[JulienDarve]

Summary

Let vllm-rs, the Rust OpenAI-compatible front end, compute each request's prompt prefix-cache block hashes and send them
with the request, as the Python front end already does with FEH=1 (#72) and VLLM_FEH_MEMO=1 (#141). Without this,
requests from vllm-rs reach EngineCore unhashed and EngineCore hashes every prompt itself. The Dynamo vLLM sidecar
deployment of Qwen3.6 depends on it.

Design

• EngineCore reports the hashing it accepts in its ready response through a new optional field,
  native_frontend_hash_config: algorithm sha256-pickle5-feh1, block size and none_hash. It is set only when FEH is
  enabled, prefix caching uses sha256 with pickle protocol 5, and the hash block size equals FEH's block
  (vllm/v1/engine/frontend_block_hashing.py, vllm/v1/engine/core.py).
• vllm-rs (rust/src/engine-core-client/src/frontend_hashing.rs) hashes only with VLLM_NATIVE_FEH=1 and when every
  engine reports the same config. Otherwise it logs why and sends requests unhashed, as before.
• Hashes are the same pickle-5 SHA-256 block hashes as the Python front end, including cache salt and partial blocks.
  A content-keyed memo with 1024-token checkpoints reuses hashes across turns that share a prefix, like VLLM_FEH_MEMO.
  Hashing runs on a blocking thread, off the async request path.
• vllm/utils/hashing.py gains a SHA256_PICKLE_PROTOCOL constant. The edits to
  tests/v1/engine/test_frontend_block_hashing.py and the blank-line moves in vllm/v1/engine/core.py are
  ruff formatting only.

Configuration

Env key	Read by	Default	Effect
VLLM_NATIVE_FEH	vllm-rs	unset (off)	1 enables native hashing when EngineCore advertises a matching config.

FEH and VLLM_FEH_MEMO are unchanged.

Numerics

Exact by design. The hashes are bit-identical to the Python front end's, which unit tests check against CPython pickle-5
output. Prefix-cache behavior is unchanged.

Testing

• cargo test -p vllm-engine-core-client: 120 passed, including 6 frontend_hashing tests (the rust/ tree is
  identical to the qwen-v2 PR).
• pytest tests/v1/engine/test_frontend_block_hashing.py: 18 passed.
• pre-commit run --files <changed files>: no new findings. Pre-existing on the base: ruff E501 at core.py:268,
  ruff format, mypy (MainThreadRepin assignments, frontend_block_hashing.py:237) and check-root-lazy-imports
  (vllm.lcd_pdl).
• End to end: based on the current head 47a8eced, so it includes the built-in GB300 serving defaults. End-to-end
  validation of this exact commit behind Dynamo vLLM sidecars is in progress; the same change without the defaults
  commit (JulienDarve/vllm ef527ad0) was validated in our Qwen3.6 GB300 serving runs. Results are not shared here.

Not a duplicate

#72 and #141 add this hashing and its memo to the Python front end; this ports the same hashing to vllm-rs. No open
CentML PR touches rust/src/engine-core-client.

Merging

Please merge with a merge commit, not squash or rebase. Our benchmark harness pins this commit ID (f9e88e1f).

AI assistance

AI assistance was used: Codex for the implementation and Claude for preparing this PR. I reviewed every changed line
and the test results above.