From 1b96fae1d4a8425c44d4d3bd60acd818d05bf4f6 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 19 Jul 2025 01:55:23 -0700 Subject: [PATCH 01/24] Add nested style of dual cfg to DualCFGGuider node. (#8965) --- comfy_extras/nodes_custom_sampler.py | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 33bc41842..d17737e1a 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -683,9 +683,10 @@ class CFGGuider: return (guider,) class Guider_DualCFG(comfy.samplers.CFGGuider): - def set_cfg(self, cfg1, cfg2): + def set_cfg(self, cfg1, cfg2, nested=False): self.cfg1 = cfg1 self.cfg2 = cfg2 + self.nested = nested def set_conds(self, positive, middle, negative): middle = node_helpers.conditioning_set_values(middle, {"prompt_type": "negative"}) @@ -695,14 +696,20 @@ class Guider_DualCFG(comfy.samplers.CFGGuider): negative_cond = self.conds.get("negative", None) middle_cond = self.conds.get("middle", None) positive_cond = self.conds.get("positive", None) - if model_options.get("disable_cfg1_optimization", False) == False: - if math.isclose(self.cfg2, 1.0): - negative_cond = None - if math.isclose(self.cfg1, 1.0): - middle_cond = None - out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options) - return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1 + if self.nested: + out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options) + pred_text = comfy.samplers.cfg_function(self.inner_model, out[2], out[1], self.cfg1, x, timestep, model_options=model_options, cond=positive_cond, uncond=middle_cond) + return out[0] + self.cfg2 * (pred_text - out[0]) + else: + if model_options.get("disable_cfg1_optimization", False) == False: + if math.isclose(self.cfg2, 1.0): + negative_cond = None + if math.isclose(self.cfg1, 1.0): + middle_cond = None + + out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options) + return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1 class DualCFGGuider: @classmethod @@ -714,6 +721,7 @@ class DualCFGGuider: "negative": ("CONDITIONING", ), "cfg_conds": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), "cfg_cond2_negative": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), + "style": (["regular", "nested"],), } } @@ -722,10 +730,10 @@ class DualCFGGuider: FUNCTION = "get_guider" CATEGORY = "sampling/custom_sampling/guiders" - def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative): + def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style): guider = Guider_DualCFG(model) guider.set_conds(cond1, cond2, negative) - guider.set_cfg(cfg_conds, cfg_cond2_negative) + guider.set_cfg(cfg_conds, cfg_cond2_negative, nested=(style == "nested")) return (guider,) class DisableNoise: From 1da5639e865a50f921d870a92c7c87110ce20c48 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 19 Jul 2025 18:08:00 +0800 Subject: [PATCH 02/24] Update template to 0.1.37 (#8967) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7705918a8..a7e44095f 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.23.4 -comfyui-workflow-templates==0.1.36 +comfyui-workflow-templates==0.1.37 comfyui-embedded-docs==0.2.4 torch torchsde From 100c2478eaba71ab735539fdc00c9d0de49bc224 Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Sun, 20 Jul 2025 11:09:11 +0800 Subject: [PATCH 03/24] Add SamplingPercentToSigma node (#8963) It's helpful to adjust start_percent or end_percent based on the corresponding sigma. --- comfy_extras/nodes_custom_sampler.py | 30 ++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index d17737e1a..d011f433b 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -301,6 +301,35 @@ class ExtendIntermediateSigmas: return (extended_sigmas,) + +class SamplingPercentToSigma: + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "model": (IO.MODEL, {}), + "sampling_percent": (IO.FLOAT, {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.0001}), + "return_actual_sigma": (IO.BOOLEAN, {"default": False, "tooltip": "Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."}), + } + } + + RETURN_TYPES = (IO.FLOAT,) + RETURN_NAMES = ("sigma_value",) + CATEGORY = "sampling/custom_sampling/sigmas" + + FUNCTION = "get_sigma" + + def get_sigma(self, model, sampling_percent, return_actual_sigma): + model_sampling = model.get_model_object("model_sampling") + sigma_val = model_sampling.percent_to_sigma(sampling_percent) + if return_actual_sigma: + if sampling_percent == 0.0: + sigma_val = model_sampling.sigma_max.item() + elif sampling_percent == 1.0: + sigma_val = model_sampling.sigma_min.item() + return (sigma_val,) + + class KSamplerSelect: @classmethod def INPUT_TYPES(s): @@ -887,6 +916,7 @@ NODE_CLASS_MAPPINGS = { "FlipSigmas": FlipSigmas, "SetFirstSigma": SetFirstSigma, "ExtendIntermediateSigmas": ExtendIntermediateSigmas, + "SamplingPercentToSigma": SamplingPercentToSigma, "CFGGuider": CFGGuider, "DualCFGGuider": DualCFGGuider, From a0c0785635a9f4d2da64b58fef063825f386d8da Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 19 Jul 2025 22:24:09 -0700 Subject: [PATCH 04/24] Document what the fast_fp16_accumulation is in the portable. (#8973) --- .ci/windows_base_files/README_VERY_IMPORTANT.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/.ci/windows_base_files/README_VERY_IMPORTANT.txt b/.ci/windows_base_files/README_VERY_IMPORTANT.txt index d46acbcbf..8ab70c890 100755 --- a/.ci/windows_base_files/README_VERY_IMPORTANT.txt +++ b/.ci/windows_base_files/README_VERY_IMPORTANT.txt @@ -4,6 +4,9 @@ if you have a NVIDIA gpu: run_nvidia_gpu.bat +if you want to enable the fast fp16 accumulation (faster for fp16 models with slightly less quality): + +run_nvidia_gpu_fast_fp16_accumulation.bat To run it in slow CPU mode: From 7d627f764c2137d816a39adbc358cb28c1718a47 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Mon, 21 Jul 2025 03:58:35 +0800 Subject: [PATCH 05/24] Update template to 0.1.39 (#8981) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index a7e44095f..8f6a6d112 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.23.4 -comfyui-workflow-templates==0.1.37 +comfyui-workflow-templates==0.1.39 comfyui-embedded-docs==0.2.4 torch torchsde From 9a470e073e2742d4edd6e7ea1ce28d861a77d9c4 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 21 Jul 2025 14:05:43 -0400 Subject: [PATCH 06/24] ComfyUI version 0.3.45 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 7981fbaca..180ecaf8a 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.44" +__version__ = "0.3.45" diff --git a/pyproject.toml b/pyproject.toml index 96ead2157..b1d6d9df6 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.44" +version = "0.3.45" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 54a45b996729b361ea12f473de760e481dcf1f0a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 21 Jul 2025 11:19:14 -0700 Subject: [PATCH 07/24] Replace torchaudio.load with pyav. (#8989) --- comfy_extras/nodes_audio.py | 58 ++++++++++++++++++++++++++++++++++++- 1 file changed, 57 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 8cd647846..38697240e 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -278,6 +278,62 @@ class PreviewAudio(SaveAudio): "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, } +def f32_pcm(wav: torch.Tensor) -> torch.Tensor: + """Convert audio to float 32 bits PCM format.""" + if wav.dtype.is_floating_point: + return wav + elif wav.dtype == torch.int16: + return wav.float() / (2 ** 15) + elif wav.dtype == torch.int32: + return wav.float() / (2 ** 31) + raise ValueError(f"Unsupported wav dtype: {wav.dtype}") + +def load(filepath: str, frame_offset: int = 0, num_frames: int = -1) -> tuple[torch.Tensor, int]: + with av.open(filepath) as af: + if not af.streams.audio: + raise ValueError("No audio stream found in the file.") + + stream = af.streams.audio[0] + sr = stream.codec_context.sample_rate + n_channels = stream.channels + + seek_time = frame_offset / sr if frame_offset > 0 else 0.0 + duration = num_frames / sr if num_frames > 0 else -1.0 + + sample_offset = int(sr * seek_time) + num_samples = int(sr * duration) if duration >= 0 else -1 + + # Small negative offset for MP3 artifacts, NOTE: this is LLM code so idk if it's actually necessary' + seek_sec = max(0, seek_time - 0.1) if filepath.lower().endswith('.mp3') else