[Bugfix][MiniMax-H3] Use Hopper-safe modulation precision
Signed-off-by: david6666666 <530634352@qq.com>
This commit is contained in:
1 parent
3b84cc4a7f
commit
49fa488951
6 files changed
+101
-57
No files matched your search
@@ -1953,22 +1953,7 @@ class MiniMaxH3Pipeline(
|
||||
|
||||
def _encode_local_media(self, media: MiniMaxH3EncoderMediaInput | None) -> MiniMaxH3EncoderMediaConditioning:
|
||||
group, rank, world_size = _dit_rank_world()
|
||||
# Keep image/keyframe encoding on the output rank. The legacy
|
||||
# single-stage path encoded images locally and only used the tiled VAE
|
||||
# collective for reference videos. Treating every visual input as a
|
||||
# distributed video after the unified encoder path was introduced
|
||||
# changes the H100 keyframe latent and therefore the I2VA trajectory.
|
||||
distributed_vae = self.video_vae.is_distributed_enabled()
|
||||
distributed_video = False
|
||||
if distributed_vae:
|
||||
has_video = torch.tensor(
|
||||
[int(media is not None and bool(media.videos))],
|
||||
dtype=torch.long,
|
||||
device=self.device,
|
||||
)
|
||||
if world_size > 1:
|
||||
dist.broadcast(has_video, src=0, group=group)
|
||||
distributed_video = bool(has_video.item())
|
||||
distributed_video = self.video_vae.is_distributed_enabled()
|
||||
if distributed_video:
|
||||
media = self._distribute_media_inputs(media)
|
||||
conditioning = None
|
||||
@@ -1981,7 +1966,6 @@ class MiniMaxH3Pipeline(
|
||||
video_vae=self.video_vae,
|
||||
audio_vae=self.audio_vae,
|
||||
emit_conditioning=rank == 0,
|
||||
encode_images=rank == 0,
|
||||
component_scope=self._component_on_device,
|
||||
)
|
||||
except ValueError as exc:
|
||||
|
||||
@@ -602,13 +602,8 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
|
||||
if getattr(self, "decode_only", False):
|
||||
raise RuntimeError("MiniMax H3 decode-only video VAE cannot encode images")
|
||||
previous_parallel = self.model.parallel_tiling
|
||||
parallel_size = int(getattr(self, "parallel_size", 1))
|
||||
if parallel_size <= 1:
|
||||
if int(getattr(self, "parallel_size", 1)) <= 1:
|
||||
self.model.parallel_tiling = False
|
||||
# Keyframe conditioning follows the official single-image contract.
|
||||
# Do not let it enter the tiled VAE collective even when reference
|
||||
# videos use the same VAE instance with patch parallelism enabled.
|
||||
tiling_context = self._rank_local_tiling() if parallel_size > 1 else nullcontext()
|
||||
parameter = next(self.parameters())
|
||||
previous_dtype = parameter.dtype
|
||||
if previous_dtype != torch.float32:
|
||||
@@ -616,7 +611,7 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
|
||||
devices = [parameter.device] if parameter.device.type != "cpu" else []
|
||||
try:
|
||||
with (
|
||||
tiling_context,
|
||||
self._encoder_tiling_context(image.height, image.width),
|
||||
torch.random.fork_rng(
|
||||
devices=devices,
|
||||
device_type=parameter.device.type,
|
||||
|
||||
Reference in new issue
Block a user