Make old fp8 system use new mixed quant system.

Upgrade the internal mixed quant format to be easier to deal with.

Author: comfyanonymous

comfy/model_base.py

@@ -57,6 +57,7 @@
 import comfy.latent_formats
 import comfy.model_sampling
 import math
+import json
 from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from comfy.model_patcher import ModelPatcher
@@ -134,7 +135,7 @@ def __init__(self, model_config, model_type=ModelType.EPS, device=None, unet_mod
         if not unet_config.get("disable_unet_model_creation", False):
             if model_config.custom_operations is None:
                 fp8 = model_config.optimizations.get("fp8", False)
-                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config)
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, model_config=model_config)
             else:
                 operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
@@ -329,18 +330,6 @@ def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_
             extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
 
         unet_state_dict = self.diffusion_model.state_dict()
-
-        if self.model_config.scaled_fp8 is not None:
-            unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8)
-
-        # Save mixed precision metadata
-        if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config:
-            metadata = {
-                "format_version": "1.0",
-                "layers": self.model_config.layer_quant_config
-            }
-            unet_state_dict["_quantization_metadata"] = metadata
-
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
 
         if self.model_type == ModelType.V_PREDICTION:

comfy/model_detection.py

@@ -6,20 +6,6 @@
 import logging
 import torch
 
-
-def detect_layer_quantization(metadata):
-    quant_key = "_quantization_metadata"
-    if metadata is not None and quant_key in metadata:
-        quant_metadata = metadata.pop(quant_key)
-        quant_metadata = json.loads(quant_metadata)
-        if isinstance(quant_metadata, dict) and "layers" in quant_metadata:
-            logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})")
-            return quant_metadata["layers"]
-        else:
-            raise ValueError("Invalid quantization metadata format")
-    return None
-
-
 def count_blocks(state_dict_keys, prefix_string):
     count = 0
     while True:
@@ -767,22 +753,11 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     if model_config is None and use_base_if_no_match:
         model_config = comfy.supported_models_base.BASE(unet_config)
 
-    scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix)
-    if scaled_fp8_key in state_dict:
-        scaled_fp8_weight = state_dict.pop(scaled_fp8_key)
-        model_config.scaled_fp8 = scaled_fp8_weight.dtype
-        if model_config.scaled_fp8 == torch.float32:
-            model_config.scaled_fp8 = torch.float8_e4m3fn
-        if scaled_fp8_weight.nelement() == 2:
-            model_config.optimizations["fp8"] = False
-        else:
-            model_config.optimizations["fp8"] = True
-
     # Detect per-layer quantization (mixed precision)
-    layer_quant_config = detect_layer_quantization(metadata)
-    if layer_quant_config:
-        model_config.layer_quant_config = layer_quant_config
-        logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized")
+    quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix)
+    if quant_config:
+        model_config.quant_config = quant_config
+        logging.info("Detected mixed precision quantization")
 
     return model_config
 

comfy/model_patcher.py

@@ -126,27 +126,11 @@ class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
         self.patches = patches
-        self.convert_func = convert_func
+        self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
 
     def __call__(self, weight):
-        intermediate_dtype = weight.dtype
-        if self.convert_func is not None:
-            weight = self.convert_func(weight, inplace=False)
-
-        if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
-            intermediate_dtype = torch.float32
-            out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype)
-            if self.set_func is None:
-                return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key))
-            else:
-                return self.set_func(out, seed=string_to_seed(self.key), return_weight=True)
-
-        out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
-        if self.set_func is not None:
-            return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype)
-        else:
-            return out
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
 
 #The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3

comfy/ops.py

@@ -23,6 +23,7 @@
 import comfy.float
 import comfy.rmsnorm
 import contextlib
+import json
 
 def run_every_op():
     if torch.compiler.is_compiling():
@@ -422,22 +423,12 @@ def fp8_linear(self, input):
 
     if input.ndim == 3 or input.ndim == 2:
         w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
+        scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
 
