add video support for language models, improve torch compile for transformers models, make it easier to load videos, fix progess messages from language nodes

comfy/cmd/server.py

@@ -1161,8 +1161,12 @@ class PromptServer(ExecutorToClientProgress):
             await self.send_image(data, sid=sid)
         elif event == BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA:
             # data is (preview_image, metadata)
+
             data: PreviewImageWithMetadataMessage
             preview_image, metadata = data
+            if isinstance(preview_image, dict):
+            # todo: this has to be fixed from transformers loader for previewing tokens in real time
+                return
             await self.send_image_with_metadata(preview_image, metadata, sid=sid)
         elif isinstance(data, (bytes, bytearray)):
             await self.send_bytes(event, data, sid)

comfy/language/language_types.py

@@ -92,14 +92,13 @@ class LanguageModel(Protocol):
 
     def generate(self, tokens: TOKENS_TYPE = None,
                  max_new_tokens: int = 512,
-                 repetition_penalty: float = 0.0,
                  seed: int = 0,
                  sampler: Optional[GENERATION_KWARGS_TYPE] = None,
                  *args,
                  **kwargs) -> str:
         ...
 
-    def tokenize(self, prompt: str | LanguagePrompt, images: RGBImageBatch | None, chat_template: str | None = None) -> ProcessorResult:
+    def tokenize(self, prompt: str | LanguagePrompt, images: RGBImageBatch | None, videos: list[torch.Tensor] | None, chat_template: str | None = None) -> ProcessorResult:
         ...
 
     @property

comfy/language/transformers_model_management.py

@@ -19,7 +19,7 @@ from transformers import PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixi
 from huggingface_hub import hf_api
 from huggingface_hub.file_download import hf_hub_download
 from transformers.models.auto.modeling_auto import MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES, \
-    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES, MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES, MODEL_FOR_CAUSAL_LM_MAPPING_NAMES, AutoModelForImageTextToText
 
 from .chat_templates import KNOWN_CHAT_TEMPLATES
 from .language_types import ProcessorResult, TOKENS_TYPE, GENERATION_KWARGS_TYPE, TransformerStreamedProgress, \
@@ -122,7 +122,7 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
             for i, kwargs_to_try in enumerate(kwargses_to_try):
                 try:
                     if model_type in MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES:
-                        model = AutoModelForVision2Seq.from_pretrained(**from_pretrained_kwargs, **kwargs_to_try)
+                        model = AutoModelForImageTextToText.from_pretrained(**from_pretrained_kwargs, **kwargs_to_try)
                     elif model_type in MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES:
                         model = AutoModelForSeq2SeqLM.from_pretrained(**from_pretrained_kwargs, **kwargs_to_try)
                     elif model_type in _OVERRIDDEN_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
@@ -176,7 +176,6 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
 
     def generate(self, tokens: TOKENS_TYPE = None,
                  max_new_tokens: int = 512,
-                 repetition_penalty: float = 0.0,
                  seed: int = 0,
                  sampler: Optional[GENERATION_KWARGS_TYPE] = None,
                  *args,
@@ -257,7 +256,6 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
                     **inputs,
                     streamer=text_streamer if num_beams <= 1 else None,
                     max_new_tokens=max_new_tokens,
-                    repetition_penalty=repetition_penalty if repetition_penalty != 0 else None,
                     **generate_kwargs
                 )
 
@@ -364,7 +362,7 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
         if processor is not None and hasattr(processor, "image_processor") and hasattr(processor.image_processor, "do_rescale"):
             processor.image_processor.do_rescale = False
 
-    def tokenize(self, prompt: str | LanguagePrompt, images: RGBImageBatch | None, chat_template: str | None = None) -> ProcessorResult:
+    def tokenize(self, prompt: str | LanguagePrompt, images: RGBImageBatch | None, videos: list[torch.Tensor] | None = None, chat_template: str | None = None) -> ProcessorResult:
         tokenizer = self.processor if self.processor is not None else self.tokenizer
         assert tokenizer is not None
         assert hasattr(tokenizer, "decode")
@@ -391,16 +389,18 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
                 messages: LanguagePrompt
                 if isinstance(prompt, list) and len(prompt) > 0 and isinstance(prompt[0], dict):
                     messages = prompt
-                elif "content[" in chat_template:
+                elif images is not None and len(images) > 0 or videos is not None and len(videos) > 0:
                     messages = [
                         {"role": "user",
                          "content": [
                                         {
                                             "type": "text",
-                                            "text": prompt
+                                            "text": prompt if isinstance(prompt, str) else ""
                                         }
                                     ] + [
                                         {"type": "image"} for _ in range(len(images))
+                                    ] + [
+                                        {"type": "video"} for _ in range(len(videos))
                                     ]
 
