PARC
A deployment optimization framework for on-device multimodal agents. Through parallel-aligned structured pruning and turn-aware KV cache lifetime management, it addresses both hardware core alignment and cache bloat.
Timeline (last 90 days)
PARC (Parallelism-Aligned Pruning and Round-Aware Context Reuse) is a research method from rss aimed at resolving the mismatch in on-device multimodal Agent deployment at both the model structure and runtime cache levels.
PARC's structured pruning constraint requires that the pruned attention dimension be divisible by the target core count while keeping the query-to-key/value ratio and the per-head dimension unchanged.
PARC's parameter-matched control experiments show that, at equal parameter counts, only the pruned version aligned to the core count obtains speed gains, indicating that the performance improvement comes mainly from core alignment rather than mere parameter reduction.
PARC splits KV cache content into three categories by lifetime: system prompt, persistent text history, and transient visual observations.
PARC synchronously updates text history when executing tool calls and moves historical images out of the persistent cache, thereby increasing the number of interaction rounds that fit within a fixed memory budget.
The study evaluated the incremental interface on 2,062 GUI actions and observed no accuracy regression.
Task-aware recovery training restored the action-type accuracy gap to 98.7% and recovered the coordinate hit rate to its original level.
Under seven-core deployment conditions, PARC's hardware-aligned pruning reduces vision encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers peak memory per stack by 12.7%.
Recovery training recovers most of the accuracy loss caused by pruning, indicating that structured pruning combined with targeted fine-tuning is a viable path for on-device deployment.
All Facts (9)
PARC (Parallelism-Aligned Pruning and Round-Aware Context Reuse) is a research method from rss aimed at resolving the mismatch in on-device multimodal Agent deployment at both the model structure and runtime cache levels.
50%UnverifiedPARC's structured pruning constraint requires that the pruned attention dimension be divisible by the target core count while keeping the query-to-key/value ratio and the per-head dimension unchanged.
50%UnverifiedPARC's parameter-matched control experiments show that, at equal parameter counts, only the pruned version aligned to the core count obtains speed gains, indicating that the performance improvement comes mainly from core alignment rather than mere parameter reduction.
50%UnverifiedPARC splits KV cache content into three categories by lifetime: system prompt, persistent text history, and transient visual observations.
50%UnverifiedPARC synchronously updates text history when executing tool calls and moves historical images out of the persistent cache, thereby increasing the number of interaction rounds that fit within a fixed memory budget.
50%UnverifiedThe study evaluated the incremental interface on 2,062 GUI actions and observed no accuracy regression.
50%UnverifiedTask-aware recovery training restored the action-type accuracy gap to 98.7% and recovered the coordinate hit rate to its original level.
50%UnverifiedUnder seven-core deployment conditions, PARC's hardware-aligned pruning reduces vision encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers peak memory per stack by 12.7%.
50%UnverifiedRecovery training recovers most of the accuracy loss caused by pruning, indicating that structured pruning combined with targeted fine-tuning is a viable path for on-device deployment.
50%