Merge 4cebbc50f7 into 0904cc3fe5

This is an experimental WIP option that might not work in your workflow but
should lower memory usage if it does.

Currently only the VAE and the load image node will output in fp16 when
this option is turned on.

comfy/cli_args.py

@@ -83,6 +83,8 @@ fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
 fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
+parser.add_argument("--fp16-intermediates", action="store_true", help="Experimental: Use fp16 for intermediate tensors between nodes instead of fp32.")
+
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")

comfy/float.py

@@ -209,3 +209,39 @@ def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=
         output_block[i:i + slice_size].copy_(block)
 
     return output_fp4, to_blocked(output_block, flatten=False)
+
+
+def stochastic_round_quantize_mxfp8_by_block(x, pad_32x, seed=0):
+    def roundup(x_val, multiple):
+        return ((x_val + multiple - 1) // multiple) * multiple
+
+    if pad_32x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 32)
+        padded_cols = roundup(cols, 32)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+
+    F8_E4M3_MAX = 448.0
+    E8M0_BIAS = 127
+    BLOCK_SIZE = 32
+
+    rows, cols = x.shape
+    x_blocked = x.reshape(rows, -1, BLOCK_SIZE)
+    max_abs = torch.amax(torch.abs(x_blocked), dim=-1)
+
+    # E8M0 block scales (power-of-2 exponents)
+    scale_needed = torch.clamp(max_abs.float() / F8_E4M3_MAX, min=2**(-127))
+    exp_biased = torch.clamp(torch.ceil(torch.log2(scale_needed)).to(torch.int32) + E8M0_BIAS, 0, 254)
+    block_scales_e8m0 = exp_biased.to(torch.uint8)
+
+    zero_mask = (max_abs == 0)
+    block_scales_f32 = (block_scales_e8m0.to(torch.int32) << 23).view(torch.float32)
+    block_scales_f32 = torch.where(zero_mask, torch.ones_like(block_scales_f32), block_scales_f32)
+
+    # Scale per-block then stochastic round
+    data_scaled = (x_blocked.float() / block_scales_f32.unsqueeze(-1)).reshape(rows, cols)
+    output_fp8 = stochastic_rounding(data_scaled, torch.float8_e4m3fn, seed=seed)
+
+    block_scales_e8m0 = torch.where(zero_mask, torch.zeros_like(block_scales_e8m0), block_scales_e8m0)
+    return output_fp8, to_blocked(block_scales_e8m0, flatten=False).view(torch.float8_e8m0fnu)

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -11,6 +11,7 @@ from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+import comfy.model_management
 from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
@@ -536,7 +537,7 @@ class Decoder(nn.Module):
                     mark_conv3d_ended(self.conv_out)
                 sample = self.conv_out(sample, causal=self.causal)
                 if sample is not None and sample.shape[2] > 0:
-                    output.append(sample)
+                    output.append(sample.to(comfy.model_management.intermediate_device()))
                 return
 
             up_block = self.up_blocks[idx]

comfy/model_management.py

@@ -1050,6 +1050,12 @@ def intermediate_device():
     else:
         return torch.device("cpu")
 
+def intermediate_dtype():
+    if args.fp16_intermediates:
+        return torch.float16
+    else:
+        return torch.float32
+
 def vae_device():
     if args.cpu_vae:
         return torch.device("cpu")
@@ -1712,6 +1718,19 @@ def supports_nvfp4_compute(device=None):
 
     return True
 
+def supports_mxfp8_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    if torch_version_numeric < (2, 10):
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/ops.py

@@ -766,6 +766,71 @@ from .quant_ops import (
 )
 
