On-device multimodal agents are limited by two deployment mismatches: pretrained attention structures may not divide evenly across the compute cores of an edge accelerator, causing irregular partitions; and standard prefix caching keeps historical visual tokens in the persistent KV cache, whose capacity grows with the number of rounds. We present PARC, which combines parallelism-aligned structured pruning with lifetime-aware KV-cache reorganization. At the model level, it uses first-order Taylor importance and grouped-query-attention-aware selection to make attention dimensions divisible by the target core count while preserving the query-to-key/value ratio and head dimension. At runtime, it separates the system prompt, persistent text history, and transient visual observations by lifetime, updating text history during tool execution so that historical images leave the persistent cache and more interaction rounds fit a fixed budget. On 2,062 GUI actions, the incremental interface shows no accuracy regression, and task-aware recovery training restores 98.7% of the action-type accuracy gap and the coordinate hit rate to the original level. Under seven-core deployment, hardware-aligned pruning reduces vision-encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers per-heap high-water memory by 12.7%; a parameter-matched control confirms that the speedup stems mainly from core alignment rather than parameter reduction.
Figures & tables
Figure 1: Overview of PARC. Model dimensions are aligned with available cores, while fixed prompts and round-specific inputs are managed according to their lifetimes.
Figure 2: Illustrative non-divisible and aligned mappings. The method generalizes to other core counts and partition granularities.
Figure 3: Alternating visible inference and masked text-only KV-cache reconstruction.
Model
NPU Cores
Interface
Type Acc.
Coord. Hit
Exact Match
Original
8
Baseline
83.4
85.2 (757)
77.9
Original
7
One-shot
83.1
85.7 (759)
77.3
Original
7
Multi-round GUI
83.1
85.7 (759)
77.4
Pruned
7
One-shot
67.1
36.9 (759)
43.8
Pruned
7
Multi-round GUI
67.4
36.4 (758)
43.8
Recovered
7
One-shot
83.0
86.3 (759)
71.5
Table 1: Accuracy on 2,062 GUI Actions
Figure 4: Incremental-interface accuracy before pruning, after direct pruning, and after recovery training. Coordinate hit rate is conditional on a valid coordinate prediction.
Model
NPU
Params
System prompt
Vision
ViT
Prefill
Decode
cores
(s)
input
(ms)
(tok/s)
(tok/s)
Original
8
4.4378B
0.848
640×352
184.84
729.59
79.17
Original
7
4.4378B
1.24
640×352
223.67
650.59
62.32
Aligned pruned
7
4.31B
1.04
640×352
198.28
651.04
77.12
FFN-pruned (non-aligned)
7
4.3140B
1.23
640×352
225.72
655.00
62.89
Table 2: RKNN Model Inference Performance
Heap
Original (MB)
Pruned (MB)
Reduction (%)
n0
524.93
470.39
10.4
n1
476.71
416.72
12.6
n2
480.98
415.42
13.6
n3
481.64
410.54
14.8
n4
474.45
410.56
13.5
n5
468.04
409.93
12.4
Table 3: NPU Heap High-Water Usage During Inference