MacOS Support: tests + devShell + mujoco

README.md

@@ -54,6 +54,22 @@ We are shipping a first look at the DIMOS x Unitree Go2 integration, allowing fo
 - **DimOS Interface / Development Tools**
   - Local development interface to control your robot, orchestrate agents, visualize camera/lidar streams, and debug your dimensional agentive application.
 
+## MacOS Installation
+
+```sh
+# Install Nix
+curl --proto '=https' --tlsv1.2 -sSf -L https://install.determinate.systems/nix | sh -s -- install
+
+# clone the repository
+git clone --branch dev --single-branch https://github.com/dimensionalOS/dimos.git
+
+# setup the environment (follow the prompts after nix develop)
+cd dimos
+nix develop
+
+# You should be able to follow the instructions below as well for a more manual installation
+```
+
 ---
 ## Python Installation
 Tested on Ubuntu 22.04/24.04
@@ -83,10 +99,10 @@ pip install torch==2.0.1 torchvision torchaudio --index-url https://download.pyt
 #### Install dependencies
 ```bash
 # CPU only (reccomended to attempt first)
-pip install -e .[cpu,dev]
+pip install -e '.[cpu,dev]'
 
 # CUDA install
-pip install -e .[cuda,dev]
+pip install -e '.[cuda,dev]'
 
 # Copy and configure environment variables
 cp default.env .env
@@ -99,27 +115,27 @@ pytest -s dimos/
 
 #### Test Dimensional with a replay UnitreeGo2 stream (no robot required)
 ```bash
-CONNECTION_TYPE=replay python dimos/robot/unitree_webrtc/unitree_go2.py
+dimos --replay run unitree-go2
 ```
 
 #### Test Dimensional with a simulated UnitreeGo2 in MuJoCo (no robot required)
 ```bash
-pip install -e .[sim]
+pip install -e '.[sim]'
 export DISPLAY=:1 # Or DISPLAY=:0 if getting GLFW/OpenGL X11 errors
-CONNECTION_TYPE=mujoco python dimos/robot/unitree_webrtc/unitree_go2.py
+dimos --simulation run unitree-go2
 ```
 
 #### Test Dimensional with a real UnitreeGo2 over WebRTC
 ```bash
 export ROBOT_IP=192.168.X.XXX # Add the robot IP address
-python dimos/robot/unitree_webrtc/unitree_go2.py
+dimos run unitree-go2
 ```
 
 #### Test Dimensional with a real UnitreeGo2 running Agents
 *OpenAI / Alibaba keys required*
 ```bash
 export ROBOT_IP=192.168.X.XXX # Add the robot IP address
-python dimos/robot/unitree_webrtc/run_agents2.py
+dimos run unitree-go2-agentic
 ```
 ---
 

dimos/agents/memory/chroma_impl.py

@@ -145,10 +145,18 @@ def __init__(
     def create(self) -> None:
         """Create local embedding model and initialize the ChromaDB client."""
         # Load the sentence transformer model
-        # Use CUDA if available, otherwise fall back to CPU
-        device = "cuda" if torch.cuda.is_available() else "cpu"
-        print(f"Using device: {device}")
-        self.model = SentenceTransformer(self.model_name, device=device)  # type: ignore[name-defined]
+
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
+            self.device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            self.device = "mps"
+        else:
+            self.device = "cpu"
+
+        print(f"Using device: {self.device}")
+        self.model = SentenceTransformer(self.model_name, device=self.device)  # type: ignore[name-defined]
 
         # Create a custom embedding class that implements the embed_query method
         class SentenceTransformerEmbeddings:

dimos/agents/memory/image_embedding.py

@@ -22,6 +22,7 @@
 import base64
 import io
 import os
+import sys
 
 import cv2
 import numpy as np
@@ -76,6 +77,12 @@ def _initialize_model(self):  # type: ignore[no-untyped-def]
                 processor_id = "openai/clip-vit-base-patch32"
 
                 providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+                if sys.platform == "darwin":
+                    # 2025-11-17 12:36:47.877215 [W:onnxruntime:, helper.cc:82 IsInputSupported] CoreML does not support input dim > 16384. Input:text_model.embeddings.token_embedding.weight, shape: {49408,512}
+                    # 2025-11-17 12:36:47.878496 [W:onnxruntime:, coreml_execution_provider.cc:107 GetCapability] CoreMLExecutionProvider::GetCapability, number of partitions supported by CoreML: 88 number of nodes in the graph: 1504 number of nodes supported by CoreML: 933
+                    providers = ["CoreMLExecutionProvider"] + [
+                        each for each in providers if each != "CUDAExecutionProvider"
+                    ]
 
                 self.model = ort.InferenceSession(str(model_id), providers=providers)
 

dimos/models/Detic/tools/dump_clip_features.py

@@ -40,7 +40,14 @@
     if args.use_underscore:
         cat_names = [x.strip().replace("/ ", "/").replace(" ", "_") for x in cat_names]
     print("cat_names", cat_names)
-    device = "cuda" if torch.cuda.is_available() else "cpu"
+    # Use GPU if available, otherwise fall back to CPU
