When the requested (mode, scale) is mathematically equivalent to vanilla
sampling -- centered_exp/value_bias scale=0, mul scale<=0 or scale==1,
or other expectation modes at scale==1 -- _req_wants_value_guidance now
returns False so the entire LVM path is skipped for that request.

The paper's neutral config (centered_exp scale=0) is exactly such a
no-op tilt: exp(0 * value) == 1 for all candidates, so the LVM forward
runs but contributes nothing to token selection. Skipping it should
drop wall_clock from ~72 s back to vanilla (~20 s) at scale=0, with
zero change to token-selection behavior.

Hard-constraint requests (target_value/target_length/value_constraint/
cmp/op) always need LVM, regardless of scale. Per-req result is cached
on the Req via the existing _lvm_wants_guidance slot.

Adds _get_req_value_mode_and_scale helper that caches (mode, scale)
parsing per req, keyed on custom_params identity + entries.

Co-Authored-By: Zhen Zhang <namezhenzhang@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Cherry-picked from upstream PR UCSB-AI#3 (namezhenzhang):

* lvm_value_utils.get_eos_token_ids: cache result on req._lvm_eos_token_ids;
  was re-walking sets every candidate filter call.
* qwen2_lvm.py / qwen3_lvm.py / qwen2_5_vl_lvm.py: drop two
  forward_batch.{extend_seq_lens,extend_prefix_lens}.tolist() calls per
  LVM forward by carrying tree_value_cached_prefix_lens in the
  TreeValueSpecInput. With ~1300 LVM forwards in the paper run, that
  removes ~2600 GPU->CPU syncs.
* tree_value_spec.py: vectorize the candidate self-attention diagonal
  using numpy fancy indexing instead of a per-token Python loop.

No algorithm change; same forward outputs.

Co-Authored-By: Zhen Zhang <namezhenzhang@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…CSB-AI#3)

lvm_inproc_runner.py:
  - eval_candidates_batch_gpu now accepts candidate_lens_per_req so
    the runner can skip re-walking candidate_ids_per_req lists.
  - Add extend_and_eval_candidates_batch_gpu(): in steady-state decode
    each request adds at most 1 prefix token, so we can run prefix-extend
    + candidate-scoring in ONE forward (Q_len = 1 + k) with the right
    tree mask. Returns None for unsupported configurations (VLM, page_size
    != 1, or any per-req prefix delta > 1 token) so callers can fall back.

lvm_guided_sampling.py:
  - PendingLvmResult gains candidate_lens_send so build_pending can pass
    per-row valid counts down without recomputing.
  - _Inproc adds tree_value_extend_and_launch_gpu wrapper that schedules
    the fused kernel on lvm_stream.
  - apply() GPU path now tries the fused kernel first; on None fall-back
    runs the original two-phase tree_value_extend + tree_value_launch_gpu.
  - timer.set_meta(lvm_fused_path=int(fused)) so we can quantify how
    often fusion actually triggers.

Our build_pending CPU/sync trim + vectorized scatter + per-req cache
are all preserved; the fused path is opt-in on the GPU branch only.

Co-Authored-By: Zhen Zhang <namezhenzhang@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>