[single file] Cosmos (#11801)

* update

* update

* update docs

docs/source/en/api/pipelines/cosmos.md

@@ -24,6 +24,31 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers)
 
 </Tip>
 
+## Loading original format checkpoints
+
+Original format checkpoints that have not been converted to diffusers-expected format can be loaded using the `from_single_file` method.
+
+```python
+import torch
+from diffusers import Cosmos2TextToImagePipeline, CosmosTransformer3DModel
+
+model_id = "nvidia/Cosmos-Predict2-2B-Text2Image"
+transformer = CosmosTransformer3DModel.from_single_file(
+    "https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image/blob/main/model.pt",
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+pipe = Cosmos2TextToImagePipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+prompt = "A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess."
+negative_prompt = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."
+
+output = pipe(
+    prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
+).images[0]
+output.save("output.png")
+```
+
 ## CosmosTextToWorldPipeline
 
 [[autodoc]] CosmosTextToWorldPipeline

scripts/convert_cosmos_to_diffusers.py

@@ -95,7 +95,6 @@ TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0 = {
     "mlp.layer1": "ff.net.0.proj",
     "mlp.layer2": "ff.net.2",
     "x_embedder.proj.1": "patch_embed.proj",
-    # "extra_pos_embedder": "learnable_pos_embed",
     "final_layer.adaln_modulation.1": "norm_out.linear_1",
     "final_layer.adaln_modulation.2": "norm_out.linear_2",
     "final_layer.linear": "proj_out",

src/diffusers/loaders/single_file_model.py

@@ -31,6 +31,7 @@ from .single_file_utils import (
     convert_autoencoder_dc_checkpoint_to_diffusers,
     convert_chroma_transformer_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
+    convert_cosmos_transformer_checkpoint_to_diffusers,
     convert_flux_transformer_checkpoint_to_diffusers,
     convert_hidream_transformer_to_diffusers,
     convert_hunyuan_video_transformer_to_diffusers,
@@ -143,6 +144,10 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": convert_hidream_transformer_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "CosmosTransformer3DModel": {
+        "checkpoint_mapping_fn": convert_cosmos_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -127,6 +127,16 @@ CHECKPOINT_KEY_NAMES = {
     "wan": ["model.diffusion_model.head.modulation", "head.modulation"],
     "wan_vae": "decoder.middle.0.residual.0.gamma",
     "hidream": "double_stream_blocks.0.block.adaLN_modulation.1.bias",
+    "cosmos-1.0": [
+        "net.x_embedder.proj.1.weight",
+        "net.blocks.block1.blocks.0.block.attn.to_q.0.weight",
+        "net.extra_pos_embedder.pos_emb_h",
+    ],
+    "cosmos-2.0": [
+        "net.x_embedder.proj.1.weight",
+        "net.blocks.0.self_attn.q_proj.weight",
+        "net.pos_embedder.dim_spatial_range",
+    ],
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -193,6 +203,14 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "wan-t2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-14B-Diffusers"},
     "wan-i2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"},
     "hidream": {"pretrained_model_name_or_path": "HiDream-ai/HiDream-I1-Dev"},
+    "cosmos-1.0-t2w-7B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-7B-Text2World"},
+    "cosmos-1.0-t2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-14B-Text2World"},
+    "cosmos-1.0-v2w-7B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-7B-Video2World"},
+    "cosmos-1.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-14B-Video2World"},
+    "cosmos-2.0-t2i-2B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-2B-Text2Image"},
+    "cosmos-2.0-t2i-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Text2Image"},
+    "cosmos-2.0-v2w-2B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-2B-Video2World"},
+    "cosmos-2.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Video2World"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -704,11 +722,32 @@ def infer_diffusers_model_type(checkpoint):
             model_type = "wan-t2v-14B"
         else:
             model_type = "wan-i2v-14B"
+
     elif CHECKPOINT_KEY_NAMES["wan_vae"] in checkpoint:
         # All Wan models use the same VAE so we can use the same default model repo to fetch the config
         model_type = "wan-t2v-14B"
+
     elif CHECKPOINT_KEY_NAMES["hidream"] in checkpoint:
         model_type = "hidream"
+
+    elif all(key in checkpoint for key in CHECKPOINT_KEY_NAMES["cosmos-1.0"]):
+        x_embedder_shape = checkpoint[CHECKPOINT_KEY_NAMES["cosmos-1.0"][0]].shape
+        if x_embedder_shape[1] == 68:
+            model_type = "cosmos-1.0-t2w-7B" if x_embedder_shape[0] == 4096 else "cosmos-1.0-t2w-14B"
+        elif x_embedder_shape[1] == 72:
+            model_type = "cosmos-1.0-v2w-7B" if x_embedder_shape[0] == 4096 else "cosmos-1.0-v2w-14B"
+        else:
+            raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 1.0 model.")
+
+    elif all(key in checkpoint for key in CHECKPOINT_KEY_NAMES["cosmos-2.0"]):
+        x_embedder_shape = checkpoint[CHECKPOINT_KEY_NAMES["cosmos-2.0"][0]].shape
+        if x_embedder_shape[1] == 68:
+            model_type = "cosmos-2.0-t2i-2B" if x_embedder_shape[0] == 2048 else "cosmos-2.0-t2i-14B"
+        elif x_embedder_shape[1] == 72:
+            model_type = "cosmos-2.0-v2w-2B" if x_embedder_shape[0] == 2048 else "cosmos-2.0-v2w-14B"
+        else:
+            raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 2.0 model.")
+
     else:
         model_type = "v1"
 
