Add AudioLDM (#2232)

* Add AudioLDM

* up

* add vocoder

* start unet

* unconditional unet

* clap, vocoder and vae

* clean-up: conversion scripts

* fix: conversion script token_type_ids

* clean-up: pipeline docstring

* tests: from SD

* clean-up: cpu offload vocoder instead of safety checker

* feat: adapt tests to audioldm

* feat: add docs

* clean-up: amend pipeline docstrings

* clean-up: make style

* clean-up: make fix-copies

* fix: add doc path to toctree

* clean-up: args for conversion script

* clean-up: paths to checkpoints

* fix: use conditional unet

* clean-up: make style

* fix: type hints for UNet

* clean-up: docstring for UNet

* clean-up: make style

* clean-up: remove duplicate in docstring

* clean-up: make style

* clean-up: make fix-copies

* clean-up: move imports to start in code snippet

* fix: pass cross_attention_dim as a list/tuple to unet

* clean-up: make fix-copies

* fix: update checkpoint path

* fix: unet cross_attention_dim in tests

* film embeddings -> class embeddings

* Apply suggestions from code review
Co-authored-by: Will Berman <wlbberman@gmail.com>

* fix: unet film embed to use existing args

* fix: unet tests to use existing args

* fix: make style

* fix: transformers import and version in init

* clean-up: make style

* Revert "clean-up: make style"

This reverts commit 5d6d1f8b324f5583e7805dc01e2c86e493660d66.

* clean-up: make style

* clean-up: use pipeline tester mixin tests where poss

* clean-up: skip attn slicing test

* fix: add torch dtype to docs

* fix: remove conversion script out of src

* fix: remove .detach from 1d waveform

* fix: reduce default num inf steps

* fix: swap height/width -> audio_length_in_s

* clean-up: make style

* fix: remove nightly tests

* fix: imports in conversion script

* clean-up: slim-down to two slow tests

* clean-up: slim-down fast tests

* fix: batch consistent tests

* clean-up: make style

* clean-up: remove vae slicing fast test

* clean-up: propagate changes to doc

* fix: increase test tol to 1e-2

* clean-up: finish docs

* clean-up: make style

* feat: vocoder / VAE compatibility check

* feat: possibly expand / cut audio waveform

* fix: pipeline call signature test

* fix: slow tests output len

* clean-up: make style

* make style

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: William Berman <WLBberman@gmail.com>

docs/source/en/_toctree.yml

@@ -134,6 +134,8 @@
       title: AltDiffusion
     - local: api/pipelines/audio_diffusion
       title: Audio Diffusion
+    - local: api/pipelines/audioldm
+      title: AudioLDM
     - local: api/pipelines/cycle_diffusion
       title: Cycle Diffusion
     - local: api/pipelines/dance_diffusion

docs/source/en/api/pipelines/audioldm.mdx

+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# AudioLDM
+
+## Overview
+
+AudioLDM was proposed in [AudioLDM: Text-to-Audio Generation with Latent Diffusion Models](https://arxiv.org/abs/2301.12503) by Haohe Liu et al.
+
+Inspired by [Stable Diffusion](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/overview), AudioLDM
+is a text-to-audio _latent diffusion model (LDM)_ that learns continuous audio representations from [CLAP](https://huggingface.co/docs/transformers/main/model_doc/clap)
+latents. AudioLDM takes a text prompt as input and predicts the corresponding audio. It can generate text-conditional
+sound effects, human speech and music.
+
+This pipeline was contributed by [sanchit-gandhi](https://huggingface.co/sanchit-gandhi). The original codebase can be found [here](https://github.com/haoheliu/AudioLDM).
+
+## Text-to-Audio
+
+The [`AudioLDMPipeline`] can be used to load pre-trained weights from [cvssp/audioldm](https://huggingface.co/cvssp/audioldm) and generate text-conditional audio outputs:
+
+```python
+from diffusers import AudioLDMPipeline
+import torch
+import scipy
+
+repo_id = "cvssp/audioldm"
+pipe = AudioLDMPipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
+pipe = pipe.to("cuda")
+
+prompt = "Techno music with a strong, upbeat tempo and high melodic riffs"
+audio = pipe(prompt, num_inference_steps=10, audio_length_in_s=5.0).audios[0]
+
+# save the audio sample as a .wav file
+scipy.io.wavfile.write("techno.wav", rate=16000, data=audio)
+```
+
+### Tips
+
+Prompts:
+* Descriptive prompt inputs work best: you can use adjectives to describe the sound (e.g. "high quality" or "clear") and make the prompt context specific (e.g., "water stream in a forest" instead of "stream").
+* It's best to use general terms like 'cat' or 'dog' instead of specific names or abstract objects that the model may not be familiar with.
+
+Inference:
+* The _quality_ of the predicted audio sample can be controlled by the `num_inference_steps` argument: higher steps give higher quality audio at the expense of slower inference.
+* The _length_ of the predicted audio sample can be controlled by varying the `audio_length_in_s` argument.
+
+### How to load and use different schedulers
+
+The AudioLDM pipeline uses [`DDIMScheduler`] scheduler by default. But `diffusers` provides many other schedulers
+that can be used with the AudioLDM pipeline such as [`PNDMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], 
+[`EulerAncestralDiscreteScheduler`] etc. We recommend using the [`DPMSolverMultistepScheduler`] as it's currently the fastest
+scheduler there is.
+
+To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`]
+method, or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the
+[`DPMSolverMultistepScheduler`], you can do the following:
+
+```python
+>>> from diffusers import AudioLDMPipeline, DPMSolverMultistepScheduler
+>>> import torch
+
+>>> pipeline = AudioLDMPipeline.from_pretrained("cvssp/audioldm", torch_dtype=torch.float16)
+>>> pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
+
+>>> # or
+>>> dpm_scheduler = DPMSolverMultistepScheduler.from_pretrained("cvssp/audioldm", subfolder="scheduler")
+>>> pipeline = AudioLDMPipeline.from_pretrained("cvssp/audioldm", scheduler=dpm_scheduler, torch_dtype=torch.float16)
+```
+
+## AudioLDMPipeline
+[[autodoc]] AudioLDMPipeline
+	- all
+	- __call__

scripts/convert_original_audioldm_to_diffusers.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Conversion script for the AudioLDM checkpoints."""
+
+import argparse
+import re
+
+import torch
+from transformers import (
+    AutoTokenizer,
+    ClapTextConfig,
+    ClapTextModelWithProjection,
+    SpeechT5HifiGan,
+    SpeechT5HifiGanConfig,
+)
+
+from diffusers import (
+    AudioLDMPipeline,
+    AutoencoderKL,
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    HeunDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    UNet2DConditionModel,
+)
+from diffusers.utils import is_omegaconf_available, is_safetensors_available
+from diffusers.utils.import_utils import BACKENDS_MAPPING
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.shave_segments
+def shave_segments(path, n_shave_prefix_segments=1):
+    """
+    Removes segments. Positive values shave the first segments, negative shave the last segments.
+    """
+    if n_shave_prefix_segments >= 0:
+        return ".".join(path.split(".")[n_shave_prefix_segments:])
+    else:
+        return ".".join(path.split(".")[:n_shave_prefix_segments])
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_resnet_paths
+def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item.replace("in_layers.0", "norm1")
+        new_item = new_item.replace("in_layers.2", "conv1")
+
+        new_item = new_item.replace("out_layers.0", "norm2")
+        new_item = new_item.replace("out_layers.3", "conv2")
+
+        new_item = new_item.replace("emb_layers.1", "time_emb_proj")
+        new_item = new_item.replace("skip_connection", "conv_shortcut")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_vae_resnet_paths
+def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("nin_shortcut", "conv_shortcut")
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_attention_paths
+def renew_attention_paths(old_list):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        #         new_item = new_item.replace('norm.weight', 'group_norm.weight')
+        #         new_item = new_item.replace('norm.bias', 'group_norm.bias')
+
+        #         new_item = new_item.replace('proj_out.weight', 'proj_attn.weight')
+        #         new_item = new_item.replace('proj_out.bias', 'proj_attn.bias')
+
+        #         new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_vae_attention_paths
+def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("norm.weight", "group_norm.weight")
+        new_item = new_item.replace("norm.bias", "group_norm.bias")
+
+        new_item = new_item.replace("q.weight", "query.weight")
+        new_item = new_item.replace("q.bias", "query.bias")
+
+        new_item = new_item.replace("k.weight", "key.weight")
+        new_item = new_item.replace("k.bias", "key.bias")
+
+        new_item = new_item.replace("v.weight", "value.weight")
+        new_item = new_item.replace("v.bias", "value.bias")
+
+        new_item = new_item.replace("proj_out.weight", "proj_attn.weight")
+        new_item = new_item.replace("proj_out.bias", "proj_attn.bias")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.assign_to_checkpoint
+def assign_to_checkpoint(
+    paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None
+):
+    """
+    This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
+    attention layers, and takes into account additional replacements that may arise.
+
+    Assigns the weights to the new checkpoint.
+    """
+    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
+
+    # Splits the attention layers into three variables.
+    if attention_paths_to_split is not None:
+        for path, path_map in attention_paths_to_split.items():
+            old_tensor = old_checkpoint[path]
+            channels = old_tensor.shape[0] // 3
+
+            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
+
+            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3
+
+            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
+            query, key, value = old_tensor.split(channels // num_heads, dim=1)
+
+            checkpoint[path_map["query"]] = query.reshape(target_shape)
+            checkpoint[path_map["key"]] = key.reshape(target_shape)
+            checkpoint[path_map["value"]] = value.reshape(target_shape)
+
+    for path in paths:
+        new_path = path["new"]
+
+        # These have already been assigned
+        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
+            continue
+
+        # Global renaming happens here
+        new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
+        new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
+        new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")
+
+        if additional_replacements is not None:
+            for replacement in additional_replacements:
+                new_path = new_path.replace(replacement["old"], replacement["new"])
+
+        # proj_attn.weight has to be converted from conv 1D to linear
+        if "proj_attn.weight" in new_path:
+            checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
+        else:
+            checkpoint[new_path] = old_checkpoint[path["old"]]
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.conv_attn_to_linear
+def conv_attn_to_linear(checkpoint):
+    keys = list(checkpoint.keys())
+    attn_keys = ["query.weight", "key.weight", "value.weight"]
+    for key in keys:
+        if ".".join(key.split(".")[-2:]) in attn_keys:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0, 0]
+        elif "proj_attn.weight" in key:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def create_unet_diffusers_config(original_config, image_size: int):
+    """
+    Creates a UNet config for diffusers based on the config of the original AudioLDM model.
+    """
+    unet_params = original_config.model.params.unet_config.params
+    vae_params = original_config.model.params.first_stage_config.params.ddconfig
+
+    block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult]
+
+    down_block_types = []
+    resolution = 1
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnDownBlock2D" if resolution in unet_params.attention_resolutions else "DownBlock2D"
+        down_block_types.append(block_type)
+        if i != len(block_out_channels) - 1:
+            resolution *= 2
+
+    up_block_types = []
+    for i in range(len(block_out_channels)):
+        block_type = "CrossAttnUpBlock2D" if resolution in unet_params.attention_resolutions else "UpBlock2D"
+        up_block_types.append(block_type)
+        resolution //= 2
+
+    vae_scale_factor = 2 ** (len(vae_params.ch_mult) - 1)
+
+    cross_attention_dim = (
+        unet_params.cross_attention_dim if "cross_attention_dim" in unet_params else block_out_channels
+    )
+
+    class_embed_type = "simple_projection" if "extra_film_condition_dim" in unet_params else None
+    projection_class_embeddings_input_dim = (
+        unet_params.extra_film_condition_dim if "extra_film_condition_dim" in unet_params else None
+    )
+    class_embeddings_concat = unet_params.extra_film_use_concat if "extra_film_use_concat" in unet_params else None
+
+    config = dict(
+        sample_size=image_size // vae_scale_factor,
+        in_channels=unet_params.in_channels,
+        out_channels=unet_params.out_channels,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        layers_per_block=unet_params.num_res_blocks,
+        cross_attention_dim=cross_attention_dim,
+        class_embed_type=class_embed_type,
+        projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+        class_embeddings_concat=class_embeddings_concat,
+    )
+
+    return config
+
+
+# Adapted from diffusers.pipelines.stable_diffusion.convert_from_ckpt.create_vae_diffusers_config
+def create_vae_diffusers_config(original_config, checkpoint, image_size: int):
+    """
+    Creates a VAE config for diffusers based on the config of the original AudioLDM model. Compared to the original
+    Stable Diffusion conversion, this function passes a *learnt* VAE scaling factor to the diffusers VAE.
+    """
+    vae_params = original_config.model.params.first_stage_config.params.ddconfig
+    _ = original_config.model.params.first_stage_config.params.embed_dim
+
+    block_out_channels = [vae_params.ch * mult for mult in vae_params.ch_mult]
+    down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
+    up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
+
+    scaling_factor = checkpoint["scale_factor"] if "scale_by_std" in original_config.model.params else 0.18215
+
+    config = dict(
+        sample_size=image_size,
+        in_channels=vae_params.in_channels,
+        out_channels=vae_params.out_ch,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        latent_channels=vae_params.z_channels,
+        layers_per_block=vae_params.num_res_blocks,
+        scaling_factor=float(scaling_factor),
+    )
+    return config
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.create_diffusers_schedular
+def create_diffusers_schedular(original_config):
+    schedular = DDIMScheduler(
+        num_train_timesteps=original_config.model.params.timesteps,
+        beta_start=original_config.model.params.linear_start,
+        beta_end=original_config.model.params.linear_end,
+        beta_schedule="scaled_linear",
+    )
+    return schedular
+
+
+# Adapted from diffusers.pipelines.stable_diffusion.convert_from_ckpt.convert_ldm_unet_checkpoint
+def convert_ldm_unet_checkpoint(checkpoint, config, path=None, extract_ema=False):
+    """
+    Takes a state dict and a config, and returns a converted checkpoint. Compared to the original Stable Diffusion
+    conversion, this function additionally converts the learnt film embedding linear layer.
+    """
+
+    # extract state_dict for UNet
+    unet_state_dict = {}
+    keys = list(checkpoint.keys())
+
+    unet_key = "model.diffusion_model."
+    # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
+    if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
+        print(f"Checkpoint {path} has both EMA and non-EMA weights.")
+        print(
+            "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
+            " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
+        )
+        for key in keys:
+            if key.startswith("model.diffusion_model"):
+                flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
+                unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
+    else:
+        if sum(k.startswith("model_ema") for k in keys) > 100:
+            print(
+                "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
+                " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
+            )
+
+        for key in keys:
+            if key.startswith(unet_key):
+                unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
+
+    new_checkpoint = {}
+
+    new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
+    new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"]
+    new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
+    new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]
+
+    new_checkpoint["class_embedding.weight"] = unet_state_dict["film_emb.weight"]
+    new_checkpoint["class_embedding.bias"] = unet_state_dict["film_emb.bias"]
+
+    new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
+    new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
+
+    new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
+    new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
+    new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
+    new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]
+
+    # Retrieves the keys for the input blocks only
+    num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
+    input_blocks = {
+        layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
+        for layer_id in range(num_input_blocks)
+    }
+
+    # Retrieves the keys for the middle blocks only
+    num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
+    middle_blocks = {
+        layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
+        for layer_id in range(num_middle_blocks)
+    }
+
+    # Retrieves the keys for the output blocks only
+    num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
+    output_blocks = {
+        layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
+        for layer_id in range(num_output_blocks)
+    }
+
+    for i in range(1, num_input_blocks):
+        block_id = (i - 1) // (config["layers_per_block"] + 1)
+        layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+        resnets = [
+            key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+        ]
+        attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+
+        if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.op.weight"
+            )
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
+                f"input_blocks.{i}.0.op.bias"
+            )
+
+        paths = renew_resnet_paths(resnets)
+        meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
+        assign_to_checkpoint(
+            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+        )
+
+        if len(attentions):
+            paths = renew_attention_paths(attentions)
+            meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+    resnet_0 = middle_blocks[0]
+    attentions = middle_blocks[1]
+    resnet_1 = middle_blocks[2]
+
+    resnet_0_paths = renew_resnet_paths(resnet_0)
+    assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config)
+
+    resnet_1_paths = renew_resnet_paths(resnet_1)
+    assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config)
+
+    attentions_paths = renew_attention_paths(attentions)
+    meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(
+        attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+    )
+
+    for i in range(num_output_blocks):
+        block_id = i // (config["layers_per_block"] + 1)
+        layer_in_block_id = i % (config["layers_per_block"] + 1)
+        output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
+        output_block_list = {}
+
+        for layer in output_block_layers:
+            layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
+            if layer_id in output_block_list:
+                output_block_list[layer_id].append(layer_name)
+            else:
+                output_block_list[layer_id] = [layer_name]
+
+        if len(output_block_list) > 1:
+            resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
+            attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]
+
+            resnet_0_paths = renew_resnet_paths(resnets)
+            paths = renew_resnet_paths(resnets)
+
+            meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+            output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
+            if ["conv.bias", "conv.weight"] in output_block_list.values():
+                index = list(output_block_list.values()).index(["conv.bias", "conv.weight"])
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.weight"
+                ]
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.bias"
+                ]
+
+                # Clear attentions as they have been attributed above.
+                if len(attentions) == 2:
+                    attentions = []
+
+            if len(attentions):
+                paths = renew_attention_paths(attentions)
+                meta_path = {
+                    "old": f"output_blocks.{i}.1",
+                    "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
+                }
+                assign_to_checkpoint(
+                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+                )
+        else:
+            resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
+            for path in resnet_0_paths:
+                old_path = ".".join(["output_blocks", str(i), path["old"]])
+                new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])
+
+                new_checkpoint[new_path] = unet_state_dict[old_path]
+
+    return new_checkpoint
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.convert_ldm_vae_checkpoint
+def convert_ldm_vae_checkpoint(checkpoint, config):
+    # extract state dict for VAE
+    vae_state_dict = {}
+    vae_key = "first_stage_model."
+    keys = list(checkpoint.keys())
+    for key in keys:
+        if key.startswith(vae_key):
+            vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)
+
+    new_checkpoint = {}
+
+    new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
+    new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
+    new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
+    new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
+    new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
+    new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]
+
+    new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
+    new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
+    new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
+    new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
+    new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
+    new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]
+
+    new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
+    new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
+    new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
+    new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]
+
+    # Retrieves the keys for the encoder down blocks only
+    num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
+    down_blocks = {
+        layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+    }
+
+    # Retrieves the keys for the decoder up blocks only
+    num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
+    up_blocks = {
+        layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+    }
+
+    for i in range(num_down_blocks):
+        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+
+        if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.weight"
+            )
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.bias"
+            )
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+    conv_attn_to_linear(new_checkpoint)
+
+    for i in range(num_up_blocks):
+        block_id = num_up_blocks - 1 - i
+        resnets = [
+            key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+        ]
+
+        if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.weight"
+            ]
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.bias"
+            ]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+    mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+    conv_attn_to_linear(new_checkpoint)
+    return new_checkpoint
+
+
+CLAP_KEYS_TO_MODIFY_MAPPING = {
+    "text_branch": "text_model",
+    "attn": "attention.self",
+    "self.proj": "output.dense",
+    "attention.self_mask": "attn_mask",
+    "mlp.fc1": "intermediate.dense",
+    "mlp.fc2": "output.dense",
+    "norm1": "layernorm_before",
+    "norm2": "layernorm_after",
+    "bn0": "batch_norm",
+}
+
+CLAP_KEYS_TO_IGNORE = ["text_transform"]
+
+CLAP_EXPECTED_MISSING_KEYS = ["text_model.embeddings.token_type_ids"]
+
+
+def convert_open_clap_checkpoint(checkpoint):
+    """
+    Takes a state dict and returns a converted CLAP checkpoint.
+    """
+    # extract state dict for CLAP text embedding model, discarding the audio component
+    model_state_dict = {}
+    model_key = "cond_stage_model.model.text_"
+    keys = list(checkpoint.keys())
+    for key in keys:
+        if key.startswith(model_key):
+            model_state_dict[key.replace(model_key, "text_")] = checkpoint.get(key)
+
+    new_checkpoint = {}
+
+    sequential_layers_pattern = r".*sequential.(\d+).*"
+    text_projection_pattern = r".*_projection.(\d+).*"
+
+    for key, value in model_state_dict.items():
+        # check if key should be ignored in mapping
+        if key.split(".")[0] in CLAP_KEYS_TO_IGNORE:
+            continue
+
+        # check if any key needs to be modified
+        for key_to_modify, new_key in CLAP_KEYS_TO_MODIFY_MAPPING.items():
+            if key_to_modify in key:
+                key = key.replace(key_to_modify, new_key)
+
+        if re.match(sequential_layers_pattern, key):
+            # replace sequential layers with list
+            sequential_layer = re.match(sequential_layers_pattern, key).group(1)
+
+            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer)//3}.linear.")
+        elif re.match(text_projection_pattern, key):
+            projecton_layer = int(re.match(text_projection_pattern, key).group(1))
+
+            # Because in CLAP they use `nn.Sequential`...
+            transformers_projection_layer = 1 if projecton_layer == 0 else 2
+
+            key = key.replace(f"_projection.{projecton_layer}.", f"_projection.linear{transformers_projection_layer}.")
+
+        if "audio" and "qkv" in key:
+            # split qkv into query key and value
+            mixed_qkv = value
+            qkv_dim = mixed_qkv.size(0) // 3
+
+            query_layer = mixed_qkv[:qkv_dim]
+            key_layer = mixed_qkv[qkv_dim : qkv_dim * 2]
+            value_layer = mixed_qkv[qkv_dim * 2 :]
+
+            new_checkpoint[key.replace("qkv", "query")] = query_layer
+            new_checkpoint[key.replace("qkv", "key")] = key_layer
+            new_checkpoint[key.replace("qkv", "value")] = value_layer
+        else:
+            new_checkpoint[key] = value
+
+    return new_checkpoint
+
+
+def create_transformers_vocoder_config(original_config):
+    """
+    Creates a config for transformers SpeechT5HifiGan based on the config of the vocoder model.
+    """
+    vocoder_params = original_config.model.params.vocoder_config.params
+
+    config = dict(
+        model_in_dim=vocoder_params.num_mels,
+        sampling_rate=vocoder_params.sampling_rate,
+        upsample_initial_channel=vocoder_params.upsample_initial_channel,
+        upsample_rates=list(vocoder_params.upsample_rates),
+        upsample_kernel_sizes=list(vocoder_params.upsample_kernel_sizes),
+        resblock_kernel_sizes=list(vocoder_params.resblock_kernel_sizes),
+        resblock_dilation_sizes=[
+            list(resblock_dilation) for resblock_dilation in vocoder_params.resblock_dilation_sizes
+        ],
+        normalize_before=False,
+    )
+
+    return config
+
+
+def convert_hifigan_checkpoint(checkpoint, config):
+    """
+    Takes a state dict and config, and returns a converted HiFiGAN vocoder checkpoint.
+    """
+    # extract state dict for vocoder
+    vocoder_state_dict = {}
+    vocoder_key = "first_stage_model.vocoder."
+    keys = list(checkpoint.keys())
+    for key in keys:
+        if key.startswith(vocoder_key):
+            vocoder_state_dict[key.replace(vocoder_key, "")] = checkpoint.get(key)
+
+    # fix upsampler keys, everything else is correct already
+    for i in range(len(config.upsample_rates)):
+        vocoder_state_dict[f"upsampler.{i}.weight"] = vocoder_state_dict.pop(f"ups.{i}.weight")
+        vocoder_state_dict[f"upsampler.{i}.bias"] = vocoder_state_dict.pop(f"ups.{i}.bias")
+
+    if not config.normalize_before:
+        # if we don't set normalize_before then these variables are unused, so we set them to their initialised values
+        vocoder_state_dict["mean"] = torch.zeros(config.model_in_dim)
+        vocoder_state_dict["scale"] = torch.ones(config.model_in_dim)
+
+    return vocoder_state_dict
+
+
+# Adapted from https://huggingface.co/spaces/haoheliu/audioldm-text-to-audio-generation/blob/84a0384742a22bd80c44e903e241f0623e874f1d/audioldm/utils.py#L72-L73
+DEFAULT_CONFIG = {
+    "model": {
+        "params": {
+            "linear_start": 0.0015,
+            "linear_end": 0.0195,
+            "timesteps": 1000,
+            "channels": 8,
+            "scale_by_std": True,
+            "unet_config": {
+                "target": "audioldm.latent_diffusion.openaimodel.UNetModel",
+                "params": {
+                    "extra_film_condition_dim": 512,
+                    "extra_film_use_concat": True,
+                    "in_channels": 8,
+                    "out_channels": 8,
+                    "model_channels": 128,
+                    "attention_resolutions": [8, 4, 2],
+                    "num_res_blocks": 2,
+                    "channel_mult": [1, 2, 3, 5],
+                    "num_head_channels": 32,
+                },
+            },
+            "first_stage_config": {
+                "target": "audioldm.variational_autoencoder.autoencoder.AutoencoderKL",
+                "params": {
+                    "embed_dim": 8,
+                    "ddconfig": {
+                        "z_channels": 8,
+                        "resolution": 256,
+                        "in_channels": 1,
+                        "out_ch": 1,
+                        "ch": 128,
+                        "ch_mult": [1, 2, 4],
+                        "num_res_blocks": 2,
+                    },
+                },
+            },
+            "vocoder_config": {
+                "target": "audioldm.first_stage_model.vocoder",
+                "params": {
+                    "upsample_rates": [5, 4, 2, 2, 2],
+                    "upsample_kernel_sizes": [16, 16, 8, 4, 4],
+                    "upsample_initial_channel": 1024,
+                    "resblock_kernel_sizes": [3, 7, 11],
+                    "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    "num_mels": 64,
+                    "sampling_rate": 16000,
+                },
+            },
+        },
+    },
+}
+
+
+def load_pipeline_from_original_audioldm_ckpt(
+    checkpoint_path: str,
+    original_config_file: str = None,
+    image_size: int = 512,
+    prediction_type: str = None,
+    extract_ema: bool = False,
+    scheduler_type: str = "ddim",
+    num_in_channels: int = None,
+    device: str = None,
+    from_safetensors: bool = False,
+) -> AudioLDMPipeline:
+    """
+    Load an AudioLDM pipeline object from a `.ckpt`/`.safetensors` file and (ideally) a `.yaml` config file.
+
+    Although many of the arguments can be automatically inferred, some of these rely on brittle checks against the
+    global step count, which will likely fail for models that have undergone further fine-tuning. Therefore, it is
+    recommended that you override the default values and/or supply an `original_config_file` wherever possible.
+
+    :param checkpoint_path: Path to `.ckpt` file. :param original_config_file: Path to `.yaml` config file
+    corresponding to the original architecture.
+            If `None`, will be automatically instantiated based on default values.
+    :param image_size: The image size that the model was trained on. Use 512 for original AudioLDM checkpoints. :param
+    prediction_type: The prediction type that the model was trained on. Use `'epsilon'` for original
+            AudioLDM checkpoints.
+    :param num_in_channels: The number of input channels. If `None` number of input channels will be automatically
+            inferred.
+    :param scheduler_type: Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler",
+            "euler-ancestral", "dpm", "ddim"]`.
+    :param extract_ema: Only relevant for checkpoints that have both EMA and non-EMA weights. Whether to extract
+            the EMA weights or not. Defaults to `False`. Pass `True` to extract the EMA weights. EMA weights usually
+            yield higher quality images for inference. Non-EMA weights are usually better to continue fine-tuning.
+    :param device: The device to use. Pass `None` to determine automatically. :param from_safetensors: If
+    `checkpoint_path` is in `safetensors` format, load checkpoint with safetensors
+            instead of PyTorch.
+    :return: An AudioLDMPipeline object representing the passed-in `.ckpt`/`.safetensors` file.
+    """
+
+    if not is_omegaconf_available():
+        raise ValueError(BACKENDS_MAPPING["omegaconf"][1])
+
+    from omegaconf import OmegaConf
+
+    if from_safetensors:
+        if not is_safetensors_available():
+            raise ValueError(BACKENDS_MAPPING["safetensors"][1])
+
+        from safetensors import safe_open
+
+        checkpoint = {}
+        with safe_open(checkpoint_path, framework="pt", device="cpu") as f:
+            for key in f.keys():
+                checkpoint[key] = f.get_tensor(key)
+    else:
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            checkpoint = torch.load(checkpoint_path, map_location=device)
+        else:
+            checkpoint = torch.load(checkpoint_path, map_location=device)
+
+    if "state_dict" in checkpoint:
+        checkpoint = checkpoint["state_dict"]
+
+    if original_config_file is None:
+        original_config = DEFAULT_CONFIG
+        original_config = OmegaConf.create(original_config)
+    else:
+        original_config = OmegaConf.load(original_config_file)
+
+    if num_in_channels is not None:
+        original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
+
+    if (
+        "parameterization" in original_config["model"]["params"]
+        and original_config["model"]["params"]["parameterization"] == "v"
+    ):
+        if prediction_type is None:
+            prediction_type = "v_prediction"
+    else:
+        if prediction_type is None:
+            prediction_type = "epsilon"
+
+    if image_size is None:
+        image_size = 512
+
+    num_train_timesteps = original_config.model.params.timesteps
+    beta_start = original_config.model.params.linear_start
+    beta_end = original_config.model.params.linear_end
+
+    scheduler = DDIMScheduler(
+        beta_end=beta_end,
+        beta_schedule="scaled_linear",
+        beta_start=beta_start,
+        num_train_timesteps=num_train_timesteps,
+        steps_offset=1,
+        clip_sample=False,
+        set_alpha_to_one=False,
+        prediction_type=prediction_type,
+    )
+    # make sure scheduler works correctly with DDIM
+    scheduler.register_to_config(clip_sample=False)
+
+    if scheduler_type == "pndm":
+        config = dict(scheduler.config)
+        config["skip_prk_steps"] = True
+        scheduler = PNDMScheduler.from_config(config)
+    elif scheduler_type == "lms":
+        scheduler = LMSDiscreteScheduler.from_config(scheduler.config)
+    elif scheduler_type == "heun":
+        scheduler = HeunDiscreteScheduler.from_config(scheduler.config)
+    elif scheduler_type == "euler":
+        scheduler = EulerDiscreteScheduler.from_config(scheduler.config)
+    elif scheduler_type == "euler-ancestral":
+        scheduler = EulerAncestralDiscreteScheduler.from_config(scheduler.config)
+    elif scheduler_type == "dpm":
+        scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config)
+    elif scheduler_type == "ddim":
+        scheduler = scheduler
+    else:
+        raise ValueError(f"Scheduler of type {scheduler_type} doesn't exist!")
+
+    # Convert the UNet2DModel
+    unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
+    unet = UNet2DConditionModel(**unet_config)
+
+    converted_unet_checkpoint = convert_ldm_unet_checkpoint(
+        checkpoint, unet_config, path=checkpoint_path, extract_ema=extract_ema
+    )
+
+    unet.load_state_dict(converted_unet_checkpoint)
+
+    # Convert the VAE model
+    vae_config = create_vae_diffusers_config(original_config, checkpoint=checkpoint, image_size=image_size)
+    converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+    vae = AutoencoderKL(**vae_config)
+    vae.load_state_dict(converted_vae_checkpoint)
+
+    # Convert the text model
+    # AudioLDM uses the same configuration and tokenizer as the original CLAP model
+    config = ClapTextConfig.from_pretrained("laion/clap-htsat-unfused")
+    tokenizer = AutoTokenizer.from_pretrained("laion/clap-htsat-unfused")
+
+    converted_text_model = convert_open_clap_checkpoint(checkpoint)
+    text_model = ClapTextModelWithProjection(config)
+
+    missing_keys, unexpected_keys = text_model.load_state_dict(converted_text_model, strict=False)
+    # we expect not to have token_type_ids in our original state dict so let's ignore them
+    missing_keys = list(set(missing_keys) - set(CLAP_EXPECTED_MISSING_KEYS))
+
+    if len(unexpected_keys) > 0:
+        raise ValueError(f"Unexpected keys when loading CLAP model: {unexpected_keys}")
+
+    if len(missing_keys) > 0:
+        raise ValueError(f"Missing keys when loading CLAP model: {missing_keys}")
+
+    # Convert the vocoder model
+    vocoder_config = create_transformers_vocoder_config(original_config)
+    vocoder_config = SpeechT5HifiGanConfig(**vocoder_config)
+    converted_vocoder_checkpoint = convert_hifigan_checkpoint(checkpoint, vocoder_config)
+
+    vocoder = SpeechT5HifiGan(vocoder_config)
+    vocoder.load_state_dict(converted_vocoder_checkpoint)
+
+    # Instantiate the diffusers pipeline
+    pipe = AudioLDMPipeline(
+        vae=vae,
+        text_encoder=text_model,
+        tokenizer=tokenizer,
+        unet=unet,
+        scheduler=scheduler,
+        vocoder=vocoder,
+    )
+
+    return pipe
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--original_config_file",
+        default=None,
+        type=str,
+        help="The YAML config file corresponding to the original architecture.",
+    )
+    parser.add_argument(
+        "--num_in_channels",
+        default=None,
+        type=int,
+        help="The number of input channels. If `None` number of input channels will be automatically inferred.",
+    )
+    parser.add_argument(
+        "--scheduler_type",
+        default="ddim",
+        type=str,
+        help="Type of scheduler to use. Should be one of ['pndm', 'lms', 'ddim', 'euler', 'euler-ancestral', 'dpm']",
+    )
+    parser.add_argument(
+        "--image_size",
+        default=None,
+        type=int,
+        help=("The image size that the model was trained on."),
+    )
+    parser.add_argument(
+        "--prediction_type",
+        default=None,
+        type=str,
+        help=("The prediction type that the model was trained on."),
+    )
+    parser.add_argument(
+        "--extract_ema",
+        action="store_true",
+        help=(
+            "Only relevant for checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights"
+            " or not. Defaults to `False`. Add `--extract_ema` to extract the EMA weights. EMA weights usually yield"
+            " higher quality images for inference. Non-EMA weights are usually better to continue fine-tuning."
+        ),
+    )
+    parser.add_argument(
+        "--from_safetensors",
+        action="store_true",
+        help="If `--checkpoint_path` is in `safetensors` format, load checkpoint with safetensors instead of PyTorch.",
+    )
+    parser.add_argument(
+        "--to_safetensors",
+        action="store_true",
+        help="Whether to store pipeline in safetensors format or not.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
+    parser.add_argument("--device", type=str, help="Device to use (e.g. cpu, cuda:0, cuda:1, etc.)")
+    args = parser.parse_args()
+
+    pipe = load_pipeline_from_original_audioldm_ckpt(
+        checkpoint_path=args.checkpoint_path,
+        original_config_file=args.original_config_file,
+        image_size=args.image_size,
+        prediction_type=args.prediction_type,
+        extract_ema=args.extract_ema,
+        scheduler_type=args.scheduler_type,
+        num_in_channels=args.num_in_channels,
+        from_safetensors=args.from_safetensors,
+        device=args.device,
+    )
+    pipe.save_pretrained(args.dump_path, safe_serialization=args.to_safetensors)

src/diffusers/__init__.py

@@ -112,6 +112,7 @@ else:
     from .pipelines import (
         AltDiffusionImg2ImgPipeline,
         AltDiffusionPipeline,
+        AudioLDMPipeline,
         CycleDiffusionPipeline,
         LDMTextToImagePipeline,
         PaintByExamplePipeline,

src/diffusers/models/unet_2d_condition.py

@@ -86,13 +86,14 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
             If `None`, it will skip the normalization and activation layers in post-processing
         norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
-        cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
         attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
         resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
             for resnet blocks, see [`~models.resnet.ResnetBlock2D`]. Choose from `default` or `scale_shift`.
         class_embed_type (`str`, *optional*, defaults to None):
             The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
-            `"timestep"`, `"identity"`, or `"projection"`.
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
         num_class_embeds (`int`, *optional*, defaults to None):
             Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
             class conditioning with `class_embed_type` equal to `None`.
@@ -106,6 +107,8 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
         projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
             using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+        embeddings with the class embeddings.
     """
 
     _supports_gradient_checkpointing = True
@@ -135,7 +138,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         act_fn: str = "silu",
         norm_num_groups: Optional[int] = 32,
         norm_eps: float = 1e-5,
-        cross_attention_dim: int = 1280,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
         attention_head_dim: Union[int, Tuple[int]] = 8,
         dual_cross_attention: bool = False,
         use_linear_projection: bool = False,
@@ -149,6 +152,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         conv_in_kernel: int = 3,
         conv_out_kernel: int = 3,
         projection_class_embeddings_input_dim: Optional[int] = None,
+        class_embeddings_concat: bool = False,
     ):
         super().__init__()
 
@@ -175,6 +179,11 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
             )
 
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
         # input
         conv_in_padding = (conv_in_kernel - 1) // 2
         self.conv_in = nn.Conv2d(
@@ -228,6 +237,12 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
             # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
             self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
         else:
             self.class_embedding = None
 
@@ -240,6 +255,17 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         if isinstance(attention_head_dim, int):
             attention_head_dim = (attention_head_dim,) * len(down_block_types)
 
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
         # down
         output_channel = block_out_channels[0]
         for i, down_block_type in enumerate(down_block_types):
@@ -252,12 +278,12 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 num_layers=layers_per_block,
                 in_channels=input_channel,
                 out_channels=output_channel,
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 add_downsample=not is_final_block,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[i],
                 attn_num_head_channels=attention_head_dim[i],
                 downsample_padding=downsample_padding,
                 dual_cross_attention=dual_cross_attention,
@@ -272,12 +298,12 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         if mid_block_type == "UNetMidBlock2DCrossAttn":
             self.mid_block = UNetMidBlock2DCrossAttn(
                 in_channels=block_out_channels[-1],
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 output_scale_factor=mid_block_scale_factor,
                 resnet_time_scale_shift=resnet_time_scale_shift,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[-1],
                 attn_num_head_channels=attention_head_dim[-1],
                 resnet_groups=norm_num_groups,
                 dual_cross_attention=dual_cross_attention,
@@ -287,11 +313,11 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
             self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                 in_channels=block_out_channels[-1],
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 output_scale_factor=mid_block_scale_factor,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[-1],
                 attn_num_head_channels=attention_head_dim[-1],
                 resnet_groups=norm_num_groups,
                 resnet_time_scale_shift=resnet_time_scale_shift,
@@ -307,6 +333,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         # up
         reversed_block_out_channels = list(reversed(block_out_channels))
         reversed_attention_head_dim = list(reversed(attention_head_dim))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
         only_cross_attention = list(reversed(only_cross_attention))
 
         output_channel = reversed_block_out_channels[0]
@@ -330,12 +357,12 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 in_channels=input_channel,
                 out_channels=output_channel,
                 prev_output_channel=prev_output_channel,
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 add_upsample=add_upsample,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=reversed_cross_attention_dim[i],
                 attn_num_head_channels=reversed_attention_head_dim[i],
                 dual_cross_attention=dual_cross_attention,
                 use_linear_projection=use_linear_projection,
@@ -571,7 +598,11 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 class_labels = self.time_proj(class_labels)
 
             class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
-            emb = emb + class_emb
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
 
         # 2. pre-process
         sample = self.conv_in(sample)

src/diffusers/pipelines/__init__.py

@@ -44,6 +44,7 @@ except OptionalDependencyNotAvailable:
     from ..utils.dummy_torch_and_transformers_objects import *  # noqa F403
 else:
     from .alt_diffusion import AltDiffusionImg2ImgPipeline, AltDiffusionPipeline
+    from .audioldm import AudioLDMPipeline
     from .latent_diffusion import LDMTextToImagePipeline
     from .paint_by_example import PaintByExamplePipeline
     from .semantic_stable_diffusion import SemanticStableDiffusionPipeline

src/diffusers/pipelines/audioldm/__init__.py

+from ...utils import (
+    OptionalDependencyNotAvailable,
+    is_torch_available,
+    is_transformers_available,
+    is_transformers_version,
+)
+
+
+try:
+    if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils.dummy_torch_and_transformers_objects import (
+        AudioLDMPipeline,
+    )
+else:
+    from .pipeline_audioldm import AudioLDMPipeline

src/diffusers/pipelines/audioldm/pipeline_audioldm.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from transformers import ClapTextModelWithProjection, RobertaTokenizer, RobertaTokenizerFast, SpeechT5HifiGan
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
+from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import AudioLDMPipeline
+
+        >>> pipe = AudioLDMPipeline.from_pretrained("cvssp/audioldm", torch_dtype=torch.float16)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "A hammer hitting a wooden surface"
+        >>> audio = pipe(prompt).audio[0]
+        ```
+"""
+
+
+class AudioLDMPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-audio generation using AudioLDM.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode audios to and from latent representations.
+        text_encoder ([`ClapTextModelWithProjection`]):
+            Frozen text-encoder. AudioLDM uses the text portion of
+            [CLAP](https://huggingface.co/docs/transformers/main/model_doc/clap#transformers.ClapTextModelWithProjection),
+            specifically the [RoBERTa HSTAT-unfused](https://huggingface.co/laion/clap-htsat-unfused) variant.
+        tokenizer ([`PreTrainedTokenizer`]):
+            Tokenizer of class
+            [RobertaTokenizer](https://huggingface.co/docs/transformers/model_doc/roberta#transformers.RobertaTokenizer).
+        unet ([`UNet2DConditionModel`]): U-Net architecture to denoise the encoded audio latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded audio latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        vocoder ([`SpeechT5HifiGan`]):
+            Vocoder of class
+            [SpeechT5HifiGan](https://huggingface.co/docs/transformers/main/en/model_doc/speecht5#transformers.SpeechT5HifiGan).
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: ClapTextModelWithProjection,
+        tokenizer: Union[RobertaTokenizer, RobertaTokenizerFast],
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        vocoder: SpeechT5HifiGan,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            vocoder=vocoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and vocoder have their state dicts saved to CPU and then are moved to a `torch.device('meta')
+        and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.vocoder]:
+            cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_waveforms_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device (`torch.device`):
+                torch device
+            num_waveforms_per_prompt (`int`):
+                number of waveforms that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the audio generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+        """
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            attention_mask = text_inputs.attention_mask
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLAP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            prompt_embeds = self.text_encoder(
+                text_input_ids.to(device),
+                attention_mask=attention_mask.to(device),
+            )
+            prompt_embeds = prompt_embeds.text_embeds
+            # additional L_2 normalization over each hidden-state
+            prompt_embeds = F.normalize(prompt_embeds, dim=-1)
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        (
+            bs_embed,
+            seq_len,
+        ) = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_waveforms_per_prompt)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_waveforms_per_prompt, seq_len)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            uncond_input_ids = uncond_input.input_ids.to(device)
+            attention_mask = uncond_input.attention_mask.to(device)
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input_ids,
+                attention_mask=attention_mask,
+            )
+            negative_prompt_embeds = negative_prompt_embeds.text_embeds
+            # additional L_2 normalization over each hidden-state
+            negative_prompt_embeds = F.normalize(negative_prompt_embeds, dim=-1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_waveforms_per_prompt)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_waveforms_per_prompt, seq_len)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        return prompt_embeds
+
+    def decode_latents(self, latents):
+        latents = 1 / self.vae.config.scaling_factor * latents
+        mel_spectrogram = self.vae.decode(latents).sample
+        return mel_spectrogram
+
+    def mel_spectrogram_to_waveform(self, mel_spectrogram):
+        if mel_spectrogram.dim() == 4:
+            mel_spectrogram = mel_spectrogram.squeeze(1)
+
+        waveform = self.vocoder(mel_spectrogram)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        waveform = waveform.cpu()
+        return waveform
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        audio_length_in_s,
+        vocoder_upsample_factor,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        min_audio_length_in_s = vocoder_upsample_factor * self.vae_scale_factor
+        if audio_length_in_s < min_audio_length_in_s:
+            raise ValueError(
+                f"`audio_length_in_s` has to be a positive value greater than or equal to {min_audio_length_in_s}, but "
+                f"is {audio_length_in_s}."
+            )
+
+        if self.vocoder.config.model_in_dim % self.vae_scale_factor != 0:
+            raise ValueError(
+                f"The number of frequency bins in the vocoder's log-mel spectrogram has to be divisible by the "
+                f"VAE scale factor, but got {self.vocoder.config.model_in_dim} bins and a scale factor of "
+                f"{self.vae_scale_factor}."
+            )
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents with width->self.vocoder.config.model_in_dim
+    def prepare_latents(self, batch_size, num_channels_latents, height, dtype, device, generator, latents=None):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            self.vocoder.config.model_in_dim // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        audio_length_in_s: Optional[float] = None,
+        num_inference_steps: int = 10,
+        guidance_scale: float = 2.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_waveforms_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        output_type: Optional[str] = "np",
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the audio generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            audio_length_in_s (`int`, *optional*, defaults to 5.12):
+                The length of the generated audio sample in seconds.
+            num_inference_steps (`int`, *optional*, defaults to 10):
+                The number of denoising steps. More denoising steps usually lead to a higher quality audio at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 2.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate audios that are closely linked to the text `prompt`,
+                usually at the expense of lower sound quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the audio generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            num_waveforms_per_prompt (`int`, *optional*, defaults to 1):
+                The number of waveforms to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for audio
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generate image. Choose between:
+                - `"np"`: Return Numpy `np.ndarray` objects.
+                - `"pt"`: Return PyTorch `torch.Tensor` objects.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated audios.
+        """
+        # 0. Convert audio input length from seconds to spectrogram height
+        vocoder_upsample_factor = np.prod(self.vocoder.config.upsample_rates) / self.vocoder.config.sampling_rate
+
+        if audio_length_in_s is None:
+            audio_length_in_s = self.unet.config.sample_size * self.vae_scale_factor * vocoder_upsample_factor
+
+        height = int(audio_length_in_s / vocoder_upsample_factor)
+
+        original_waveform_length = int(audio_length_in_s * self.vocoder.config.sampling_rate)
+        if height % self.vae_scale_factor != 0:
+            height = int(np.ceil(height / self.vae_scale_factor)) * self.vae_scale_factor
+            logger.info(
+                f"Audio length in seconds {audio_length_in_s} is increased to {height * vocoder_upsample_factor} "
+                f"so that it can be handled by the model. It will be cut to {audio_length_in_s} after the "
+                f"denoising process."
+            )
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            audio_length_in_s,
+            vocoder_upsample_factor,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_waveforms_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_waveforms_per_prompt,
+            num_channels_latents,
+            height,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=None,
+                    class_labels=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                ).sample
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        # 8. Post-processing
+        mel_spectrogram = self.decode_latents(latents)
+
+        audio = self.mel_spectrogram_to_waveform(mel_spectrogram)
+
+        audio = audio[:, :original_waveform_length]
+
+        if output_type == "np":
+            audio = audio.numpy()
+
+        if not return_dict:
+            return (audio,)
+
+        return AudioPipelineOutput(audios=audio)

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -167,13 +167,14 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
             If `None`, it will skip the normalization and activation layers in post-processing
         norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
-        cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
         attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
         resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
             for resnet blocks, see [`~models.resnet.ResnetBlockFlat`]. Choose from `default` or `scale_shift`.
         class_embed_type (`str`, *optional*, defaults to None):
             The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
-            `"timestep"`, `"identity"`, or `"projection"`.
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
         num_class_embeds (`int`, *optional*, defaults to None):
             Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
             class conditioning with `class_embed_type` equal to `None`.
@@ -187,6 +188,8 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
         projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
             using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+        embeddings with the class embeddings.
     """
 
     _supports_gradient_checkpointing = True
@@ -221,7 +224,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         act_fn: str = "silu",
         norm_num_groups: Optional[int] = 32,
         norm_eps: float = 1e-5,
-        cross_attention_dim: int = 1280,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
         attention_head_dim: Union[int, Tuple[int]] = 8,
         dual_cross_attention: bool = False,
         use_linear_projection: bool = False,
@@ -235,6 +238,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         conv_in_kernel: int = 3,
         conv_out_kernel: int = 3,
         projection_class_embeddings_input_dim: Optional[int] = None,
+        class_embeddings_concat: bool = False,
     ):
         super().__init__()
 
@@ -265,6 +269,12 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 f" {attention_head_dim}. `down_block_types`: {down_block_types}."
             )
 
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`:"
+                f" {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
         # input
         conv_in_padding = (conv_in_kernel - 1) // 2
         self.conv_in = LinearMultiDim(
@@ -318,6 +328,12 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
             # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
             # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
             self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
         else:
             self.class_embedding = None
 
@@ -330,6 +346,17 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         if isinstance(attention_head_dim, int):
             attention_head_dim = (attention_head_dim,) * len(down_block_types)
 
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
         # down
         output_channel = block_out_channels[0]
         for i, down_block_type in enumerate(down_block_types):
@@ -342,12 +369,12 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 num_layers=layers_per_block,
                 in_channels=input_channel,
                 out_channels=output_channel,
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 add_downsample=not is_final_block,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[i],
                 attn_num_head_channels=attention_head_dim[i],
                 downsample_padding=downsample_padding,
                 dual_cross_attention=dual_cross_attention,
@@ -362,12 +389,12 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         if mid_block_type == "UNetMidBlockFlatCrossAttn":
             self.mid_block = UNetMidBlockFlatCrossAttn(
                 in_channels=block_out_channels[-1],
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 output_scale_factor=mid_block_scale_factor,
                 resnet_time_scale_shift=resnet_time_scale_shift,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[-1],
                 attn_num_head_channels=attention_head_dim[-1],
                 resnet_groups=norm_num_groups,
                 dual_cross_attention=dual_cross_attention,
@@ -377,11 +404,11 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         elif mid_block_type == "UNetMidBlockFlatSimpleCrossAttn":
             self.mid_block = UNetMidBlockFlatSimpleCrossAttn(
                 in_channels=block_out_channels[-1],
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 output_scale_factor=mid_block_scale_factor,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim[-1],
                 attn_num_head_channels=attention_head_dim[-1],
                 resnet_groups=norm_num_groups,
                 resnet_time_scale_shift=resnet_time_scale_shift,
@@ -397,6 +424,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
         # up
         reversed_block_out_channels = list(reversed(block_out_channels))
         reversed_attention_head_dim = list(reversed(attention_head_dim))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
         only_cross_attention = list(reversed(only_cross_attention))
 
         output_channel = reversed_block_out_channels[0]
@@ -420,12 +448,12 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 in_channels=input_channel,
                 out_channels=output_channel,
                 prev_output_channel=prev_output_channel,
-                temb_channels=time_embed_dim,
+                temb_channels=blocks_time_embed_dim,
                 add_upsample=add_upsample,
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
-                cross_attention_dim=cross_attention_dim,
+                cross_attention_dim=reversed_cross_attention_dim[i],
                 attn_num_head_channels=reversed_attention_head_dim[i],
                 dual_cross_attention=dual_cross_attention,
                 use_linear_projection=use_linear_projection,
@@ -661,7 +689,11 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 class_labels = self.time_proj(class_labels)
 
             class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
-            emb = emb + class_emb
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
 
         # 2. pre-process
         sample = self.conv_in(sample)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -32,6 +32,21 @@ class AltDiffusionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class AudioLDMPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class CycleDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/models/test_models_unet_2d_condition.py

@@ -199,6 +199,74 @@ class UNet2DConditionModelTests(ModelTesterMixin, unittest.TestCase):
         expected_shape = inputs_dict["sample"].shape
         self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
 
+    def test_model_with_cross_attention_dim_tuple(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        init_dict["cross_attention_dim"] = (32, 32)
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            output = model(**inputs_dict)
+
+            if isinstance(output, dict):
+                output = output.sample
+
+        self.assertIsNotNone(output)
+        expected_shape = inputs_dict["sample"].shape
+        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
+
+    def test_model_with_simple_projection(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        batch_size, _, _, sample_size = inputs_dict["sample"].shape
+
+        init_dict["class_embed_type"] = "simple_projection"
+        init_dict["projection_class_embeddings_input_dim"] = sample_size
+
+        inputs_dict["class_labels"] = floats_tensor((batch_size, sample_size)).to(torch_device)
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            output = model(**inputs_dict)
+
+            if isinstance(output, dict):
+                output = output.sample
+
+        self.assertIsNotNone(output)
+        expected_shape = inputs_dict["sample"].shape
+        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
+
+    def test_model_with_class_embeddings_concat(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        batch_size, _, _, sample_size = inputs_dict["sample"].shape
+
+        init_dict["class_embed_type"] = "simple_projection"
+        init_dict["projection_class_embeddings_input_dim"] = sample_size
+        init_dict["class_embeddings_concat"] = True
+
+        inputs_dict["class_labels"] = floats_tensor((batch_size, sample_size)).to(torch_device)
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            output = model(**inputs_dict)
+
+            if isinstance(output, dict):
+                output = output.sample
+
+        self.assertIsNotNone(output)
+        expected_shape = inputs_dict["sample"].shape
+        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
+
     def test_model_attention_slicing(self):
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 

tests/pipeline_params.py

@@ -103,6 +103,19 @@ UNCONDITIONAL_AUDIO_GENERATION_PARAMS = frozenset(["batch_size"])
 
 UNCONDITIONAL_AUDIO_GENERATION_BATCH_PARAMS = frozenset([])
 
+TEXT_TO_AUDIO_PARAMS = frozenset(
+    [
+        "prompt",
+        "audio_length_in_s",
+        "guidance_scale",
+        "negative_prompt",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "cross_attention_kwargs",
+    ]
+)
+
+TEXT_TO_AUDIO_BATCH_PARAMS = frozenset(["prompt", "negative_prompt"])
 TOKENS_TO_AUDIO_GENERATION_PARAMS = frozenset(["input_tokens"])
 
 TOKENS_TO_AUDIO_GENERATION_BATCH_PARAMS = frozenset(["input_tokens"])

tests/pipelines/audioldm/test_audioldm.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import gc
+import unittest
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from transformers import (
+    ClapTextConfig,
+    ClapTextModelWithProjection,
+    RobertaTokenizer,
+    SpeechT5HifiGan,
+    SpeechT5HifiGanConfig,
+)
+
+from diffusers import (
+    AudioLDMPipeline,
+    AutoencoderKL,
+    DDIMScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    UNet2DConditionModel,
+)
+from diffusers.utils import slow, torch_device
+
+from ...pipeline_params import TEXT_TO_AUDIO_BATCH_PARAMS, TEXT_TO_AUDIO_PARAMS
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+class AudioLDMPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = AudioLDMPipeline
+    params = TEXT_TO_AUDIO_PARAMS
+    batch_params = TEXT_TO_AUDIO_BATCH_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "num_waveforms_per_prompt",
+            "generator",
+            "latents",
+            "output_type",
+            "return_dict",
+            "callback",
+            "callback_steps",
+        ]
+    )
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=(32, 64),
+            class_embed_type="simple_projection",
+            projection_class_embeddings_input_dim=32,
+            class_embeddings_concat=True,
+        )
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=1,
+            out_channels=1,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = ClapTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            projection_dim=32,
+        )
+        text_encoder = ClapTextModelWithProjection(text_encoder_config)
+        tokenizer = RobertaTokenizer.from_pretrained("hf-internal-testing/tiny-random-roberta", model_max_length=77)
+
+        vocoder_config = SpeechT5HifiGanConfig(
+            model_in_dim=8,
+            sampling_rate=16000,
+            upsample_initial_channel=16,
+            upsample_rates=[2, 2],
+            upsample_kernel_sizes=[4, 4],
+            resblock_kernel_sizes=[3, 7],
+            resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5]],
+            normalize_before=False,
+        )
+
+        vocoder = SpeechT5HifiGan(vocoder_config)
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "vocoder": vocoder,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "A hammer hitting a wooden surface",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+        }
+        return inputs
+
+    def test_audioldm_ddim(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+
+        components = self.get_dummy_components()
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        output = audioldm_pipe(**inputs)
+        audio = output.audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) == 256
+
+        audio_slice = audio[:10]
+        expected_slice = np.array(
+            [-0.0050, 0.0050, -0.0060, 0.0033, -0.0026, 0.0033, -0.0027, 0.0033, -0.0028, 0.0033]
+        )
+
+        assert np.abs(audio_slice - expected_slice).max() < 1e-2
+
+    def test_audioldm_prompt_embeds(self):
+        components = self.get_dummy_components()
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs["prompt"] = 3 * [inputs["prompt"]]
+
+        # forward
+        output = audioldm_pipe(**inputs)
+        audio_1 = output.audios[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        prompt = 3 * [inputs.pop("prompt")]
+
+        text_inputs = audioldm_pipe.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=audioldm_pipe.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_inputs = text_inputs["input_ids"].to(torch_device)
+
+        prompt_embeds = audioldm_pipe.text_encoder(
+            text_inputs,
+        )
+        prompt_embeds = prompt_embeds.text_embeds
+        # additional L_2 normalization over each hidden-state
+        prompt_embeds = F.normalize(prompt_embeds, dim=-1)
+
+        inputs["prompt_embeds"] = prompt_embeds
+
+        # forward
+        output = audioldm_pipe(**inputs)
+        audio_2 = output.audios[0]
+
+        assert np.abs(audio_1 - audio_2).max() < 1e-2
+
+    def test_audioldm_negative_prompt_embeds(self):
+        components = self.get_dummy_components()
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        inputs["negative_prompt"] = negative_prompt
+        inputs["prompt"] = 3 * [inputs["prompt"]]
+
+        # forward
+        output = audioldm_pipe(**inputs)
+        audio_1 = output.audios[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        prompt = 3 * [inputs.pop("prompt")]
+
+        embeds = []
+        for p in [prompt, negative_prompt]:
+            text_inputs = audioldm_pipe.tokenizer(
+                p,
+                padding="max_length",
+                max_length=audioldm_pipe.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_inputs = text_inputs["input_ids"].to(torch_device)
+
+            text_embeds = audioldm_pipe.text_encoder(
+                text_inputs,
+            )
+            text_embeds = text_embeds.text_embeds
+            # additional L_2 normalization over each hidden-state
+            text_embeds = F.normalize(text_embeds, dim=-1)
+
+            embeds.append(text_embeds)
+
+        inputs["prompt_embeds"], inputs["negative_prompt_embeds"] = embeds
+
+        # forward
+        output = audioldm_pipe(**inputs)
+        audio_2 = output.audios[0]
+
+        assert np.abs(audio_1 - audio_2).max() < 1e-2
+
+    def test_audioldm_negative_prompt(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        components["scheduler"] = PNDMScheduler(skip_prk_steps=True)
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        negative_prompt = "egg cracking"
+        output = audioldm_pipe(**inputs, negative_prompt=negative_prompt)
+        audio = output.audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) == 256
+
+        audio_slice = audio[:10]
+        expected_slice = np.array(
+            [-0.0051, 0.0050, -0.0060, 0.0034, -0.0026, 0.0033, -0.0027, 0.0033, -0.0028, 0.0032]
+        )
+
+        assert np.abs(audio_slice - expected_slice).max() < 1e-2
+
+    def test_audioldm_num_waveforms_per_prompt(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        components["scheduler"] = PNDMScheduler(skip_prk_steps=True)
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A hammer hitting a wooden surface"
+
+        # test num_waveforms_per_prompt=1 (default)
+        audios = audioldm_pipe(prompt, num_inference_steps=2).audios
+
+        assert audios.shape == (1, 256)
+
+        # test num_waveforms_per_prompt=1 (default) for batch of prompts
+        batch_size = 2
+        audios = audioldm_pipe([prompt] * batch_size, num_inference_steps=2).audios
+
+        assert audios.shape == (batch_size, 256)
+
+        # test num_waveforms_per_prompt for single prompt
+        num_waveforms_per_prompt = 2
+        audios = audioldm_pipe(prompt, num_inference_steps=2, num_waveforms_per_prompt=num_waveforms_per_prompt).audios
+
+        assert audios.shape == (num_waveforms_per_prompt, 256)
+
+        # test num_waveforms_per_prompt for batch of prompts
+        batch_size = 2
+        audios = audioldm_pipe(
+            [prompt] * batch_size, num_inference_steps=2, num_waveforms_per_prompt=num_waveforms_per_prompt
+        ).audios
+
+        assert audios.shape == (batch_size * num_waveforms_per_prompt, 256)
+
+    def test_audioldm_audio_length_in_s(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+        vocoder_sampling_rate = audioldm_pipe.vocoder.config.sampling_rate
+
+        inputs = self.get_dummy_inputs(device)
+        output = audioldm_pipe(audio_length_in_s=0.016, **inputs)
+        audio = output.audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) / vocoder_sampling_rate == 0.016
+
+        output = audioldm_pipe(audio_length_in_s=0.032, **inputs)
+        audio = output.audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) / vocoder_sampling_rate == 0.032
+
+    def test_audioldm_vocoder_model_in_dim(self):
+        components = self.get_dummy_components()
+        audioldm_pipe = AudioLDMPipeline(**components)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        prompt = ["hey"]
+
+        output = audioldm_pipe(prompt, num_inference_steps=1)
+        audio_shape = output.audios.shape
+        assert audio_shape == (1, 256)
+
+        config = audioldm_pipe.vocoder.config
+        config.model_in_dim *= 2
+        audioldm_pipe.vocoder = SpeechT5HifiGan(config).to(torch_device)
+        output = audioldm_pipe(prompt, num_inference_steps=1)
+        audio_shape = output.audios.shape
+        # waveform shape is unchanged, we just have 2x the number of mel channels in the spectrogram
+        assert audio_shape == (1, 256)
+
+    def test_attention_slicing_forward_pass(self):
+        self._test_attention_slicing_forward_pass(test_mean_pixel_difference=False)
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(test_mean_pixel_difference=False)
+
+
+@slow
+# @require_torch_gpu
+class AudioLDMPipelineSlowTests(unittest.TestCase):
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
+        latents = np.random.RandomState(seed).standard_normal((1, 8, 128, 16))
+        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
+        inputs = {
+            "prompt": "A hammer hitting a wooden surface",
+            "latents": latents,
+            "generator": generator,
+            "num_inference_steps": 3,
+            "guidance_scale": 2.5,
+        }
+        return inputs
+
+    def test_audioldm(self):
+        audioldm_pipe = AudioLDMPipeline.from_pretrained("cvssp/audioldm")
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 25
+        audio = audioldm_pipe(**inputs).audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) == 81920
+
+        audio_slice = audio[77230:77240]
+        expected_slice = np.array(
+            [-0.4884, -0.4607, 0.0023, 0.5007, 0.5896, 0.5151, 0.3813, -0.0208, -0.3687, -0.4315]
+        )
+        max_diff = np.abs(expected_slice - audio_slice).max()
+        assert max_diff < 1e-2
+
+    def test_audioldm_lms(self):
+        audioldm_pipe = AudioLDMPipeline.from_pretrained("cvssp/audioldm")
+        audioldm_pipe.scheduler = LMSDiscreteScheduler.from_config(audioldm_pipe.scheduler.config)
+        audioldm_pipe = audioldm_pipe.to(torch_device)
+        audioldm_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_inputs(torch_device)
+        audio = audioldm_pipe(**inputs).audios[0]
+
+        assert audio.ndim == 1
+        assert len(audio) == 81920
+
+        audio_slice = audio[27780:27790]
+        expected_slice = np.array([-0.2131, -0.0873, -0.0124, -0.0189, 0.0569, 0.1373, 0.1883, 0.2886, 0.3297, 0.2212])
+        max_diff = np.abs(expected_slice - audio_slice).max()
+        assert max_diff < 1e-2