seek_time + af.seek(int(seek_sec / stream.time_base), stream=stream) + + frames = [] + length = 0 + for frame in af.decode(streams=stream.index): + current_offset = int(frame.rate * frame.pts * frame.time_base) + strip = max(0, sample_offset - current_offset) + + buf = torch.from_numpy(frame.to_ndarray()) + if buf.shape[0] != n_channels: + buf = buf.view(-1, n_channels).t() + + buf = buf[:, strip:] + frames.append(buf) + length += buf.shape[1] + + if num_samples > 0 and length >= num_samples: + break + + if not frames: + raise ValueError("No audio frames decoded.") + + wav = torch.cat(frames, dim=1) + if num_samples > 0: + wav = wav[:, :num_samples] + + wav = f32_pcm(wav) + return wav, sr + class LoadAudio: @classmethod def INPUT_TYPES(s): @@ -292,7 +348,7 @@ class LoadAudio: def load(self, audio): audio_path = folder_paths.get_annotated_filepath(audio) - waveform, sample_rate = torchaudio.load(audio_path) + waveform, sample_rate = load(audio_path) audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate} return (audio, ) From 5249e45a1c7d91656ebefdebe3815005ec3d39d7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 21 Jul 2025 12:23:41 -0700 Subject: [PATCH 08/24] Add hidream e1.1 example to readme. (#8990) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 0e021a687..d004364ee 100644 --- a/README.md +++ b/README.md @@ -69,6 +69,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - Image Editing Models - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/) - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model) + - [HiDream E1.1](https://comfyanonymous.github.io/ComfyUI_examples/hidream/#hidream-e11) - Video Models - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/) - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/) From 0aa1c58b04b27311c6ba38b1d9949e7e20037d00 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 21 Jul 2025 13:48:25 -0700 Subject: [PATCH 09/24] This is not needed. (#8991) --- comfy_extras/nodes_audio.py | 22 +--------------------- 1 file changed, 1 insertion(+), 21 deletions(-) diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 38697240e..a90b31779 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -288,7 +288,7 @@ def f32_pcm(wav: torch.Tensor) -> torch.Tensor: return wav.float() / (2 ** 31) raise ValueError(f"Unsupported wav dtype: {wav.dtype}") -def load(filepath: str, frame_offset: int = 0, num_frames: int = -1) -> tuple[torch.Tensor, int]: +def load(filepath: str) -> tuple[torch.Tensor, int]: with av.open(filepath) as af: if not af.streams.audio: raise ValueError("No audio stream found in the file.") @@ -297,40 +297,20 @@ def load(filepath: str, frame_offset: int = 0, num_frames: int = -1) -> tuple[to sr = stream.codec_context.sample_rate n_channels = stream.channels - seek_time = frame_offset / sr if frame_offset > 0 else 0.0 - duration = num_frames / sr if num_frames > 0 else -1.0 - - sample_offset = int(sr * seek_time) - num_samples = int(sr * duration) if duration >= 0 else -1 - - # Small negative offset for MP3 artifacts, NOTE: this is LLM code so idk if it's actually necessary' - seek_sec = max(0, seek_time - 0.1) if filepath.lower().endswith('.mp3') else seek_time - af.seek(int(seek_sec / stream.time_base), stream=stream) - frames = [] length = 0 for frame in af.decode(streams=stream.index): - current_offset = int(frame.rate * frame.pts * frame.time_base) - strip = max(0, sample_offset - current_offset) - buf = torch.from_numpy(frame.to_ndarray()) if buf.shape[0] != n_channels: buf = buf.view(-1, n_channels).t() - buf = buf[:, strip:] frames.append(buf) length += buf.shape[1] - if num_samples > 0 and length >= num_samples: - break - if not frames: raise ValueError("No audio frames decoded.") wav = torch.cat(frames, dim=1) - if num_samples > 0: - wav = wav[:, :num_samples] - wav = f32_pcm(wav) return wav, sr From 5ac9ec214ba3ef1632701416f27948a57ec60919 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 22 Jul 2025 01:07:51 -0700 Subject: [PATCH 10/24] Try to fix line endings workflow. (#9001) --- .github/workflows/check-line-endings.yml | 1 + 1 file changed, 1 insertion(+) diff --git a/.github/workflows/check-line-endings.yml b/.github/workflows/check-line-endings.yml index f20dca565..03b3e3ced 100644 --- a/.github/workflows/check-line-endings.yml +++ b/.github/workflows/check-line-endings.yml @@ -17,6 +17,7 @@ jobs: - name: Check for Windows line endings (CRLF) run: | # Get the list of changed files in the PR + git merge origin/${{ github.base_ref }} --no-edit CHANGED_FILES=$(git diff --name-only origin/${{ github.base_ref }}..HEAD) # Flag to track if CRLF is found From 255f1398638b265a47d0e74fb4759fe6cfc3b3d4 Mon Sep 17 00:00:00 2001 From: Simon Lui <502929+simonlui@users.noreply.github.com> Date: Tue, 22 Jul 2025 12:20:09 -0700 Subject: [PATCH 11/24] Add xpu version for async offload and some other things. (#9004) --- comfy/model_management.py | 41 +++++++++++++++++++++++++++++++++------ 1 file changed, 35 insertions(+), 6 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 816caf18f..ab1e9bf3a 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -101,7 +101,7 @@ if args.directml is not None: lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default. try: - import intel_extension_for_pytorch as ipex + import intel_extension_for_pytorch as ipex # noqa: F401 _ = torch.xpu.device_count() xpu_available = xpu_available or torch.xpu.is_available() except: @@ -186,8 +186,12 @@ def get_total_memory(dev=None, torch_total_too=False): elif is_intel_xpu(): stats = torch.xpu.memory_stats(dev) mem_reserved = stats['reserved_bytes.all.current'] + if torch_version_numeric < (2, 6): + mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory + else: + _, mem_total_xpu = torch.xpu.mem_get_info(dev) mem_total_torch = mem_reserved - mem_total = torch.xpu.get_device_properties(dev).total_memory + mem_total = mem_total_xpu elif is_ascend_npu(): stats = torch.npu.memory_stats(dev) mem_reserved = stats['reserved_bytes.all.current'] @@ -929,7 +933,7 @@ def device_supports_non_blocking(device): if is_device_mps(device): return False #pytorch bug? mps doesn't support non blocking if is_intel_xpu(): - return False + return True if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews) return False if directml_enabled: @@ -968,6 +972,8 @@ def get_offload_stream(device): stream_counter = (stream_counter + 1) % len(ss) if is_device_cuda(device): ss[stream_counter].wait_stream(torch.cuda.current_stream()) + elif is_device_xpu(device): + ss[stream_counter].wait_stream(torch.xpu.current_stream()) stream_counters[device] = stream_counter return s elif is_device_cuda(device): @@ -979,6 +985,15 @@ def get_offload_stream(device): stream_counter = (stream_counter + 1) % len(ss) stream_counters[device] = stream_counter return s + elif is_device_xpu(device): + ss = [] + for k in range(NUM_STREAMS): + ss.append(torch.xpu.Stream(device=device, priority=0)) + STREAMS[device] = ss + s = ss[stream_counter] + stream_counter = (stream_counter + 1) % len(ss) + stream_counters[device] = stream_counter + return s return None def sync_stream(device, stream): @@ -986,6 +1001,8 @@ def sync_stream(device, stream): return if is_device_cuda(device): torch.cuda.current_stream().wait_stream(stream) + elif is_device_xpu(device): + torch.xpu.current_stream().wait_stream(stream) def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None): if device is None or weight.device == device: @@ -1092,8 +1109,11 @@ def get_free_memory(dev=None, torch_free_too=False): stats = torch.xpu.memory_stats(dev) mem_active = stats['active_bytes.all.current'] mem_reserved = stats['reserved_bytes.all.current'] + if torch_version_numeric < (2, 6): + mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved + else: + mem_free_xpu, _ = torch.xpu.mem_get_info(dev) mem_free_torch = mem_reserved - mem_active - mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved mem_free_total = mem_free_xpu + mem_free_torch elif is_ascend_npu(): stats = torch.npu.memory_stats(dev) @@ -1142,6 +1162,9 @@ def is_device_cpu(device): def is_device_mps(device): return is_device_type(device, 'mps') +def is_device_xpu(device): + return is_device_type(device, 'xpu') + def is_device_cuda(device): return is_device_type(device, 'cuda') @@ -1173,7 +1196,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma return False if is_intel_xpu(): - return True + if torch_version_numeric < (2, 3): + return True + else: + return torch.xpu.get_device_properties(device).has_fp16 if is_ascend_npu(): return True @@ -1236,7 +1262,10 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma return False if is_intel_xpu(): - return True + if torch_version_numeric < (2, 6): + return True + else: + return torch.xpu.get_device_capability(device)['has_bfloat16_conversions'] if is_ascend_npu(): return True From 5ad33787dee43d36f8d054c590818b3153b55370 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 23 Jul 2025 11:20:49 -0700 Subject: [PATCH 12/24] Add default device argument. (#9023) --- comfy/cli_args.py | 3 ++- comfy/model_management.py | 1 + main.py | 9 +++++++++ 3 files changed, 12 insertions(+), 1 deletion(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index ef0d4337e..0d760d524 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -49,7 +49,8 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.") parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.") parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.") -parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.") +parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.") +parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.") cm_group = parser.add_mutually_exclusive_group() cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).") cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.") diff --git a/comfy/model_management.py b/comfy/model_management.py index ab1e9bf3a..346673895 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -880,6 +880,7 @@ def vae_dtype(device=None, allowed_dtypes=[]): return d # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32 + # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3 if d == torch.bfloat16 and (not is_amd()) and should_use_bf16(device): return d diff --git a/main.py b/main.py index 2b4ffafd4..e8ca8152a 100644 --- a/main.py +++ b/main.py @@ -115,6 +115,15 @@ if os.name == "nt": logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage()) if __name__ == "__main__": + if args.default_device is not None: + default_dev = args.default_device + devices = list(range(32)) + devices.remove(default_dev) + devices.insert(0, default_dev) + devices = ','.join(map(str, devices)) + os.environ['CUDA_VISIBLE_DEVICES'] = str(devices) + os.environ['HIP_VISIBLE_DEVICES'] = str(devices) + if args.cuda_device is not None: os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device) From 39dda1d40d1f2f18ccda8ade860932d0b8a07af4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 23 Jul 2025 15:10:59 -0700 Subject: [PATCH 13/24] Fix xpu function not implemented. (#9026) --- comfy/model_management.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 346673895..746b063ed 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -186,10 +186,7 @@ def get_total_memory(dev=None, torch_total_too=False): elif is_intel_xpu(): stats = torch.xpu.memory_stats(dev) mem_reserved = stats['reserved_bytes.all.current'] - if torch_version_numeric < (2, 6): - mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory - else: - _, mem_total_xpu = torch.xpu.mem_get_info(dev) + mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory mem_total_torch = mem_reserved mem_total = mem_total_xpu elif is_ascend_npu(): From a86a58c308c2423e86054462a8c9f1125536a034 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 23 Jul 2025 15:18:20 -0700 Subject: [PATCH 14/24] Fix xpu function not implemented p2. (#9027) --- comfy/model_management.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 746b063ed..42873d09b 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1107,10 +1107,7 @@ def get_free_memory(dev=None, torch_free_too=False): stats = torch.xpu.memory_stats(dev) mem_active = stats['active_bytes.all.current'] mem_reserved = stats['reserved_bytes.all.current'] - if torch_version_numeric < (2, 6): - mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved - else: - mem_free_xpu, _ = torch.xpu.mem_get_info(dev) + mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved mem_free_torch = mem_reserved - mem_active mem_free_total = mem_free_xpu + mem_free_torch elif is_ascend_npu(): From d3504e1778c0cc8992b04fe30dc0fae239c13713 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 23 Jul 2025 16:21:29 -0700 Subject: [PATCH 15/24] Enable pytorch attention by default for gfx1201 on torch 2.8 (#9029) --- comfy/model_management.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 42873d09b..e8b9b5c81 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -308,7 +308,10 @@ try: logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much - if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx1201 and gfx950 + if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 + ENABLE_PYTORCH_ATTENTION = True + if torch_version_numeric >= (2, 8): + if any((a in arch) for a in ["gfx1201"]): ENABLE_PYTORCH_ATTENTION = True if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4): if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]): # TODO: more arches From e78d2304966b6265fa2320b4d87dca534ea15642 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 23 Jul 2025 16:37:43 -0700 Subject: [PATCH 16/24] Only enable cuda malloc on cuda torch. (#9031) --- cuda_malloc.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/cuda_malloc.py b/cuda_malloc.py index eb2857c5f..c1d9ae3ca 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -74,7 +74,8 @@ if not args.cuda_malloc: module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) version = module.__version__ - if int(version[0]) >= 2: #enable by default for torch version 2.0 and up + + if int(version[0]) >= 2 and "+cu" in version: #enable by default for torch version 2.0 and up only on cuda torch args.cuda_malloc = cuda_malloc_supported() except: pass From e729a5cc1157bc0ece7daae9583c3a5a3ba95fbb Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Thu, 24 Jul 2025 07:47:05 +0800 Subject: [PATCH 17/24] Separate denoised and noise estimation in Euler CFG++ (#9008) This will change their behavior with the sampling CONST type. It also combines euler_cfg_pp and euler_ancestral_cfg_pp into one main function. --- comfy/k_diffusion/sampling.py | 64 +++++++++++++++++------------------ 1 file changed, 32 insertions(+), 32 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 2ed415b1f..a2bc492fd 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1210,39 +1210,21 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None, return x_next -@torch.no_grad() -def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None): - extra_args = {} if extra_args is None else extra_args - - temp = [0] - def post_cfg_function(args): - temp[0] = args["uncond_denoised"] - return args["denoised"] - - model_options = extra_args.get("model_options", {}).copy() - extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True) - - s_in = x.new_ones([x.shape[0]]) - for i in trange(len(sigmas) - 1, disable=disable): - sigma_hat = sigmas[i] - denoised = model(x, sigma_hat * s_in, **extra_args) - d = to_d(x, sigma_hat, temp[0]) - if callback is not None: - callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) - # Euler method - x = denoised + d * sigmas[i + 1] - return x - @torch.no_grad() def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None): - """Ancestral sampling with Euler method steps.""" + """Ancestral sampling with Euler method steps (CFG++).""" extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler - temp = [0] + model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling") + lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) + + uncond_denoised = None + def post_cfg_function(args): - temp[0] = args["uncond_denoised"] + nonlocal uncond_denoised + uncond_denoised = args["uncond_denoised"] return args["denoised"] model_options = extra_args.get("model_options", {}).copy() @@ -1251,15 +1233,33 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No s_in = x.new_ones([x.shape[0]]) for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) - sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) - d = to_d(x, sigmas[i], temp[0]) - # Euler method - x = denoised + d * sigma_down - if sigmas[i + 1] > 0: - x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up + if sigmas[i + 1] == 0: + # Denoising step + x = denoised + else: + alpha_s = sigmas[i] * lambda_fn(sigmas[i]).exp() + alpha_t = sigmas[i + 1] * lambda_fn(sigmas[i + 1]).exp() + d = to_d(x, sigmas[i], alpha_s * uncond_denoised) # to noise + + # DDIM stochastic sampling + sigma_down, sigma_up = get_ancestral_step(sigmas[i] / alpha_s, sigmas[i + 1] / alpha_t, eta=eta) + sigma_down = alpha_t * sigma_down + + # Euler method + x = alpha_t * denoised + sigma_down * d + if eta > 0 and s_noise > 0: + x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up return x + + +@torch.no_grad() +def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None): + """Euler method steps (CFG++).""" + return sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None) + + @torch.no_grad() def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None): """Ancestral sampling with DPM-Solver++(2S) second-order steps.""" From eb2f78b4e09b1970e2fc51fc5d2e062f1a826399 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 24 Jul 2025 08:57:27 +0800 Subject: [PATCH 18/24] [Training Node] algo support, grad acc, optional grad ckpt (#9015) * Add factorization utils for lokr * Add lokr train impl * Add loha train impl * Add adapter map for algo selection * Add optional grad ckpt and algo selection * Update __init__.py * correct key name for loha * Use custom fwd/bwd func and better init for loha * Support gradient accumulation * Fix bugs of loha * use more stable init * Add OFT training * linting --- comfy/weight_adapter/__init__.py | 13 ++- comfy/weight_adapter/base.py | 40 +++++++++ comfy/weight_adapter/loha.py | 134 ++++++++++++++++++++++++++++++- comfy/weight_adapter/lokr.py | 86 +++++++++++++++++++- comfy/weight_adapter/oft.py | 67 +++++++++++++++- comfy_extras/nodes_train.py | 47 ++++++++--- 6 files changed, 372 insertions(+), 15 deletions(-) diff --git a/comfy/weight_adapter/__init__.py b/comfy/weight_adapter/__init__.py index 560b82be3..b40f920e4 100644 --- a/comfy/weight_adapter/__init__.py +++ b/comfy/weight_adapter/__init__.py @@ -15,9 +15,20 @@ adapters: list[type[WeightAdapterBase]] = [ OFTAdapter, BOFTAdapter, ] +adapter_maps: dict[str, type[WeightAdapterBase]] = { + "LoRA": LoRAAdapter, + "LoHa": LoHaAdapter, + "LoKr": LoKrAdapter, + "OFT": OFTAdapter, + ## We disable not implemented algo for now + # "GLoRA": GLoRAAdapter, + # "BOFT": BOFTAdapter, +} + __all__ = [ "WeightAdapterBase", "WeightAdapterTrainBase", - "adapters" + "adapters", + "adapter_maps", ] + [a.__name__ for a in adapters] diff --git a/comfy/weight_adapter/base.py b/comfy/weight_adapter/base.py index b5c7db423..43644b106 100644 --- a/comfy/weight_adapter/base.py +++ b/comfy/weight_adapter/base.py @@ -133,3 +133,43 @@ def tucker_weight_from_conv(up, down, mid): def tucker_weight(wa, wb, t): temp = torch.einsum("i j ..., j r -> i r ...", t, wb) return torch.einsum("i j ..., i r -> r j ...", temp, wa) + + +def factorization(dimension: int, factor: int = -1) -> tuple[int, int]: + """ + return a tuple of two value of input dimension decomposed by the number closest to factor + second value is higher or equal than first value. + + examples) + factor + -1 2 4 8 16 ... + 127 -> 1, 127 127 -> 1, 127 127 -> 1, 127 127 -> 1, 127 127 -> 1, 127 + 128 -> 8, 16 128 -> 2, 64 128 -> 4, 32 128 -> 8, 16 128 -> 8, 16 + 250 -> 10, 25 250 -> 2, 125 250 -> 2, 125 250 -> 5, 50 250 -> 10, 25 + 360 -> 8, 45 360 -> 2, 180 360 -> 4, 90 360 -> 8, 45 360 -> 12, 30 + 512 -> 16, 32 512 -> 2, 256 512 -> 4, 128 512 -> 8, 64 512 -> 16, 32 + 1024 -> 32, 32 1024 -> 2, 512 1024 -> 4, 256 1024 -> 8, 128 1024 -> 16, 64 + """ + + if factor > 0 and (dimension % factor) == 0 and dimension >= factor**2: + m = factor + n = dimension // factor + if m > n: + n, m = m, n + return m, n + if factor < 0: + factor = dimension + m, n = 1, dimension + length = m + n + while m < n: + new_m = m + 1 + while dimension % new_m != 0: + new_m += 1 + new_n = dimension // new_m + if new_m + new_n > length or new_m > factor: + break + else: + m, n = new_m, new_n + if m > n: + n, m = m, n + return m, n diff --git a/comfy/weight_adapter/loha.py b/comfy/weight_adapter/loha.py index ce79abad5..55c97a3af 100644 --- a/comfy/weight_adapter/loha.py +++ b/comfy/weight_adapter/loha.py @@ -3,7 +3,120 @@ from typing import Optional import torch import comfy.model_management -from .base import WeightAdapterBase, weight_decompose +from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose + + +class HadaWeight(torch.autograd.Function): + @staticmethod + def forward(ctx, w1u, w1d, w2u, w2d, scale=torch.tensor(1)): + ctx.save_for_backward(w1d, w1u, w2d, w2u, scale) + diff_weight = ((w1u @ w1d) * (w2u @ w2d)) * scale + return diff_weight + + @staticmethod + def backward(ctx, grad_out): + (w1d, w1u, w2d, w2u, scale) = ctx.saved_tensors + grad_out = grad_out * scale + temp = grad_out * (w2u @ w2d) + grad_w1u = temp @ w1d.T + grad_w1d = w1u.T @ temp + + temp = grad_out * (w1u @ w1d) + grad_w2u = temp @ w2d.T + grad_w2d = w2u.T @ temp + + del temp + return grad_w1u, grad_w1d, grad_w2u, grad_w2d, None + + +class HadaWeightTucker(torch.autograd.Function): + @staticmethod + def forward(ctx, t1, w1u, w1d, t2, w2u, w2d, scale=torch.tensor(1)): + ctx.save_for_backward(t1, w1d, w1u, t2, w2d, w2u, scale) + + rebuild1 = torch.einsum("i j ..., j r, i p -> p r ...", t1, w1d, w1u) + rebuild2 = torch.einsum("i j ..., j r, i p -> p r ...", t2, w2d, w2u) + + return rebuild1 * rebuild2 * scale + + @staticmethod + def backward(ctx, grad_out): + (t1, w1d, w1u, t2, w2d, w2u, scale) = ctx.saved_tensors + grad_out = grad_out * scale + + temp = torch.einsum("i j ..., j r -> i r ...", t2, w2d) + rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w2u) + + grad_w = rebuild * grad_out + del rebuild + + grad_w1u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w) + grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w1u.T) + del grad_w, temp + + grad_w1d = torch.einsum("i r ..., i j ... -> r j", t1, grad_temp) + grad_t1 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w1d.T) + del grad_temp + + temp = torch.einsum("i j ..., j r -> i r ...", t1, w1d) + rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w1u) + + grad_w = rebuild * grad_out + del rebuild + + grad_w2u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w) + grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w2u.T) + del grad_w, temp + + grad_w2d = torch.einsum("i r ..., i j ... -> r j", t2, grad_temp) + grad_t2 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w2d.T) + del grad_temp + return grad_t1, grad_w1u, grad_w1d, grad_t2, grad_w2u, grad_w2d, None + + +class LohaDiff(WeightAdapterTrainBase): + def __init__(self, weights): + super().__init__() + # Unpack weights tuple from LoHaAdapter + w1a, w1b, alpha, w2a, w2b, t1, t2, _ = weights + + # Create trainable parameters + self.hada_w1_a = torch.nn.Parameter(w1a) + self.hada_w1_b = torch.nn.Parameter(w1b) + self.hada_w2_a = torch.nn.Parameter(w2a) + self.hada_w2_b = torch.nn.Parameter(w2b) + + self.use_tucker = False + if t1 is not None and t2 is not None: + self.use_tucker = True + self.hada_t1 = torch.nn.Parameter(t1) + self.hada_t2 = torch.nn.Parameter(t2) + else: + # Keep the attributes for consistent access + self.hada_t1 = None + self.hada_t2 = None + + # Store rank and non-trainable alpha + self.rank = w1b.shape[0] + self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False) + + def __call__(self, w): + org_dtype = w.dtype + + scale = self.alpha / self.rank + if self.use_tucker: + diff_weight = HadaWeightTucker.apply(self.hada_t1, self.hada_w1_a, self.hada_w1_b, self.hada_t2, self.hada_w2_a, self.hada_w2_b, scale) + else: + diff_weight = HadaWeight.apply(self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale) + + # Add the scaled difference to the original weight + weight = w.to(diff_weight) + diff_weight.reshape(w.shape) + + return weight.to(org_dtype) + + def passive_memory_usage(self): + """Calculates memory usage of the trainable