ASPIRE: Asynchronous Batch Self-Speculative Decoding

简体中文

L1 batch-aware scheduler and refresh gate, not an integrated vLLM serving engine.

Paper information

Field	Details
Paper	ASPIRE: Asynchronous Batch Self-Speculative Decoding
Institution	University of Southern California (first author).
First public release	2026-09-16 (arXiv v1)
Author code	No public author repository found as of 2026-09-19.
Local adapter or method key	aspire
Local reproduction code	src/auto_research/foundation_latest_20260919.py

Paper method and evidence

Choose each request's draft length from online acceptance and current batch cost. Periodic full-attention refresh corrects the self-speculative state.

flowchart LR
  A[Online acceptance and batch cost] --> D[Per-request draft length]
  D --> R[Refresh gate]
  R --> F[Periodic full-attention correction]

Original paper figure

Figure 2 shows request-level draft scheduling and refresh. From the original paper; copyright remains with the authors.

Core equation

The reference exposes schedule, gate, and state updates in the metrics; variables and invariants are tested.

Reported results

Paper serving, benchmark, and training results are not local measurements.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry. The CUDA path ran on a real NVIDIA A100; see sanitized receipts under docs/gpu-validations/. No vLLM engine modification or paper throughput claim.