[Performance] Optimize MammothModa2 AR → DiT hidden-state transfer (#7102)

Signed-off-by: kunkunblueberry <1833921874@qq.com>
Signed-off-by: kunkun <72174834+kunkunblueberry@users.noreply.github.com>
This commit is contained in:
kunkun authored and GitHub committed 2026-09-14 15:25:08 -07:00
1 parent 9f85b3722e
commit c68eadbbe0
4 files changed
+212 -5

No files matched your search

@@ -33,6 +33,18 @@ def test_tensor_serialization():
assert torch.equal(tensor, deserialized)
def test_bfloat16_tensor_serialization():
"""The raw tensor serializer preserves the Mammoth BF16 wire dtype."""
import torch
tensor = torch.tensor([[1.0, -2.5], [3.25, 4.5]], dtype=torch.bfloat16)
serialized = OmniSerializer.serialize({"text_prompt_embeds": tensor})
deserialized = OmniSerializer.deserialize(serialized)
assert deserialized["text_prompt_embeds"].dtype == torch.bfloat16
assert torch.equal(deserialized["text_prompt_embeds"], tensor)
def test_noncontiguous_tensor_serialization():
"""Test torch.Tensor serialization for strided views."""
import torch