 
+class QuantLinearFunc(torch.autograd.Function):
+    """Custom autograd function for quantized linear: quantized forward, compute_dtype backward.
+    Handles any input rank by flattening to 2D for matmul and restoring shape after.
+    """
+
+    @staticmethod
+    def forward(ctx, input_float, weight, bias, layout_type, input_scale, compute_dtype):
+        input_shape = input_float.shape
+        inp = input_float.detach().flatten(0, -2)  # zero-cost view to 2D
+
+        # Quantize input (same as inference path)
+        if layout_type is not None:
+            q_input = QuantizedTensor.from_float(inp, layout_type, scale=input_scale)
+        else:
+            q_input = inp
+
+        w = weight.detach() if weight.requires_grad else weight
+        b = bias.detach() if bias is not None and bias.requires_grad else bias
+
+        output = torch.nn.functional.linear(q_input, w, b)
+
+        # Restore original input shape
+        if len(input_shape) > 2:
+            output = output.unflatten(0, input_shape[:-1])
+
+        ctx.save_for_backward(input_float, weight)
+        ctx.input_shape = input_shape
+        ctx.has_bias = bias is not None
+        ctx.compute_dtype = compute_dtype
+        ctx.weight_requires_grad = weight.requires_grad
+
+        return output
+
+    @staticmethod
+    @torch.autograd.function.once_differentiable
+    def backward(ctx, grad_output):
+        input_float, weight = ctx.saved_tensors
+        compute_dtype = ctx.compute_dtype
+        grad_2d = grad_output.flatten(0, -2).to(compute_dtype)
+
+        # Dequantize weight to compute dtype for backward matmul
+        if isinstance(weight, QuantizedTensor):
+            weight_f = weight.dequantize().to(compute_dtype)
+        else:
+            weight_f = weight.to(compute_dtype)
+
+        # grad_input = grad_output @ weight
+        grad_input = torch.mm(grad_2d, weight_f)
+        if len(ctx.input_shape) > 2:
+            grad_input = grad_input.unflatten(0, ctx.input_shape[:-1])
+
+        # grad_weight (only if weight requires grad, typically frozen for quantized training)
+        grad_weight = None
+        if ctx.weight_requires_grad:
+            input_f = input_float.flatten(0, -2).to(compute_dtype)
+            grad_weight = torch.mm(grad_2d.t(), input_f)
+
+        # grad_bias
+        grad_bias = None
+        if ctx.has_bias:
+            grad_bias = grad_2d.sum(dim=0)
+
+        return grad_input, grad_weight, grad_bias, None, None, None
+
+
 def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
     class MixedPrecisionOps(manual_cast):
         _quant_config = quant_config
@@ -857,6 +922,22 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                             orig_shape=(self.out_features, self.in_features),
                         )
 
+                    elif self.quant_format == "mxfp8":
+                        # MXFP8: E8M0 block scales stored as uint8 in safetensors
+                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
+                                                             dtype=torch.uint8)
+
+                        if block_scale is None:
+                            raise ValueError(f"Missing MXFP8 block scales for layer {layer_name}")
+
+                        block_scale = block_scale.view(torch.float8_e8m0fnu)
+
+                        params = layout_cls.Params(
+                            scale=block_scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+
                     elif self.quant_format == "nvfp4":
                         # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
                         tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
@@ -944,10 +1025,37 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 #If cast needs to apply lora, it should be done in the compute dtype
                 compute_dtype = input.dtype
 
-                if (getattr(self, 'layout_type', None) is not None and
+                _use_quantized = (
+                    getattr(self, 'layout_type', None) is not None and
                     not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
                     not getattr(self, 'comfy_force_cast_weights', False) and
-                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0
+                )
+
+                # Training path: quantized forward with compute_dtype backward via autograd function
+                if (input.requires_grad and _use_quantized):
+
+                    weight, bias, offload_stream = cast_bias_weight(
+                        self,
+                        input,
+                        offloadable=True,
+                        compute_dtype=compute_dtype,
+                        want_requant=True
+                    )
+
+                    scale = getattr(self, 'input_scale', None)
+                    if scale is not None:
+                        scale = comfy.model_management.cast_to_device(scale, input.device, None)
+
+                    output = QuantLinearFunc.apply(
+                        input, weight, bias, self.layout_type, scale, compute_dtype
+                    )
+
+                    uncast_bias_weight(self, weight, bias, offload_stream)
+                    return output
+
+                # Inference path (unchanged)
+                if _use_quantized:
 
                     # 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
@@ -995,7 +1103,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 for key, param in self._parameters.items():
                     if param is None:
                         continue
-                    self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False))
+                    p = fn(param)
+                    if p.is_inference():
+                        p = p.clone()
+                    self.register_parameter(key, torch.nn.Parameter(p, requires_grad=False))
                 for key, buf in self._buffers.items():
                     if buf is not None:
                         self._buffers[key] = fn(buf)
@@ -1006,12 +1117,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 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
     nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
+    mxfp8_compute = comfy.model_management.supports_mxfp8_compute(load_device)
 
     if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
         logging.info("Using mixed precision operations")
         disabled = set()
         if not nvfp4_compute:
             disabled.add("nvfp4")
+        if not mxfp8_compute:
+            disabled.add("mxfp8")
         if not fp8_compute:
             disabled.add("float8_e4m3fn")
             disabled.add("float8_e5m2")

comfy/quant_ops.py

@@ -43,6 +43,18 @@ except ImportError as e:
     def get_layout_class(name):
         return None
 
