On-Device Multimodal Runtime Verification Audit
Audit on-device vision-language models for Android NPU execution, memory budgets, and thermal stability.
Use this checklist when deploying compact multimodal foundation models onto consumer Android devices via LiteRT or ExecuTorch. It validates NPU delegate compatibility, sensor buffer allocation, and memory pressure limits.
Role: Principal On-Device ML Android Architect specializing in edge vision-language models and hardware acceleration.
Context
- Target NPU and SoC architecture: {{target_npu_soc}}
- Vision-language model variant: {{vision_model_variant}}
- Maximum runtime memory headroom: {{memory_headroom_mb}}
- Model quantization and weight format: {{quantization_format}}
- Target camera ingestion throughput: {{min_frame_rate_fps}}
- Android CameraX image stream format: {{camera_stream_format}}
Task
Deliver an exhaustive technical verification checklist to evaluate the on-device inference pipeline of {{vision_model_variant}} on {{target_npu_soc}}, ensuring zero frame drops, strict memory compliance within {{memory_headroom_mb}}, and stable hardware acceleration.
Method
- Analyze NPU operator mapping and delegate fallbacks for {{vision_model_variant}} under {{quantization_format}}.
- Verify zero-copy buffer sharing between {{camera_stream_format}} and the vision encoder tensor input.
- Profile native memory allocations against the designated limit of {{memory_headroom_mb}} during concurrent token generation.
- Validate KV-cache management and paging strategies under prolonged conversational turns.
- Audit thermal throttling curve and sustained performance at {{min_frame_rate_fps}} over 15-minute operational sessions.
- Evaluate multi-threading and CPU affinity configurations for asynchronous tokenizer execution.
- Formulate pass/fail criteria for fallback mechanisms when NPU acceleration is unavailable.
Constraints
- Checkpoints MUST specify explicit memory offsets, API boundaries, or native NDK flags.
- Verification items MUST NOT accept synthetic desktop benchmark approximations.
- All items must differentiate between fatal runtime blockers and non-critical performance optimizations.
- The checklist must strictly cover edge inference without dependencies on remote cloud fallbacks.
Output format
- Section 1: Hardware Acceleration and Operator Support (5 items)
- Section 2: Memory Footprint and Direct Buffer Management (5 items)
- Section 3: CameraX Ingestion and Tokenization Pipeline (4 items)
- Section 4: Thermal, Power, and Graceful Degradation (4 items)
Each item must follow the format:
[ ] [CRITICALITY] Verification Objective: Concrete test procedure and pass criteria.
Self-review
- Confirm all 6 input variables are directly referenced across the criteria.
- Verify every checklist item has explicit, measurable pass/fail thresholds.
- Ensure zero ambiguous guidance or generalized Android platform suggestions.
Explicit role, a named task, and discrete steps the model can follow.
Background, inputs and variables the model needs before it starts.
Hard boundaries — what the model must and must not do.
A named, field-level shape for the response.
Ordered work items that force analysis before an answer.
Length and structure that travel across frontier models.
Signal density — instruction weight without padding.
Documented variables so the scaffold adapts to new inputs.
Quality bar, assumptions and behaviour when inputs are thin.
How much real usage the template has behind it.