Initial commit

hyvideo/constants.py

+import os
+import torch
+
+__all__ = [
+    "C_SCALE",
+    "PROMPT_TEMPLATE",
+    "MODEL_BASE",
+    "PRECISIONS",
+    "NORMALIZATION_TYPE",
+    "ACTIVATION_TYPE",
+    "VAE_PATH",
+    "TEXT_ENCODER_PATH",
+    "TOKENIZER_PATH",
+    "TEXT_PROJECTION",
+    "DATA_TYPE",
+    "NEGATIVE_PROMPT",
+]
+
+PRECISION_TO_TYPE = {
+    'fp32': torch.float32,
+    'fp16': torch.float16,
+    'bf16': torch.bfloat16,
+}
+
+# =================== Constant Values =====================
+# Computation scale factor, 1P = 1_000_000_000_000_000. Tensorboard will display the value in PetaFLOPS to avoid
+# overflow error when tensorboard logging values.
+C_SCALE = 1_000_000_000_000_000
+
+# When using decoder-only models, we must provide a prompt template to instruct the text encoder
+# on how to generate the text.
+# --------------------------------------------------------------------
+PROMPT_TEMPLATE_ENCODE = (
+    "<|start_header_id|>system<|end_header_id|>\n\nDescribe the image by detailing the color, shape, size, texture, "
+    "quantity, text, spatial relationships of the objects and background:<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+) 
+PROMPT_TEMPLATE_ENCODE_VIDEO = (
+    "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
+    "1. The main content and theme of the video."
+    "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
+    "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
+    "4. background environment, light, style and atmosphere."
+    "5. camera angles, movements, and transitions used in the video:<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+)  
+
+NEGATIVE_PROMPT = "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion"
+
+PROMPT_TEMPLATE = {
+    "dit-llm-encode": {
+        "template": PROMPT_TEMPLATE_ENCODE,
+        "crop_start": 36,
+    },
+    "dit-llm-encode-video": {
+        "template": PROMPT_TEMPLATE_ENCODE_VIDEO,
+        "crop_start": 95,
+    },
+}
+
+# ======================= Model ======================
+PRECISIONS = {"fp32", "fp16", "bf16"}
+NORMALIZATION_TYPE = {"layer", "rms"}
+ACTIVATION_TYPE = {"relu", "silu", "gelu", "gelu_tanh"}
+
+# =================== Model Path =====================
+MODEL_BASE = os.getenv("MODEL_BASE", "./ckpts")
+
+# =================== Data =======================
+DATA_TYPE = {"image", "video", "image_video"}
+
+# 3D VAE
+VAE_PATH = {"884-16c-hy": f"{MODEL_BASE}/hunyuan-video-t2v-720p/vae"}
+
+# Text Encoder
+TEXT_ENCODER_PATH = {
+    "clipL": f"{MODEL_BASE}/text_encoder_2",
+    "llm": f"{MODEL_BASE}/text_encoder",
+}
+
+# Tokenizer
+TOKENIZER_PATH = {
+    "clipL": f"{MODEL_BASE}/text_encoder_2",
+    "llm": f"{MODEL_BASE}/text_encoder",
+}
+
+TEXT_PROJECTION = {
+    "linear",  # Default, an nn.Linear() layer
+    "single_refiner",  # Single TokenRefiner. Refer to LI-DiT
+}

hyvideo/diffusion/__init__.py

+from .pipelines import HunyuanVideoPipeline
+from .schedulers import FlowMatchDiscreteScheduler

