[Bug] MPS crash in SAM3 Text Segmentation (grid_sample) on macOS - Placeholder tensor is empty

[HM1579]

[Bug] MPS crash in SAM3 Text Segmentation (grid_sample) on macOS - Placeholder tensor is empty

Description

When running SAM3 Text Segmentation on macOS (Apple Silicon, MPS backend), the node crashes with a PyTorch internal assertion error:

RuntimeError: [srcBuf length] > 0 INTERNAL ASSERT FAILED ...
Placeholder tensor is empty!

The error occurs inside torch.nn.functional.grid_sample during the grounding phase.

---

Environment

• OS: macOS 14.8.2 (Apple Silicon)
• Device: MPS (Metal)
• PyTorch: 2.10.0
• Python: 3.12
• ComfyUI: 0.18.2
• comfyui-sam3: latest (as of 2026-04)

---

Full Error Trace

RuntimeError: [srcBuf length] > 0 INTERNAL ASSERT FAILED at ...
Placeholder tensor is empty!

File ".../sam3/model.py", line 2344:
sampled = torch.nn.functional.grid_sample(img_feats, grid, align_corners=False)

---

Root Cause (Analysis)

From code inspection:

# load_model.py
load_device = comfy.model_management.get_torch_device()

# utils.py
if device is None:
    device = comfy.model_management.get_torch_device()

This means:

• The model always follows ComfyUI’s global device
• On macOS → this becomes mps

The crash happens specifically in:

grid_sample (MPS backend)

This suggests:

PyTorch MPS backend + SAM3 grounding pipeline is unstable or unsupported in this code path.

---

Expected Behavior

• SAM3 Text Segmentation should run correctly on MPS
• OR fallback to CPU automatically if unsupported operations are used

---

Actual Behavior

• Node crashes during grounding
• No output produced
• Error occurs consistently on MPS

---

Workaround (Confirmed Working)

Forcing the model to run on CPU resolves the issue completely.

Fix 1 — nodes/load_model.py

import torch
load_device = torch.device("cpu")

---

Fix 2 — nodes/_model_cache.py

import torch
load_device = torch.device("cpu")
offload_device = torch.device("cpu")

---

Fix 3 (optional, safer) — nodes/sam3/utils.py

if device is None:
    device = torch.device("cpu")

---

Result After Fix

• No crash
• Segmentation works correctly
• Multiple instance masks are generated as expected
• Only downside: slower inference (CPU)

---

Suggestion

Consider:

1. Detecting MPS and disabling problematic ops (e.g. grid_sample)
2. Automatically falling back to CPU when running grounding
3. Adding a device override option in the node UI

---

Additional Notes

• This issue is independent of prompt correctness
• Once running on CPU, results are stable and correct
• Likely related to PyTorch MPS limitations rather than model logic

[epwr]

Running into the same issue on Macbook with M2 chipset.

[PozzettiAndrea]

@epwr @HM1579 Hi guys, I have since improved the MPS support on SAM3. Could you try installing the latest version and letting me know how it goes?