                          }
@@ -409,6 +409,7 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
                     messages = [
                         {"role": "user", "content": prompt},
                     ]
+
                 prompt = tokenizer.apply_chat_template(messages, chat_template=chat_template, add_generation_prompt=True, tokenize=False)
         except Exception as exc:
             logger.debug("Could not apply chat template", exc_info=exc)
@@ -422,7 +423,14 @@ class TransformersManagedModel(ModelManageableStub, LanguageModel):
             # convert tuple to list from images.unbind() for paligemma workaround
             image_tensor_list = list(images.unbind()) if images is not None and len(images) > 0 else None
             try:
-                batch_feature: BatchFeature = self.processor(text=[prompt], images=image_tensor_list, return_tensors="pt", padding=True)
+                batch_feature: BatchFeature = self.processor(
+                    text=[prompt],
+                    images=image_tensor_list,
+                    videos=None if videos is not None and len(videos) == 0 or (hasattr(videos, "shape") and videos.shape[0]) == 0 else videos,
+                    return_tensors="pt",
+                    padding=True,
+                    input_data_format="channels_last"  # Ensure this is set for Qwen
+                )
             except TypeError as exc_info:
                 logger.warning(f"Exception while trying to run processor. Your transformers package is version {transformers.__version__} and may need to be updated")
                 raise exc_info

comfy_execution/progress.py

@@ -65,7 +65,7 @@ class ProgressHandler(ABC):
             max_value: float,
             state: NodeProgressState,
             prompt_id: str,
-            image: PreviewImageTuple | None = None,
+            image: PreviewImageTuple | dict | None = None,
     ):
         """Called when a node's progress is updated"""
         pass
@@ -210,7 +210,7 @@ class WebUIProgressHandler(ProgressHandler):
             max_value: float,
             state: NodeProgressState,
             prompt_id: str,
-            image: PreviewImageTuple | None = None,
+            image: PreviewImageTuple | dict | None = None,
     ):
         # Send progress state of all nodes
         if self.registry:
@@ -233,6 +233,7 @@ class WebUIProgressHandler(ProgressHandler):
                     ),
                     "real_node_id": self.registry.dynprompt.get_real_node_id(node_id),
                 }
+                # todo: image can be a dict because of transformers loader, do we have to deal with this specially? things just work
                 message: PreviewImageWithMetadataMessage = (image, metadata)
                 self.server_instance.send_sync(
                     BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA,
@@ -300,7 +301,7 @@ class ProgressRegistry(AbstractProgressRegistry):
                 handler.start_handler(node_id, entry, self.prompt_id)
 
     def update_progress(
-            self, node_id: str, value: float, max_value: float, image: PreviewImageTuple | None = None
+            self, node_id: str, value: float, max_value: float, image: PreviewImageTuple | dict | None = None
     ) -> None:
         """Update progress for a node"""
         entry = self.ensure_entry(node_id)

comfy_extras/nodes/nodes_language.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import operator
 import os.path
+import re
 from abc import ABC, abstractmethod
 from functools import reduce
 from typing import Optional, List
@@ -332,6 +333,7 @@ class OneShotInstructTokenize(CustomNode):
             },
             "optional": {
                 "images": ("IMAGE", {}),
+                "videos": ("VIDEO", {}),
                 "system_prompt": ("STRING", {"multiline": True, "default": ""})
             }
         }
@@ -340,9 +342,8 @@ class OneShotInstructTokenize(CustomNode):
     RETURN_TYPES = (TOKENS_TYPE_NAME,)
     FUNCTION = "execute"
 
-    def execute(self, model: LanguageModel, prompt: str, images: List[torch.Tensor] | torch.Tensor = None, chat_template: Optional[str] = _AUTO_CHAT_TEMPLATE, system_prompt: str = "") -> ValidatedNodeResult:
+    def execute(self, model: LanguageModel, prompt: str, images: List[torch.Tensor] | torch.Tensor = None, videos: list | object = None, chat_template: Optional[str] = _AUTO_CHAT_TEMPLATE, system_prompt: str = "") -> ValidatedNodeResult:
         if chat_template == _AUTO_CHAT_TEMPLATE:
-            # use an exact match
             model_name = os.path.basename(model.repo_id)
             if model_name in KNOWN_CHAT_TEMPLATES:
                 chat_template = KNOWN_CHAT_TEMPLATES[model_name]
@@ -351,22 +352,43 @@ class OneShotInstructTokenize(CustomNode):
         elif chat_template is not None:
             chat_template = KNOWN_CHAT_TEMPLATES[chat_template]
 