hyvideo/diffusion/pipelines/__init__.py

+from .pipeline_hunyuan_video import HunyuanVideoPipeline

hyvideo/diffusion/schedulers/__init__.py

+from .scheduling_flow_match_discrete import FlowMatchDiscreteScheduler

hyvideo/diffusion/schedulers/scheduling_flow_match_discrete.py

+# Copyright 2024 Stability AI, Katherine Crowson and 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.
+# ==============================================================================
+#
+# Modified from diffusers==0.29.2
+#
+# ==============================================================================
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class FlowMatchDiscreteSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+
+    prev_sample: torch.FloatTensor
+
+
+class FlowMatchDiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Euler scheduler.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        timestep_spacing (`str`, defaults to `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        shift (`float`, defaults to 1.0):
+            The shift value for the timestep schedule.
+        reverse (`bool`, defaults to `True`):
+            Whether to reverse the timestep schedule.
+    """
+
+    _compatibles = []
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        shift: float = 1.0,
+        reverse: bool = True,
+        solver: str = "euler",
+        n_tokens: Optional[int] = None,
+    ):
+        sigmas = torch.linspace(1, 0, num_train_timesteps + 1)
+
+        if not reverse:
+            sigmas = sigmas.flip(0)
+
+        self.sigmas = sigmas
+        # the value fed to model
+        self.timesteps = (sigmas[:-1] * num_train_timesteps).to(dtype=torch.float32)
+
+        self._step_index = None
+        self._begin_index = None
+
+        self.supported_solver = ["euler"]
+        if solver not in self.supported_solver:
+            raise ValueError(
+                f"Solver {solver} not supported. Supported solvers: {self.supported_solver}"
+            )
+
+    @property
+    def step_index(self):
+        """
+        The index counter for current timestep. It will increase 1 after each scheduler step.
+        """
+        return self._step_index
+
+    @property
+    def begin_index(self):
+        """
+        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+        """
+        return self._begin_index
+
+    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+        Args:
+            begin_index (`int`):
+                The begin index for the scheduler.
+        """
+        self._begin_index = begin_index
+
+    def _sigma_to_t(self, sigma):
+        return sigma * self.config.num_train_timesteps
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device] = None,
+        n_tokens: int = None,
+    ):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+            n_tokens (`int`, *optional*):
+                Number of tokens in the input sequence.
+        """
+        self.num_inference_steps = num_inference_steps
+        
+        sigmas = torch.linspace(1, 0, num_inference_steps + 1)
+        sigmas = self.sd3_time_shift(sigmas)
+
+        if not self.config.reverse:
+            sigmas = 1 - sigmas
+
+        self.sigmas = sigmas
+        self.timesteps = (sigmas[:-1] * self.config.num_train_timesteps).to(
+            dtype=torch.float32, device=device
+        )
+
+        # Reset step index
+        self._step_index = None
+
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
+        if schedule_timesteps is None:
+            schedule_timesteps = self.timesteps
+
+        indices = (schedule_timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        pos = 1 if len(indices) > 1 else 0
+
+        return indices[pos].item()
+
+    def _init_step_index(self, timestep):
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def scale_model_input(
+        self, sample: torch.Tensor, timestep: Optional[int] = None
+    ) -> torch.Tensor:
+        return sample
+
+    def sd3_time_shift(self, t: torch.Tensor):
+        return (self.config.shift * t) / (1 + (self.config.shift - 1) * t)
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        return_dict: bool = True,
+    ) -> Union[FlowMatchDiscreteSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            n_tokens (`int`, *optional*):
+                Number of tokens in the input sequence.
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
+                tuple.
+
+        Returns:
+            [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
+                returned, otherwise a tuple is returned where the first element is the sample tensor.
+        """
+
+        if (
+            isinstance(timestep, int)
+            or isinstance(timestep, torch.IntTensor)
+            or isinstance(timestep, torch.LongTensor)
+        ):
+            raise ValueError(
+                (
+                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                    " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+                    " one of the `scheduler.timesteps` as a timestep."
+                ),
+            )
+
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+
+        dt = self.sigmas[self.step_index + 1] - self.sigmas[self.step_index]
+
+        if self.config.solver == "euler":
+            prev_sample = sample + model_output.to(torch.float32) * dt
+        else:
+            raise ValueError(
+                f"Solver {self.config.solver} not supported. Supported solvers: {self.supported_solver}"
+            )
+
+        # upon completion increase step index by one
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return FlowMatchDiscreteSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self):
+        return self.config.num_train_timesteps

hyvideo/modules/__init__.py

+from .models import HYVideoDiffusionTransformer, HUNYUAN_VIDEO_CONFIG
+
+
+def load_model(args, in_channels, out_channels, factor_kwargs):
+    """load hunyuan video model
+
+    Args:
+        args (dict): model args
+        in_channels (int): input channels number
+        out_channels (int): output channels number
+        factor_kwargs (dict): factor kwargs
+
+    Returns:
+        model (nn.Module): The hunyuan video model
+    """
+    if args.model in HUNYUAN_VIDEO_CONFIG.keys():
+        model = HYVideoDiffusionTransformer(
+            args,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            **HUNYUAN_VIDEO_CONFIG[args.model],
+            **factor_kwargs,
+        )
+        return model
+    else:
+        raise NotImplementedError()

hyvideo/modules/activation_layers.py

+import torch.nn as nn
+
+
+def get_activation_layer(act_type):
+    """get activation layer
+
+    Args:
+        act_type (str): the activation type
+
+    Returns:
+        torch.nn.functional: the activation layer
+    """
+    if act_type == "gelu":
+        return lambda: nn.GELU()
+    elif act_type == "gelu_tanh":
+        # Approximate `tanh` requires torch >= 1.13
+        return lambda: nn.GELU(approximate="tanh")
+    elif act_type == "relu":
+        return nn.ReLU
+    elif act_type == "silu":
+        return nn.SiLU
+    else:
+        raise ValueError(f"Unknown activation type: {act_type}")

hyvideo/modules/attenion.py

+import importlib.metadata
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    import flash_attn
+    from flash_attn.flash_attn_interface import _flash_attn_forward
+    from flash_attn.flash_attn_interface import flash_attn_varlen_func
+except ImportError:
+    flash_attn = None
+    flash_attn_varlen_func = None
+    _flash_attn_forward = None
+
+
+MEMORY_LAYOUT = {
+    "flash": (
+        lambda x: x.view(x.shape[0] * x.shape[1], *x.shape[2:]),
+        lambda x: x,
+    ),
+    "torch": (
+        lambda x: x.transpose(1, 2),
+        lambda x: x.transpose(1, 2),
+    ),
+    "vanilla": (
+        lambda x: x.transpose(1, 2),
+        lambda x: x.transpose(1, 2),
+    ),
+}
+
+
+def get_cu_seqlens(text_mask, img_len):
+    """Calculate cu_seqlens_q, cu_seqlens_kv using text_mask and img_len
+
+    Args:
+        text_mask (torch.Tensor): the mask of text
+        img_len (int): the length of image
+
+    Returns:
+        torch.Tensor: the calculated cu_seqlens for flash attention
+    """
+    batch_size = text_mask.shape[0]
+    text_len = text_mask.sum(dim=1)
+    max_len = text_mask.shape[1] + img_len
+
+    cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device="cuda")
+
+    for i in range(batch_size):
+        s = text_len[i] + img_len
+        s1 = i * max_len + s
+        s2 = (i + 1) * max_len
+        cu_seqlens[2 * i + 1] = s1
+        cu_seqlens[2 * i + 2] = s2
+
+    return cu_seqlens
+
+
+def attention(
+    q,
+    k,
+    v,
+    mode="flash",
+    drop_rate=0,
+    attn_mask=None,
+    causal=False,
+    cu_seqlens_q=None,
+    cu_seqlens_kv=None,
+    max_seqlen_q=None,
+    max_seqlen_kv=None,
+    batch_size=1,
+):
+    """
+    Perform QKV self attention.
+
+    Args:
+        q (torch.Tensor): Query tensor with shape [b, s, a, d], where a is the number of heads.
+        k (torch.Tensor): Key tensor with shape [b, s1, a, d]
+        v (torch.Tensor): Value tensor with shape [b, s1, a, d]
+        mode (str): Attention mode. Choose from 'self_flash', 'cross_flash', 'torch', and 'vanilla'.
+        drop_rate (float): Dropout rate in attention map. (default: 0)
+        attn_mask (torch.Tensor): Attention mask with shape [b, s1] (cross_attn), or [b, a, s, s1] (torch or vanilla).
+            (default: None)
+        causal (bool): Whether to use causal attention. (default: False)
+        cu_seqlens_q (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
+            used to index into q.
+        cu_seqlens_kv (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
+            used to index into kv.
+        max_seqlen_q (int): The maximum sequence length in the batch of q.
+        max_seqlen_kv (int): The maximum sequence length in the batch of k and v.
+
+    Returns:
+        torch.Tensor: Output tensor after self attention with shape [b, s, ad]
+    """
+    pre_attn_layout, post_attn_layout = MEMORY_LAYOUT[mode]
+    q = pre_attn_layout(q)
+    k = pre_attn_layout(k)
+    v = pre_attn_layout(v)
+
+    if mode == "torch":
+        if attn_mask is not None and attn_mask.dtype != torch.bool:
+            attn_mask = attn_mask.to(q.dtype)
+        if cu_seqlens_q is None:
+            x = F.scaled_dot_product_attention(
+                q, k, v, attn_mask=attn_mask, dropout_p=drop_rate, is_causal=causal