@@ -3479,3 +3518,116 @@ def convert_chroma_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
     converted_state_dict["proj_out.bias"] = checkpoint.pop("final_layer.linear.bias")
 
     return converted_state_dict
+
+
+def convert_cosmos_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys())}
+
+    def remove_keys_(key: str, state_dict):
+        state_dict.pop(key)
+
+    def rename_transformer_blocks_(key: str, state_dict):
+        block_index = int(key.split(".")[1].removeprefix("block"))
+        new_key = key
+        old_prefix = f"blocks.block{block_index}"
+        new_prefix = f"transformer_blocks.{block_index}"
+        new_key = new_prefix + new_key.removeprefix(old_prefix)
+        state_dict[new_key] = state_dict.pop(key)
+
+    TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0 = {
+        "t_embedder.1": "time_embed.t_embedder",
+        "affline_norm": "time_embed.norm",
+        ".blocks.0.block.attn": ".attn1",
+        ".blocks.1.block.attn": ".attn2",
+        ".blocks.2.block": ".ff",
+        ".blocks.0.adaLN_modulation.1": ".norm1.linear_1",
+        ".blocks.0.adaLN_modulation.2": ".norm1.linear_2",
+        ".blocks.1.adaLN_modulation.1": ".norm2.linear_1",
+        ".blocks.1.adaLN_modulation.2": ".norm2.linear_2",
+        ".blocks.2.adaLN_modulation.1": ".norm3.linear_1",
+        ".blocks.2.adaLN_modulation.2": ".norm3.linear_2",
+        "to_q.0": "to_q",
+        "to_q.1": "norm_q",
+        "to_k.0": "to_k",
+        "to_k.1": "norm_k",
+        "to_v.0": "to_v",
+        "layer1": "net.0.proj",
+        "layer2": "net.2",
+        "proj.1": "proj",
+        "x_embedder": "patch_embed",
+        "extra_pos_embedder": "learnable_pos_embed",
+        "final_layer.adaLN_modulation.1": "norm_out.linear_1",
+        "final_layer.adaLN_modulation.2": "norm_out.linear_2",
+        "final_layer.linear": "proj_out",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0 = {
+        "blocks.block": rename_transformer_blocks_,
+        "logvar.0.freqs": remove_keys_,
+        "logvar.0.phases": remove_keys_,
+        "logvar.1.weight": remove_keys_,
+        "pos_embedder.seq": remove_keys_,
+    }
+
+    TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0 = {
+        "t_embedder.1": "time_embed.t_embedder",
+        "t_embedding_norm": "time_embed.norm",
+        "blocks": "transformer_blocks",
+        "adaln_modulation_self_attn.1": "norm1.linear_1",
+        "adaln_modulation_self_attn.2": "norm1.linear_2",
+        "adaln_modulation_cross_attn.1": "norm2.linear_1",
+        "adaln_modulation_cross_attn.2": "norm2.linear_2",
+        "adaln_modulation_mlp.1": "norm3.linear_1",
+        "adaln_modulation_mlp.2": "norm3.linear_2",
+        "self_attn": "attn1",
+        "cross_attn": "attn2",
+        "q_proj": "to_q",
+        "k_proj": "to_k",
+        "v_proj": "to_v",
+        "output_proj": "to_out.0",
+        "q_norm": "norm_q",
+        "k_norm": "norm_k",
+        "mlp.layer1": "ff.net.0.proj",
+        "mlp.layer2": "ff.net.2",
+        "x_embedder.proj.1": "patch_embed.proj",
+        "final_layer.adaln_modulation.1": "norm_out.linear_1",
+        "final_layer.adaln_modulation.2": "norm_out.linear_2",
+        "final_layer.linear": "proj_out",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0 = {
+        "accum_video_sample_counter": remove_keys_,
+        "accum_image_sample_counter": remove_keys_,
+        "accum_iteration": remove_keys_,
+        "accum_train_in_hours": remove_keys_,
+        "pos_embedder.seq": remove_keys_,
+        "pos_embedder.dim_spatial_range": remove_keys_,
+        "pos_embedder.dim_temporal_range": remove_keys_,
+        "_extra_state": remove_keys_,
+    }
+
+    PREFIX_KEY = "net."
+    if "net.blocks.block1.blocks.0.block.attn.to_q.0.weight" in checkpoint:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0
+    else:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0
+
+    state_dict_keys = list(converted_state_dict.keys())
+    for key in state_dict_keys:
+        new_key = key[:]
+        if new_key.startswith(PREFIX_KEY):
+            new_key = new_key.removeprefix(PREFIX_KEY)
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    state_dict_keys = list(converted_state_dict.keys())
+    for key in state_dict_keys:
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict

src/diffusers/models/transformers/transformer_cosmos.py

@@ -20,6 +20,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin
 from ...utils import is_torchvision_available
 from ..attention import FeedForward
 from ..attention_processor import Attention
@@ -377,7 +378,7 @@ class CosmosLearnablePositionalEmbed(nn.Module):
         return (emb / norm).type_as(hidden_states)
 
 
-class CosmosTransformer3DModel(ModelMixin, ConfigMixin):
+class CosmosTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     r"""
     A Transformer model for video-like data used in [Cosmos](https://github.com/NVIDIA/Cosmos).