-        scale_weight = self.scale_weight
-        scale_input = self.scale_input
-        if scale_weight is None:
-            scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
-        else:
-            scale_weight = scale_weight.to(input.device)
-
-        if scale_input is None:
-            scale_input = torch.ones((), device=input.device, dtype=torch.float32)
-            input = torch.clamp(input, min=-448, max=448, out=input)
-            layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
-            quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
-        else:
-            scale_input = scale_input.to(input.device)
-            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
+        scale_input = torch.ones((), device=input.device, dtype=torch.float32)
+        input = torch.clamp(input, min=-448, max=448, out=input)
+        layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
+        quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
 
         # Wrap weight in QuantizedTensor - this enables unified dispatch
         # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
@@ -471,59 +462,6 @@ def forward_comfy_cast_weights(self, input):
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
-def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
-    logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
-    class scaled_fp8_op(manual_cast):
-        class Linear(manual_cast.Linear):
-            def __init__(self, *args, **kwargs):
-                if override_dtype is not None:
-                    kwargs['dtype'] = override_dtype
-                super().__init__(*args, **kwargs)
-
-            def reset_parameters(self):
-                if not hasattr(self, 'scale_weight'):
-                    self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
-
-                if not scale_input:
-                    self.scale_input = None
-
-                if not hasattr(self, 'scale_input'):
-                    self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
-                return None
-
-            def forward_comfy_cast_weights(self, input):
-                if fp8_matrix_mult:
-                    out = fp8_linear(self, input)
-                    if out is not None:
-                        return out
-
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-
-                if weight.numel() < input.numel(): #TODO: optimize
-                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
-                else:
-                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
-                uncast_bias_weight(self, weight, bias, offload_stream)
-                return x
-
-            def convert_weight(self, weight, inplace=False, **kwargs):
-                if inplace:
-                    weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype)
-                    return weight
-                else:
-                    return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32)
-
-            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
-                weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
-                if return_weight:
-                    return weight
-                if inplace_update:
-                    self.weight.data.copy_(weight)
-                else:
-                    self.weight = torch.nn.Parameter(weight, requires_grad=False)
-
-    return scaled_fp8_op
-
 CUBLAS_IS_AVAILABLE = False
 try:
     from cublas_ops import CublasLinear
@@ -550,9 +488,9 @@ def forward(self, *args, **kwargs):
 from .quant_ops import QuantizedTensor, QUANT_ALGOS
 
 
-def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
+def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
     class MixedPrecisionOps(manual_cast):
-        _layer_quant_config = layer_quant_config
+        _quant_config = quant_config
         _compute_dtype = compute_dtype
         _full_precision_mm = full_precision_mm
 
@@ -595,27 +533,36 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
 
                 manually_loaded_keys = [weight_key]
 
-                if layer_name not in MixedPrecisionOps._layer_quant_config:
+                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
+                if layer_conf is not None:
+                    layer_conf = json.loads(layer_conf.numpy().tobytes())
+
+                if layer_conf is None:
                     self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
                 else:
-                    quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None)
-                    if quant_format is None:
+                    self.quant_format = layer_conf.get("format", None)
+                    if not self._full_precision_mm:
+                        self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False)
+
+                    if self.quant_format is None:
                         raise ValueError(f"Unknown quantization format for layer {layer_name}")
 
-                    qconfig = QUANT_ALGOS[quant_format]
+                    qconfig = QUANT_ALGOS[self.quant_format]
                     self.layout_type = qconfig["comfy_tensor_layout"]
 
                     weight_scale_key = f"{prefix}weight_scale"
+                    scale = state_dict.pop(weight_scale_key, None)
                     layout_params = {
-                        'scale': state_dict.pop(weight_scale_key, None),
+                        'scale': scale,
                         'orig_dtype': MixedPrecisionOps._compute_dtype,
                         'block_size': qconfig.get("group_size", None),
                     }
-                    if layout_params['scale'] is not None:
+
+                    if scale is not None:
                         manually_loaded_keys.append(weight_scale_key)
 
                     self.weight = torch.nn.Parameter(
-                        QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
+                        QuantizedTensor(weight.to(device=device, dtype=qconfig.get("storage_t", None)), self.layout_type, layout_params),
                         requires_grad=False
                     )
 
