Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,19 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+examples/01_hanfu.png filter=lfs diff=lfs merge=lfs -text
+examples/02_astronaut.png filter=lfs diff=lfs merge=lfs -text
+examples/03_taipei.png filter=lfs diff=lfs merge=lfs -text
+examples/04_shiba.png filter=lfs diff=lfs merge=lfs -text
+examples/05_ink.png filter=lfs diff=lfs merge=lfs -text
+outputs/01_hanfu_1024.png filter=lfs diff=lfs merge=lfs -text
+outputs/01_hanfu_512.png filter=lfs diff=lfs merge=lfs -text
+outputs/02_astronaut_1024.png filter=lfs diff=lfs merge=lfs -text
+outputs/02_astronaut_512.png filter=lfs diff=lfs merge=lfs -text
+outputs/03_taipei_1024.png filter=lfs diff=lfs merge=lfs -text
+outputs/03_taipei_512.png filter=lfs diff=lfs merge=lfs -text
+outputs/04_shiba_1024.png filter=lfs diff=lfs merge=lfs -text
+outputs/04_shiba_512.png filter=lfs diff=lfs merge=lfs -text
+outputs/05_ink_1024.png filter=lfs diff=lfs merge=lfs -text
+outputs/05_ink_512.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,159 @@
+---
+license: other
+license_name: stabilityai-ai-community
+license_link: https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/blob/main/LICENSE.md
+base_model: stabilityai/stable-diffusion-3.5-medium
+library_name: optimum-intel
+pipeline_tag: text-to-image
+tags:
+- text-to-image
+- stable-diffusion
+- stable-diffusion-3
+- diffusers
+- OpenVINO
+- optimum-intel
+- int4
+- nncf
+- sd3
+---
+
+# SD3.5-medium OpenVINO INT4
+
+## SD3.5-medium OpenVINO INT4
+
+OpenVINO + NNCF weight-only INT4 conversion of `stabilityai/stable-diffusion-3.5-medium`.
+
+`transformer` + `text_encoder_3` (T5-XXL):
+`OVWeightQuantizationConfig(bits=4, sym=False, group_size=128, group_size_fallback="adjust", ratio=1.0)`
+
+rest: default INT8 (`OVWeightQuantizationConfig()` bits=8)
+
+via `OVQuantizer` + `ov_config=OVConfig(quantization_config=OVPipelineQuantizationConfig(...))`
+
+FP16 export: `optimum-cli export openvino -m stabilityai/stable-diffusion-3.5-medium --task text-to-image --library diffusers --weight-format fp16 sd3.5-medium-ov-fp16`
+
+Inference params follow the **original SD3.5-medium model card**: 1024x1024, `num_inference_steps=40`, `guidance_scale=4.5`.
+(The sibling `HelloSun/FLUX.2-klein-4B-OpenVINO-INT4` uses 4 steps / guidance 1.0 because FLUX.2-klein-4B is a 4-step distilled model. Do not copy those values here.)
+
+Pinned env: `diffusers==0.37.1 transformers==4.57.6 tokenizers==0.22.0 huggingface-hub==0.35.1 optimum-intel==2.2.0 optimum==2.3.0 openvino==2026.4.0 nncf==3.4.0 torch pillow psutil sentencepiece accelerate`
+
+> **License.** This is a derivative of `stabilityai/stable-diffusion-3.5-medium`, which is released under the
+> **Stability AI Community License** (`license: other`), *not* Apache-2.0. Anyone redistributing or serving this
+> repo must keep complying with the original Stability AI terms. See
+> [LICENSE.md](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/blob/main/LICENSE.md).
+> The base model is gated: accept the license on the Hub page before running the scripts.
+
+---
+
+## Examples (1024x1024, INT4, CPU)
+
+### 01_hanfu (seed 42)
+![01_hanfu](examples/01_hanfu.png)
+`Young Chinese woman in red Hanfu, intricate embroidery, impeccable makeup, red floral forehead pattern, elaborate high bun, golden phoenix headdress, soft-lit outdoor night background, silhouetted tiered pagoda, blurred colorful distant lights, photorealistic, ultra detailed, 8k`
+
+### 02_astronaut (seed 43)
+![02_astronaut](examples/02_astronaut.png)
+`Astronaut in a jungle, cold color palette, muted colors, detailed, 8k, photorealistic, cinematic lighting`
+
+### 03_taipei (seed 44)
+![03_taipei](examples/03_taipei.png)
+`Cyberpunk street in Taipei at night, heavy rain, neon signs with text 'TAIPEI' and Chinese characters '台北', reflections on wet asphalt, crowded night market, cinematic, ultra detailed`
+
+### 04_shiba (seed 45)
+![04_shiba](examples/04_shiba.png)
+`Cute Shiba Inu wearing a tiny astronaut helmet, sitting in a field of sunflowers under a starry sky, dreamy illustration, vibrant colors, high quality`
+
+### 05_ink (seed 46)
+![05_ink](examples/05_ink.png)
+`Traditional Chinese ink wash landscape, misty mountains, a small pagoda on a cliff, cranes flying, minimalist, elegant, high aesthetic quality`
+
+512px resized controls are in `outputs/` (`*_512.png`), full timing in `outputs/benchmark.json`.
+
+---
+
+## 實驗數據
+
+HW: Intel Xeon Platinum 8559C, 192 logical / 96 physical, OpenVINO 2026.4.0. Load+compile 8.72s, final RSS 3276.2MB.
+
+| name | seed | steps | guidance | 總時間(s) | 平均單步(s) | peak RSS(MB) |
+|---|---|---|---|---|---|---|
+| 01_hanfu | 42 | 40 | 4.5 | 228.41 | 5.71 | 3276.2 |
+| 02_astronaut | 43 | 40 | 4.5 | 228.58 | 5.71 | 3276.2 |
+| 03_taipei | 44 | 40 | 4.5 | 224.70 | 5.62 | 3276.2 |
+| 04_shiba | 45 | 40 | 4.5 | 222.18 | 5.55 | 3276.2 |
+| 05_ink | 46 | 40 | 4.5 | 226.27 | 5.66 | 3276.2 |
+
+Quantize: 68.55s. Sizes: FP16 transformer 4.60GB / text_encoder_3 8.87GB (15.17GB total) -> INT4 transformer 1.51GB / text_encoder_3 2.38GB (4.76GB total, ~3.19x).
+詳見 [REPORT.md](REPORT.md) + `outputs/benchmark.json` + `outputs/prompts.txt`.
+
+---
+
+## 用法
+
+```bash
+pip install diffusers==0.37.1 transformers==4.57.6 tokenizers==0.22.0 huggingface-hub==0.35.1 \
+            optimum-intel==2.2.0 optimum==2.3.0 openvino==2026.4.0 nncf==3.4.0 torch pillow psutil sentencepiece accelerate
+```
+
+```python
+from optimum.intel import OVDiffusionPipeline
+import torch
+
+pipe = OVDiffusionPipeline.from_pretrained("HelloSun/SD3.5-medium-OpenVINO-INT4", compile=True)
+img = pipe(
+    prompt="Young Chinese woman in red Hanfu, intricate embroidery, ...",
+    num_inference_steps=40,
+    guidance_scale=4.5,
+    height=1024,
+    width=1024,
+    generator=torch.Generator().manual_seed(42),
+).images[0]
+img.save("out.png")
+```
+
+或用本 repo 腳本:
+
+```bash
+python inference_int4.py --prompt "Astronaut in a jungle, cold color palette, ..." --seed 43 --output out.png
+python generate5.py            # 複現 5 張 + benchmark.json/prompts.txt
+python quantize_int4.py        # FP16 導出 + INT4 量化（見 REPORT.md）
+```
+
+**Speed note.** 40 steps at 1024x1024 is the original model card's setting, and it is slow on CPU. For a faster
+preview: `--steps 28`, or `--width/--height 512`. Lowering steps below ~28 visibly degrades SD3.5-medium; it is not a
+distilled few-step model like FLUX.2-klein-4B.
+
+---
+
+## 檔案結構
+
+```
+./ (INT4 模型: model_index.json, transformer/, text_encoder/, text_encoder_2/, text_encoder_3/,
+     vae_decoder/, vae_encoder/, scheduler/, tokenizer/, tokenizer_2/, tokenizer_3/, openvino_config.json)
+README.md / REPORT.md
+inference_int4.py / quantize_int4.py / generate5.py
+examples/01_hanfu.png ... 05_ink.png   (展示用, 同 outputs 1024)
+outputs/benchmark.json / prompts.txt / benchmark_quantization.json / *_1024.png / *_512.png
+```
+
+---
+
+## 轉換細節
+
+The base model is a `StableDiffusion3Pipeline` with **three** text encoders:
+
+| component | class | approx. params | precision used here |
+|---|---|---|---|
+| `transformer` | `SD3Transformer2DModel` (MMDiT, 24 layers) | 2.5B | **INT4** |
+| `text_encoder_3` | `T5EncoderModel` (T5-XXL) | 4.7B | **INT4** |
+| `text_encoder` | `CLIPTextModelWithProjection` (CLIP-L) | 123M | INT8 |
+| `text_encoder_2` | `CLIPTextModelWithProjection` (CLIP-G) | 694M | INT8 |
+| `vae_decoder` / `vae_encoder` | `AutoencoderKL` | 84M | INT8 |
+
+Only `text_encoder_3` goes to INT4, unlike the FLUX.2 reference where the single text encoder is T5. The two CLIP
+encoders stay at INT8 on purpose: they are small, they dominate prompt fidelity, and per the optimum-intel docs,
+"quantizing the rest of the diffusion pipeline does not significantly improve inference performance but could
+potentially lead to substantial accuracy degradation."
+
+`group_size_fallback="adjust"` matters for SD3: several MMDiT projections have shapes that are not divisible by 128,
+and without it NNCF would silently skip those layers and leave them in FP16. See [REPORT.md](REPORT.md).

REPORT.md

@@ -0,0 +1,143 @@
+# SD3.5-medium → OpenVINO INT4 轉換報告
+
+目標模型：`stabilityai/stable-diffusion-3.5-medium`（`StableDiffusion3Pipeline`）
+目標倉庫：`HelloSun/SD3.5-medium-OpenVINO-INT4`
+方法參考：`HelloSun/FLUX.2-klein-4B-OpenVINO-INT4`
+
+---
+
+## 1. 為什麼不能直接沿用 FLUX.2 的設定
+
+| | FLUX.2-klein-4B | SD3.5-medium |
+|---|---|---|
+| pipeline | FluxPipeline | StableDiffusion3Pipeline |
+| text encoder | 1 個（T5-XXL） | **3 個**：CLIP-L、CLIP-G、T5-XXL |
+| transformer | FLUX transformer | `SD3Transformer2DModel`（MMDiT，24 層，joint_attention_dim 4096） |
+| 原模型推論參數 | 1024, 4 steps, guidance 1.0 | 1024, **40 steps**, **guidance 4.5** |
+| 架構 | distilled 4-step | flow-matching，非 distilled |
+| 授權 | Apache-2.0 | **Stability AI Community License（`license: other`）** |
+
+兩點特別注意：
+
+1. **推論參數不能照抄。** FLUX.2-klein-4B 的 4 steps 是因為它本身是蒸餾後的少步模型。SD3.5-medium 是完整
+   flow-matching 模型，原 model card 用 40 steps。降到 28 以下畫質會明顯變差。
+2. **授權不能沿用 apache-2.0。** SD3.5-medium 是 `license: other` /
+   `stabilityai-ai-community`，衍生倉庫必須保留並遵守 Stability AI 條款。model card 的 YAML 與內文都必須反映這點。
+
+## 2. 元件與精度分配
+
+原始 pipeline 有三個 text encoder，所以 INT4 只打在最大的那一個：
+
+| component | class | 約略參數量 | 精度 |
+|---|---|---|---|
+| `transformer` | SD3Transformer2DModel | 2.5B | **INT4** |
+| `text_encoder_3` | T5EncoderModel (T5-XXL) | 4.7B | **INT4** |
+| `text_encoder` | CLIPTextModelWithProjection (CLIP-L) | 123M | INT8 |
+| `text_encoder_2` | CLIPTextModelWithProjection (CLIP-G) | 694M | INT8 |
+| `vae_decoder` / `vae_encoder` | AutoencoderKL | 84M | INT8 |
+
+理由：T5-XXL FP16 約 9.5 GB，不壓 INT4 的話整體體積降不下來；而兩個 CLIP 很小，卻直接影響 prompt
+遵循度。optimum-intel 官方文件也註明擴散 pipeline 中其餘元件量化「不會顯著提升推論效能，反而可能造成
+實質的準確率退化」。
+
+## 3. 量化參數
+
+```python
+OVPipelineQuantizationConfig(
+    quantization_configs={
+        "transformer":     OVWeightQuantizationConfig(
+            bits=4, sym=False, group_size=128, group_size_fallback="adjust", ratio=1.0),
+        "text_encoder_3":  OVWeightQuantizationConfig(
+            bits=4, sym=False, group_size=128, group_size_fallback="adjust", ratio=1.0),
+    },
+    default_config=OVWeightQuantizationConfig(bits=8),
+)
+```
+
+透過 `OVQuantizer` + `ov_config=OVConfig(quantization_config=...)` 套用。
+
+`group_size_fallback="adjust"` 對 SD3 特別重要。MMDiT 內有部分 projection 的維度無法被 128 整除；預設值是
+`"ignore"`，那樣 NNCF 會**靜默跳過**那些層、讓它們留在 FP16，導致實際壓縮率不如預期。`"adjust"` 會把
+group size 自動調到可整除的最大值（若連 32 都無法整除則退回 backup precision，預設 int8_asym）。
+
+`ratio=1.0` 代表全部符合條件的層都壓到 INT4，不做混合精度。
+
+## 4. 執行步驟
+
+```bash
+export HF_TOKEN=...            # SD3.5-medium 是 gated，需先在 Hub 頁面接受授權
+
+# 階段一：FP16 導出（約 16 GB，耗時最久）
+python quantize_int4.py --export-only
+
+# 階段二：INT4 量化
+python quantize_int4.py --quantize-only
+
+# 驗證 + 產生展示圖與 benchmark
+python generate5.py
+python inference_int4.py --prompt "Astronaut in a jungle, ..." --seed 43 --output out.png
+```
+
+`quantize_int4.py` 會寫出 `benchmark_quantization.json`（導出耗時、量化耗時、FP16/INT4 體積、壓縮倍率）。
+`generate5.py` 會寫出 `outputs/benchmark.json` + `outputs/prompts.txt`。把這兩個檔案裡的數字填回
+`README.md` 的「實驗數據」表格。
+
+### 記憶體
+
+實測（Intel Xeon Platinum 8559C, 192 logical / 96 physical）：
+
+```
+FP16 IR: transformer 4.60GB / text_encoder_3 8.87GB / text_encoder 0.23GB /
+         text_encoder_2 1.30GB / vae 0.15GB  (合計 15.17GB)
+INT4:    transformer 1.51GB / text_encoder_3 2.38GB / text_encoder 0.12GB /
+         text_encoder_2 0.65GB / vae 0.08GB  (合計 4.76GB, ~3.19x)
+量化耗時 68.55s, 推論 peak RSS 3276.2MB (VmHWM), load+compile 8.72s
+```
+
+加上 `optimum-cli` 匯出時的轉換開銷，實務上需要 **32 GB 以上**可用 RAM。`quantize_int4.py` 傳入
+`immediate_save=True`，讓每個子模型量化完立刻落盤並釋放，可以明顯降低尖峰記憶體，但量化完成後
+in-memory pipeline 就不能用了（此腳本流程中不需要再用）。注意 `immediate_save=True` 會就地覆寫
+輸出目錄，腳本會自動把 FP16 IR 先複製過去；且 `TMPDIR` 必須與輸出目錄在同一檔案系統，
+否則 `rename()` 會報 `EXDEV Invalid cross-device link`。
+
+### 建議硬體
+
+| | 最低 | 建議 |
+|---|---|---|
+| RAM（匯出 + 量化） | 32 GB | 64 GB |
+| 磁碟 | 40 GB 可用 | 60 GB 可用 |
+| 推論（40 steps, 1024²） | 很慢 | 96 實體核心以上（對照 FLUX 卡片：4 steps 約 35s） |
+
+## 5. 推論
+
+```python
+from optimum.intel import OVDiffusionPipeline
+import torch
+
+pipe = OVDiffusionPipeline.from_pretrained("HelloSun/SD3.5-medium-OpenVINO-INT4", compile=True)
+img = pipe(prompt=..., num_inference_steps=40, guidance_scale=4.5,
+           height=1024, width=1024,
+           generator=torch.Generator().manual_seed(42)).images[0]
+```
+
+scheduler 為 `FlowMatchEulerDiscreteScheduler`，`shift=3.0` 直接取自原 repo 的
+`scheduler/scheduler_config.json`，載入時會自動套用，不需要手動設定。
+
+`inference_int4.py` 另外提供 `--stepsize`（預設 8）用來微調 shift，可在速度與畫質間取捨。
+
+## 6. 上傳
+
+```bash
+export HF_TOKEN=...        # 或 hf auth login；token 絕不寫進檔案
+huggingface-cli upload HelloSun/SD3.5-medium-OpenVINO-INT4 SD3.5-medium-OpenVINO-INT4 --repo-type model
+```
+
+`.gitignore` 已排除本地产物（`sd3.5-medium-ov-fp16/`、INT4 输出目录、`outputs/`），避免把十幾 GB 的
+權重誤commit進任何 git 倉庫。
+
+## 7. 待填項目
+
+- [x] `README.md` 「Examples」區段的 5 張圖（`examples/*.png`, 2026-09-30 實測產生）
+- [x] `README.md` 「實驗數據」的 HW 描述、Load+compile 時間、最終 RSS（8559C / 8.72s / 3276.2MB）
+- [x] `README.md` 「實驗數據」的 5 列量測表（見 `outputs/benchmark.json`）
+- [x] `README.md` / `REPORT.md` 的 FP16/INT4 體積與壓縮倍率、Quantize 耗時（15.17GB -> 4.76GB ~3.19x / 68.55s）

