resolve rebase conflict

comfy/quant_ops.py

@@ -3,22 +3,6 @@ import logging
 
 from comfy.cli_args import args
 
-
-def _rocm_kitchen_arch_supported():
-    """comfy-kitchen's INT8 Triton kernels compile tl.dot to matrix-core instructions.
-    RDNA3/3.5/4 (gfx11xx/gfx12xx) have WMMA and CDNA (gfx9xx) has MFMA; RDNA1/RDNA2
-    (gfx10xx) have neither, so the INT8 path hangs the GPU there. Gates the automatic
-    ROCm default so those cards stay on the eager fallback (an explicit
-    --enable-triton-backend still forces it on any arch)."""
-    try:
-        arch = torch.cuda.get_device_properties(torch.cuda.current_device()).gcnArchName.split(":")[0]
-    except Exception:
-        return False
-    if arch.startswith(("gfx11", "gfx12")):
-        return True
-    return arch in ("gfx908", "gfx90a", "gfx940", "gfx941", "gfx942", "gfx950")
-
-
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -42,13 +26,9 @@ try:
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
     # On ROCm/AMD the CUDA backend is unavailable, so Triton is the only accelerated
-    # comfy-kitchen backend. Enable it by default there, but only on Triton >= 3.7 AND a
+    # comfy-kitchen backend. Enable it by default there, but only on Triton >= 3.7:
-    # matrix-core GPU (RDNA3+ WMMA gfx11xx/gfx12xx, CDNA MFMA gfx9xx). RDNA1/RDNA2
-    # (gfx10xx) have no WMMA -> the INT8 tl.dot path hangs the GPU, so they stay eager.
     # older Triton lacks libdevice.rint on the HIP backend and hard-crashes the INT8 path.
-    if args.disable_triton_backend:
+    if args.enable_triton_backend or torch.version.hip is not None:
-        ck.registry.disable("triton")
-    elif args.enable_triton_backend: # or (torch.version.hip is not None and _rocm_kitchen_arch_supported()):
         try:
             import triton
             triton_version = tuple(int(v) for v in triton.__version__.split(".")[:2])

comfy/quantization/torchao/__init__.py

@@ -1,19 +1,24 @@
 """
 ComfyUI TorchAO INT4 weight-only quantization backend.
-
+...
-W4A16 scheme: weights quantized to INT4 via torchao,
-activations remain FP16. Supports CUDA, Intel XPU, and CPU.
-
-Includes:
-- TINT4Linear: INT4 linear layer with LoRA forward injection + QuaRot
-- quantize_model: torchao INT4 quantization entry point
-- reconstruct_int4_state_dict: rebuild TINT4Linear from safetensors
-- build_hadamard / rotate_weight: QuaRot Hadamard rotation
 """
+import torch
 
-from .linear import TINT4Linear
+_AVAILABLE = False
-from .quantize import quantize_model, reconstruct_int4_state_dict
+try:
-from .quarot import build_hadamard, rotate_weight
+    import torchao
+    from .linear import TINT4Linear
+    from .quantize import quantize_model, reconstruct_int4_state_dict
+    from .quarot import build_hadamard, rotate_weight
+    _AVAILABLE = True
+except ImportError:
+    pass
+
+if not _AVAILABLE:
+    raise ImportError(
+        "torchao is required for INT4 quantization. "
+        "Install: pip install torchao>=0.17.0"
+    )
 
 __all__ = [
     "TINT4Linear",