+        video_tensors = []
+        if videos is not None:
+            if not isinstance(videos, list):
+                videos_list = [videos]
+            else:
+                videos_list = videos
+
+            for vid in videos_list:
+                if hasattr(vid, "get_components"):
+                    components = vid.get_components()
+                    video_tensors.append(components.images)
+                elif isinstance(vid, torch.Tensor):
+                    video_tensors.append(vid)
+
         messages: LanguagePrompt | str
-        if system_prompt != "":
+
-            messages: LanguagePrompt = [
+        has_images = images is not None and len(images) > 0
-                {"role": "system",
+        has_videos = len(video_tensors) > 0
-                 "content": system_prompt},
+
-                {"role": "user",
+        if system_prompt != "" or has_images or has_videos:
-                 "content": [
+            user_content = [{"type": "text", "text": prompt}]
-                                {"type": "text",
+            if has_images:
-                                 "text": prompt}
+                user_content += [{"type": "image"} for _ in range(len(images))]
-                            ] + [
+
-                                {"type": "image"} for _ in range(len(images) if images is not None else 0)
+            if has_videos:
-                            ], }
+                user_content += [{"type": "video"} for _ in range(len(video_tensors))]
+
+            messages = [
+                {"role": "user", "content": user_content}
             ]
+
+            if system_prompt.strip() != "":
+                messages.insert(0, {"role": "system", "content": system_prompt})
         else:
-            messages: str = prompt
+            messages = prompt
-        return model.tokenize(messages, images, chat_template),
+
+        return model.tokenize(messages, images, video_tensors, chat_template),
 
 
 class TransformersGenerate(CustomNode):
@@ -376,8 +398,7 @@ class TransformersGenerate(CustomNode):
             "required": {
                 "model": ("MODEL",),
                 "tokens": (TOKENS_TYPE_NAME, {}),
-                "max_new_tokens": ("INT", {"default": 512, "min": 1}),
+                "max_new_tokens": ("INT", {"default": 512, "min": 1, "max": 0xffffffff}),
-                "repetition_penalty": ("FLOAT", {"default": 0.0, "min": 0}),
                 "seed": Seed,
             },
             "optional": {
@@ -393,11 +414,10 @@ class TransformersGenerate(CustomNode):
                 model: Optional[LanguageModel] = None,
                 tokens: TOKENS_TYPE = None,
                 max_new_tokens: int = 512,
-                repetition_penalty: float = 0.0,
                 seed: int = 0,
                 sampler: Optional[GENERATION_KWARGS_TYPE] = None,
                 ):
-        return model.generate(tokens, max_new_tokens, repetition_penalty, seed, sampler),
+        return model.generate(tokens, max_new_tokens, seed, sampler),
 
 
 class PreviewString(CustomNode):
@@ -427,7 +447,7 @@ class SaveString(CustomNode):
                 "filename_prefix": ("STRING", {"default": "ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."})
             },
             "optional": {
-                "extension": ("STRING", {"default": ".json"})
+                "extension": ("STRING", {"default": ".txt"})
             }
         }
 
@@ -439,17 +459,53 @@ class SaveString(CustomNode):
     def get_save_path(self, filename_prefix) -> SaveImagePathTuple:
         return folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory(), 0, 0)
 
-    def execute(self, value: str | list[str], filename_prefix: str, extension: str = ".json"):
+    def execute(self, value: str | list[str] = "", filename_prefix: str = "ComfyUI", extension: str = ".txt"):
         full_output_folder, filename, counter, subfolder, filename_prefix = self.get_save_path(filename_prefix)
         if isinstance(value, str):
             value = [value]
 
         for i, value_i in enumerate(value):
             # roughly matches the behavior of save image, but does not support batch numbers
-            with open(os.path.join(full_output_folder, f"{filename}_{counter:05d}_{extension}" if len(value) == 1 else f"{filename}_{counter:05d}_{i:02d}_{extension}"), "wt+") as f:
+            with open(os.path.join(full_output_folder, f"{filename}_{counter:05d}{extension}" if len(value) == 1 else f"{filename}_{counter:05d}_{i:02d}{extension}"), "wt+") as f:
                 f.write(value_i)
         return {"ui": {"string": value}}
 
