github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix audio loss when doing do_guidance_loss

Browse Source
This commit is contained in:
Jaret Burkett
2026-08-04 21:43:31 -06:00
parent d870e9b68a
commit 0f9094db95
5 changed files with 132 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
extensions_built_in/diffusion_models/ltx2/ltx2.py
View File

@@ -852,6 +852,14 @@ class LTX2Model(BaseModel):
batch: "DataLoaderBatchDTO" = None,
**kwargs,
):
# the primary (loss carrying) prediction is the first one made with grad
# enabled. Prior/cfg/guidance passes run under no_grad, and the
# preservation pass (diff_output_preservation, blank_prompt_preservation)
# runs with grad but after the loss, so it must not restate the audio
# prediction the primary pass stored on the batch.
is_primary_pred = (
torch.is_grad_enabled() and batch is not None and batch.audio_pred is None
)
with torch.no_grad():
if self.model.device == torch.device("cpu"):
self.model.to(self.device_torch)
@@ -945,8 +953,21 @@ class LTX2Model(BaseModel):
audio_num_frames = raw_audio_latents.shape[1]
# add the audio targets to the batch for loss calculation later
audio_noise = torch.randn_like(raw_audio_latents)
batch.audio_target = (audio_noise - raw_audio_latents).detach()
# the audio noise is drawn once per step and shared by every
# pass (prior, primary, cfg/guidance, preservation) so they all
# see the same soundtrack and the stored target keeps matching
if (
batch.audio_noise is not None
and batch.audio_noise.shape == raw_audio_latents.shape
):
audio_noise = batch.audio_noise.to(
raw_audio_latents.device, dtype=raw_audio_latents.dtype
)
else:
audio_noise = torch.randn_like(raw_audio_latents)
batch.audio_noise = audio_noise
if batch.audio_target is None:
batch.audio_target = (audio_noise - raw_audio_latents).detach()
audio_latents = self.add_noise(
raw_audio_latents,
audio_noise,
@@ -1040,7 +1061,10 @@ class LTX2Model(BaseModel):
# add audio latent to batch if we had audio
if batch.audio_target is not None:
batch.audio_pred = noise_pred_audio
if is_primary_pred:
batch.audio_pred = noise_pred_audio
else:
batch.audio_pred_uncond = noise_pred_audio
unpacked_output = self.pipeline._unpack_latents(
latents=noise_pred_video,

