecutionContext,
+    device_buffers: dict,
+    host_buffers: dict,
+    input_names: list,
+    output_names: list,
+    cuda_stream: cuda.Stream
+):
+    """
+    1) Copy inputs (host->device)
+    2) context.set_tensor_address(...)
+    3) context.execute_async_v3(cuda_stream.handle)
+    4) stream.synchronize()
+    5) Copy outputs (device->host)
+    6) return output in the same order as output_names
+    """
+    # 1) Copy inputs to device
+    for tname in input_names:
+        cuda.memcpy_htod_async(device_buffers[tname], host_buffers[tname], stream=cuda_stream)
+
+    # 2) Set addresses in context
+    for tname in input_names + output_names:
+        context.set_tensor_address(tname, int(device_buffers[tname]))
+
+    # 3) Execute
+    success = context.execute_async_v3(cuda_stream.handle)
+    if not success:
+        raise RuntimeError("TensorRT execute_async_v3() returned False.")
+
+    # 4) Synchronize to ensure inference is done before copying back
+    cuda_stream.synchronize()
+
+    # 5) Copy outputs back to host
+    for tname in output_names:
+        cuda.memcpy_dtoh_async(host_buffers[tname], device_buffers[tname], stream=cuda_stream)
+
+    # Wait until all D->H copies are done
+    cuda_stream.synchronize()
+
+    # 6) Gather outputs
+    outputs = [host_buffers[t].copy() for t in output_names]
+    return outputs
+
+def run_encoder_async(
+    image_bgr: np.ndarray,
+    context: trt.IExecutionContext,
+    device_buffers: dict,
+    host_buffers: dict,
+    input_names: list,
+    output_names: list,
+    image_size: tuple = (1024, 1024),
+    cuda_stream: cuda.Stream = None
+):
+    """
+    Run the SAM encoder in async mode.
+    """
+    if cuda_stream is None:
+        cuda_stream = cuda.Stream()
+
+    # Preprocess image
+    H, W = image_size
+    resized = cv2.resize(image_bgr, (W, H))
+    rgb = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
+    nchw = np.transpose(rgb, (2, 0, 1))[None, ...]  # (1,3,H,W)
+
+    if len(input_names) != 1:
+        raise RuntimeError(f"Expected exactly 1 input for encoder, got {input_names}")
+
+    in_name = input_names[0]
+    host_buffers[in_name][:] = nchw.flatten()
+
+    # Async inference
+    outputs = do_inference_async(
+        context,
+        device_buffers,
+        host_buffers,
+        input_names,
+        output_names,
+        cuda_stream
+    )
+
+    # Expect 3 outputs
+    if len(outputs) != 3:
+        raise RuntimeError(f"Encoder expected 3 outputs, got {len(outputs)}")
+
+    hr0_flat, hr1_flat, emb_flat = outputs
+    # Reshape according to model's output shapes
+    # Shapes for sam2_hiera_tiny:
+    #   hr0: (1, 32, 256, 256)
+    #   hr1: (1, 64, 128, 128)
+    #   emb: (1, 256, 64, 64)
+    hr0 = hr0_flat.reshape((1, 32, 256, 256))
+    hr1 = hr1_flat.reshape((1, 64, 128, 128))
+    emb = emb_flat.reshape((1, 256, 64, 64))
+
+    print("[DEBUG] Encoder outputs shapes:", hr0.shape, hr1.shape, emb.shape)
+    return hr0, hr1, emb
+
+def run_decoder_async(
+    high_res_feats_0: np.ndarray,
+    high_res_feats_1: np.ndarray,
+    image_embed: np.ndarray,
+    point_coords: np.ndarray,
+    point_labels: np.ndarray,
+    mask_input: np.ndarray,
+    has_mask_input: np.ndarray,
+    context: trt.IExecutionContext,
+    device_buffers: dict,
+    host_buffers: dict,
+    input_names: list,
+    output_names: list,
+    cuda_stream: cuda.Stream = None
+):
+    """
+    Run the SAM decoder in async mode with 7 inputs:
+      0) image_embed
+      1) high_res_feats_0
+      2) high_res_feats_1
+      3) point_coords
+      4) point_labels
+      5) mask_input
+      6) has_mask_input
+    """
+    if cuda_stream is None:
+        cuda_stream = cuda.Stream()
+
+    if len(input_names) != 7:
+        raise RuntimeError(f"Decoder expects 7 inputs, got {len(input_names)}: {input_names}")
+
+    # Flatten & place data
+    host_buffers[input_names[0]][:] = image_embed.flatten()
+    host_buffers[input_names[1]][:] = high_res_feats_0.flatten()
+    host_buffers[input_names[2]][:] = high_res_feats_1.flatten()
+    host_buffers[input_names[3]][:] = point_coords.flatten()
+    host_buffers[input_names[4]][:] = point_labels.flatten()
+    host_buffers[input_names[5]][:] = mask_input.flatten()
+    host_buffers[input_names[6]][:] = has_mask_input.flatten()
+
+    outputs = do_inference_async(
+        context,
+        device_buffers,
+        host_buffers,
+        input_names,
+        output_names,
+        cuda_stream
+    )
+
+    # Typically 2 outputs: masks, iou_predictions
+    if len(outputs) != 2:
+        raise RuntimeError(f"Decoder expected 2 outputs, got {len(outputs)}")
+
+    masks_flat, iou_flat = outputs
+    # Reshape as needed for your model
+    # Example: (1,3,256,256) for masks and (1,3) for iou_preds
+    masks = masks_flat.reshape((1, 3, 256, 256))
+    iou_preds = iou_flat.reshape((1, 3))
+    print("Decoder outputs shapes:", masks.shape, iou_preds.shape)
+    return masks, iou_preds
+
+def prepare_prompts(
+    batch_size: int = 1,
+    image_size: tuple = (1024, 1024),
+):
+    """
+    Create minimal prompt inputs for the decoder:
+      - point_coords (B, 1, 2)
+      - point_labels (B, 1)
+      - mask_input   (B, 1, 256, 256) (for sam2_hiera_tiny)
+      - has_mask_input (B,)
+    """
+    H, W = image_size
+    point_coords = np.array([[[W // 2, H // 2]]], dtype=np.float32)  # shape (B,1,2)
+    point_labels = np.array([[1]], dtype=np.float32)  # shape (B,1)
+
+    # For sam2_hiera_tiny => final embedding is 64×64, so mask_input is 4× that => 256×256
+    mask_input = np.zeros((batch_size, 1, 256, 256), dtype=np.float32)
+    has_mask_input = np.zeros((batch_size,), dtype=np.float32)
+
+    return point_coords, point_labels, mask_input, has_mask_input

sam2_realtime/sam2_tensorrt_predictor.py

@@ -0,0 +1,203 @@
+import numpy as np
+import cv2
+
+from sam2_realtime.sam2_tensorrt import (
+    load_engine,
+    allocate_io_tensors,
+    run_encoder_async,
+    run_decoder_async,
+    prepare_prompts
+)
+
+class Sam2TensorrtPredictor:
+    """
+    A simplified single-object predictor that:
+      - Uses the TensorRT-based SAM2 encoder/decoder.
+      - Tracks across frames by feeding the previously predicted mask as a prompt.
+      - Allows resetting the tracking with a new point prompt (i.e., ignoring prior mask).
+    """
+
+    def __init__(
+        self,
+        encoder_engine_path: str,
+        decoder_engine_path: str,
+        image_size=(1024, 1024)
+    ):
+        """
+        1) Load TensorRT engines
+        2) Create execution contexts
+        3) Allocate device+host buffers
+        """
+        self.image_size = image_size  # (height, width)
+
+        print("[Sam2TensorrtPredictor] Loading TensorRT engines...")
+        self.encoder_engine = load_engine(encoder_engine_path)
+        self.decoder_engine = load_engine(decoder_engine_path)
+
+        # Create execution contexts
+        self.encoder_context = self.encoder_engine.create_execution_context()
+        self.decoder_context = self.decoder_engine.create_execution_context()
+
+        # Allocate buffers for encoder
+        (
+            self.enc_device_bufs,
+            self.enc_host_bufs,
+            self.enc_in_names,
+            self.enc_out_names
+        ) = allocate_io_tensors(self.encoder_engine)
+
+        # Allocate buffers for decoder
+        (
+            self.dec_device_bufs,
+            self.dec_host_bufs,
+            self.dec_in_names,
+            self.dec_out_names
+        ) = allocate_io_tensors(self.decoder_engine)
+
+        print("[Sam2TensorrtPredictor] Engines loaded and buffers allocated.")
+
+        # Internal states
+        self._initialized = False
+        self.prev_mask  = None
+        self.prev_feats0 = None
+        self.prev_feats1 = None
+        self.prev_embed  = None
+
+    def load_first_frame_and_prompt(self, frame_bgr: np.ndarray, point_coord, point_label=1):
+        """
+        1) Encode the first frame to get feats0, feats1, embed.
+        2) Use the user-provided point prompt to decode an initial mask.
+        3) Store that mask in self.prev_mask (for subsequent frames).
+        """
+        self._initialized = False
+        self.prev_mask  = None
+        self.prev_feats0 = None
+        self.prev_feats1 = None
+        self.prev_embed  = None
+
+        # -- (1) run the encoder
+        feats0, feats1, embed = run_encoder_async(
+            image_bgr=frame_bgr,
+            context=self.encoder_context,
+            device_buffers=self.enc_device_bufs,
+            host_buffers=self.enc_host_bufs,
+            input_names=self.enc_in_names,
+            output_names=self.enc_out_names,
+            image_size=self.image_size,
+        )
+
+        # -- (2) build the prompt arrays
+        batch_size = 1
+        pt_coords, pt_labels, zero_mask, zero_has_mask = prepare_prompts(
+            batch_size=batch_size,
+            image_size=self.image_size
+        )
+
+        # Overwrite the default center prompt with the user-provided point
+        #   Note: point_coord = (x, y)
+        pt_coords[0, 0, 0] = float(point_coord[0])  # x
+        pt_coords[0, 0, 1] = float(point_coord[1])  # y
+        pt_labels[0, 0]    = float(point_label)
+
+        # decode (prompt-based)
+        masks, iou_preds = run_decoder_async(
+            high_res_feats_0=feats0,
+            high_res_feats_1=feats1,
+            image_embed=embed,
+            point_coords=pt_coords,
+            point_labels=pt_labels,
+            mask_input=zero_mask,      # no prior mask for the first frame
+            has_mask_input=zero_has_mask,
+            context=self.decoder_context,
+            device_buffers=self.dec_device_bufs,
+            host_buffers=self.dec_host_bufs,
+            input_names=self.dec_in_names,
+            output_names=self.dec_out_names,
+        )
+
+        # pick best mask
+        best_idx = np.argmax(iou_preds[0])  # choose the highest IOU channel out of 3
+        mask_logit = masks[0, best_idx]     # shape: (256, 256)
+        mask_prob = 1 / (1 + np.exp(-mask_logit))
+        mask_bin = (mask_prob > 0.5).astype(np.uint8)
+
+        # -- (3) store for next step
+        self.prev_mask  = mask_bin
+        self.prev_feats0 = feats0
+        self.prev_feats1 = feats1
+        self.prev_embed  = embed
+        self._initialized = True
+
+        print("[load_first_frame_and_prompt] First mask computed from point prompt.")
+        return mask_bin
+
+    def track(self, frame_bgr: np.ndarray):
+        """
+        For each subsequent frame:
+          1) Encode the new frame
+          2) Provide the previously predicted mask as 'mask_input' prompt
+          3) Pick best mask => new self.prev_mask
+        """
+        if not self._initialized or self.prev_mask is None:
+            raise RuntimeError(
+                "Sam2TensorrtPredictor not initialized with a first frame/prompt. "
+                "Call load_first_frame_and_prompt(...) first."
+            )
+
+        # 1) Encode the new frame
+        feats0, feats1, embed = run_encoder_async(
+            image_bgr=frame_bgr,
+            context=self.encoder_context,
+            device_buffers=self.enc_device_bufs,
+            host_buffers=self.enc_host_bufs,
+            input_names=self.enc_in_names,
+            output_names=self.enc_out_names,
+            image_size=self.image_size,
+        )
+
+        # 2) Prepare mask_input from the previously predicted binary mask
+        if self.prev_mask.shape != (256, 256):
+            # if stored differently, resize to 256x256
+            pm = cv2.resize(
+                self.prev_mask.astype(np.float32),
+                (256, 256),
+                interpolation=cv2.INTER_LINEAR
+            )[None, None]  # shape: (1,1,256,256)
+        else:
+            pm = self.prev_mask[None, None].astype(np.float32)
+
+        has_mask_input = np.array([1.0], dtype=np.float32)  # shape: (1,)
+
+        # 3) We want no new user points, we set them to dummy zero
+        point_coords = np.zeros((1, 1, 2), dtype=np.float32)
+        point_labels = np.zeros((1, 1), dtype=np.float32)
+
+        masks, iou_preds = run_decoder_async(
+            high_res_feats_0=feats0,
+            high_res_feats_1=feats1,
+            image_embed=embed,
+            point_coords=point_coords,
+            point_labels=point_labels,
+            mask_input=pm,
+            has_mask_input=has_mask_input,
+            context=self.decoder_context,
+            device_buffers=self.dec_device_bufs,
+            host_buffers=self.dec_host_bufs,
+            input_names=self.dec_in_names,
+            output_names=self.dec_out_names,
+        )
+
+        # 3.5) pick best mask
+        best_idx = np.argmax(iou_preds[0])  # shape: (3,)
+        mask_logit = masks[0, best_idx]     # shape: (256, 256)
+        mask_prob = 1 / (1 + np.exp(-mask_logit))
+        mask_bin = (mask_prob > 0.5).astype(np.uint8)
+
+        # 4) update internal states
+        self.prev_mask  = mask_bin
+        self.prev_feats0 = feats0
+        self.prev_feats1 = feats1
+        self.prev_embed  = embed
+
+        print("[track_next_frame] Tracking updated with previous mask as prompt.")
+        return mask_bin

setup.py

@@ -36,6 +36,7 @@ def get_extensions():
     compile_args = {
         "cxx": [],
         "nvcc": [
+            "--compiler-bindir=/usr/bin/gcc-12",
             "-DCUDA_HAS_FP16=1",
             "-D__CUDA_NO_HALF_OPERATORS__",
             "-D__CUDA_NO_HALF_CONVERSIONS__",