 
+class OmitThink(CustomNode):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypes:
+        return {
+            "required": {
+                "value": ("STRING", {"forceInput": True}),
+            },
+        }
+
+    CATEGORY = "strings"
+    FUNCTION = "execute"
+    OUTPUT_NODE = True
+    RETURN_TYPES = ("STRING",)
+
+    def execute(self, value: str | list[str] = "") -> tuple[list[str]]:
+        pattern_explicit = r"<think>.*?</think>"
+        pattern_missing_start = r"^.*?</think>"
+
+        if isinstance(value, str):
+            values = [value]
+        else:
+            values = value
+
+        result = []
+        for value in values:
+            if "<think>" in value:
+                cleaned_text = re.sub(pattern_explicit, "", value, flags=re.DOTALL)
+            elif "</think>" in value:
+                cleaned_text = re.sub(pattern_missing_start, "", value, flags=re.DOTALL)
+            else:
+                cleaned_text = value
+            result.append(cleaned_text.strip())
+
+        return result,
+
+
 export_custom_nodes()
 export_package_as_web_directory("comfy_extras.language_web")

comfy_extras/nodes/nodes_open_api.py

@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 import dataclasses
-import io
 import json
 import logging
 import os
@@ -16,6 +15,7 @@ from typing import Sequence, Optional, TypedDict, List, Literal, Tuple, Any, Dic
 
 import PIL
 import aiohttp
+import av
 import certifi
 import cv2
 import fsspec
@@ -872,4 +872,162 @@ class ImageRequestParameter(CustomNode):
         return ImageMaskTuple(output_images_batched, output_masks_batched)
 