48
extensions_built_in/diffusion_models/minimax_h3/minimax_h3.py
View File

@@ -141,6 +141,13 @@ class MiniMaxH3VaeBundle(torch.nn.Module):
def dtype(self):
return self.video_vae.dtype
def enable_gradient_checkpointing(self, enable: bool = True):
self.video_vae.enable_gradient_checkpointing(enable)
self.audio_vae.enable_gradient_checkpointing(enable)
def disable_gradient_checkpointing(self):
self.enable_gradient_checkpointing(False)
class MinimaxH3Model(BaseModel):
arch = "minimax_h3"
@@ -712,6 +719,15 @@ class MinimaxH3Model(BaseModel):
if self.model.device == torch.device("cpu"):
self.model.to(device)
# the primary (loss carrying) prediction is the first one made with grad
# enabled. Prior/cfg/guidance passes run under no_grad, and the
# preservation pass (diff_output_preservation, blank_prompt_preservation)
# runs with grad but after the loss, so it must not restate the audio
# prediction the primary pass stored on the batch.
is_primary_pred = (
torch.is_grad_enabled() and batch is not None and batch.audio_pred is None
)
batch_size, _, t_lat, h_lat, w_lat = latent_model_input.shape
with torch.no_grad():
@@ -778,16 +794,27 @@ class MinimaxH3Model(BaseModel):
raw_audio = torch.nn.functional.pad(
raw_audio, (0, 0, 0, expected_rows - raw_audio.shape[1])
)
audio_noise = torch.randn_like(raw_audio)
# model predicts clean - noise; audio_pred is negated below so
# the stored target follows ai-toolkit's noise - clean
batch.audio_target = (audio_noise - raw_audio).detach()
# the audio noise is drawn once per step and shared by every
# pass (prior, primary, cfg/guidance, preservation) so they all
# see the same soundtrack and the stored target keeps matching
if (
batch.audio_noise is not None
and batch.audio_noise.shape == raw_audio.shape
):
audio_noise = batch.audio_noise.to(device, torch.float32)
else:
audio_noise = torch.randn_like(raw_audio)
batch.audio_noise = audio_noise
audio_rows = (1.0 - sa) * raw_audio + sa * audio_noise
# expose what audio perceptual losses need to rebuild the
# clean estimate (x0 = noisy - sigma_a * pred) and its target
batch.audio_latents = raw_audio
batch.audio_noisy = audio_rows
batch.audio_sigma = sigma_a
if batch.audio_target is None:
# model predicts clean - noise; audio_pred is negated below
# so the stored target follows ai-toolkit's noise - clean
batch.audio_target = (audio_noise - raw_audio).detach()
# expose what audio perceptual losses need to rebuild the
# clean estimate (x0 = noisy - sigma_a * pred)
batch.audio_noisy = audio_rows
batch.audio_sigma = sigma_a
else:
# no soundtrack: silence (zeros) noised at the audio sigma
# rides along without contributing to the loss
@@ -863,7 +890,10 @@ class MinimaxH3Model(BaseModel):
if batch is not None and batch.audio_target is not None:
# flip to ai-toolkit's noise - clean convention
batch.audio_pred = -audio_pred
if is_primary_pred:
batch.audio_pred = -audio_pred
else:
batch.audio_pred_uncond = -audio_pred
video_pred = video_pred[:, num_cond:]
noise_pred = unpatchify_video_tokens(video_pred, t_lat, h_lat, w_lat)

31
extensions_built_in/diffusion_models/minimax_h3/src/audio_vae.py
View File

@@ -20,6 +20,7 @@ import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from torch.utils.checkpoint import checkpoint
# fmt: off
# Per-channel latent statistics from the released FL2VA/audio_vae/config.json.
@@ -395,15 +396,25 @@ class BigVGANDecoder(nn.Module):
self.activation_post = AliasFreeActivation1d(SnakeBeta(channels))
self.conv_post = nn.Conv1d(channels, 1, 7, padding=3, bias=False)
self.gradient_checkpointing = True
def _stage(self, i: int, x: Tensor) -> Tensor:
x = self.ups[i][0](x)
acc = None
for j in range(self.num_kernels):
y = self.resblocks[i * self.num_kernels + j](x)
acc = y if acc is None else acc + y
return acc / self.num_kernels
def forward(self, x: Tensor) -> Tensor:
x = self.conv_pre(x)
for i, up in enumerate(self.ups):
x = up[0](x)
acc = None
for j in range(self.num_kernels):
y = self.resblocks[i * self.num_kernels + j](x)
acc = y if acc is None else acc + y
x = acc / self.num_kernels
for i in range(len(self.ups)):
# gated on is_grad_enabled (not train mode): the VAE stays eval
# but differentiable decodes during training should checkpoint
if torch.is_grad_enabled() and self.gradient_checkpointing:
x = checkpoint(self._stage, i, x, use_reentrant=False)
else:
x = self._stage(i, x)
x = self.activation_post(x)
x = self.conv_post(x)
return torch.clamp(x, min=-1.0, max=1.0)
@@ -462,6 +473,12 @@ class MiniMaxH3AudioVAE(nn.Module):
def downsampling_ratio(self) -> int:
return self.HOP_LENGTH
def enable_gradient_checkpointing(self, enable: bool = True):
self.decoder.gradient_checkpointing = enable
def disable_gradient_checkpointing(self):
self.enable_gradient_checkpointing(False)
def _apply(self, fn, recurse=True):
# This VAE is pinned to fp32 (bf16 decodes are audibly degraded).
# Device moves pass through, but any cast that would land a float