@@ -624,7 +571,7 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                         _v = state_dict.pop(param_key, None)
                         if _v is None:
                             continue
-                        setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
+                        self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
                         manually_loaded_keys.append(param_key)
 
                 super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
@@ -633,6 +580,16 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                     if key in missing_keys:
                         missing_keys.remove(key)
 
+            def state_dict(self, *args, destination=None, prefix="", **kwargs):
+                sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs)
+                if isinstance(self.weight, QuantizedTensor):
+                    sd["{}weight_scale".format(prefix)] = self.weight._layout_params['scale']
+                    quant_conf = {"format": self.quant_format}
+                    if self._full_precision_mm:
+                        quant_conf["full_precision_matrix_mult"] = True
+                    sd["{}comfy_quant".format(prefix)] = torch.frombuffer(json.dumps(quant_conf).encode('utf-8'), dtype=torch.uint8)
+                return sd
+
             def _forward(self, input, weight, bias):
                 return torch.nn.functional.linear(input, weight, bias)
 
@@ -648,9 +605,8 @@ def forward(self, input, *args, **kwargs):
                 if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                     return self.forward_comfy_cast_weights(input, *args, **kwargs)
                 if (getattr(self, 'layout_type', None) is not None and
-                    getattr(self, 'input_scale', None) is not None and
                     not isinstance(input, QuantizedTensor)):
-                    input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype)
+                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None), dtype=self.weight.dtype)
                 return self._forward(input, self.weight, self.bias)
 
             def convert_weight(self, weight, inplace=False, **kwargs):
@@ -661,7 +617,7 @@ def convert_weight(self, weight, inplace=False, **kwargs):
 
             def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 if getattr(self, 'layout_type', None) is not None:
-                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale=None, dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True)
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True)
                 else:
                     weight = weight.to(self.weight.dtype)
                 if return_weight:
@@ -672,15 +628,12 @@ def set_weight(self, weight, inplace_update=False, seed=None, return_weight=Fals
 
     return MixedPrecisionOps
 
-def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
+def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
 
-    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
-        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
-        return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute)
-
-    if scaled_fp8 is not None:
-        return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)
+    if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
+        logging.info("Using mixed precision operations")
+        return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute)
 
     if (
         fp8_compute and

comfy/quant_ops.py

@@ -238,6 +238,12 @@ def is_pinned(self):
     def is_contiguous(self, *arg, **kwargs):
         return self._qdata.is_contiguous(*arg, **kwargs)
 
+    def storage(self):
+        return self._qdata.storage()
+
+    def untyped_storage(self):
+        return self._qdata.untyped_storage()
+
 # ==============================================================================
 # Generic Utilities (Layout-Agnostic Operations)
 # ==============================================================================
@@ -397,17 +403,20 @@ class TensorCoreFP8Layout(QuantizedLayout):
     def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False):
         orig_dtype = tensor.dtype
 
-        if scale is None:
+        if scale == "recalculate":
             scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
 
-        if not isinstance(scale, torch.Tensor):
-            scale = torch.tensor(scale)
-        scale = scale.to(device=tensor.device, dtype=torch.float32)
+        if scale is not None:
+            if not isinstance(scale, torch.Tensor):
+                scale = torch.tensor(scale)
+            scale = scale.to(device=tensor.device, dtype=torch.float32)
 
-        if inplace_ops:
-            tensor *= (1.0 / scale).to(tensor.dtype)
+            if inplace_ops:
+                tensor *= (1.0 / scale).to(tensor.dtype)
+            else:
+                tensor = tensor * (1.0 / scale).to(tensor.dtype)
         else:
-            tensor = tensor * (1.0 / scale).to(tensor.dtype)
+            scale = torch.ones((), device=tensor.device, dtype=torch.float32)
 
         if stochastic_rounding > 0:
             tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding)