MinerU2.5-Pro -> Core AI: vision (fp16) + decoder (int8lin S=1) + tokenizer

README.md

@@ -38,11 +38,13 @@ Apache-2.0.
 
 ## Verified (M4 Max, GPU, Core AI pipelined engine)
 
-- **End-to-end real generation on the engine: the sample page read verbatim** — GPU vision `.aimodel`
-  → `image_embeds` → AOT-compiled (h16c) int8lin S=1 decoder, autoregressive greedy: title +
-  paragraph + the **full HTML-less table reconstructed row-by-row** (*"Quarterly Report / On-device
-  inference shipped across all product lines this quarter… / Whisper 809M 0.18 s/token / …"*),
-  matching the fp32 reference. **211.7 tok/s** decode (int8lin S=1, AOT h16c GPU).
+- **End-to-end real generation on the engine (shipped portrait 768 config): the sample page read
+  verbatim** — letterbox → CLIP-norm → non-square patchify → GPU vision `.aimodel` → `image_embeds`
+  → AOT-compiled (h16c) int8lin S=1 decoder, autoregressive greedy: title + paragraph + the **full
+  table reconstructed row-by-row** (*"Quarterly Report / On-device inference shipped across all
+  product lines this quarter… / Whisper 809M 0.18 s/token / …"*), matching the fp32 reference.
+  **211.7 tok/s** decode (int8lin S=1, AOT h16c GPU). Portrait vision vs HF: per-token cos min
+  **0.9975**.
 - Torch ladder vs HF fp32: text-only + vision + full-VLM argmax **706/706 exact**, max logit diff
   **0.0001**, the generation-driving token bit-identical.
 - Engine gate (Mac GPU): vision `image_embeds` cos **1.0002**; AOT int8lin decoder teacher-forced over
@@ -50,23 +52,37 @@ Apache-2.0.
 - **int8lin vs fp16: 13 / 706 argmax flips, all at visual-token positions (0 in the text region)** —
   the OCR text is preserved. Greedy generation byte-identical to fp16.
 
-## Pipeline (host side, 2-stage whole-page)
+## Run in app — `KitMineruReader` (iPhone + Mac)
+
+The shipped bundle rides the kit's VL rope-shift runtime (`VLRuntime` + `VLArchitecture.mineru`).
+The page is letterboxed (aspect-fit + white pad) into a **fixed 672×896 portrait canvas**, CLIP-
+normalized, and non-square-patchified into a **32×24 merged grid (768 visual tokens)** — the
+`VLImagePreprocessor` non-square path added for Qwen2-VL. Two calls:
+
+```swift
+let reader = try await KitMineruReader(catalog: "mineru2.5-pro")
+let markdown = try await reader.read(imageAt: documentURL)   // "Text Recognition:" whole-page
+```
+
+v1 is a single whole-page `"Text Recognition:"` pass (correct for clean single-column pages). The
+2-stage layout→region pipeline is a follow-up.
+
+## Pipeline (host side)
 
 ```
-page image
-  ├─ layout stage:  resize → vision .aimodel → image_embeds → decoder ["…<image>×N, "Layout Detection:"]
-  │                 → parse boxes {type, bbox(0–1), angle} + reading order
-  └─ per region (in reading order):
-        crop(bbox) → vision .aimodel → image_embeds
-        → decoder [ …<image>×N, prompt ]   (image ids → V+slot; shift_start=img_start+N,
-                                            shift_amount=N−max(gh,gw))
-        prompt by type: text → "Text Recognition:" · table → "Table Recognition:"
-                        formula → "Formula Recognition:" · figure → "Image Analysis:"
-  → assemble content_list (type + text) in reading order → json2md → Markdown / JSON
+page image → letterbox into 672×896 (aspect-fit + white pad) → CLIP-normalize
+           → non-square patchify [3072, 1176] (block-major, 64×48 patches)
+           → vision .aimodel → image_embeds [768, 896]
+           → prompt: [ <|vision_start|>, <image>×768, <|vision_end|>, "Text Recognition:" ]
+             (image ids → V+slot; shift_start = img_start+768, shift_amount = 768−32 = 736)
+           → decoder S=1 pipelined greedy decode → tokens → detokenize → markdown / <table>HTML / LaTeX
 ```
 
-Prompt/box formats follow [`opendatalab/mineru-vl-utils`](https://github.com/opendatalab/mineru-vl-utils)
-(`MinerUClient.two_step_extract` + `post_process.json2md`).
+The full 2-stage whole-page mode (layout boxes → per-region crop → recognition → `json2md`) follows
+[`opendatalab/mineru-vl-utils`](https://github.com/opendatalab/mineru-vl-utils)
+(`MinerUClient.two_step_extract` + `post_process.json2md`) — prompts by type: text →
+`"Text Recognition:"` · table → `"Table Recognition:"` · formula → `"Formula Recognition:"` ·
+figure → `"Image Analysis:"`.
 
 ## Use / reproduce
 

decoder/metadata.json

@@ -22,7 +22,7 @@
   },
   "compression": null,
   "compilation": {
-    "date": "2026-07-04T09:08:45.208212+00:00",
+    "date": "2026-07-04T12:29:12.769162+00:00",
     "targets": []
   }
 }

decoder/mineru_decode_int8lin_s1.aimodel/main.hash

@@ -1 +1 @@
-.m���~ۯ��UV}�A�L@N�};�?�
+�|��U�Пv8Yΰ������t�`�-�C

decoder/mineru_decode_int8lin_s1.aimodel/main.mlirb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2e6ddaf482ca7edbafe58ee71a5556187da0418e160f4c404ee47d3bbe3f8112
-size 652928624
+oid sha256:e77cc08b55e8d09f763859ceb08adffe07f7bec374d01d60d82de9430804181d
+size 652928607

vision/mineru_vision.aimodel/main.hash

@@ -1 +1 @@
-�O-<w�YeOd4���r�u6�AC�e{�'�gY
+%@d��蠹(Z�7�VS��G�U�Q����R�!

vision/mineru_vision.aimodel/main.mlirb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bd4f2d3c77be59654f6434d9c2ee72c67536c94110438f657bd127f18cac6759
-size 1325089300
+oid sha256:254064daf8e8a0b90f285aeb37d55653abc847ae1a559551bb81dcecb0529a21
+size 1325199338