+            )
+        else:
+            attn1 = F.scaled_dot_product_attention(
+                q[:, :, :cu_seqlens_q[1]],
+                k[:, :, :cu_seqlens_kv[1]],
+                v[:, :, :cu_seqlens_kv[1]],
+                attn_mask=attn_mask,
+                dropout_p=drop_rate,
+                is_causal=causal
+            )
+            attn2 = F.scaled_dot_product_attention(
+                q[:, :, cu_seqlens_q[1]:],
+                k[:, :, cu_seqlens_kv[1]:],
+                v[:, :, cu_seqlens_kv[1]:],
+                attn_mask=None,
+                dropout_p=drop_rate,
+                is_causal=False
+            )
+            x = torch.cat([attn1, attn2], dim=2)
+    elif mode == "flash":
+        x = flash_attn_varlen_func(
+            q,
+            k,
+            v,
+            cu_seqlens_q,
+            cu_seqlens_kv,
+            max_seqlen_q,
+            max_seqlen_kv,
+        )
+        # x with shape [(bxs), a, d]
+        x = x.view(
+            batch_size, max_seqlen_q, x.shape[-2], x.shape[-1]
+        )  # reshape x to [b, s, a, d]
+    elif mode == "vanilla":
+        scale_factor = 1 / math.sqrt(q.size(-1))
+
+        b, a, s, _ = q.shape
+        s1 = k.size(2)
+        attn_bias = torch.zeros(b, a, s, s1, dtype=q.dtype, device=q.device)
+        if causal:
+            # Only applied to self attention
+            assert (
+                attn_mask is None
+            ), "Causal mask and attn_mask cannot be used together"
+            temp_mask = torch.ones(b, a, s, s, dtype=torch.bool, device=q.device).tril(
+                diagonal=0
+            )
+            attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
+            attn_bias.to(q.dtype)
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
+            else:
+                attn_bias += attn_mask
+
+        # TODO: Maybe force q and k to be float32 to avoid numerical overflow
+        attn = (q @ k.transpose(-2, -1)) * scale_factor
+        attn += attn_bias
+        attn = attn.softmax(dim=-1)
+        attn = torch.dropout(attn, p=drop_rate, train=True)
+        x = attn @ v
+    else:
+        raise NotImplementedError(f"Unsupported attention mode: {mode}")
+
+    x = post_attn_layout(x)
+    b, s, a, d = x.shape
+    out = x.reshape(b, s, -1)
+    return out
+
+
+def parallel_attention(
+    hybrid_seq_parallel_attn,
+    q,
+    k,
+    v,
+    img_q_len,
+    img_kv_len,
+    cu_seqlens_q,
+    cu_seqlens_kv
+):
+    attn1 = hybrid_seq_parallel_attn(
+        None,
+        q[:, :img_q_len, :, :],
+        k[:, :img_kv_len, :, :],
+        v[:, :img_kv_len, :, :],
+        dropout_p=0.0,
+        causal=False,
+        joint_tensor_query=q[:,img_q_len:cu_seqlens_q[1]],
+        joint_tensor_key=k[:,img_kv_len:cu_seqlens_kv[1]],
+        joint_tensor_value=v[:,img_kv_len:cu_seqlens_kv[1]],
+        joint_strategy="rear",
+    )
+    if flash_attn.__version__ >= '2.7.0':
+        attn2, *_ = _flash_attn_forward(
+            q[:,cu_seqlens_q[1]:],
+            k[:,cu_seqlens_kv[1]:],
+            v[:,cu_seqlens_kv[1]:],
+            dropout_p=0.0,
+            softmax_scale=q.shape[-1] ** (-0.5),
+            causal=False,
+            window_size_left=-1,
+            window_size_right=-1,
+            softcap=0.0,
+            alibi_slopes=None,
+            return_softmax=False,
+        )
+    else:
+        attn2, *_ = _flash_attn_forward(
+            q[:,cu_seqlens_q[1]:],
+            k[:,cu_seqlens_kv[1]:],
+            v[:,cu_seqlens_kv[1]:],
+            dropout_p=0.0,
+            softmax_scale=q.shape[-1] ** (-0.5),
+            causal=False,
+            window_size=(-1, -1),
+            softcap=0.0,
+            alibi_slopes=None,
+            return_softmax=False,
+        )
+    attn = torch.cat([attn1, attn2], dim=1)
+    b, s, a, d = attn.shape
+    attn = attn.reshape(b, s, -1)
+
+    return attn