 
+class LoadImageFromURL(ImageRequestParameter):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypes:
+        return {
+            "required": {
+                "value": ("STRING", {"default": ""})
+            },
+            "optional": {
+                "default_if_empty": ("IMAGE",),
+                "alpha_is_transparency": ("BOOLEAN", {"default": False}),
+            }
+        }
+
+    def execute(self, value: str = "", default_if_empty=None, alpha_is_transparency=False, *args, **kwargs) -> ImageMaskTuple:
+        return super().execute(value, default_if_empty, alpha_is_transparency, *args, **kwargs)
+
+
+class VideoRequestParameter(CustomNode):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypes:
+        return {
+            "required": {
+                "value": ("STRING", {"default": ""})
+            },
+            "optional": {
+                **_open_api_common_schema,
+                "default_if_empty": ("VIDEO",),
+                "frame_load_cap": ("INT", {"default": 0, "min": 0, "step": 1, "tooltip": "0 for no limit, otherwise stop loading after N frames"}),
+                "skip_first_frames": ("INT", {"default": 0, "min": 0, "step": 1}),
+                "select_every_nth": ("INT", {"default": 1, "min": 1, "step": 1}),
+            }
+        }
+
+    RETURN_TYPES = ("VIDEO", "MASK", "INT", "FLOAT")
+    RETURN_NAMES = ("VIDEO", "MASK", "frame_count", "fps")
+    FUNCTION = "execute"
+    CATEGORY = "api/openapi"
+
+    def execute(self, value: str = "", default_if_empty=None, frame_load_cap=0, skip_first_frames=0, select_every_nth=1, *args, **kwargs) -> tuple[Tensor, Tensor, int, float]:
+        if value.strip() == "":
+            if default_if_empty is None:
+                return (torch.zeros((0, 1, 1, 3)), torch.zeros((0, 1, 1)), 0, 0.0)
+
+            frames = default_if_empty.shape[0] if isinstance(default_if_empty, torch.Tensor) else 0
+            height = default_if_empty.shape[1] if frames > 0 else 1
+            width = default_if_empty.shape[2] if frames > 0 else 1
+
+            default_mask = torch.ones((frames, height, width), dtype=torch.float32)
+            return (default_if_empty, default_mask, frames, 0.0)
+
+        output_videos = []
+        output_masks = []
+        total_frames_loaded = 0
+        fps = 0.0
+
+        fsspec_kwargs = {}
+        if value.startswith('http'):
+            fsspec_kwargs.update({
+                "headers": {
+                    'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/113.0.5672.64 Safari/537.36'
+                },
+                'get_client': get_client
+            })
+
+        with fsspec.open_files(value, mode="rb", **fsspec_kwargs) as files:
+            for f in files:
+                try:
+                    container = av.open(f)
+                except Exception as e:
+                    logger.error(f"VideoRequestParameter: Failed to open video container for {value}: {e}")
+                    continue
+
+                if len(container.streams.video) == 0:
+                    continue
+
+                stream = container.streams.video[0]
+                stream.thread_type = "AUTO"
+
+                if fps == 0.0:
+                    fps = float(stream.average_rate)
+
+                frames_list = []
+                masks_list = []
+                frames_processed = 0
+                frames_kept = 0
+
+                for frame in container.decode(stream):
+                    frames_processed += 1
+
+                    if frames_processed <= skip_first_frames:
+                        continue
+
+                    if (frames_processed - skip_first_frames - 1) % select_every_nth != 0:
+                        continue
+
+                    np_frame = frame.to_ndarray(format="rgba")
+                    tensor_img = torch.from_numpy(np_frame[..., :3]).float() / 255.0
+                    frames_list.append(tensor_img)
+                    tensor_mask = torch.from_numpy(np_frame[..., 3]).float() / 255.0
+                    masks_list.append(tensor_mask)
+
+                    frames_kept += 1
+
+                    if frame_load_cap > 0 and frames_kept >= frame_load_cap:
+                        break
+
+                container.close()
+
+                if frames_list:
+                    video_tensor = torch.stack(frames_list)
+                    mask_tensor = torch.stack(masks_list)
+
+                    output_videos.append(video_tensor)
+                    output_masks.append(mask_tensor)
+                    total_frames_loaded += frames_kept
+
+        if not output_videos:
+            if default_if_empty is not None:
+                frames = default_if_empty.shape[0]
+                height = default_if_empty.shape[1]
+                width = default_if_empty.shape[2]
+                return (default_if_empty, torch.ones((frames, height, width), dtype=torch.float32), frames, 0.0)
+            return (torch.zeros((0, 1, 1, 3)), torch.zeros((0, 1, 1)), 0, 0.0)
+
+        try:
+            final_video = torch.cat(output_videos, dim=0)
+            final_mask = torch.cat(output_masks, dim=0)
+        except RuntimeError:
+            logger.warning("VideoRequestParameter: Video resolutions mismatch in input list. Returning only the first video.")
+            final_video = output_videos[0]
+            final_mask = output_masks[0]
+            total_frames_loaded = final_video.shape[0]
+
+        return (final_video, final_mask, total_frames_loaded, fps)
+
+
+class LoadVideoFromURL(VideoRequestParameter):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypes:
+        return {
+            "required": {
+                "value": ("STRING", {"default": ""})
+            },
+            "optional": {
+                "default_if_empty": ("VIDEO",),
+                "frame_load_cap": ("INT", {"default": 0, "min": 0, "step": 1}),
+                "skip_first_frames": ("INT", {"default": 0, "min": 0, "step": 1}),
+                "select_every_nth": ("INT", {"default": 1, "min": 1, "step": 1}),
+            }
+        }
+
+    RETURN_TYPES = ("VIDEO", "MASK", "INT", "FLOAT")
+    RETURN_NAMES = ("VIDEO", "MASK", "frame_count", "fps")
+
+    def execute(self, value: str = "", default_if_empty=None, frame_load_cap=0, skip_first_frames=0, select_every_nth=1, *args, **kwargs):
+        return super().execute(value, default_if_empty, frame_load_cap, skip_first_frames, select_every_nth, *args, **kwargs)
+
+
 export_custom_nodes()

comfy_extras/nodes/nodes_torch_compile.py

@@ -46,6 +46,8 @@ def write_atomic(
 
 
 torch._inductor.codecache.write_atomic = write_atomic
+
+
 # torch._inductor.utils.is_big_gpu = lambda *args: True
 
 
@@ -98,7 +100,18 @@ class TorchCompileModel(CustomNode):
                     "make_refittable": True,
                 }
                 del compile_kwargs["mode"]
-            if isinstance(model, (ModelPatcher, TransformersManagedModel, VAE)):
+            if isinstance(model, TransformersManagedModel):
+                to_return = model.clone()
+                model = to_return.model
+
+                model_management.unload_all_models()
+                model.to(device=model_management.get_torch_device())
+                res = torch.compile(model=model, **compile_kwargs),
+                model.to(device=model_management.unet_offload_device())
+
+                to_return.add_object_patch("model", res)
+                return to_return,
+            elif isinstance(model, (ModelPatcher, VAE)):
                 to_return = model.clone()
                 object_patches = [p.strip() for p in object_patch.split(",")]
                 patcher: ModelPatcher