support hybrid attention for vllm connector && add integration tests - #257

Open

lpdink wants to merge 56 commits into

mainfrom

feature/vllm-hybrid-attention-rewrite

Commits on Jul 29, 2026

[py_connector] rewrite vLLM connector around per-group transfer for hybrid attention
lpdink
committed
3bd6ea0
[integration_test] add vLLM e2e KV verification for full + hybrid attention
lpdink
committed
991d0b5
[py_connector] cap full-prompt external hit and fix canceled-save KeyError
lpdink
committed
f702cd6
[py_connector] add unit tests for translation, transfer results and scheduler state
lpdink
committed
e0718d9
[py_connector] fix load-failure retry loop and null mamba state transfer
lpdink
committed
d462dcb
[integration_test] harden vllm e2e harness and fix hybrid prompt coverage
lpdink
committed
645e58a
[integration_test] add mutation meta-test and four vllm e2e scenarios
lpdink
committed
f909e82
[py_connector] make connector unit tests runnable without torch
lpdink
committed
4052b4c
[integration_test] drop dev-machine assumptions from the vllm e2e harness
lpdink
committed
729e95f
[py_connector] restore pre-0.26 vLLM KV layouts via shape-based detection
lpdink
committed
9dc5961

Commits on Jul 30, 2026

Commits on Jul 31, 2026

Commits on Aug 17, 2026

Commits on Aug 18, 2026

Commits on Aug 19, 2026

[py_connector] dedupe in-flight getCacheLocation re-asks
lpdink
committed
05714af

Commits on Aug 20, 2026

Commits on Aug 25, 2026

Commits on Aug 26, 2026

Commits on Aug 27, 2026

[py_connector] make the test stubs survive requests subclassing (CI red fix)
lpdink
committed
10bf0e8