generate5.py

@@ -0,0 +1,267 @@
+#!/usr/bin/env python3
+"""Reproduce the 5 example images and write outputs/benchmark.json + outputs/prompts.txt.
+
+Same layout as HelloSun/FLUX.2-klein-4B-OpenVINO-INT4 so the two model cards can be
+compared side by side. Parameters follow the SD3.5-medium original model card:
+
+    1024x1024, num_inference_steps=40, guidance_scale=4.5, seeds 42-46
+
+Warning: 40 steps on CPU is the expensive part. On a 16 vCPU box expect minutes
+per image, not seconds. Use --steps 28 for a quick pass, or --size 512 to
+pre-flight the pipeline, then re-run at full settings for the published numbers.
+
+Usage
+-----
+    python generate5.py
+    python generate5.py --steps 28 --size 768 --outdir outputs_preview
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import platform
+import threading
+import time
+from pathlib import Path
+
+from PIL import Image
+
+REPO_ID = "HelloSun/SD3.5-medium-OpenVINO-INT4"
+
+# Original model card parameters.
+DEFAULT_STEPS = 40
+DEFAULT_GUIDANCE = 4.5
+DEFAULT_SIZE = 1024
+
+PROMPTS = [
+    (
+        "01_hanfu",
+        42,
+        "Young Chinese woman in red Hanfu, intricate embroidery, impeccable makeup, "
+        "red floral forehead pattern, elaborate high bun, golden phoenix headdress, "
+        "soft-lit outdoor night background, silhouetted tiered pagoda, blurred colorful "
+        "distant lights, photorealistic, ultra detailed, 8k",
+    ),
+    (
+        "02_astronaut",
+        43,
+        "Astronaut in a jungle, cold color palette, muted colors, detailed, "
+        "8k, photorealistic, cinematic lighting",
+    ),
+    (
+        "03_taipei",
+        44,
+        "Cyberpunk street in Taipei at night, heavy rain, neon signs with text 'TAIPEI' "
+        "and Chinese characters '台北', reflections on wet asphalt, crowded night market, "
+        "cinematic, ultra detailed",
+    ),
+    (
+        "04_shiba",
+        45,
+        "Cute Shiba Inu wearing a tiny astronaut helmet, sitting in a field of sunflowers "
+        "under a starry sky, dreamy illustration, vibrant colors, high quality",
+    ),
+    (
+        "05_ink",
+        46,
+        "Traditional Chinese ink wash landscape, misty mountains, a small pagoda on a cliff, "
+        "cranes flying, minimalist, elegant, high aesthetic quality",
+    ),
+]
+
+
+def _vmhwm_mb() -> float:
+    """Peak RSS high-water mark in MB from /proc. Exact, and dependency-free.
+
+    VmHWM is monotonic, so a single read at the end gives the true peak for the
+    whole process lifetime. Returns 0.0 if unavailable (non-Linux).
+    """
+    try:
+        for line in Path("/proc/self/status").read_text().splitlines():
+            if line.startswith("VmHWM:"):
+                return int(line.split()[1]) / 1024
+    except (OSError, ValueError, IndexError):
+        pass
+    return 0.0
+
+
+class PeakRss:
+    """Tracks peak resident memory for this process.
+
+    Prefers the kernel's VmHWM counter (exact high-water mark, no sampling
+    error, no extra dependency). Falls back to sampling RSS via psutil only if
+    VmHWM is unavailable, and to 0.0 if neither works - in which case
+    `peak_rss_mb` is reported as 0 rather than being silently wrong.
+    """
+
+    def __init__(self, interval: float = 0.2) -> None:
+        self.interval = interval
+        self.peak_mb = 0.0
+        self.source = "none"
+        self._stop = threading.Event()
+        self._thread: threading.Thread | None = None
+
+    def _sample_psutil(self) -> None:
+        import psutil
+
+        proc = psutil.Process(os.getpid())
+        while not self._stop.is_set():
+            self.peak_mb = max(self.peak_mb, proc.memory_info().rss / 1024**2)
+            self._stop.wait(self.interval)
+
+    def _start(self) -> None:
+        if _vmhwm_mb() > 0.0:
+            self.source = "/proc/self/status VmHWM"
+            return
+        try:
+            import psutil  # noqa: F401
+        except ImportError:
+            print(
+                "WARNING: no VmHWM and no psutil; peak_rss_mb will be reported as 0. "
+                "Install psutil to get memory numbers.",
+                flush=True,
+            )
+            return
+        self.source = "psutil sampled"
+        self._thread = threading.Thread(target=self._sample_psutil, daemon=True)
+        self._thread.start()
+
+    def _refresh(self) -> None:
+        # VmHWM is monotonic and authoritative, so always fold it in even when
+        # psutil sampling is the active source.
+        hwm = _vmhwm_mb()
+        if hwm > 0.0:
+            self.peak_mb = max(self.peak_mb, hwm)
+        elif self.source == "none":
+            self.source = "unavailable"
+
+    def __enter__(self) -> PeakRss:
+        self._start()
+        self._refresh()
+        return self
+
+    def __exit__(self, *exc) -> None:
+        self._stop.set()
+        if self._thread:
+            self._thread.join(timeout=2 * self.interval)
+        self._refresh()
+
+
+def cpu_name() -> str:
+    try:
+        for line in Path("/proc/cpuinfo").read_text().splitlines():
+            if line.startswith("model name"):
+                return line.split(":", 1)[1].strip()
+    except OSError:
+        pass
+    return platform.processor() or "unknown"
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--model", default=REPO_ID, help="local path or Hub id")
+    parser.add_argument("--steps", type=int, default=DEFAULT_STEPS)
+    parser.add_argument("--guidance", type=float, default=DEFAULT_GUIDANCE)
+    parser.add_argument("--size", type=int, default=DEFAULT_SIZE)
+    parser.add_argument("--preview-size", type=int, default=512, help="0 disables the downscaled copies")
+    parser.add_argument("--outdir", type=Path, default=Path("outputs"))
+    parser.add_argument("--device", default="CPU")
+    parser.add_argument("--no-compile", action="store_true")
+    args = parser.parse_args()
+
+    import torch
+    from optimum.intel import OVDiffusionPipeline
+
+    args.outdir.mkdir(parents=True, exist_ok=True)
+
+    print(f"loading {args.model} ...", flush=True)
+    t0 = time.perf_counter()
+    pipe = OVDiffusionPipeline.from_pretrained(args.model, device=args.device, compile=not args.no_compile)
+    load_seconds = time.perf_counter() - t0
+    print(f"load+compile {load_seconds:.2f}s", flush=True)
+
+    try:
+        import openvino
+
+        ov_version = openvino.__version__
+    except Exception:
+        ov_version = "unknown"
+
+    results = []
+    with PeakRss() as rss:
+        for name, seed, prompt in PROMPTS:
+            print(f"--- {name} (seed {seed}) ---", flush=True)
+            started = time.perf_counter()
+            image = pipe(
+                prompt=prompt,
+                num_inference_steps=args.steps,
+                guidance_scale=args.guidance,
+                width=args.size,
+                height=args.size,
+                generator=torch.Generator().manual_seed(seed),
+            ).images[0]
+            elapsed = time.perf_counter() - started
+
+            image.save(args.outdir / f"{name}_{args.size}.png")
+            if args.preview_size:
+                try:
+                    resample = Image.Resampling.LANCZOS  # Pillow >= 9.1
+                except AttributeError:
+                    resample = Image.LANCZOS  # Pillow < 9.1
+                image.resize((args.preview_size, args.preview_size), resample).save(
+                    args.outdir / f"{name}_{args.preview_size}.png"
+                )
+
+            entry = {
+                "name": name,
+                "seed": seed,
+                "steps": args.steps,
+                "guidance": args.guidance,
+                "size": f"{args.size}x{args.size}",
+                "total_seconds": round(elapsed, 2),
+                "seconds_per_step": round(elapsed / max(args.steps, 1), 2),
+                "peak_rss_mb": round(rss.peak_mb, 1),
+                "prompt": prompt,
+            }
+            results.append(entry)
+            print(json.dumps(entry, ensure_ascii=False), flush=True)
+
+    benchmark = {
+        "model_id": REPO_ID,
+        "hardware": {
+            "cpu": cpu_name(),
+            "logical_cores": os.cpu_count(),
+            "device": args.device,
+        },
+        "software": {
+            "openvino": ov_version,
+            "python": platform.python_version(),
+            "platform": platform.platform(),
+        },
+        "load_compile_seconds": round(load_seconds, 2),
+        "params": {
+            "num_inference_steps": args.steps,
+            "guidance_scale": args.guidance,
+            "height": args.size,
+            "width": args.size,
+        },
+        "results": results,
+        "peak_rss_mb": round(max(r["peak_rss_mb"] for r in results), 1),
+        "peak_rss_source": rss.source,
+    }
+    (args.outdir / "benchmark.json").write_text(
+        json.dumps(benchmark, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
+    )
+    (args.outdir / "prompts.txt").write_text(
+        "".join(f"{n}\tseed={s}\t{p}\n\n" for n, s, p in PROMPTS), encoding="utf-8"
+    )
+
+    print(f"\nwrote {args.outdir / 'benchmark.json'} and {args.outdir / 'prompts.txt'}", flush=True)
+    print("paste the table into README.md 實驗數據 when publishing.", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

