smdesai/GLiNER25-Decide-FP16-CoreAI

GLiNER2.5-Decide — Core AI (iOS 27+ / macOS 27+)

File	Purpose
GLiNER25-Decide-FP16-CoreAI.aimodel/	Core AI source asset, FP16 (~873 MB), three functions sharing one set of weights
tokenizer.json, tokenizer_config.json, special_tokens_map.json	Checkpoint tokenizer
model-card.md	Upstream model card (fastino/GLiNER2.5-Decide, Apache-2.0)

Source: checkpoint revision bbe10ff77ebb238777c17d3a8ac9260e30929057, classification-only (DeBERTa encoder + classifier). Exported with coreai-torch 0.4.3 / coreai-core 1.0.0b3 / Torch 2.11.0. The graph uses constant relative-position buckets, gather-free relative-shift bias and fused SDPA. Learned weights are unchanged.

Interface

Functions context128, context256, context512:

• inputs: input_ids int32 [1, L], attention_mask int32 [1, L] (1 = real token, 0 = padding)
• output: logits [1, L], one score per token position. Classification reads the positions of the schema's label markers ([L]).

Choose the smallest function that fits the request and pad to its length. Reject longer requests; do not truncate.

Required: GPU placement

let model = try await AIModel(
  contentsOf: url, options: SpecializationOptions(preferredComputeUnitKind: .gpu))

Default placement was unsafe for this graph on iOS 27.2 with the previous export (coreai-torch 0.4.1). It crashed with an MPSGraph ANE-region assertion, and earlier assets showed per-launch nondeterminism at 128. Default placement has not been retested with this export; use GPU.

Validation (iPhone 17 Pro, iOS 27.2, GPU)

Strict gate: every decision matches the FP32 PyTorch oracle and max probability error ≤ 0.005. Two launches, all functions: pass every launch, max error 0.0010, no decision mismatches. Corpora: 75 (128), 86 (256), 95 (512, including 9 requests of 296–512 tokens).

Function	Median, launch 1	Median, launch 2
context128	34.6 ms	36.0 ms
context256	49.3 ms	53.6 ms
context512	111.6 ms	115.9 ms

The first load specializes the asset for the device (~8 s); later loads hit the system cache. After ~2 s idle, add ≈ 45–50 ms (GPU wake-up).

SHA-256

038e5b01e0cbeb42661b12abde9ae20482b849e41a50dd477db264895ee5e708  GLiNER25-Decide-FP16-CoreAI.aimodel/main.mlirb
3ad87d9ffe669147063e70850927dd2da90249e2acc5c8527f1eb65df467bcc8  tokenizer.json