+_CK_MXFP8_AVAILABLE = False
+if _CK_AVAILABLE:
+    try:
+        from comfy_kitchen.tensor import TensorCoreMXFP8Layout as _CKMxfp8Layout
+        _CK_MXFP8_AVAILABLE = True
+    except ImportError:
+        logging.warning("comfy_kitchen does not support MXFP8, please update comfy_kitchen.")
+
+if not _CK_MXFP8_AVAILABLE:
+    class _CKMxfp8Layout:
+        pass
+
 import comfy.float
 
 # ==============================================================================
@@ -84,6 +96,31 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout):
         return qdata, params
 
 
+class TensorCoreMXFP8Layout(_CKMxfp8Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"MXFP8 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
+
+        if stochastic_rounding > 0:
+            qdata, block_scale = comfy.float.stochastic_round_quantize_mxfp8_by_block(tensor, pad_32x=needs_padding, seed=stochastic_rounding)
+        else:
+            qdata, block_scale = ck.quantize_mxfp8(tensor, pad_32x=needs_padding)
+
+        params = cls.Params(
+            scale=block_scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+        )
+        return qdata, params
+
+
 class TensorCoreNVFP4Layout(_CKNvfp4Layout):
     @classmethod
     def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
@@ -137,6 +174,8 @@ register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
 register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
 register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
 register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
+if _CK_MXFP8_AVAILABLE:
+    register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
@@ -157,6 +196,14 @@ QUANT_ALGOS = {
     },
 }
 
+if _CK_MXFP8_AVAILABLE:
+    QUANT_ALGOS["mxfp8"] = {
+        "storage_t": torch.float8_e4m3fn,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreMXFP8Layout",
+        "group_size": 32,
+    }
+
 
 # ==============================================================================
 # Re-exports for backward compatibility

comfy/sd.py

@@ -871,13 +871,16 @@ class VAE:
                 pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
         return pixels
 
+    def vae_output_dtype(self):
+        return model_management.intermediate_dtype()
+
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
         steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         output = self.process_output(
             (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
             comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
@@ -887,16 +890,16 @@ class VAE:
 
     def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
-            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         else:
             og_shape = samples.shape
             samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
-            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
 
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, index_formulas=self.upscale_index_formula, output_device=self.output_device))
 
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
@@ -905,7 +908,7 @@ class VAE:
         steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
@@ -914,7 +917,7 @@ class VAE:
 
     def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
         if self.latent_dim == 1:
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
             out_channels = self.latent_channels
             upscale_amount = 1 / self.downscale_ratio
         else:
@@ -923,7 +926,7 @@ class VAE:
             tile_x = tile_x // extra_channel_size
             overlap = overlap // extra_channel_size
             upscale_amount = 1 / self.downscale_ratio
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).to(dtype=self.vae_output_dtype())
 
         out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
         if self.latent_dim == 1:
@@ -932,7 +935,7 @@ class VAE:
             return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
     def decode(self, samples_in, vae_options={}):
@@ -950,9 +953,9 @@ class VAE:
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).to(dtype=self.vae_output_dtype()))
                 if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 pixel_samples[x:x+batch_number] = out
         except Exception as e:
             model_management.raise_non_oom(e)
@@ -1025,9 +1028,9 @@ class VAE:
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
                 pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
-                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
 
         except Exception as e:

comfy/utils.py

@@ -897,6 +897,10 @@ def set_attr(obj, attr, value):
     return prev
 
 def set_attr_param(obj, attr, value):
+    # Clone inference tensors (created under torch.inference_mode) since
+    # their version counter is frozen and nn.Parameter() cannot wrap them.
+    if value.is_inference():
+        value = value.clone()
     return set_attr(obj, attr, torch.nn.Parameter(value, requires_grad=False))
 
 def set_attr_buffer(obj, attr, value):

comfy_extras/nodes_train.py

@@ -15,6 +15,7 @@ import comfy.sampler_helpers
 import comfy.sd
 import comfy.utils
 import comfy.model_management
+from comfy.cli_args import args, PerformanceFeature
 import comfy_extras.nodes_custom_sampler
 import folder_paths
 import node_helpers
@@ -138,6 +139,7 @@ class TrainSampler(comfy.samplers.Sampler):
         training_dtype=torch.bfloat16,
         real_dataset=None,
         bucket_latents=None,
+        use_grad_scaler=False,
     ):
         self.loss_fn = loss_fn
         self.optimizer = optimizer
@@ -152,6 +154,8 @@ class TrainSampler(comfy.samplers.Sampler):
         self.bucket_latents: list[torch.Tensor] | None = (
             bucket_latents  # list of (Bi, C, Hi, Wi)
         )
+        # GradScaler for fp16 training
+        self.grad_scaler = torch.amp.GradScaler() if use_grad_scaler else None
         # Precompute bucket offsets and weights for sampling
         if bucket_latents is not None:
             self._init_bucket_data(bucket_latents)
@@ -204,10 +208,13 @@ class TrainSampler(comfy.samplers.Sampler):
                 batch_sigmas.requires_grad_(True),
                 **batch_extra_args,
             )
-            loss = self.loss_fn(x0_pred, x0)
+            loss = self.loss_fn(x0_pred.float(), x0.float())
         if bwd:
             bwd_loss = loss / self.grad_acc
-            bwd_loss.backward()
+            if self.grad_scaler is not None:
+                self.grad_scaler.scale(bwd_loss).backward()
+            else:
+                bwd_loss.backward()
         return loss
 
     def _generate_batch_sigmas(self, model_wrap, batch_size, device):
@@ -307,7 +314,10 @@ class TrainSampler(comfy.samplers.Sampler):
             )
             total_loss += loss
         total_loss = total_loss / self.grad_acc / len(indicies)
-        total_loss.backward()
+        if self.grad_scaler is not None:
+            self.grad_scaler.scale(total_loss).backward()
+        else:
+            total_loss.backward()
         if self.loss_callback:
             self.loss_callback(total_loss.item())
         pbar.set_postfix({"loss": f"{total_loss.item():.4f}"})
@@ -348,12 +358,18 @@ class TrainSampler(comfy.samplers.Sampler):
                 self._train_step_multires_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar)
 
             if (i + 1) % self.grad_acc == 0:
+                if self.grad_scaler is not None:
+                    self.grad_scaler.unscale_(self.optimizer)
                 for param_groups in self.optimizer.param_groups:
                     for param in param_groups["params"]:
                         if param.grad is None:
                             continue
                         param.grad.data = param.grad.data.to(param.data.dtype)
-                self.optimizer.step()
+                if self.grad_scaler is not None:
+                    self.grad_scaler.step(self.optimizer)
+                    self.grad_scaler.update()
+                else:
+                    self.optimizer.step()
                 self.optimizer.zero_grad()
             ui_pbar.update(1)
         torch.cuda.empty_cache()
@@ -1004,9 +1020,9 @@ class TrainLoraNode(io.ComfyNode):
                 ),
                 io.Combo.Input(
                     "training_dtype",
-                    options=["bf16", "fp32"],
+                    options=["bf16", "fp32", "none"],
                     default="bf16",
-                    tooltip="The dtype to use for training.",
+                    tooltip="The dtype to use for training. 'none' preserves the model's native compute dtype instead of overriding it. For fp16 models, GradScaler is automatically enabled.",
                 ),
                 io.Combo.Input(
                     "lora_dtype",
@@ -1035,7 +1051,7 @@ class TrainLoraNode(io.ComfyNode):
                 io.Boolean.Input(
                     "offloading",
                     default=False,
-                    tooltip="Offload the Model to RAM. Requires Bypass Mode.",
+                    tooltip="Offload model weights to CPU during training to save GPU memory.",
                 ),
                 io.Combo.Input(
                     "existing_lora",
@@ -1120,22 +1136,32 @@ class TrainLoraNode(io.ComfyNode):
 
