Skip to content

Support Torch Cpp Runtime for HSTU#294

Draft
geoffreyQiu wants to merge 2 commits intoNVIDIA:mainfrom
geoffreyQiu:cpp_runtime
Draft

Support Torch Cpp Runtime for HSTU#294
geoffreyQiu wants to merge 2 commits intoNVIDIA:mainfrom
geoffreyQiu:cpp_runtime

Conversation

@geoffreyQiu
Copy link
Collaborator

Description

Support Torch Cpp Runtime for HSTU

Checklist

  • I am familiar with the Contributing Guidelines.
  • New or existing tests cover these changes.
  • The documentation is up to date with these changes.

@geoffreyQiu geoffreyQiu changed the title Init apply_inference_embedding implementation Support Torch Cpp Runtime for HSTU Feb 5, 2026
for idx, embc in enumerate(embc_modulelist):
emb_backend = embc_configs[idx].backend
create_embedding_collection = select_embedding(emb_backend)
embc_modulelist[idx] = create_embedding_collection(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

do we need to uset setattr here? like reference

@shijieliu
Copy link
Collaborator

the overall arch looks good to me. I think you can create a inference directory under dynamicemb and move the inference embedding related code under it.

embedding_configs,
embedding_backend,
sparse_shareables,
def create_torchrec_embedding(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maybe we dont need to support this.

)
model_state_dict = torch.load(model_state_dict_path)["model_state_dict"]
self.load_state_dict(model_state_dict, strict=False)
pass
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

for load dump, we can reuse DynamicEmbLoad/DynamicEmbDump for dynamicemb checkpoint

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants