Add multi-GPU vLLM CAA steering hook

[linmou]

Summary

This PR adds a lightweight CAA activation-add hook for vLLM inference, including tensor-parallel worker support.

The hook lets EasyEdit apply already-computed CAA vectors to vLLM-loaded models by installing activation additions on selected decoder layers inside each worker process.

What Changed

• Added steer/vllm_caa_hooks.py

  • install and clear CAA vectors on vLLM model layers
  • worker RPC helpers for tensor-parallel vLLM engines
  • hook call/configuration stats
  • support for common vLLM worker model layouts

• Added tests/test_vllm_caa_hooks.py

  • focused hook-only unit tests with fake vLLM-style models/workers
  • no dataset dependency
  • no LLM judge dependency
  • no generated artifacts

• Added examples/vllm_caa_gpu_e2e.py

  • lightweight optional real-GPU smoke test
  • records baseline, steered, and restored outputs
  • records worker install/clear results and hook stats

• Added docs/vllm_caa_multigpu_hook.md

  • documents runtime API and lightweight validation commands

Validation

pytest -q tests/test_vllm_caa_hooks.py

Result:

9 passed

python -m compileall steer/vllm_caa_hooks.py examples/vllm_caa_gpu_e2e.py tests/test_vllm_caa_hooks.py

Result: passed.

git diff --check main...HEAD

Result: passed.

Optional GPU smoke command:

CUDA_VISIBLE_DEVICES=0,1 \
VLLM_USE_FLASHINFER_SAMPLER=0 \
VLLM_ALLOW_INSECURE_SERIALIZATION=1 \
python examples/vllm_caa_gpu_e2e.py \
  --model /path/to/model \
  --tensor-parallel-size 2 \
  --layer 12 \
  --multiplier 0.0 \
  --vector-value 0.0 \
  --output /tmp/vllm_caa_gpu_e2e.json \
  --monitor-output /tmp/vllm_caa_gpu_e2e.gpu.csv

Notes

The larger steering-effect consistency experiments used during development are not included in this PR. This keeps the runtime contribution focused on the multi-GPU vLLM CAA hook and lightweight validation.

Related Issue

Closes #695