[XPU] comfy_kitchen FP8 quantized tensors segfault on .to('xpu') during clone()

[tinafengfun]

[XPU] comfy_kitchen FP8 quantized tensors segfault on .to('xpu') during clone()

Summary

The comfy_kitchen FP8 quantized tensor implementation segfaults on Intel XPU when a model containing quantized parameters is moved to the XPU device via .to('xpu'). The crash occurs inside comfy_kitchen.tensor.base.clone() during the _quantized_apply path triggered by nn.Module._apply(). This blocks all FP8 quantized text encoders (e.g., Qwen 2.5 VL 7B FP8) from running on Intel GPU.

Environment

• ComfyUI version: v0.22.0-124-gc87384f2 (commit c87384f2b8f9e9a7f717e1a269b860698b6634bf, 2026-06-02)
• comfy-kitchen: 0.2.8
• PyTorch: 2.11.0+xpu
• Python: 3.13.3
• GPU: Intel(R) Graphics [0xe211] (PCI 8086:E211), 160 CUs, 24.4 GB VRAM
• Driver: xe kernel driver, Linux 6.14.0-1008-intel

Bug Details

Code path

clip.patcher.model.to('xpu')
→ torch.nn.Module._apply() (recursive)
→ comfy.ops._quantized_apply()              # ops.py:1315
→ comfy.ops._quantized_apply._quantized_apply_fn()  # ops.py:1064
→ torch.nn.Parameter.__new__()
→ comfy_kitchen QTensor.__torch_dispatch__()
→ QTensor._handle_detach()                  # base.py:396
→ QTensor._copy_with()                      # base.py:252
→ QTensor.clone()                           # base.py:73 → SEGFAULT

When nn.Module.to(device) is called, PyTorch iterates all parameters and calls _apply() on each. For quantized parameters managed by comfy_kitchen, ComfyUI routes through _quantized_apply (comfy/ops.py:1315), which calls the quantized tensor's clone() to copy data to the new device. The clone() implementation in comfy_kitchen/tensor/base.py:73 crashes on Intel XPU.

Root cause

The comfy_kitchen quantized tensor clone() operation attempts to clone FP8 (e4m3fn) quantized data on the XPU device. The underlying operation (likely a raw memory copy or type-specific clone) lacks a working XPU/SYCL kernel and segfaults.

Reproducer

import comfy.sd

clip = comfy.sd.load_clip(
    ["qwen_2.5_vl_7b_fp8_scaled.safetensors"],
    clip_type=comfy.sd.CLIPType.QWEN_IMAGE,
)
# SEGFAULT here — during .to('xpu') on quantized parameters
clip.patcher.model.to('xpu')

Full reproducer: repro2_clip_quantized_xpu.py

Results

3/3 deterministic segfaults (exit code 139):

Stack trace from faulthandler:

Fatal Python error: Segmentation fault

Current thread (most recent call first):
  File "comfy_kitchen/tensor/base.py", line 73 in clone
  File "comfy_kitchen/tensor/base.py", line 252 in _copy_with
  File "comfy_kitchen/tensor/base.py", line 396 in _handle_detach
  File "comfy_kitchen/tensor/base.py", line 340 in __torch_dispatch__
  File "torch/nn/parameter.py", line 60 in __new__
  File "comfy/ops.py", line 1064 in _quantized_apply
  File "comfy/ops.py", line 1315 in _apply
  File "torch/nn/modules/module.py", line 934 in _apply
  ... (recursive _apply calls)
  File "torch/nn/modules/module.py", line 1384 in to
  File "repro2_clip_quantized_xpu.py", line 69 in main

Impact

• All FP8 quantized text encoders (Qwen 2.5 VL 7B FP8, etc.) must run on CPU on Intel XPU systems
• This blocks end-to-end XPU pipelines — the text encoder becomes a bottleneck
• CLIP encoding on CPU adds ~3-10 seconds depending on prompt length and model size

Suggested investigation

1. comfy_kitchen: Check if the clone() method in tensor/base.py:73 has a device-specific code path. The FP8 quantized tensor likely needs a dequantize→copy→requantize path for cross-device operations, or a raw buffer clone that works on XPU.
2. PyTorch XPU: The crash may be in a low-level memory copy kernel for FP8 tensors on the SYCL backend. Check if torch._C._nn._copy_from or similar functions support FP8 on XPU.
3. Workaround: Dequantize the CLIP model to bf16 before moving to XPU. This increases memory usage but avoids the clone path.

Related

• ComfyUI's comfy/ops.py has a _quantized_apply function (line 1064) specifically for handling quantized tensors during .to() and .half() operations
• The comfy_kitchen package (v0.2.8) is ComfyUI's quantization support library
• This issue was discovered during Intel XPU migration of the Qwen-Image-2512 workflow

[coderabbitai]

🔗 Related PRs

#13801 - fix: Fix bug when mask not on same device (CORE-181) [merged]
#14052 - fixup threaded loader with worksplit multi-gpu [merged]
#14077 - feat: Microsoft Lens support (CORE-248) [merged]
#14116 - Threaded Loader performance fixes / improvements (+ Aimdo 0.4.6) [merged]
#14359 - Revert "Add SeedVR2 support (CORE-6)" [merged]

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!