[Bugfix][MiniMax-H3] Stabilize keyframe VAE encoding (#7191)

Signed-off-by: david6666666 <530634352@qq.com>
Signed-off-by: David Chen <530634352@qq.com>
This commit is contained in:
WeiQing Chen authored and GitHub committed 2026-09-07 21:52:54 +08:00
1 parent 4b96c57730
commit 30d6a0b4e6
2 files changed
+118 -4

No files matched your search

+26 -4
View File
@@ -38,6 +38,27 @@ MINIMAX_H3_AUDIO_CHANNELS = 2
logger = init_logger(__name__)
@contextmanager
def _minimax_h3_keyframe_encode_context(
device: torch.device,
) -> Iterator[None]:
if device.type != "cuda":
yield
return
# The official keyframe latent uses cuDNN's TF32 convolution path. The
# default non-deterministic algorithm can select numerically different
# reductions on H100s, and the difference is amplified by the denoiser.
# Pin both the algorithm and math mode for this sensitive encode only.
with torch.backends.cudnn.flags(
enabled=True,
benchmark=False,
deterministic=True,
allow_tf32=True,
):
yield
def _load_component_config(component_path: str) -> dict[str, Any]:
config_path = Path(component_path) / "config.json"
config = json.loads(config_path.read_text(encoding="utf-8"))
@@ -280,10 +301,11 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
for device in devices:
with self.device_module.device(device):
self.device_module.manual_seed(MINIMAX_H3_KEYFRAME_ENCODE_SEED)
latent = self.model.encode_images(
image,
use_fp16_latent=True,
)[0]
with _minimax_h3_keyframe_encode_context(parameter.device):
latent = self.model.encode_images(
image,
use_fp16_latent=True,
)[0]
finally:
self.model.parallel_tiling = previous_parallel
if previous_dtype != torch.float32: