Support asym w4a8_int (#15308)

* support asym w4a8_int

* Simplify

* Fixes

comfy/quant_ops.py

@@ -28,6 +28,7 @@ try:
         TensorCoreNVFP4Layout as _CKNvfp4Layout,
         TensorCoreConvRotW4A4Layout as _CKTensorCoreConvRotW4A4Layout,
         TensorWiseINT8Layout as _CKTensorWiseINT8Layout,
+        AsymW4A8Int8Layout as _CKAsymW4A8Int8Layout,
         register_layout_op,
         register_layout_class,
         get_layout_class,
@@ -83,6 +84,9 @@ except ImportError as e:
     class _CKTensorCoreConvRotW4A4Layout:
         pass
 
+    class _CKAsymW4A8Int8Layout:
+        pass
+
     def register_layout_class(name, cls):
         pass
 
@@ -212,7 +216,7 @@ class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
 TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
 TensorWiseINT8Layout = _CKTensorWiseINT8Layout
 TensorCoreConvRotW4A4Layout = _CKTensorCoreConvRotW4A4Layout
-
+AsymW4A8Int8Layout = _CKAsymW4A8Int8Layout
 
 # ==============================================================================
 # Registry
@@ -226,6 +230,7 @@ register_layout_class("TensorWiseINT8Layout", _CKTensorWiseINT8Layout)
 register_layout_class("TensorCoreConvRotW4A4Layout", _CKTensorCoreConvRotW4A4Layout)
 if _CK_MXFP8_AVAILABLE:
     register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
+register_layout_class("AsymW4A8Int8Layout", _CKAsymW4A8Int8Layout)
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
@@ -268,6 +273,13 @@ QUANT_ALGOS["convrot_w4a4"] = {
     "quantize_input": False,
 }
 
+QUANT_ALGOS["asym_w4a8_int8"] = {
+    "storage_t": torch.int8,
+    "parameters": {"weight_scale"},
+    "comfy_tensor_layout": "AsymW4A8Int8Layout",
+    "quantize_input": False,
+}
+
 
 # ==============================================================================
 # Re-exports for backward compatibility
@@ -282,6 +294,7 @@ __all__ = [
     "TensorCoreNVFP4Layout",
     "TensorCoreConvRotW4A4Layout",
     "TensorWiseINT8Layout",
+    "AsymW4A8Int8Layout",
     "QUANT_ALGOS",
     "register_layout_op",
 ]