Merge 69d3bfa391 into e84a200a3c

If a subclass BYO _load_from_state_dict and doesnt call the super() the
needed default init of these weights is missed and can lead to problems
for uninitialized weights.

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/comfy_types/node_typing.py

@@ -193,6 +193,8 @@ class HiddenInputTypeDict(TypedDict):
     """EXTRA_PNGINFO is a dictionary that will be copied into the metadata of any .png files saved. Custom nodes can store additional information in this dictionary for saving (or as a way to communicate with a downstream node)."""
     dynprompt: NotRequired[Literal["DYNPROMPT"]]
     """DYNPROMPT is an instance of comfy_execution.graph.DynamicPrompt. It differs from PROMPT in that it may mutate during the course of execution in response to Node Expansion."""
+    prompt_id: NotRequired[Literal["PROMPT_ID"]]
+    """PROMPT_ID is the unique identifier of the current prompt/job being executed. Useful for associating progress updates with specific jobs."""
 
 
 class InputTypeDict(TypedDict):

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

@@ -336,7 +336,10 @@ class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 super().__init__(in_features, out_features, bias, device, dtype)
                 return
 
@@ -357,7 +360,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(
@@ -564,7 +569,10 @@ class disable_weight_init:
         def __init__(self, num_embeddings, embedding_dim, padding_idx=None, max_norm=None,
                      norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None,
                      _freeze=False, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 super().__init__(num_embeddings, embedding_dim, padding_idx, max_norm,
                                  norm_type, scale_grad_by_freq, sparse, _weight,
                                  _freeze, device, dtype)
@@ -590,7 +598,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(
@@ -857,6 +867,22 @@ 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)
@@ -1006,12 +1032,15 @@ 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_api/feature_flags.py

@@ -12,6 +12,7 @@ from comfy.cli_args import args
 # Default server capabilities
 SERVER_FEATURE_FLAGS: dict[str, Any] = {
     "supports_preview_metadata": True,
+    "supports_progress_text_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
     "extension": {"manager": {"supports_v4": True}},
     "node_replacements": True,

comfy_api/latest/_io.py

@@ -1282,9 +1282,16 @@ class V3Data(TypedDict):
     'When True, the value of the dynamic input will be in the format (value, path_key).'
 
 class HiddenHolder:
+    """Holds hidden input values resolved during node execution.
+
+    Hidden inputs are special values automatically provided by the execution
+    engine (e.g., node ID, prompt data, authentication tokens) rather than
+    being connected by the user in the graph.
+    """
     def __init__(self, unique_id: str, prompt: Any,
                  extra_pnginfo: Any, dynprompt: Any,
-                 auth_token_comfy_org: str, api_key_comfy_org: str, **kwargs):
+                 auth_token_comfy_org: str, api_key_comfy_org: str,
+                 prompt_id: str = None, **kwargs):
         self.unique_id = unique_id
         """UNIQUE_ID is the unique identifier of the node, and matches the id property of the node on the client side. It is commonly used in client-server communications (see messages)."""
         self.prompt = prompt
@@ -1297,6 +1304,8 @@ class HiddenHolder:
         """AUTH_TOKEN_COMFY_ORG is a token acquired from signing into a ComfyOrg account on frontend."""
         self.api_key_comfy_org = api_key_comfy_org
         """API_KEY_COMFY_ORG is an API Key generated by ComfyOrg that allows skipping signing into a ComfyOrg account on frontend."""
+        self.prompt_id = prompt_id
+        """PROMPT_ID is the unique identifier of the current prompt/job being executed."""
 
     def __getattr__(self, key: str):
         '''If hidden variable not found, return None.'''
@@ -1304,6 +1313,14 @@ class HiddenHolder:
 
     @classmethod
     def from_dict(cls, d: dict | None):
+        """Create a HiddenHolder from a dictionary of hidden input values.
+
+        Args:
+            d: Dictionary mapping Hidden enum values to their resolved values.
+
+        Returns:
+            A new HiddenHolder instance with values populated from the dict.
+        """
         if d is None:
             d = {}
         return cls(
@@ -1313,6 +1330,7 @@ class HiddenHolder:
             dynprompt=d.get(Hidden.dynprompt, None),
             auth_token_comfy_org=d.get(Hidden.auth_token_comfy_org, None),
             api_key_comfy_org=d.get(Hidden.api_key_comfy_org, None),
+            prompt_id=d.get(Hidden.prompt_id, None),
         )
 
     @classmethod
@@ -1335,6 +1353,8 @@ class Hidden(str, Enum):
     """AUTH_TOKEN_COMFY_ORG is a token acquired from signing into a ComfyOrg account on frontend."""
     api_key_comfy_org = "API_KEY_COMFY_ORG"
     """API_KEY_COMFY_ORG is an API Key generated by ComfyOrg that allows skipping signing into a ComfyOrg account on frontend."""
+    prompt_id = "PROMPT_ID"
+    """PROMPT_ID is the unique identifier of the current prompt/job being executed. Useful for associating progress updates with specific jobs."""
 
