mlboydaisuke commited on
Commit
87924d3
·
verified ·
1 Parent(s): 3f0e133

card: Neural Engine bundle section

Browse files
Files changed (1) hide show
  1. README.md +15 -0
README.md CHANGED
@@ -86,6 +86,21 @@ conversation history; `streamResponse(to:)` yields tokens as they decode.
86
  - Measure in Release — Debug is ~3× slower on per-token host work
87
  <!-- gen-cards:use-it end -->
88
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
  ## Measured
90
 
91
  | | decode | prefill | numerics | size |
 
86
  - Measure in Release — Debug is ~3× slower on per-token host work
87
  <!-- gen-cards:use-it end -->
88
 
89
+ ## Neural Engine bundle (iOS static export, AOT h18p) — 2026-09-15
90
+
91
+ `ios-ane-h18p/` is Apple's stock `coreai.llm.export --platform iOS` static export of this checkpoint
92
+ (k-means **8-bit palettized, group 32** (`conversion/minicpm5_pal8_g32.yaml`); embeddings int8; static graphs `prompt_opt`/`extend` ×
93
+ contexts {256, 512, 1024, 2048, 4096} × query {8, 16, 64}), AOT-compiled with
94
+ `xcrun coreai-build compile --platform iOS --preferred-compute neural-engine --architecture h18p`
95
+ (31/31 ANE regions, 1.3 GB). It loads through Apple's `EngineFactory` → `StaticShapeEngine` unchanged
96
+ (iPhone 17-class devices). `ios-static/` is the same export before AOT — the portable IR; compile it
97
+ for another chip yourself. Apple's default 4-bit preset **fails** the same gate on this checkpoint (two margin-clear flips on a chat turn, unchanged with fp16 embeddings; `gate-minicpm5-1b-ane-device-4bit-FAIL.json`), so the 1B ships at 8 bits.
98
+
99
+ Gated on an iPhone 17 Pro (iOS 27.0) against the fp32 HF oracle: teacher-forced single-step sweep 24/24 + 10/10 + 16/16 and free-run greedy token-exact including the stop — transcript in the zoo,
100
+ `models/minicpm5-1b/gate-minicpm5-1b-ane-device.json`. Speed on the same phone (Apple llm-benchmark
101
+ method, 512-token prompt, 1024 generated, 5 trials, two back-to-back runs): decode **69.6 / 58.6 tok/s**, prefill 2710 / 2364 tok/s; footprint 1.3 GB (p128 / g256: decode 62.3). The Neural Engine is the power-efficient lane; these are
102
+ not a GPU comparison (different protocol from the `int8/` row above).
103
+
104
  ## Measured
105
 
106
  | | decode | prefill | numerics | size |