+    if torch.cuda.is_available():
+        device = "cuda"
+    # MacOS Metal performance shaders
+    elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+        device = "mps"
+    else:
+        device = "cpu"
 
     if args.prompt == "a":
         sentences = ["a " + x for x in cat_names]

dimos/models/embedding/clip.py

@@ -43,7 +43,15 @@ def __init__(
             device: Device to run on (cuda/cpu), auto-detects if None
             normalize: Whether to L2 normalize embeddings
         """
-        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
+            self.device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            self.device = "mps"
+        else:
+            self.device = "cpu"
+
         self.normalize = normalize
 
         # Load model and processor

dimos/models/embedding/embedding_models_disabled_tests.py

@@ -298,7 +298,7 @@ def test_gpu_query_performance(embedding_model, test_image) -> None:  # type: ig
 
     assert len(results) == 5, "Should return top-5 results"
     # All should be high similarity (same image, allow some variation for image preprocessing)
-    for idx, sim in results:
+    for _, sim in results:
         assert sim > 0.90, f"Same images should have high similarity, got {sim}"
 
 

dimos/models/embedding/mobileclip.py

@@ -51,7 +51,15 @@ def __init__(
                 "Install it with: pip install open-clip-torch"
             )
 
-        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
+            self.device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            self.device = "mps"
+        else:
+            self.device = "cpu"
+
         self.normalize = normalize
 
         # Load model

dimos/models/embedding/treid.py

@@ -51,7 +51,15 @@ def __init__(
                 "torchreid is required for TorchReIDModel. Install it with: pip install torchreid"
             )
 
-        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
+            self.device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            self.device = "mps"
+        else:
+            self.device = "cpu"
+
         self.normalize = normalize
 
         # Load model using torchreid's FeatureExtractor

dimos/models/vl/moondream.py

@@ -23,7 +23,15 @@ def __init__(
         dtype: torch.dtype = torch.bfloat16,
     ) -> None:
         self._model_name = model_name
-        self._device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
+            self._device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            self._device = "mps"
+        else:
+            self._device = "cpu"
+
         self._dtype = dtype
 
     @cached_property

dimos/perception/segmentation/sam_2d_seg.py

@@ -31,7 +31,6 @@
     plot_results,
 )
 from dimos.utils.data import get_data
-from dimos.utils.gpu_utils import is_cuda_available
 from dimos.utils.logging_config import setup_logger
 
 logger = setup_logger()
@@ -48,14 +47,20 @@ def __init__(
         use_rich_labeling: bool = False,
         use_filtering: bool = True,
     ) -> None:
-        if is_cuda_available():  # type: ignore[no-untyped-call]
+        # Use GPU if available, otherwise fall back to CPU
+        if torch.cuda.is_available():
             logger.info("Using CUDA for SAM 2d segmenter")
             if hasattr(onnxruntime, "preload_dlls"):  # Handles CUDA 11 / onnxruntime-gpu<=1.18
                 onnxruntime.preload_dlls(cuda=True, cudnn=True)
             self.device = "cuda"
+        # MacOS Metal performance shaders
+        elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
+            logger.info("Using Metal for SAM 2d segmenter")
+            self.device = "mps"
         else:
             logger.info("Using CPU for SAM 2d segmenter")
             self.device = "cpu"
+
         # Core components
         self.model = FastSAM(get_data(model_path) / model_name)
         self.use_tracker = use_tracker

dimos/protocol/pubsub/shmpubsub.py

@@ -233,8 +233,9 @@ def _ensure_topic(self, topic: str) -> _TopicState:
             cap = int(self.config.default_capacity)
 
             def _names_for_topic(topic: str, capacity: int) -> tuple[str, str]:
-                # Python’s SharedMemory requires names without a leading '/'
-                h = hashlib.blake2b(f"{topic}:{capacity}".encode(), digest_size=12).hexdigest()
+                # Python's SharedMemory requires names without a leading '/'
+                # Use shorter digest to avoid macOS shared memory name length limits
+                h = hashlib.blake2b(f"{topic}:{capacity}".encode(), digest_size=8).hexdigest()
                 return f"psm_{h}_data", f"psm_{h}_ctrl"
 
             data_name, ctrl_name = _names_for_topic(topic, cap)