 
 @dataclass

comfy_api_nodes/util/client.py

@@ -17,6 +17,7 @@ from pydantic import BaseModel
 
 from comfy import utils
 from comfy_api.latest import IO
+from comfy_execution.utils import get_executing_context
 from server import PromptServer
 
 from . import request_logger
@@ -440,6 +441,17 @@ def _display_text(
     status: str | int | None = None,
     price: float | None = None,
 ) -> None:
+    """Send a progress text message to the client for display on a node.
+
+    Assembles status, price, and text lines, then sends them via WebSocket.
+    Automatically retrieves the current prompt_id from the execution context.
+
+    Args:
+        node_cls: The ComfyNode class sending the progress text.
+        text: Optional text content to display.
+        status: Optional status string or code to display.
+        price: Optional price in dollars to display as credits.
+    """
     display_lines: list[str] = []
     if status:
         display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}")
@@ -450,7 +462,9 @@ def _display_text(
     if text is not None:
         display_lines.append(text)
     if display_lines:
-        PromptServer.instance.send_progress_text("\n".join(display_lines), get_node_id(node_cls))
+        ctx = get_executing_context()
+        prompt_id = ctx.prompt_id if ctx is not None else None
+        PromptServer.instance.send_progress_text("\n".join(display_lines), get_node_id(node_cls), prompt_id=prompt_id)
 
 
 def _display_time_progress(

comfy_extras/nodes_images.py

@@ -566,7 +566,7 @@ class GetImageSize(IO.ComfyNode):
                 IO.Int.Output(display_name="height"),
                 IO.Int.Output(display_name="batch_size"),
             ],
-            hidden=[IO.Hidden.unique_id],
+            hidden=[IO.Hidden.unique_id, IO.Hidden.prompt_id],
         )
 
     @classmethod
@@ -577,7 +577,7 @@ class GetImageSize(IO.ComfyNode):
 
         # Send progress text to display size on the node
         if cls.hidden.unique_id:
-            PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id)
+            PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id, prompt_id=cls.hidden.prompt_id)
 
         return IO.NodeOutput(width, height, batch_size)
 

execution.py

@@ -150,7 +150,7 @@ class CacheSet:
 
 SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org")
 
-def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}):
+def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}, prompt_id=None):
     is_v3 = issubclass(class_def, _ComfyNodeInternal)
     v3_data: io.V3Data = {}
     hidden_inputs_v3 = {}
@@ -197,6 +197,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=
                 hidden_inputs_v3[io.Hidden.auth_token_comfy_org] = extra_data.get("auth_token_comfy_org", None)
             if io.Hidden.api_key_comfy_org.name in hidden:
                 hidden_inputs_v3[io.Hidden.api_key_comfy_org] = extra_data.get("api_key_comfy_org", None)
+            if io.Hidden.prompt_id.name in hidden:
+                hidden_inputs_v3[io.Hidden.prompt_id] = prompt_id
     else:
         if "hidden" in valid_inputs:
             h = valid_inputs["hidden"]
@@ -213,6 +215,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=
                     input_data_all[x] = [extra_data.get("auth_token_comfy_org", None)]
                 if h[x] == "API_KEY_COMFY_ORG":
                     input_data_all[x] = [extra_data.get("api_key_comfy_org", None)]
+                if h[x] == "PROMPT_ID":
+                    input_data_all[x] = [prompt_id]
     v3_data["hidden_inputs"] = hidden_inputs_v3
     return input_data_all, missing_keys, v3_data
 
@@ -470,7 +474,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
             has_subgraph = False
         else:
             get_progress_state().start_progress(unique_id)
-            input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data)
+            input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data, prompt_id=prompt_id)
             if server.client_id is not None:
                 server.last_node_id = display_node_id
                 server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id)

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b4
+comfyui_manager==4.1b5

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
@@ -1230,13 +1230,47 @@ class PromptServer():
         return json_data
 
     def send_progress_text(
-        self, text: Union[bytes, bytearray, str], node_id: str, sid=None
+        self,
+        text: Union[bytes, bytearray, str],
+        node_id: str,
+        prompt_id: Optional[str] = None,
+        sid=None,
     ):
+        """Send a progress text message to the client via WebSocket.
+
+        Encodes the text as a binary message with length-prefixed node_id. When
+        the client supports the ``supports_progress_text_metadata`` feature flag,
+        the prompt_id is always prepended as a length-prefixed field (empty string
+        when None) to ensure consistent binary framing.
+
+        Args:
+            text: The progress text content to send.
+            node_id: The unique identifier of the node sending the progress.
+            prompt_id: Optional prompt/job identifier to associate the message with.
+            sid: Optional session ID to target a specific client.
+        """
         if isinstance(text, str):
             text = text.encode("utf-8")
         node_id_bytes = str(node_id).encode("utf-8")
 
-        # Pack the node_id length as a 4-byte unsigned integer, followed by the node_id bytes
-        message = struct.pack(">I", len(node_id_bytes)) + node_id_bytes + text
+        # Auto-resolve sid to the currently executing client
+        target_sid = sid if sid is not None else self.client_id
 
-        self.send_sync(BinaryEventTypes.TEXT, message, sid)
+        # When client supports the new format, always send
+        # [prompt_id_len][prompt_id][node_id_len][node_id][text]
+        # even when prompt_id is None (encoded as zero-length string)
+        if feature_flags.supports_feature(
+            self.sockets_metadata, target_sid, "supports_progress_text_metadata"
+        ):
+            prompt_id_bytes = (prompt_id or "").encode("utf-8")
+            message = (
+                struct.pack(">I", len(prompt_id_bytes))
+                + prompt_id_bytes
+                + struct.pack(">I", len(node_id_bytes))
+                + node_id_bytes
+                + text
+            )
+        else:
+            message = struct.pack(">I", len(node_id_bytes)) + node_id_bytes + text
+
+        self.send_sync(BinaryEventTypes.TEXT, message, target_sid)

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