         # Setup model and dtype
         mp = model.clone()
-        dtype = node_helpers.string_to_torch_dtype(training_dtype)
+        use_grad_scaler = False
+        if training_dtype != "none":
+            dtype = node_helpers.string_to_torch_dtype(training_dtype)
+            mp.set_model_compute_dtype(dtype)
+        else:
+            # Detect model's native dtype for autocast
+            model_dtype = mp.model.get_dtype()
+            if model_dtype == torch.float16:
+                dtype = torch.float16
+                use_grad_scaler = True
+                # Warn about fp16 accumulation instability during training
+                if PerformanceFeature.Fp16Accumulation in args.fast:
+                    logging.warning(
+                        "WARNING: FP16 model detected with fp16_accumulation enabled. "
+                        "This combination can be numerically unstable during training and may cause NaN values. "
+                        "Suggested fixes: 1) Set training_dtype to 'bf16', or 2) Disable fp16_accumulation (remove from --fast flags)."
+                    )
+            else:
+                # For fp8, bf16, or other dtypes, use bf16 autocast
+                dtype = torch.bfloat16
         lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype)
-        mp.set_model_compute_dtype(dtype)
-
-        if mp.is_dynamic():
-            if not bypass_mode:
-                logging.info("Training MP is Dynamic - forcing bypass mode. Start comfy with --highvram to force weight diff mode")
-                bypass_mode = True
-            offloading = True
-        elif offloading:
-            if not bypass_mode:
-                logging.info("Training Offload selected - forcing bypass mode. Set bypass = True to remove this message")
 
         # Prepare latents and compute counts
+        latents_dtype = dtype if dtype not in (None,) else torch.bfloat16
         latents, num_images, multi_res = _prepare_latents_and_count(
-            latents, dtype, bucket_mode
+            latents, latents_dtype, bucket_mode
         )
 
         # Validate and expand conditioning
@@ -1201,6 +1227,7 @@ class TrainLoraNode(io.ComfyNode):
                     seed=seed,
                     training_dtype=dtype,
                     bucket_latents=latents,
+                    use_grad_scaler=use_grad_scaler,
                 )
             else:
                 train_sampler = TrainSampler(
@@ -1213,6 +1240,7 @@ class TrainLoraNode(io.ComfyNode):
                     seed=seed,
                     training_dtype=dtype,
                     real_dataset=latents if multi_res else None,
+                    use_grad_scaler=use_grad_scaler,
                 )
 
             # Setup guider
@@ -1337,7 +1365,7 @@ class SaveLoRA(io.ComfyNode):
                 io.Int.Input(
                     "steps",
                     optional=True,
-                    tooltip="Optional: The number of steps to LoRA has been trained for, used to name the saved file.",
+                    tooltip="Optional: The number of steps the LoRA has been trained for, used to name the saved file.",
                 ),
             ],
             outputs=[],

middleware/cache_middleware.py

@@ -32,7 +32,7 @@ async def cache_control(
     )
 
     if request.path.endswith(".js") or request.path.endswith(".css") or is_entry_point:
-        response.headers.setdefault("Cache-Control", "no-cache")
+        response.headers.setdefault("Cache-Control", "no-store")
         return response
 
     # Early return for non-image files - no cache headers needed

nodes.py

@@ -1724,6 +1724,8 @@ class LoadImage:
         output_masks = []
         w, h = None, None
 
+        dtype = comfy.model_management.intermediate_dtype()
+
         for i in ImageSequence.Iterator(img):
             i = node_helpers.pillow(ImageOps.exif_transpose, i)
 
@@ -1748,8 +1750,8 @@ class LoadImage:
                 mask = 1. - torch.from_numpy(mask)
             else:
                 mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
-            output_images.append(image)
-            output_masks.append(mask.unsqueeze(0))
+            output_images.append(image.to(dtype=dtype))
+            output_masks.append(mask.unsqueeze(0).to(dtype=dtype))
 
             if img.format == "MPO":
                 break  # ignore all frames except the first one for MPO format

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.19
+comfyui-frontend-package==1.41.20
 comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch

server.py

@@ -310,7 +310,7 @@ class PromptServer():
         @routes.get("/")
         async def get_root(request):
             response = web.FileResponse(os.path.join(self.web_root, "index.html"))
-            response.headers['Cache-Control'] = 'no-cache'
+            response.headers['Cache-Control'] = 'no-store, must-revalidate'
             response.headers["Pragma"] = "no-cache"
             response.headers["Expires"] = "0"
             return response

tests-unit/server_test/test_cache_control.py

@@ -28,31 +28,31 @@ CACHE_SCENARIOS = [
     },
     # JavaScript/CSS scenarios
     {
-        "name": "js_no_cache",
+        "name": "js_no_store",
         "path": "/script.js",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "css_no_cache",
+        "name": "css_no_store",
         "path": "/styles.css",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "index_json_no_cache",
+        "name": "index_json_no_store",
         "path": "/api/index.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "localized_index_json_no_cache",
+        "name": "localized_index_json_no_store",
         "path": "/templates/index.zh.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     # Non-matching files