Skip to content

Wzj demo - #1628

Open
hiworldwzj wants to merge 9 commits into
mainfrom
wzj_demo
Open

hiworldwzj wants to merge 9 commits into
mainfrom
wzj_demo

Conversation

@hiworldwzj

Copy link
Copy Markdown
Collaborator

No description provided.

niushengxiao and others added 9 commits October 13, 2026 17:52
Maintain request token mappings in pinned CPU memory and upload pending ranges before inference. Fuse prefill and decode padding with Triton, reduce CPU input construction overhead, and mark synchronization sites for further analysis.

Validation: 128 tests passed across padding kernels, model inputs and outputs, MTP batch alignment, CUDA Graph replay, and overlap paths.
Restore blank-line spacing from the PR base without changing code behavior.
Keep pinned CPU request indexes local and publish pending GPU indexes before DP cache transfers. Reuse GPU index tables for cache loading and offloading, and cache prompt ID buffers per CUDA stream to avoid unnecessary allocation and synchronization.

Document the lifecycle and cross-stream risks of borrowing inference index tables. Update regression tests for stream isolation, async counting, DP index publication, and cache loading.

Validation: 42 focused tests passed; git diff --check passed.
Replace linked nodes with a deque while preserving allocation order, recently freed index reuse, capacity tracking, and duplicate-release checks.

Validation: 51 request-manager tests passed; formatting and git diff --check passed.
Rely on the inference entrypoint stream dependency to order first-token writes before GPU reads. Document that CPU request updates do not require waiting for the CUDA write and that future stream changes must preserve this dependency.

Validation: 19 PD and paged-allocation tests passed; git diff --check passed.
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.

1 participant