parameters.""" + return sum(param.numel() * param.element_size() for param in self.parameters()) class LoHaAdapter(WeightAdapterBase): @@ -13,6 +126,25 @@ class LoHaAdapter(WeightAdapterBase): self.loaded_keys = loaded_keys self.weights = weights + @classmethod + def create_train(cls, weight, rank=1, alpha=1.0): + out_dim = weight.shape[0] + in_dim = weight.shape[1:].numel() + mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype) + mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype) + torch.nn.init.normal_(mat1, 0.1) + torch.nn.init.constant_(mat2, 0.0) + mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype) + mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype) + torch.nn.init.normal_(mat3, 0.1) + torch.nn.init.normal_(mat4, 0.01) + return LohaDiff( + (mat1, mat2, alpha, mat3, mat4, None, None, None) + ) + + def to_train(self): + return LohaDiff(self.weights) + @classmethod def load( cls, diff --git a/comfy/weight_adapter/lokr.py b/comfy/weight_adapter/lokr.py index 51233db2d..49b0be55f 100644 --- a/comfy/weight_adapter/lokr.py +++ b/comfy/weight_adapter/lokr.py @@ -3,7 +3,77 @@ from typing import Optional import torch import comfy.model_management -from .base import WeightAdapterBase, weight_decompose +from .base import ( + WeightAdapterBase, + WeightAdapterTrainBase, + weight_decompose, + factorization, +) + + +class LokrDiff(WeightAdapterTrainBase): + def __init__(self, weights): + super().__init__() + (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) = weights + self.use_tucker = False + if lokr_w1_a is not None: + _, rank_a = lokr_w1_a.shape[0], lokr_w1_a.shape[1] + rank_a, _ = lokr_w1_b.shape[0], lokr_w1_b.shape[1] + self.lokr_w1_a = torch.nn.Parameter(lokr_w1_a) + self.lokr_w1_b = torch.nn.Parameter(lokr_w1_b) + self.w1_rebuild = True + self.ranka = rank_a + + if lokr_w2_a is not None: + _, rank_b = lokr_w2_a.shape[0], lokr_w2_a.shape[1] + rank_b, _ = lokr_w2_b.shape[0], lokr_w2_b.shape[1] + self.lokr_w2_a = torch.nn.Parameter(lokr_w2_a) + self.lokr_w2_b = torch.nn.Parameter(lokr_w2_b) + if lokr_t2 is not None: + self.use_tucker = True + self.lokr_t2 = torch.nn.Parameter(lokr_t2) + self.w2_rebuild = True + self.rankb = rank_b + + if lokr_w1 is not None: + self.lokr_w1 = torch.nn.Parameter(lokr_w1) + self.w1_rebuild = False + + if lokr_w2 is not None: + self.lokr_w2 = torch.nn.Parameter(lokr_w2) + self.w2_rebuild = False + + self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False) + + @property + def w1(self): + if self.w1_rebuild: + return (self.lokr_w1_a @ self.lokr_w1_b) * (self.alpha / self.ranka) + else: + return self.lokr_w1 + + @property + def w2(self): + if self.w2_rebuild: + if self.use_tucker: + w2 = torch.einsum( + 'i j k l, j r, i p -> p r k l', + self.lokr_t2, + self.lokr_w2_b, + self.lokr_w2_a + ) + else: + w2 = self.lokr_w2_a @ self.lokr_w2_b + return w2 * (self.alpha / self.rankb) + else: + return self.lokr_w2 + + def __call__(self, w): + diff = torch.kron(self.w1, self.w2) + return w + diff.reshape(w.shape).to(w) + + def passive_memory_usage(self): + return sum(param.numel() * param.element_size() for param in self.parameters()) class LoKrAdapter(WeightAdapterBase): @@ -13,6 +83,20 @@ class LoKrAdapter(WeightAdapterBase): self.loaded_keys = loaded_keys self.weights = weights + @classmethod + def create_train(cls, weight, rank=1, alpha=1.0): + out_dim = weight.shape[0] + in_dim = weight.shape[1:].numel() + out1, out2 = factorization(out_dim, rank) + in1, in2 = factorization(in_dim, rank) + mat1 = torch.empty(out1, in1, device=weight.device, dtype=weight.dtype) + mat2 = torch.empty(out2, in2, device=weight.device, dtype=weight.dtype) + torch.nn.init.kaiming_uniform_(mat2, a=5**0.5) + torch.nn.init.constant_(mat1, 0.0) + return LokrDiff( + (mat1, mat2, alpha, None, None, None, None, None, None) + ) + @classmethod def load( cls, diff --git a/comfy/weight_adapter/oft.py b/comfy/weight_adapter/oft.py index 25009eca3..9d4982083 100644 --- a/comfy/weight_adapter/oft.py +++ b/comfy/weight_adapter/oft.py @@ -3,7 +3,58 @@ from typing import Optional import torch import comfy.model_management -from .base import WeightAdapterBase, weight_decompose +from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization + + +class OFTDiff(WeightAdapterTrainBase): + def __init__(self, weights): + super().__init__() + # Unpack weights tuple from LoHaAdapter + blocks, rescale, alpha, _ = weights + + # Create trainable parameters + self.oft_blocks = torch.nn.Parameter(blocks) + if rescale is not None: + self.rescale = torch.nn.Parameter(rescale) + self.rescaled = True + else: + self.rescaled = False + self.block_num, self.block_size, _ = blocks.shape + self.constraint = float(alpha) + self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False) + + def __call__(self, w): + org_dtype = w.dtype + I = torch.eye(self.block_size, device=self.oft_blocks.device) + + ## generate r + # for Q = -Q^T + q = self.oft_blocks - self.oft_blocks.transpose(1, 2) + normed_q = q + if self.constraint: + q_norm = torch.norm(q) + 1e-8 + if q_norm > self.constraint: + normed_q = q * self.constraint / q_norm + # use float() to prevent unsupported type + r = (I + normed_q) @ (I - normed_q).float().inverse() + + ## Apply chunked matmul on weight + _, *shape = w.shape + org_weight = w.to(dtype=r.dtype) + org_weight = org_weight.unflatten(0, (self.block_num, self.block_size)) + # Init R=0, so add I on it to ensure the output of step0 is original model output + weight = torch.einsum( + "k n m, k n ... -> k m ...", + r, + org_weight, + ).flatten(0, 1) + if self.rescaled: + weight = self.rescale * weight + return weight.to(org_dtype) + + def passive_memory_usage(self): + """Calculates memory usage of the trainable parameters.""" + return sum(param.numel() * param.element_size() for param in self.parameters()) class OFTAdapter(WeightAdapterBase): @@ -13,6 +64,18 @@ class OFTAdapter(WeightAdapterBase): self.loaded_keys = loaded_keys self.weights = weights + @classmethod + def create_train(cls, weight, rank=1, alpha=1.0): + out_dim = weight.shape[0] + block_size, block_num = factorization(out_dim, rank) + block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype) + return OFTDiff( + (block, None, alpha, None) + ) + + def to_train(self): + return OFTDiff(self.weights) + @classmethod def load( cls, @@ -60,6 +123,8 @@ class OFTAdapter(WeightAdapterBase): blocks = v[0] rescale = v[1] alpha = v[2] + if alpha is None: + alpha = 0 dora_scale = v[3] blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype) diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 3d05fdab5..c3aaaee9b 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -20,7 +20,7 @@ import folder_paths import node_helpers from comfy.cli_args import args from comfy.comfy_types.node_typing import IO -from comfy.weight_adapter import adapters +from comfy.weight_adapter import adapters, adapter_maps def make_batch_extra_option_dict(d, indicies, full_size=None): @@ -39,13 +39,13 @@ def make_batch_extra_option_dict(d, indicies, full_size=None): class TrainSampler(comfy.samplers.Sampler): - - def __init__(self, loss_fn, optimizer, loss_callback=None, batch_size=1, total_steps=1, seed=0, training_dtype=torch.bfloat16): + def __init__(self, loss_fn, optimizer, loss_callback=None, batch_size=1, grad_acc=1, total_steps=1, seed=0, training_dtype=torch.bfloat16): self.loss_fn = loss_fn self.optimizer = optimizer self.loss_callback = loss_callback self.batch_size = batch_size self.total_steps = total_steps + self.grad_acc = grad_acc self.seed = seed self.training_dtype = training_dtype @@ -92,8 +92,9 @@ class TrainSampler(comfy.samplers.Sampler): self.loss_callback(loss.item()) pbar.set_postfix({"loss": f"{loss.item():.4f}"}) - self.optimizer.step() - self.optimizer.zero_grad() + if (i+1) % self.grad_acc == 0: + self.optimizer.step() + self.optimizer.zero_grad() torch.cuda.empty_cache() return torch.zeros_like(latent_image) @@ -419,6 +420,16 @@ class TrainLoraNode: "tooltip": "The batch size to use for training.", }, ), + "grad_accumulation_steps": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 1024, + "step": 1, + "tooltip": "The number of gradient accumulation steps to use for training.", + } + ), "steps": ( IO.INT, { @@ -478,6 +489,17 @@ class TrainLoraNode: ["bf16", "fp32"], {"default": "bf16", "tooltip": "The dtype to use for lora."}, ), + "algorithm": ( + list(adapter_maps.keys()), + {"default": list(adapter_maps.keys())[0], "tooltip": "The algorithm to use for training."}, + ), + "gradient_checkpointing": ( + IO.BOOLEAN, + { + "default": True, + "tooltip": "Use gradient checkpointing for training.", + } + ), "existing_lora": ( folder_paths.get_filename_list("loras") + ["[None]"], { @@ -501,6 +523,7 @@ class TrainLoraNode: positive, batch_size, steps, + grad_accumulation_steps, learning_rate, rank, optimizer, @@ -508,6 +531,8 @@ class TrainLoraNode: seed, training_dtype, lora_dtype, + algorithm, + gradient_checkpointing, existing_lora, ): mp = model.clone() @@ -558,10 +583,8 @@ class TrainLoraNode: if existing_adapter is not None: break else: - # If no existing adapter found, use LoRA - # We will add algo option in the future existing_adapter = None - adapter_cls = adapters[0] + adapter_cls = adapter_maps[algorithm] if existing_adapter is not None: train_adapter = existing_adapter.to_train().to(lora_dtype) @@ -615,8 +638,9 @@ class TrainLoraNode: criterion = torch.nn.SmoothL1Loss() # setup models - for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model): - patch(m) + if gradient_checkpointing: + for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model): + patch(m) mp.model.requires_grad_(False) comfy.model_management.load_models_gpu([mp], memory_required=1e20, force_full_load=True) @@ -629,7 +653,8 @@ class TrainLoraNode: optimizer, loss_callback=loss_callback, batch_size=batch_size, - total_steps=steps, + grad_acc=grad_accumulation_steps, + total_steps=steps*grad_accumulation_steps, seed=seed, training_dtype=dtype ) From 0ccc88b03fbe190135e24ac04612565f8f0756b4 Mon Sep 17 00:00:00 2001 From: honglyua Date: Fri, 25 Jul 2025 01:57:36 +0800 Subject: [PATCH 19/24] Support Iluvatar CoreX (#8585) * Support Iluvatar CoreX Co-authored-by: mingjiang.li --- README.md | 7 +++++++ comfy/model_management.py | 23 ++++++++++++++++++++++- 2 files changed, 29 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index d004364ee..a148623cd 100644 --- a/README.md +++ b/README.md @@ -294,6 +294,13 @@ For models compatible with Cambricon Extension for PyTorch (torch_mlu). Here's a 2. Next, install the PyTorch(torch_mlu) following the instructions on the [Installation](https://www.cambricon.com/docs/sdk_1.15.0/cambricon_pytorch_1.17.0/user_guide_1.9/index.html) 3. Launch ComfyUI by running `python main.py` +#### Iluvatar Corex + +For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step guide tailored to your platform and installation method: + +1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536) +2. Launch ComfyUI by running `python main.py` + # Running ```python main.py``` diff --git a/comfy/model_management.py b/comfy/model_management.py index e8b9b5c81..9add54ceb 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -128,6 +128,11 @@ try: except: mlu_available = False +try: + ixuca_available = hasattr(torch, "corex") +except: + ixuca_available = False + if args.cpu: cpu_state = CPUState.CPU @@ -151,6 +156,12 @@ def is_mlu(): return True return False +def is_ixuca(): + global ixuca_available + if ixuca_available: + return True + return False + def get_torch_device(): global directml_enabled global cpu_state @@ -289,7 +300,7 @@ try: if torch_version_numeric[0] >= 2: if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False: ENABLE_PYTORCH_ATTENTION = True - if is_intel_xpu() or is_ascend_npu() or is_mlu(): + if is_intel_xpu() or is_ascend_npu() or is_mlu() or is_ixuca(): if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: ENABLE_PYTORCH_ATTENTION = True except: @@ -1045,6 +1056,8 @@ def xformers_enabled(): return False if is_mlu(): return False + if is_ixuca(): + return False if directml_enabled: return False return XFORMERS_IS_AVAILABLE @@ -1080,6 +1093,8 @@ def pytorch_attention_flash_attention(): return True if is_amd(): return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention + if is_ixuca(): + return True return False def force_upcast_attention_dtype(): @@ -1205,6 +1220,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma if is_mlu(): return True + if is_ixuca(): + return True + if torch.version.hip: return True @@ -1268,6 +1286,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma if is_ascend_npu(): return True + if is_ixuca(): + return True + if is_amd(): arch = torch.cuda.get_device_properties(device).gcnArchName if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]): # RDNA2 and older don't support bf16 From d03ae077b4330f58e7caba53ff94e7fd58d0dc7d Mon Sep 17 00:00:00 2001 From: SHIVANSH GUPTA <121501003+shivansh-gupta4@users.noreply.github.com> Date: Thu, 24 Jul 2025 23:35:54 +0530 Subject: [PATCH 20/24] Added parameter required_frontend_version in the /system_stats API response (#8875) * Added the parameter required_frontend_version in the /system_stats api response * Update server.py * Created a function get_required_frontend_version and wrote tests for it * Refactored the function to return currently installed frontend pacakage version * Moved required_frontend to a new function and imported that in server.py * Corrected test cases using mocking techniques * Corrected files to comply with ruff formatting --- app/frontend_management.py | 47 +++++++++++++++++--- server.py | 2 + tests-unit/app_test/frontend_manager_test.py | 35 ++++++++++++++- 3 files changed, 77 insertions(+), 7 deletions(-) diff --git a/app/frontend_management.py b/app/frontend_management.py index 001ebbecb..0bee73685 100644 --- a/app/frontend_management.py +++ b/app/frontend_management.py @@ -29,18 +29,48 @@ def frontend_install_warning_message(): This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead. """.strip() +def parse_version(version: str) -> tuple[int, int, int]: + return tuple(map(int, version.split("."))) + +def is_valid_version(version: str) -> bool: + """Validate if a string is a valid semantic version (X.Y.Z format).""" + pattern = r"^(\d+)\.(\d+)\.(\d+)$" + return bool(re.match(pattern, version)) + +def get_installed_frontend_version(): + """Get the currently installed frontend package version.""" + frontend_version_str = version("comfyui-frontend-package") + return frontend_version_str + +def get_required_frontend_version(): + """Get the required frontend version from requirements.txt.""" + try: + with open(requirements_path, "r", encoding="utf-8") as f: + for line in f: + line = line.strip() + if line.startswith("comfyui-frontend-package=="): + version_str = line.split("==")[-1] + if not is_valid_version(version_str): + logging.error(f"Invalid version format in requirements.txt: {version_str}") + return None + return version_str + logging.error("comfyui-frontend-package not found in requirements.txt") + return None + except FileNotFoundError: + logging.error("requirements.txt not found. Cannot determine required frontend version.") + return None + except Exception as e: + logging.error(f"Error reading requirements.txt: {e}") + return None def check_frontend_version(): """Check if the frontend version is up to date.""" - def parse_version(version: str) -> tuple[int, int, int]: - return tuple(map(int, version.split("."))) - try: - frontend_version_str = version("comfyui-frontend-package") + frontend_version_str = get_installed_frontend_version() frontend_version = parse_version(frontend_version_str) - with open(requirements_path, "r", encoding="utf-8") as f: - required_frontend = parse_version(f.readline().split("=")[-1]) + required_frontend_str = get_required_frontend_version() + required_frontend = parse_version(required_frontend_str) if frontend_version < required_frontend: app.logger.log_startup_warning( f""" @@ -168,6 +198,11 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None: class FrontendManager: CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions") + @classmethod + def get_required_frontend_version(cls) -> str: + """Get the required frontend package version.""" + return get_required_frontend_version() + @classmethod def default_frontend_path(cls) -> str: try: diff --git a/server.py