hyvideo/modules/fp8_optimization.py

+import os
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+def get_fp_maxval(bits=8, mantissa_bit=3, sign_bits=1):
+    _bits = torch.tensor(bits)
+    _mantissa_bit = torch.tensor(mantissa_bit)
+    _sign_bits = torch.tensor(sign_bits)
+    M = torch.clamp(torch.round(_mantissa_bit), 1, _bits - _sign_bits)
+    E = _bits - _sign_bits - M
+    bias = 2 ** (E - 1) - 1
+    mantissa = 1
+    for i in range(mantissa_bit - 1):
+        mantissa += 1 / (2 ** (i+1))
+    maxval = mantissa * 2 ** (2**E - 1 - bias)
+    return maxval
+
+def quantize_to_fp8(x, bits=8, mantissa_bit=3, sign_bits=1):
+    """
+    Default is E4M3.
+    """
+    bits = torch.tensor(bits)
+    mantissa_bit = torch.tensor(mantissa_bit)
+    sign_bits = torch.tensor(sign_bits)
+    M = torch.clamp(torch.round(mantissa_bit), 1, bits - sign_bits)
+    E = bits - sign_bits - M
+    bias = 2 ** (E - 1) - 1
+    mantissa = 1
+    for i in range(mantissa_bit - 1):
+        mantissa += 1 / (2 ** (i+1))
+    maxval = mantissa * 2 ** (2**E - 1 - bias)
+    minval = - maxval
+    minval = - maxval if sign_bits == 1 else torch.zeros_like(maxval)
+    input_clamp = torch.min(torch.max(x, minval), maxval)
+    log_scales = torch.clamp((torch.floor(torch.log2(torch.abs(input_clamp)) + bias)).detach(), 1.0)
+    log_scales = 2.0 ** (log_scales - M - bias.type(x.dtype))
+    # dequant
+    qdq_out = torch.round(input_clamp / log_scales) * log_scales
+    return qdq_out, log_scales
+
+def fp8_tensor_quant(x, scale, bits=8, mantissa_bit=3, sign_bits=1):
+    for i in range(len(x.shape) - 1):
+        scale = scale.unsqueeze(-1)
+    new_x = x / scale
+    quant_dequant_x, log_scales = quantize_to_fp8(new_x, bits=bits, mantissa_bit=mantissa_bit, sign_bits=sign_bits)
+    return quant_dequant_x, scale, log_scales
+
+def fp8_activation_dequant(qdq_out, scale, dtype):
+    qdq_out = qdq_out.type(dtype)
+    quant_dequant_x = qdq_out * scale.to(dtype)
+    return quant_dequant_x
+
+def fp8_linear_forward(cls, original_dtype, input):
+    weight_dtype = cls.weight.dtype
+    #####
+    if cls.weight.dtype != torch.float8_e4m3fn:
+        maxval = get_fp_maxval()
+        scale = torch.max(torch.abs(cls.weight.flatten())) / maxval
+        linear_weight, scale, log_scales = fp8_tensor_quant(cls.weight, scale)
+        linear_weight = linear_weight.to(torch.float8_e4m3fn)
+        weight_dtype = linear_weight.dtype
+    else:
+        scale = cls.fp8_scale.to(cls.weight.device)
+        linear_weight = cls.weight
+    #####
+
+    if weight_dtype == torch.float8_e4m3fn and cls.weight.sum() != 0:
+        if True or len(input.shape) == 3:
+            cls_dequant = fp8_activation_dequant(linear_weight, scale, original_dtype)
+            if cls.bias != None:
+                output = F.linear(input, cls_dequant, cls.bias)
+            else:
+                output = F.linear(input, cls_dequant)
+            return output
+        else:
+            return cls.original_forward(input.to(original_dtype))
+    else:
+        return cls.original_forward(input)
+
+def convert_fp8_linear(module, dit_weight_path, original_dtype, params_to_keep={}):
+    setattr(module, "fp8_matmul_enabled", True)
+
+    # loading fp8 mapping file
+    fp8_map_path = dit_weight_path.replace('.pt', '_map.pt')
+    if os.path.exists(fp8_map_path):
+        fp8_map = torch.load(fp8_map_path, map_location=lambda storage, loc: storage)
+    else:
+        raise ValueError(f"Invalid fp8_map path: {fp8_map_path}.")
+
+    fp8_layers = []
+    for key, layer in module.named_modules():
+        if isinstance(layer, nn.Linear) and ('double_blocks' in key or 'single_blocks' in key):
+            fp8_layers.append(key)
+            original_forward = layer.forward
+            layer.weight = torch.nn.Parameter(layer.weight.to(torch.float8_e4m3fn))
+            setattr(layer, "fp8_scale", fp8_map[key].to(dtype=original_dtype))
+            setattr(layer, "original_forward", original_forward)
+            setattr(layer, "forward", lambda input, m=layer: fp8_linear_forward(m, original_dtype, input))
+    
+