diff --git a/comfy/ops.py b/comfy/ops.py index 77ad1d527..6d573568d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -1109,6 +1109,51 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec orig_dtype=MixedPrecisionOps._compute_dtype, orig_shape=(self.out_features, self.in_features), ) + elif self.quant_format == "svdquant_w4a4": + # SVDQuant W4A4: per-group weight scales + low-rank correction + # (proj_down, proj_up) + activation smoothing (smooth_factor) + wscales = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) + proj_down = self._load_scale_param(state_dict, prefix, "proj_down", device, manually_loaded_keys) + proj_up = self._load_scale_param(state_dict, prefix, "proj_up", device, manually_loaded_keys) + smooth_factor = self._load_scale_param(state_dict, prefix, "smooth_factor", device, manually_loaded_keys) + act_unsigned = bool(layer_conf.get("act_unsigned", False)) + + # Early Qwen-Image conversion artifacts did not persist the + # fused GELU -> fc2 unsigned-activation flag. Those layers + # are the second linear in the feed-forward block. + if not act_unsigned and ( + layer_name.endswith(".img_mlp.net.2") or layer_name.endswith(".txt_mlp.net.2") + ): + act_unsigned = True + + if any(t is None for t in (wscales, proj_down, proj_up, smooth_factor)): + raise ValueError(f"Missing SVDQuant W4A4 parameters for layer {layer_name}") + + params = layout_cls.Params( + scale=wscales, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + proj_down=proj_down, + proj_up=proj_up, + smooth_factor=smooth_factor, + act_unsigned=act_unsigned, + ) + elif self.quant_format == "awq_w4a16": + # AWQ W4A16: int4 weight, fp16/bf16 activation. Used for + # the modulation linears (img_mod.1 / txt_mod.1) so they + # stay int4 in checkpoint + VRAM rather than getting + # dequantized to bf16 at conversion time (~10 GB saving). + wscales = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) + wzeros = self._load_scale_param(state_dict, prefix, "weight_zero", device, manually_loaded_keys) + if wscales is None or wzeros is None: + raise ValueError(f"Missing AWQ W4A16 parameters for layer {layer_name}") + params = layout_cls.Params( + scale=wscales, + zeros=wzeros, + group_size=int(layer_conf.get("group_size", qconfig.get("group_size", 64))), + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) else: raise ValueError(f"Unsupported quantization format: {self.quant_format}") @@ -1158,6 +1203,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec quant_conf = {"format": self.quant_format} if self._full_precision_mm_config: quant_conf["full_precision_matrix_mult"] = True + if bool(getattr(getattr(self.weight, "_params", None), "act_unsigned", False)): + quant_conf["act_unsigned"] = True sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) input_scale = getattr(self, 'input_scale', None) @@ -1215,18 +1262,24 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec # Inference path (unchanged) if _use_quantized: + # Some layouts (e.g. SVDQuant W4A4) do activation quantization + # inside their fused kernel and cannot pre-quantize a float + # tensor up-front. Skip the input wrapping for those. + layout_cls = get_layout_class(self.layout_type) + layout_quantizes_input = getattr(layout_cls, "QUANTIZES_INPUT", True) - # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) - input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input + if layout_quantizes_input: + # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) + input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input - # Fall back to non-quantized for non-2D tensors - if input_reshaped.ndim == 2: - reshaped_3d = input.ndim == 3 - # dtype is now implicit in the layout class - scale = getattr(self, 'input_scale', None) - if scale is not None: - scale = comfy.model_management.cast_to_device(scale, input.device, None) - input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale) + # Fall back to non-quantized for non-2D tensors + if input_reshaped.ndim == 2: + reshaped_3d = input.ndim == 3 + # dtype is now implicit in the layout class + scale = getattr(self, 'input_scale', None) + if scale is not None: + scale = comfy.model_management.cast_to_device(scale, input.device, None) + input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale) output = self.forward_comfy_cast_weights(input, compute_dtype, want_requant=isinstance(input, QuantizedTensor)) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index b90bcfd25..6e7d2bf69 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -47,6 +47,12 @@ except ImportError as e: class _CKNvfp4Layout: pass + class _CKSVDQuantW4A4Layout: + pass + + class _CKAWQW4A16Layout: + pass + def register_layout_class(name, cls): pass @@ -65,6 +71,26 @@ if not _CK_MXFP8_AVAILABLE: class _CKMxfp8Layout: pass +_CK_SVDQUANT_W4A4_AVAILABLE = False +if _CK_AVAILABLE: + try: + from comfy_kitchen.tensor import TensorCoreSVDQuantW4A4Layout as _CKSVDQuantW4A4Layout + _CK_SVDQUANT_W4A4_AVAILABLE = True + except ImportError: + logging.info("comfy_kitchen does not expose SVDQuant W4A4 layout; int4 SVDQuant checkpoints will not be supported.") + class _CKSVDQuantW4A4Layout: + pass + +_CK_AWQ_W4A16_AVAILABLE = False +if _CK_AVAILABLE: + try: + from comfy_kitchen.tensor import TensorCoreAWQW4A16Layout as _CKAWQW4A16Layout + _CK_AWQ_W4A16_AVAILABLE = True + except ImportError: + logging.info("comfy_kitchen does not expose AWQ W4A16 layout; int4 AWQ modulation checkpoints will fall back to bf16-dequantized layers.") + class _CKAWQW4A16Layout: + pass + import comfy.float # ============================================================================== @@ -172,6 +198,21 @@ class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase): FP8_DTYPE = torch.float8_e5m2 +# SVDQuant W4A4 — pre-quantized offline (no runtime quantize), pass through the +# kitchen-registered layout class unchanged. Comfy-side extension reserved in +# case per-layer input scales or other Comfy-specific metadata are added later. +class TensorCoreSVDQuantW4A4Layout(_CKSVDQuantW4A4Layout): + pass + + +# AWQ W4A16 — pre-quantized offline (no runtime quantize) via the kitchen +# eager `gemv_awq_w4a16` op. Used for modulation linears (img_mod.1 / +# txt_mod.1) on Qwen-Image-Edit and similar topologies where keeping the +# weight at int4 saves ~10 GB of VRAM vs the bf16-dequantized fallback. +class TensorCoreAWQW4A16Layout(_CKAWQW4A16Layout): + pass + + # Backward compatibility alias - default to E4M3 TensorCoreFP8Layout = TensorCoreFP8E4M3Layout @@ -186,6 +227,10 @@ register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout) register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout) if _CK_MXFP8_AVAILABLE: register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout) +if _CK_SVDQUANT_W4A4_AVAILABLE: + register_layout_class("TensorCoreSVDQuantW4A4Layout", TensorCoreSVDQuantW4A4Layout) +if _CK_AWQ_W4A16_AVAILABLE: + register_layout_class("TensorCoreAWQW4A16Layout", TensorCoreAWQW4A16Layout) QUANT_ALGOS = { "float8_e4m3fn": { @@ -214,6 +259,22 @@ if _CK_MXFP8_AVAILABLE: "group_size": 32, } +if _CK_SVDQUANT_W4A4_AVAILABLE: + QUANT_ALGOS["svdquant_w4a4"] = { + "storage_t": torch.int8, + "parameters": {"weight_scale", "proj_down", "proj_up", "smooth_factor"}, + "comfy_tensor_layout": "TensorCoreSVDQuantW4A4Layout", + "group_size": 64, + } + +if _CK_AWQ_W4A16_AVAILABLE: + QUANT_ALGOS["awq_w4a16"] = { + "storage_t": torch.int8, + "parameters": {"weight_scale", "weight_zero"}, + "comfy_tensor_layout": "TensorCoreAWQW4A16Layout", + "group_size": 64, + } + # ============================================================================== # Re-exports for backward compatibility @@ -222,10 +283,12 @@ if _CK_MXFP8_AVAILABLE: __all__ = [ "QuantizedTensor", "QuantizedLayout", + "TensorCoreAWQW4A16Layout", "TensorCoreFP8Layout", "TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreNVFP4Layout", + "TensorCoreSVDQuantW4A4Layout", "QUANT_ALGOS", "register_layout_op", ]