b/server.py index 71a58f0fa..f4de0079b 100644 --- a/server.py +++ b/server.py @@ -553,6 +553,7 @@ class PromptServer(): ram_free = comfy.model_management.get_free_memory(cpu_device) vram_total, torch_vram_total = comfy.model_management.get_total_memory(device, torch_total_too=True) vram_free, torch_vram_free = comfy.model_management.get_free_memory(device, torch_free_too=True) + required_frontend_version = FrontendManager.get_required_frontend_version() system_stats = { "system": { @@ -560,6 +561,7 @@ class PromptServer(): "ram_total": ram_total, "ram_free": ram_free, "comfyui_version": __version__, + "required_frontend_version": required_frontend_version, "python_version": sys.version, "pytorch_version": comfy.model_management.torch_version, "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded", diff --git a/tests-unit/app_test/frontend_manager_test.py b/tests-unit/app_test/frontend_manager_test.py index ce67df6c6..ce43ac564 100644 --- a/tests-unit/app_test/frontend_manager_test.py +++ b/tests-unit/app_test/frontend_manager_test.py @@ -1,7 +1,7 @@ import argparse import pytest from requests.exceptions import HTTPError -from unittest.mock import patch +from unittest.mock import patch, mock_open from app.frontend_management import ( FrontendManager, @@ -172,3 +172,36 @@ def test_init_frontend_fallback_on_error(): # Assert assert frontend_path == "/default/path" mock_check.assert_called_once() + + +def test_get_frontend_version(): + # Arrange + expected_version = "1.25.0" + mock_requirements_content = """torch +torchsde +comfyui-frontend-package==1.25.0 +other-package==1.0.0 +numpy""" + + # Act + with patch("builtins.open", mock_open(read_data=mock_requirements_content)): + version = FrontendManager.get_required_frontend_version() + + # Assert + assert version == expected_version + + +def test_get_frontend_version_invalid_semver(): + # Arrange + mock_requirements_content = """torch +torchsde +comfyui-frontend-package==1.29.3.75 +other-package==1.0.0 +numpy""" + + # Act + with patch("builtins.open", mock_open(read_data=mock_requirements_content)): + version = FrontendManager.get_required_frontend_version() + + # Assert + assert version is None From 69cb57b3426b08a82e7fb713b0b48c23725f3da7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 24 Jul 2025 12:06:25 -0700 Subject: [PATCH 21/24] Print xpu device name. (#9035) --- comfy/model_management.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 9add54ceb..232d363aa 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -392,6 +392,8 @@ def get_torch_device_name(device): except: allocator_backend = "" return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend) + elif device.type == "xpu": + return "{} {}".format(device, torch.xpu.get_device_name(device)) else: return "{}".format(device.type) elif is_intel_xpu(): From 4293e4da214f77a3fde97c15f0691307e61bc18d Mon Sep 17 00:00:00 2001 From: Eugene Fairley Date: Thu, 24 Jul 2025 17:59:19 -0700 Subject: [PATCH 22/24] Add WAN ATI support (#8874) * Add WAN ATI support * Fixes * Fix length * Remove extra functions * Fix * Fix * Ruff fix * Remove torch.no_grad * Add batch trajectory logic * Scale inputs before and after motion patch * Batch image/trajectory * Ruff fix * Clean up --- comfy/utils.py | 20 +++ comfy_extras/nodes_wan.py | 305 +++++++++++++++++++++++++++++++++++++- 2 files changed, 324 insertions(+), 1 deletion(-) diff --git a/comfy/utils.py b/comfy/utils.py index 9c076a0e0..fab28cf08 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -698,6 +698,26 @@ def resize_to_batch_size(tensor, batch_size): return output +def resize_list_to_batch_size(l, batch_size): + in_batch_size = len(l) + if in_batch_size == batch_size or in_batch_size == 0: + return l + + if batch_size <= 1: + return l[:batch_size] + + output = [] + if batch_size < in_batch_size: + scale = (in_batch_size - 1) / (batch_size - 1) + for i in range(batch_size): + output.append(l[min(round(i * scale), in_batch_size - 1)]) + else: + scale = in_batch_size / batch_size + for i in range(batch_size): + output.append(l[min(math.floor((i + 0.5) * scale), in_batch_size - 1)]) + + return output + def convert_sd_to(state_dict, dtype): keys = list(state_dict.keys()) for k in keys: diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index d6097a104..d71908f31 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -1,3 +1,4 @@ +import math import nodes import node_helpers import torch @@ -5,7 +6,9 @@ import comfy.model_management import comfy.utils import comfy.latent_formats import comfy.clip_vision - +import json +import numpy as np +from typing import Tuple class WanImageToVideo: @classmethod @@ -383,7 +386,307 @@ class WanPhantomSubjectToVideo: out_latent["samples"] = latent return (positive, cond2, negative, out_latent) +def parse_json_tracks(tracks): + """Parse JSON track data into a standardized format""" + tracks_data = [] + try: + # If tracks is a string, try to parse it as JSON + if isinstance(tracks, str): + parsed = json.loads(tracks.replace("'", '"')) + tracks_data.extend(parsed) + else: + # If tracks is a list of strings, parse each one + for track_str in tracks: + parsed = json.loads(track_str.replace("'", '"')) + tracks_data.append(parsed) + + # Check if we have a single track (dict with x,y) or a list of tracks + if tracks_data and isinstance(tracks_data[0], dict) and 'x' in tracks_data[0]: + # Single track detected, wrap it in a list + tracks_data = [tracks_data] + elif tracks_data and isinstance(tracks_data[0], list) and tracks_data[0] and isinstance(tracks_data[0][0], dict) and 'x' in tracks_data[0][0]: + # Already a list of tracks, nothing to do + pass + else: + # Unexpected format + pass + + except json.JSONDecodeError: + tracks_data = [] + return tracks_data + +def process_tracks(tracks_np: np.ndarray, frame_size: Tuple[int, int], num_frames, quant_multi: int = 8, **kwargs): + # tracks: shape [t, h, w, 3] => samples align with 24 fps, model trained with 16 fps. + # frame_size: tuple (W, H) + tracks = torch.from_numpy(tracks_np).float() + + if tracks.shape[1] == 121: + tracks = torch.permute(tracks, (1, 0, 2, 3)) + + tracks, visibles = tracks[..., :2], tracks[..., 2:3] + + short_edge = min(*frame_size) + + frame_center = torch.tensor([*frame_size]).type_as(tracks) / 2 + tracks = tracks - frame_center + + tracks = tracks / short_edge * 2 + + visibles = visibles * 2 - 1 + + trange = torch.linspace(-1, 1, tracks.shape[0]).view(-1, 1, 1, 1).expand(*visibles.shape) + + out_ = torch.cat([trange, tracks, visibles], dim=-1).view(121, -1, 4) + + out_0 = out_[:1] + + out_l = out_[1:] # 121 => 120 | 1 + a = 120 // math.gcd(120, num_frames) + b = num_frames // math.gcd(120, num_frames) + out_l = torch.repeat_interleave(out_l, b, dim=0)[1::a] # 120 => 120 * b => 120 * b / a == F + + final_result = torch.cat([out_0, out_l], dim=0) + + return final_result + +FIXED_LENGTH = 121 +def pad_pts(tr): + """Convert list of {x,y} to (FIXED_LENGTH,1,3) array, padding/truncating.""" + pts = np.array([[p['x'], p['y'], 1] for p in tr], dtype=np.float32) + n = pts.shape[0] + if n < FIXED_LENGTH: + pad = np.zeros((FIXED_LENGTH - n, 3), dtype=np.float32) + pts = np.vstack((pts, pad)) + else: + pts = pts[:FIXED_LENGTH] + return pts.reshape(FIXED_LENGTH, 1, 3) + +def ind_sel(target: torch.Tensor, ind: torch.Tensor, dim: int = 1): + """Index selection utility function""" + assert ( + len(ind.shape) > dim + ), "Index must have the target dim, but get dim: %d, ind shape: %s" % (dim, str(ind.shape)) + + target = target.expand( + *tuple( + [ind.shape[k] if target.shape[k] == 1 else -1 for k in range(dim)] + + [ + -1, + ] + * (len(target.shape) - dim) + ) + ) + + ind_pad = ind + + if len(target.shape) > dim + 1: + for _ in range(len(target.shape) - (dim + 1)): + ind_pad = ind_pad.unsqueeze(-1) + ind_pad = ind_pad.expand(*(-1,) * (dim + 1), *target.shape[(dim + 1) : :]) + + return torch.gather(target, dim=dim, index=ind_pad) + +def merge_final(vert_attr: torch.Tensor, weight: torch.Tensor, vert_assign: torch.Tensor): + """Merge vertex attributes with weights""" + target_dim = len(vert_assign.shape) - 1 + if len(vert_attr.shape) == 2: + assert vert_attr.shape[0] > vert_assign.max() + new_shape = [1] * target_dim + list(vert_attr.shape) + tensor = vert_attr.reshape(new_shape) + sel_attr = ind_sel(tensor, vert_assign.type(torch.long), dim=target_dim) + else: + assert vert_attr.shape[1] > vert_assign.max() + new_shape = [vert_attr.shape[0]] + [1] * (target_dim - 1) + list(vert_attr.shape[1:]) + tensor = vert_attr.reshape(new_shape) + sel_attr = ind_sel(tensor, vert_assign.type(torch.long), dim=target_dim) + + final_attr = torch.sum(sel_attr * weight.unsqueeze(-1), dim=-2) + return final_attr + + +def _patch_motion_single( + tracks: torch.FloatTensor, # (B, T, N, 4) + vid: torch.FloatTensor, # (C, T, H, W) + temperature: float, + vae_divide: tuple, + topk: int, +): + """Apply motion patching based on tracks""" + _, T, H, W = vid.shape + N = tracks.shape[2] + _, tracks_xy, visible = torch.split( + tracks, [1, 2, 1], dim=-1 + ) # (B, T, N, 2) | (B, T, N, 1) + tracks_n = tracks_xy / torch.tensor([W / min(H, W), H / min(H, W)], device=tracks_xy.device) + tracks_n = tracks_n.clamp(-1, 1) + visible = visible.clamp(0, 1) + + xx = torch.linspace(-W / min(H, W), W / min(H, W), W) + yy = torch.linspace(-H / min(H, W), H / min(H, W), H) + + grid = torch.stack(torch.meshgrid(yy, xx, indexing="ij")[::-1], dim=-1).to( + tracks_xy.device + ) + + tracks_pad = tracks_xy[:, 1:] + visible_pad = visible[:, 1:] + + visible_align = visible_pad.view(T - 1, 4, *visible_pad.shape[2:]).sum(1) + tracks_align = (tracks_pad * visible_pad).view(T - 1, 4, *tracks_pad.shape[2:]).sum( + 1 + ) / (visible_align + 1e-5) + dist_ = ( + (tracks_align[:, None, None] - grid[None, :, :, None]).pow(2).sum(-1) + ) # T, H, W, N + weight = torch.exp(-dist_ * temperature) * visible_align.clamp(0, 1).view( + T - 1, 1, 1, N + ) + vert_weight, vert_index = torch.topk( + weight, k=min(topk, weight.shape[-1]), dim=-1 + ) + + grid_mode = "bilinear" + point_feature = torch.nn.functional.grid_sample( + vid.permute(1, 0, 2, 3)[:1], + tracks_n[:, :1].type(vid.dtype), + mode=grid_mode, + padding_mode="zeros", + align_corners=False, + ) + point_feature = point_feature.squeeze(0).squeeze(1).permute(1, 0) # N, C=16 + + out_feature = merge_final(point_feature, vert_weight, vert_index).permute(3, 0, 1, 2) # T - 1, H, W, C => C, T - 1, H, W + out_weight = vert_weight.sum(-1) # T - 1, H, W + + # out feature -> already soft weighted + mix_feature = out_feature + vid[:, 1:] * (1 - out_weight.clamp(0, 1)) + + out_feature_full = torch.cat([vid[:, :1], mix_feature], dim=1) # C, T, H, W + out_mask_full = torch.cat([torch.ones_like(out_weight[:1]), out_weight], dim=0) # T, H, W + + return out_mask_full[None].expand(vae_divide[0], -1, -1, -1), out_feature_full + + +def patch_motion( + tracks: torch.FloatTensor, # (B, TB, T, N, 4) + vid: torch.FloatTensor, # (C, T, H, W) + temperature: float = 220.0, + vae_divide: tuple = (4, 16), + topk: int = 2, +): + B = len(tracks) + + # Process each batch separately + out_masks = [] + out_features = [] + + for b in range(B): + mask, feature = _patch_motion_single( + tracks[b], # (T, N, 4) + vid[b], # (C, T, H, W) + temperature, + vae_divide, + topk + ) + out_masks.append(mask) + out_features.append(feature) + + # Stack results: (B, C, T, H, W) + out_mask_full = torch.stack(out_masks, dim=0) + out_feature_full = torch.stack(out_features, dim=0) + + return out_mask_full, out_feature_full + +class WanTrackToVideo: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "vae": ("VAE", ), + "tracks": ("STRING", {"multiline": True, "default": "[]"}), + "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}), + "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + "temperature": ("FLOAT", {"default": 220.0, "min": 1.0, "max": 1000.0, "step": 0.1}), + "topk": ("INT", {"default": 2, "min": 1, "max": 10}), + "start_image": ("IMAGE", ), + }, + "optional": { + "clip_vision_output": ("CLIP_VISION_OUTPUT", ), + }} + + RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") + RETURN_NAMES = ("positive", "negative", "latent") + FUNCTION = "encode" + + CATEGORY = "conditioning/video_models" + + def encode(self, positive, negative, vae, tracks, width, height, length, batch_size, + temperature, topk, start_image=None, clip_vision_output=None): + + tracks_data = parse_json_tracks(tracks) + + if not tracks_data: + return WanImageToVideo().encode(positive, negative, vae, width, height, length, batch_size, start_image=start_image, clip_vision_output=clip_vision_output) + + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], + device=comfy.model_management.intermediate_device()) + + if isinstance(tracks_data[0][0], dict): + tracks_data = [tracks_data] + + processed_tracks = [] + for batch in tracks_data: + arrs = [] + for track in batch: + pts = pad_pts(track) + arrs.append(pts) + + tracks_np = np.stack(arrs, axis=0) + processed_tracks.append(process_tracks(tracks_np, (width, height), length - 1).unsqueeze(0)) + + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:batch_size].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + videos = torch.ones((start_image.shape[0], length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5 + for i in range(start_image.shape[0]): + videos[i, 0] = start_image[i] + + latent_videos = [] + videos = comfy.utils.resize_to_batch_size(videos, batch_size) + for i in range(batch_size): + latent_videos += [vae.encode(videos[i, :, :, :, :3])] + y = torch.cat(latent_videos, dim=0) + + # Scale latent since patch_motion is non-linear + y = comfy.latent_formats.Wan21().process_in(y) + + processed_tracks = comfy.utils.resize_list_to_batch_size(processed_tracks, batch_size) + res = patch_motion( + processed_tracks, y, temperature=temperature, topk=topk, vae_divide=(4, 16) + ) + + mask, concat_latent_image = res + concat_latent_image = comfy.latent_formats.Wan21().process_out(concat_latent_image) + mask = -mask + 1.0 # Invert mask to match expected format + positive = node_helpers.conditioning_set_values(positive, + {"concat_mask": mask, + "concat_latent_image": concat_latent_image}) + negative = node_helpers.conditioning_set_values(negative, + {"concat_mask": mask, + "concat_latent_image": concat_latent_image}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + out_latent = {} + out_latent["samples"] = latent + return (positive, negative, out_latent) + NODE_CLASS_MAPPINGS = { + "WanTrackToVideo": WanTrackToVideo, "WanImageToVideo": WanImageToVideo, "WanFunControlToVideo": WanFunControlToVideo, "WanFunInpaintToVideo": WanFunInpaintToVideo, From e6e5d33b351fc5ed8334d74dac77b283ecea8708 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 25 Jul 2025 01:58:28 -0700 Subject: [PATCH 23/24] Remove useless code. (#9041) This is only needed on old pytorch 2.0 and older. --- comfy/ldm/wan/vae.py | 13 ++----------- 1 file changed, 2 insertions(+), 11 deletions(-) diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index a8ebc5ec6..a83c6edfd 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -52,15 +52,6 @@ class RMS_norm(nn.Module): x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0) -class Upsample(nn.Upsample): - - def forward(self, x): - """ - Fix bfloat16 support for nearest neighbor interpolation. - """ - return super().forward(x.float()).type_as(x) - - class Resample(nn.Module): def __init__(self, dim, mode): @@ -73,11 +64,11 @@ class Resample(nn.Module): # layers if mode == 'upsample2d': self.resample = nn.Sequential( - Upsample(scale_factor=(2., 2.), mode='nearest-exact'), + nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'), ops.Conv2d(dim, dim // 2, 3, padding=1)) elif mode == 'upsample3d': self.resample = nn.Sequential( - Upsample(scale_factor=(2., 2.), mode='nearest-exact'), + nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'), ops.Conv2d(dim, dim // 2, 3, padding=1)) self.time_conv = CausalConv3d( dim, dim * 2, (3, 1, 1), padding=(1, 0, 0)) From 93bc2f8e4d5dace2328b861579df24f91684e27e Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 26 Jul 2025 01:24:23 +0800 Subject: [PATCH 24/24] Update template to 0.1.40 (#9048) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8f6a6d112..33a59b4be 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.23.4 -comfyui-workflow-templates==0.1.39 +comfyui-workflow-templates==0.1.40 comfyui-embedded-docs==0.2.4 torch torchsde