Skip to content

[Feature] Add HCCS disaggregated prefill example with 2MB-aligned PD … - #294

Open
zhaoyueqiang wants to merge 1 commit into
LMCache:mainfrom
zhaoyueqiang:feat/hccs-pd-example
Open

zhaoyueqiang wants to merge 1 commit into
LMCache:mainfrom
zhaoyueqiang:feat/hccs-pd-example

Conversation

@zhaoyueqiang

Copy link
Copy Markdown

…NPU buffer

  • Add examples/disagg_prefill/hccs: TP=1 PD example over HCCS only (no RoCE), using the hixl transfer channel with HCCS env vars.
  • Align the PD NPU transfer buffer to 2MB in the Ascend PD backend so HCCL/HIXL can register the buffer VA (errors 503900/507899 otherwise). Construct the PagedTensorMemoryAllocator directly from the aligned buffer to keep buffer/free_blocks consistent.

…NPU buffer

- Add examples/disagg_prefill/hccs: TP=1 PD example over HCCS only
  (no RoCE), using the hixl transfer channel with HCCS env vars.
- Align the PD NPU transfer buffer to 2MB in the Ascend PD backend so
  HCCL/HIXL can register the buffer VA (errors 503900/507899 otherwise).
  Construct the PagedTensorMemoryAllocator directly from the aligned
  buffer to keep buffer/free_blocks consistent.
@chloroethylene

Copy link
Copy Markdown
Collaborator

@shengyan-chen review plz

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.

2 participants