inference_int4.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""Generate one image with the INT4 OpenVINO build of stable-diffusion-3.5-medium.
+
+Inference parameters follow the ORIGINAL model card, not the FLUX.2 reference:
+
+    stabilityai/stable-diffusion-3.5-medium
+      1024x1024, num_inference_steps=40, guidance_scale=4.5
+
+(FLUX.2-klein-4B-OpenVINO-INT4 uses 4 steps / guidance 1.0 because FLUX.2 is a
+4-step distilled model. SD3.5-medium is a full flow-matching model, so the step
+count is 10x higher. The scheduler shift of 3.0 comes from the repo's
+scheduler/scheduler_config.json and is picked up automatically.)
+
+Usage
+-----
+    python inference_int4.py --prompt "Astronaut in a jungle" --seed 43 --output out.png
+    python inference_int4.py --prompt "..." --steps 28        # faster, lower quality
+    python inference_int4.py --prompt "..." --resize 512      # cheap preview
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import time
+from pathlib import Path
+
+REPO_ID = "HelloSun/SD3.5-medium-OpenVINO-INT4"
+
+# Values from the original model card.
+DEFAULT_STEPS = 40
+DEFAULT_GUIDANCE = 4.5
+DEFAULT_SIZE = 1024
+
+
+def load_pipeline(repo_id: str, device: str, compile_model: bool):
+    import torch
+    from optimum.intel import OVDiffusionPipeline
+
+    print(f"loading {repo_id} ...", flush=True)
+    t0 = time.perf_counter()
+    pipe = OVDiffusionPipeline.from_pretrained(repo_id, device=device, compile=compile_model)
+    print(f"load+compile {time.perf_counter() - t0:.2f}s", flush=True)
+    return pipe, torch
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--prompt", required=True)
+    parser.add_argument("--negative-prompt", default="")
+    parser.add_argument("--model", default=REPO_ID, help="local path or Hub id")
+    parser.add_argument("--output", type=Path, default=Path("out.png"))
+    parser.add_argument("--steps", type=int, default=DEFAULT_STEPS)
+    parser.add_argument("--guidance", type=float, default=DEFAULT_GUIDANCE)
+    parser.add_argument("--width", type=int, default=DEFAULT_SIZE)
+    parser.add_argument("--height", type=int, default=DEFAULT_SIZE)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--stepsize", type=int, default=8, help="scheduler shift; 8 is common for SD3 speed/quality")
+    parser.add_argument("--resize", type=int, default=None, help="also save a downscaled copy at this size")
+    parser.add_argument("--device", default="CPU")
+    parser.add_argument("--no-compile", action="store_true", help="skip OV graph compilation")
+    parser.add_argument("--num-images", type=int, default=1)
+    args = parser.parse_args()
+
+    pipe, torch = load_pipeline(args.model, args.device, not args.no_compile)
+
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    started = time.perf_counter()
+    image = pipe(
+        prompt=args.prompt,
+        negative_prompt=args.negative_prompt,
+        num_inference_steps=args.steps,
+        guidance_scale=args.guidance,
+        width=args.width,
+        height=args.height,
+        num_images_per_prompt=args.num_images,
+        generator=torch.Generator().manual_seed(args.seed),
+    ).images[0]
+    elapsed = time.perf_counter() - started
+
+    image.save(args.output)
+    print(f"saved {args.output}  {image.size[0]}x{image.size[1]}  {elapsed:.2f}s "
+          f"({elapsed / max(args.steps, 1):.2f}s/step)", flush=True)
+
+    if args.resize:
+        preview = args.output.with_name(f"{args.output.stem}_{args.resize}.png")
+        try:
+            from PIL import Image as _Image
+
+            _resample = _Image.Resampling.LANCZOS  # Pillow >= 9.1
+        except (ImportError, AttributeError):
+            _resample = image.LANCZOS
+        image.resize((args.resize, args.resize), _resample).save(preview)
+        print(f"saved {preview}", flush=True)
+
+    peak_mb = 0.0
+    try:
+        # Kernel high-water mark: exact and dependency-free on Linux.
+        for line in Path("/proc/self/status").read_text().splitlines():
+            if line.startswith("VmHWM:"):
+                peak_mb = int(line.split()[1]) / 1024
+                break
+        else:
+            import psutil
+
+            peak_mb = psutil.Process(os.getpid()).memory_info().rss / 1024**2
+    except (OSError, ValueError, IndexError, ImportError):
+        pass
+    if peak_mb:
+        print(f"peak RSS: {peak_mb:.0f} MB", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

model_index.json

@@ -0,0 +1,49 @@
+{
+  "_class_name": "StableDiffusion3Pipeline",
+  "_diffusers_version": "0.37.1",
+  "_name_or_path": "stabilityai/stable-diffusion-3.5-medium",
+  "feature_extractor": [
+    null,
+    null
+  ],
+  "image_encoder": [
+    null,
+    null
+  ],
+  "scheduler": [
+    "diffusers",
+    "FlowMatchEulerDiscreteScheduler"
+  ],
+  "text_encoder": [
+    "transformers",
+    "CLIPTextModelWithProjection"
+  ],
+  "text_encoder_2": [
+    "transformers",
+    "CLIPTextModelWithProjection"
+  ],
+  "text_encoder_3": [
+    "transformers",
+    "T5EncoderModel"
+  ],
+  "tokenizer": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "tokenizer_2": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "tokenizer_3": [
+    "transformers",
+    "T5TokenizerFast"
+  ],
+  "transformer": [
+    "diffusers",
+    "SD3Transformer2DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}

openvino_config.json

@@ -0,0 +1,67 @@
+{
+  "dtype": "int4_int4",
+  "input_info": null,
+  "optimum_version": "2.3.0",
+  "output_attentions": false,
+  "quantization_config": {
+    "_dataset_kwargs": {},
+    "dataset": null,
+    "default_config": {
+      "quant_method": "default"
+    },
+    "ignored_scope": null,
+    "num_samples": null,
+    "processor": null,
+    "quantization_configs": {
+      "text_encoder_3": {
+        "_dataset_kwargs": {},
+        "all_layers": null,
+        "backup_precision": null,
+        "bits": 4,
+        "dataset": null,
+        "dq_group_size": null,
+        "dtype": "int4",
+        "gptq": null,
+        "group_size": 128,
+        "group_size_fallback": "adjust",
+        "ignored_scope": null,
+        "lora_correction": null,
+        "num_samples": null,
+        "processor": null,
+        "quant_method": "default",
+        "ratio": 1.0,
+        "scale_estimation": null,
+        "sensitivity_metric": null,
+        "statistics_path": null,
+        "sym": false,
+        "tokenizer": null
+      },
+      "transformer": {
+        "_dataset_kwargs": {},
+        "all_layers": null,
+        "backup_precision": null,
+        "bits": 4,
+        "dataset": null,
+        "dq_group_size": null,
+        "dtype": "int4",
+        "gptq": null,
+        "group_size": 128,
+        "group_size_fallback": "adjust",
+        "ignored_scope": null,
+        "lora_correction": null,
+        "num_samples": null,
+        "processor": null,
+        "quant_method": "default",
+        "ratio": 1.0,
+        "scale_estimation": null,
+        "sensitivity_metric": null,
+        "statistics_path": null,
+        "sym": false,
+        "tokenizer": null
+      }
+    },
+    "tokenizer": null
+  },
+  "save_onnx_model": false,
+  "transformers_version": "4.57.6"
+}

outputs/benchmark.json

@@ -0,0 +1,79 @@
+{
+  "model_id": "HelloSun/SD3.5-medium-OpenVINO-INT4",
+  "hardware": {
+    "cpu": "Intel(R) Xeon(R) Platinum 8559C",
+    "logical_cores": 192,
+    "device": "CPU"
+  },
+  "software": {
+    "openvino": "2026.4.0-22959-99c81491cc3-releases/2026/4",
+    "python": "3.12.12",
+    "platform": "Linux-6.18.48-107.148.amzn2023.x86_64-x86_64-with-glibc2.41"
+  },
+  "load_compile_seconds": 8.72,
+  "params": {
+    "num_inference_steps": 40,
+    "guidance_scale": 4.5,
+    "height": 1024,
+    "width": 1024
+  },
+  "results": [
+    {
+      "name": "01_hanfu",
+      "seed": 42,
+      "steps": 40,
+      "guidance": 4.5,
+      "size": "1024x1024",
+      "total_seconds": 228.41,
+      "seconds_per_step": 5.71,
+      "peak_rss_mb": 3276.2,
+      "prompt": "Young Chinese woman in red Hanfu, intricate embroidery, impeccable makeup, red floral forehead pattern, elaborate high bun, golden phoenix headdress, soft-lit outdoor night background, silhouetted tiered pagoda, blurred colorful distant lights, photorealistic, ultra detailed, 8k"
+    },
+    {
+      "name": "02_astronaut",
+      "seed": 43,
+      "steps": 40,
+      "guidance": 4.5,
+      "size": "1024x1024",
+      "total_seconds": 228.58,
+      "seconds_per_step": 5.71,
+      "peak_rss_mb": 3276.2,
+      "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k, photorealistic, cinematic lighting"
+    },
+    {
+      "name": "03_taipei",
+      "seed": 44,
+      "steps": 40,
+      "guidance": 4.5,
+      "size": "1024x1024",
+      "total_seconds": 224.7,
+      "seconds_per_step": 5.62,
+      "peak_rss_mb": 3276.2,
+      "prompt": "Cyberpunk street in Taipei at night, heavy rain, neon signs with text 'TAIPEI' and Chinese characters '台北', reflections on wet asphalt, crowded night market, cinematic, ultra detailed"
+    },
+    {
+      "name": "04_shiba",
+      "seed": 45,
+      "steps": 40,
+      "guidance": 4.5,
+      "size": "1024x1024",
+      "total_seconds": 222.18,
+      "seconds_per_step": 5.55,
+      "peak_rss_mb": 3276.2,
+      "prompt": "Cute Shiba Inu wearing a tiny astronaut helmet, sitting in a field of sunflowers under a starry sky, dreamy illustration, vibrant colors, high quality"
+    },
+    {
+      "name": "05_ink",
+      "seed": 46,
+      "steps": 40,
+      "guidance": 4.5,
+      "size": "1024x1024",
+      "total_seconds": 226.27,
+      "seconds_per_step": 5.66,
+      "peak_rss_mb": 3276.2,
+      "prompt": "Traditional Chinese ink wash landscape, misty mountains, a small pagoda on a cliff, cranes flying, minimalist, elegant, high aesthetic quality"
+    }
+  ],
+  "peak_rss_mb": 3276.2,
+  "peak_rss_source": "/proc/self/status VmHWM"
+}

outputs/benchmark_quantization.json

@@ -0,0 +1,30 @@
+{
+  "model_id": "stabilityai/stable-diffusion-3.5-medium",
+  "repo_id": "HelloSun/SD3.5-medium-OpenVINO-INT4",
+  "quantize_seconds": 68.55,
+  "int4_size_gb": 4.76,
+  "int4_components_gb": {
+    "scheduler": 0.0,
+    "text_encoder": 0.12,
+    "text_encoder_2": 0.65,
+    "text_encoder_3": 2.38,
+    "tokenizer": 0.0,
+    "tokenizer_2": 0.0,
+    "tokenizer_3": 0.0,
+    "transformer": 1.51,
+    "vae_decoder": 0.05,
+    "vae_encoder": 0.03
+  },
+  "int4_config": {
+    "components": [
+      "transformer",
+      "text_encoder_3"
+    ],
+    "bits": 4,
+    "sym": false,
+    "group_size": 128,
+    "group_size_fallback": "adjust",
+    "ratio": 1.0,
+    "rest": "int8"
+  }
+}

outputs/prompts.txt

@@ -0,0 +1,10 @@
+01_hanfu	seed=42	Young Chinese woman in red Hanfu, intricate embroidery, impeccable makeup, red floral forehead pattern, elaborate high bun, golden phoenix headdress, soft-lit outdoor night background, silhouetted tiered pagoda, blurred colorful distant lights, photorealistic, ultra detailed, 8k
+
+02_astronaut	seed=43	Astronaut in a jungle, cold color palette, muted colors, detailed, 8k, photorealistic, cinematic lighting
+
+03_taipei	seed=44	Cyberpunk street in Taipei at night, heavy rain, neon signs with text 'TAIPEI' and Chinese characters '台北', reflections on wet asphalt, crowded night market, cinematic, ultra detailed
+
+04_shiba	seed=45	Cute Shiba Inu wearing a tiny astronaut helmet, sitting in a field of sunflowers under a starry sky, dreamy illustration, vibrant colors, high quality
+
+05_ink	seed=46	Traditional Chinese ink wash landscape, misty mountains, a small pagoda on a cliff, cranes flying, minimalist, elegant, high aesthetic quality
+

quantize_int4.py

@@ -0,0 +1,228 @@
+#!/usr/bin/env python3
+"""Convert stabilityai/stable-diffusion-3.5-medium to OpenVINO IR (FP16), then
+compress weights to INT4 with NNCF.
+
+Method mirrors the conversion documented in
+https://huggingface.co/HelloSun/FLUX.2-klein-4B-OpenVINO-INT4
+
+    INT4 : transformer + text_encoder_3 (T5-XXL)
+           OVWeightQuantizationConfig(bits=4, sym=False, group_size=128,
+                                      group_size_fallback="adjust", ratio=1.0)
+    rest : default INT8
+           OVWeightQuantizationConfig(bits=8)
+
+    via OVQuantizer + ov_config=OVConfig(
+            quantization_config=OVPipelineQuantizationConfig(...))
+
+SD3.5-medium has three text encoders. Only text_encoder_3 (T5-XXL, 4.7B params,
+~9.5 GB FP16) is pushed to INT4; the two small CLIP encoders (CLIP-L / CLIP-G)
+stay at INT8 because quantizing them costs little size and risks prompt fidelity.
+
+Inference parameters are NOT changed by this script: they stay at the values in
+the original model card (1024x1024, 40 steps, guidance 4.5).
+
+Usage
+-----
+    export HF_TOKEN=...            # SD3.5-medium is gated
+    python quantize_int4.py
+
+    # or split the two phases
+    python quantize_int4.py --export-only
+    python quantize_int4.py --quantize-only
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import time
+from pathlib import Path
+
+MODEL_ID = "stabilityai/stable-diffusion-3.5-medium"
+REPO_ID = "HelloSun/SD3.5-medium-OpenVINO-INT4"
+
+# INT4 keeps the two large components. group_size_fallback="adjust" downshifts to
+# the largest compatible group when a tensor's shape is not divisible by 128,
+# which is common for the small projection matrices inside MMDiT blocks.
+INT4_COMPONENTS = ("transformer", "text_encoder_3")
+
+INT4_KWARGS = {
+    "bits": 4,
+    "sym": False,
+    "group_size": 128,
+    "group_size_fallback": "adjust",
+    "ratio": 1.0,
+}
+
+
+def log(msg: str) -> None:
+    print(f"[quantize_int4] {msg}", flush=True)
+
+
+def dir_size_gb(path: Path) -> float:
+    total = 0
+    for root, _dirs, files in os.walk(path):
+        for name in files:
+            fp = Path(root) / name
+            if fp.is_file() and not fp.is_symlink():
+                total += fp.stat().st_size
+    return total / 1024**3
+
+
+def component_sizes(path: Path) -> dict[str, float]:
+    """Per-subfolder size in GB, for the components we actually quantize."""
+    out: dict[str, float] = {}
+    if not path.is_dir():
+        return out
+    for child in sorted(p for p in path.iterdir() if p.is_dir()):
+        out[child.name] = dir_size_gb(child)
+    return out
+
+
+def require_token() -> None:
+    """SD3.5-medium is gated (license must be accepted on the Hub)."""
+    if os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN"):
+        return
+    log("No HF_TOKEN in environment. If the repo is not already accepted for")
+    log("your account, run:  hf auth login")
+
+
+def export_fp16(out_dir: Path) -> None:
+    """FP16 export via optimum-cli, matching the reference card's export step."""
+    if (out_dir / "openvino_config.json").is_file():
+        log(f"FP16 export already present at {out_dir}, skipping.")
+        return
+
+    cmd = [
+        "optimum-cli",
+        "export",
+        "openvino",
+        "--model",
+        MODEL_ID,
+        "--task",
+        "text-to-image",
+        "--library",
+        "diffusers",
+        "--weight-format",
+        "fp16",
+        str(out_dir),
+    ]
+    log("running: " + " ".join(cmd))
+    log("(downloads ~16 GB of FP16 weights; this is the long part)")
+    subprocess.run(cmd, check=True)
+
+
+def quantize_int4(fp16_dir: Path, out_dir: Path) -> None:
+    import shutil
+
+    # immediate_save=True overwrites files in save_directory in place, so the
+    # FP16 IR must already be there. Seed it automatically on first run.
+    out_exists = out_dir.exists() and any(out_dir.iterdir())
+    if not out_exists:
+        if (fp16_dir / "model_index.json").is_file():
+            log(f"seeding {out_dir} from {fp16_dir} (required for immediate_save=True)")
+            shutil.copytree(fp16_dir, out_dir, dirs_exist_ok=True)
+    # TemporaryDirectory (used internally for each sub-model) must live on the
+    # same filesystem as out_dir, otherwise rename() fails with EXDEV
+    # (Invalid cross-device link). E.g.: export TMPDIR=/data-nvme/work/tmp
+    try:
+        if out_dir.exists() and Path(tempfile.gettempdir()).stat().st_dev != Path(out_dir).stat().st_dev:
+            log("WARNING: TMPDIR is on a different filesystem than out_dir; "
+                "set TMPDIR inside the same filesystem to avoid EXDEV rename errors.")
+    except OSError:
+        pass
+
+    from optimum.intel import (
+        OVConfig,
+        OVDiffusionPipeline,
+        OVPipelineQuantizationConfig,
+        OVQuantizer,
+        OVWeightQuantizationConfig,
+    )
+
+    log(f"loading FP16 OpenVINO pipeline from {fp16_dir}")
+    model = OVDiffusionPipeline.from_pretrained(str(fp16_dir))
+
+    available = sorted(c for c in ("transformer", "text_encoder_3") if getattr(model, c, None) is not None)
+    if available != list(INT4_COMPONENTS):
+        log(f"WARNING: expected components {list(INT4_COMPONENTS)}, pipeline exposes {available}")
+
+    def int4_cfg():
+        return OVWeightQuantizationConfig(**INT4_KWARGS)
+
+    quantization_config = OVPipelineQuantizationConfig(
+        quantization_configs={name: int4_cfg() for name in INT4_COMPONENTS},
+        # Everything else (CLIP encoders, VAE decoder/encoder) at INT8.
+        default_config=OVWeightQuantizationConfig(bits=8),
+    )
+
+    quantizer = OVQuantizer.from_pretrained(model)
+    log("quantizing -> INT4 (transformer, text_encoder_3) / INT8 (rest)")
+    quantizer.quantize(
+        ov_config=OVConfig(quantization_config=quantization_config),
+        save_directory=str(out_dir),
+        # Saves each sub-model as soon as it is compressed, so the FP16 copy of a
+        # component can be released. Cuts peak RSS substantially. The in-memory
+        # pipeline is unusable afterwards, which is fine here.
+        immediate_save=True,
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--fp16-dir", default="sd3.5-medium-ov-fp16", type=Path)
+    parser.add_argument("--out-dir", default=REPO_ID.split("/")[-1], type=Path)
+    parser.add_argument("--export-only", action="store_true")
+    parser.add_argument("--quantize-only", action="store_true")
+    args = parser.parse_args()
+
+    if args.export_only and args.quantize_only:
+        parser.error("--export-only and --quantize-only are mutually exclusive")
+
+    require_token()
+    report: dict[str, object] = {"model_id": MODEL_ID, "repo_id": REPO_ID}
+
+    if not args.quantize_only:
+        t0 = time.perf_counter()
+        export_fp16(args.fp16_dir)
+        report["export_seconds"] = round(time.perf_counter() - t0, 2)
+        report["fp16_size_gb"] = round(dir_size_gb(args.fp16_dir), 2)
+        report["fp16_components_gb"] = {k: round(v, 2) for k, v in component_sizes(args.fp16_dir).items()}
+        log(f"FP16 export done: {report['fp16_size_gb']} GB")
+        if args.export_only:
+            (args.out_dir).mkdir(parents=True, exist_ok=True)
+            (args.out_dir / "benchmark_quantization.json").write_text(
+                json.dumps(report, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
+            )
+            return 0
+
+    if args.out_dir.exists() and args.out_dir.is_dir() and any(args.out_dir.iterdir()):
+        log(f"WARNING: {args.out_dir} is not empty; quantize will overwrite its contents.")
+
+    t0 = time.perf_counter()
+    quantize_int4(args.fp16_dir, args.out_dir)
+    report["quantize_seconds"] = round(time.perf_counter() - t0, 2)
+
+    report["int4_size_gb"] = round(dir_size_gb(args.out_dir), 2)
+    report["int4_components_gb"] = {k: round(v, 2) for k, v in component_sizes(args.out_dir).items()}
+    if isinstance(report.get("fp16_size_gb"), float) and report["fp16_size_gb"]:
+        report["compression_ratio"] = round(report["fp16_size_gb"] / report["int4_size_gb"], 2)
+    report["int4_config"] = {"components": list(INT4_COMPONENTS), **INT4_KWARGS, "rest": "int8"}
+
+    args.out_dir.mkdir(parents=True, exist_ok=True)
+    (args.out_dir / "benchmark_quantization.json").write_text(
+        json.dumps(report, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
+    )
+    log(json.dumps(report, indent=2, ensure_ascii=False))
+
+    log("done. copy README.md / REPORT.md / *.py into the output dir, then:")
+    log(f"  huggingface-cli upload {REPO_ID} {args.out_dir} --repo-type model")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

requirements.txt

@@ -0,0 +1,15 @@
+# Pinned to the same env used for HelloSun/FLUX.2-klein-4B-OpenVINO-INT4
+# (+ sentencepiece/accelerate: required to load the SD3 T5 tokenizer & FP16 pipeline)
+diffusers==0.37.1
+transformers==4.57.6
+tokenizers==0.22.0
+huggingface-hub==0.35.1
+optimum-intel==2.2.0
+optimum==2.3.0
+openvino==2026.4.0
+nncf==3.4.0
+torch
+pillow
+psutil
+sentencepiece
+accelerate

scheduler/scheduler_config.json

@@ -0,0 +1,18 @@
+{
+  "_class_name": "FlowMatchEulerDiscreteScheduler",
+  "_diffusers_version": "0.37.1",
+  "base_image_seq_len": 256,
+  "base_shift": 0.5,
+  "invert_sigmas": false,
+  "max_image_seq_len": 4096,
+  "max_shift": 1.15,
+  "num_train_timesteps": 1000,
+  "shift": 3.0,
+  "shift_terminal": null,
+  "stochastic_sampling": false,
+  "time_shift_type": "exponential",
+  "use_beta_sigmas": false,
+  "use_dynamic_shifting": false,
+  "use_exponential_sigmas": false,
+  "use_karras_sigmas": false
+}

text_encoder/config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "CLIPTextModelWithProjection"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "dtype": "float16",
+  "eos_token_id": 2,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "output_hidden_states": true,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "transformers_version": "4.57.6",
+  "vocab_size": 49408
+}

text_encoder/openvino_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d11495c3f5811cdc97f8f3403fec66f0f6e9ffe495d64aecb5a0251e93168c41
+size 124172025

text_encoder_2/config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "CLIPTextModelWithProjection"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "dtype": "float16",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "output_hidden_states": true,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "transformers_version": "4.57.6",
+  "vocab_size": 49408
+}

text_encoder_2/openvino_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b3cc1721e60e684250207f8153520e1b4cef7f02044bc06e05e1e711d682572
+size 696453879

text_encoder_3/config.json

@@ -0,0 +1,31 @@
+{
+  "architectures": [
+    "T5EncoderModel"
+  ],
+  "classifier_dropout": 0.0,
+  "d_ff": 10240,
+  "d_kv": 64,
+  "d_model": 4096,
+  "decoder_start_token_id": 0,
+  "dense_act_fn": "gelu_new",
+  "dropout_rate": 0.1,
+  "dtype": "float16",
+  "eos_token_id": 1,
+  "feed_forward_proj": "gated-gelu",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": false,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "t5",
+  "num_decoder_layers": 24,
+  "num_heads": 64,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_max_distance": 128,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "vocab_size": 32128
+}

text_encoder_3/openvino_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5e9c9d9e8d171891c26cc5e92de2d42cffee5f0dcde691f5864924b9e6d171f
+size 2557956630

tokenizer/openvino_detokenizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d4c1a4c1bbaff35f9e5921a0c5bab412bb0e465fcf59762f32cf4ae7a01c9d7
+size 814403

tokenizer/openvino_detokenizer.xml

@@ -0,0 +1,378 @@
+<?xml version="1.0"?>
+<net name="detokenizer" version="11">
+	<layers>
+		<layer id="0" name="Parameter_244712" type="Parameter" version="opset1">
+			<data shape="?,?" element_type="i64" />
+			<output>
+				<port id="0" precision="I64" names="Parameter_244712">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1" name="Convert_244762" type="Convert" version="opset1">
+			<data destination_type="i32" />
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2" name="Constant_244714" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="0" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3" name="Constant_244716" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="197632" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="4" name="Constant_244718" type="Const" version="opset1">
+			<data element_type="u8" shape="419070" offset="395264" size="419070" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>419070</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="5" name="Slice_244723" type="Const" version="opset1">
+			<data element_type="i32" shape="2" offset="814334" size="8" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="6" name="VocabDecoder_244725" type="VocabDecoder" version="extension">
+			<data skip_tokens="" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>419070</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>2</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="7" name="FuzeRagged_244726" type="FuzeRagged" version="extension">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="8" name="UTF8Validate_244727" type="UTF8Validate" version="extension">
+			<data replace_mode="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="9" name="Constant_244729" type="Const" version="opset1">
+			<data element_type="u8" shape="4" offset="814342" size="4" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>4</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="10" name="Constant_244731" type="Const" version="opset1">
+			<data element_type="u8" shape="1" offset="814346" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="11" name="RegexNormalization_244732" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>4</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="12" name="Constant_244734" type="Const" version="opset1">
+			<data element_type="u8" shape="2" offset="814347" size="2" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="13" name="Constant_244736" type="Const" version="opset1">
+			<data element_type="u8" shape="0" offset="814349" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>0</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="14" name="RegexNormalization_244737" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>2</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>0</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="15" name="Constant_244739" type="Const" version="opset1">
+			<data element_type="u8" shape="51" offset="814350" size="51" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>51</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="16" name="Constant_244741" type="Const" version="opset1">
+			<data element_type="u8" shape="2" offset="814401" size="2" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="17" name="RegexNormalization_244742" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>51</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>2</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="18" name="StringTensorPack_244743" type="StringTensorPack" version="opset15">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="3" precision="STRING" names="Result_244744,string_output">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="19" name="Result_244744" type="Result" version="opset1" output_names="Result_244744,string_output">
+			<input>
+				<port id="0" precision="STRING">
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+	</layers>
+	<edges>
+		<edge from-layer="0" from-port="0" to-layer="1" to-port="0" />
+		<edge from-layer="1" from-port="1" to-layer="6" to-port="0" />
+		<edge from-layer="2" from-port="0" to-layer="6" to-port="1" />
+		<edge from-layer="3" from-port="0" to-layer="6" to-port="2" />
+		<edge from-layer="4" from-port="0" to-layer="6" to-port="3" />
+		<edge from-layer="5" from-port="0" to-layer="6" to-port="4" />
+		<edge from-layer="6" from-port="5" to-layer="7" to-port="0" />
+		<edge from-layer="6" from-port="6" to-layer="7" to-port="1" />
+		<edge from-layer="6" from-port="7" to-layer="7" to-port="2" />
+		<edge from-layer="6" from-port="8" to-layer="7" to-port="3" />
+		<edge from-layer="6" from-port="9" to-layer="8" to-port="2" />
+		<edge from-layer="7" from-port="4" to-layer="8" to-port="0" />
+		<edge from-layer="7" from-port="5" to-layer="8" to-port="1" />
+		<edge from-layer="8" from-port="3" to-layer="11" to-port="0" />
+		<edge from-layer="8" from-port="4" to-layer="11" to-port="1" />
+		<edge from-layer="8" from-port="5" to-layer="11" to-port="2" />
+		<edge from-layer="9" from-port="0" to-layer="11" to-port="3" />
+		<edge from-layer="10" from-port="0" to-layer="11" to-port="4" />
+		<edge from-layer="11" from-port="5" to-layer="14" to-port="0" />
+		<edge from-layer="11" from-port="6" to-layer="14" to-port="1" />
+		<edge from-layer="11" from-port="7" to-layer="14" to-port="2" />
+		<edge from-layer="12" from-port="0" to-layer="14" to-port="3" />
+		<edge from-layer="13" from-port="0" to-layer="14" to-port="4" />
+		<edge from-layer="14" from-port="5" to-layer="17" to-port="0" />
+		<edge from-layer="14" from-port="6" to-layer="17" to-port="1" />
+		<edge from-layer="14" from-port="7" to-layer="17" to-port="2" />
+		<edge from-layer="15" from-port="0" to-layer="17" to-port="3" />
+		<edge from-layer="16" from-port="0" to-layer="17" to-port="4" />
+		<edge from-layer="17" from-port="5" to-layer="18" to-port="0" />
+		<edge from-layer="17" from-port="6" to-layer="18" to-port="1" />
+		<edge from-layer="17" from-port="7" to-layer="18" to-port="2" />
+		<edge from-layer="18" from-port="3" to-layer="19" to-port="0" />
+	</edges>
+	<rt_info>
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
+		<bos_token_id value="49406" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
+		<eos_token_id value="49407" />
+		<handle_special_tokens_with_re />
+		<max_length />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2026.4.0.0-736-a04accf6282" />
+		<openvino_version value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<original_post_processor_template value="{&quot;type&quot;: &quot;RobertaProcessing&quot;, &quot;sep&quot;: [&quot;&lt;|endoftext|>&quot;, 49407], &quot;cls&quot;: [&quot;&lt;|startoftext|>&quot;, 49406], &quot;trim_offsets&quot;: false, &quot;add_prefix_space&quot;: false}" />
+		<original_tokenizer_class value="&lt;class 'transformers.models.clip.tokenization_clip_fast.CLIPTokenizerFast'>" />
+		<pad_token_id value="49407" />
+		<processed_post_processor_template value="{&quot;single&quot;: {&quot;ids&quot;: [49406, -1, 49407], &quot;type_ids&quot;: [0, 0, 0]}, &quot;pair&quot;: {&quot;ids&quot;: [49406, -1, 49407, 49407, -1, 49407], &quot;type_ids&quot;: [0, 0, 0, 0, 0, 0]}}" />
+		<sentencepiece_version value="0.2.2" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.22.0" />
+		<transformers_version value="4.57.6" />
+		<truncation value="False" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode value="replace" />
+		<with_detokenizer value="True" />
+	</rt_info>
+</net>

tokenizer/openvino_tokenizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:669e4fcb55e51ba994db4ad1eaa787883971550e1119dd93d732c70dcc24e85e
+size 2014441

tokenizer/openvino_tokenizer.xml

@@ -0,0 +1,995 @@
+<?xml version="1.0"?>
+<net name="tokenizer" version="11">
+	<layers>
+		<layer id="0" name="Parameter_244565" type="Parameter" version="opset1">
+			<data shape="?" element_type="string" />
+			<output>
+				<port id="0" precision="STRING" names="Parameter_244565">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1" name="Constant_244682" type="Const" version="opset1">
+			<data element_type="boolean" shape="" offset="0" size="1" />
+			<output>
+				<port id="0" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="2" name="ReadValue_244683" type="ReadValue" version="opset6">
+			<data variable_id="truncation" variable_type="dynamic" variable_shape="..." />
+			<input>
+				<port id="0" precision="BOOL" />
+			</input>
+			<output>
+				<port id="1" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="3" name="Constant_244693" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="4" name="Constant_244694" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="5" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="5" name="Constant_244695" type="Const" version="opset1">
+			<data element_type="i32" shape="1" offset="9" size="4" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="6" name="Constant_244571" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="7" name="StringTensorUnpack_244566" type="StringTensorUnpack" version="opset15">
+			<input>
+				<port id="0" precision="STRING">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="8" name="ShapeOf_244567" type="ShapeOf" version="opset3">
+			<data output_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="9" name="Constant_244568" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="10" name="Constant_244569" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="11" name="Gather_244570" type="Gather" version="opset8">
+			<data batch_dims="0" />
+			<input>
+				<port id="0" precision="I64">
+					<dim>1</dim>
+				</port>
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I64" />
+			</output>
+		</layer>
+		<layer id="12" name="Constant_244572" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="13" name="Range_244573" type="Range" version="opset4">
+			<data output_type="i32" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="14" name="Constant_244574" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="15" name="Constant_244575" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="16" name="Add_244576" type="Add" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+			</input>
+			<output>
+				<port id="2" precision="I64" />
+			</output>
+		</layer>
+		<layer id="17" name="Constant_244577" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="18" name="Range_244578" type="Range" version="opset4">
+			<data output_type="i32" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="19" name="Constant_244642" type="Const" version="opset1">
+			<data element_type="u8" shape="39" offset="29" size="39" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>39</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="20" name="SpecialTokensSplit_244643" type="SpecialTokensSplit" version="extension">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>39</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="21" name="CharsMapNormalization_244644" type="CharsMapNormalization" version="extension">
+			<data add_dummy_prefix="false" remove_extra_whitespaces="false" escape_whitespaces="false" normalization_form="nfc" case_fold="true" nmt="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="22" name="Constant_244646" type="Const" version="opset1">
+			<data element_type="u8" shape="3" offset="68" size="3" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="23" name="Constant_244648" type="Const" version="opset1">
+			<data element_type="u8" shape="1" offset="71" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="24" name="RegexNormalization_244649" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>3</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="25" name="Constant_244651" type="Const" version="opset1">
+			<data element_type="u8" shape="57" offset="72" size="57" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>57</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="26" name="RegexSplit_244652" type="RegexSplit" version="extension">
+			<data behaviour="remove" invert="true" max_splits="-1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>57</dim>
+				</port>
+			</input>
+			<output>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="27" name="Constant_244654" type="Const" version="opset1">
+			<data element_type="u8" shape="74" offset="129" size="74" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>74</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="28" name="RegexSplit_244655" type="RegexSplit" version="extension">
+			<data behaviour="isolate" invert="false" max_splits="-1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>74</dim>
+				</port>
+			</input>
+			<output>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="29" name="Constant_244657" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="203" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="30" name="Constant_244659" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="197835" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="31" name="Constant_244661" type="Const" version="opset1">
+			<data element_type="u8" shape="419070" offset="395467" size="419070" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>419070</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="32" name="Constant_244669" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="814537" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="33" name="Constant_244671" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1010113" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="34" name="Constant_244673" type="Const" version="opset1">
+			<data element_type="u8" shape="145154" offset="1205689" size="145154" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>145154</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="35" name="Constant_244675" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1350843" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="36" name="Constant_244677" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1546419" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="37" name="Constant_244679" type="Const" version="opset1">
+			<data element_type="u8" shape="272352" offset="1741995" size="272352" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>272352</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="38" name="Constant_244663" type="Const" version="opset1">
+			<data element_type="i32" shape="2" offset="2014347" size="8" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="39" name="Constant_244665" type="Const" version="opset1">
+			<data element_type="i32" shape="2" offset="2014355" size="8" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="40" name="Constant_244667" type="Const" version="opset1">
+			<data element_type="u8" shape="28" offset="2014363" size="28" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>28</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="41" name="Constant_244680" type="Const" version="opset1">
+			<data element_type="i32" shape="2" offset="2014391" size="8" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="42" name="BPETokenizer_244681" type="BPETokenizer" version="extension">
+			<data unk_token="&lt;|endoftext|>" fuse_unk="false" suffix_indicator="" end_suffix="&lt;/w>" byte_fallback="false" cache_capacity="20000" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>419070</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="10" precision="U8">
+					<dim>145154</dim>
+				</port>
+				<port id="11" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="12" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="13" precision="U8">
+					<dim>272352</dim>
+				</port>
+				<port id="14" precision="I32">
+					<dim>2</dim>
+				</port>
+				<port id="15" precision="I32">
+					<dim>2</dim>
+				</port>
+				<port id="16" precision="U8">
+					<dim>28</dim>
+				</port>
+				<port id="17" precision="I32">
+					<dim>2</dim>
+				</port>
+			</input>
+			<output>
+				<port id="18" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="19" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="20" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="43" name="Constant_244685" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="2014399" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="44" name="Constant_244686" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="2014403" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="45" name="Select_244687" type="Select" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="BOOL" />
+				<port id="1" precision="I32" />
+				<port id="2" precision="I32" />
+			</input>
+			<output>
+				<port id="3" precision="I32" />
+			</output>
+		</layer>
+		<layer id="46" name="Constant_244689" type="Const" version="opset1">
+			<data element_type="u8" shape="5" offset="2014407" size="5" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>5</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="47" name="Constant_244691" type="Const" version="opset1">
+			<data element_type="u8" shape="13" offset="2014412" size="13" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>13</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="48" name="Truncate_244692" type="Truncate" version="extension">
+			<data m_num_inputs="1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32" />
+				<port id="4" precision="U8">
+					<dim>5</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>13</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="49" name="Constant_244696" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="50" name="Constant_244697" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="5" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="51" name="Constant_244698" type="Const" version="opset1">
+			<data element_type="i32" shape="1" offset="2014425" size="4" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="52" name="Constant_244699" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="2014429" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="53" name="CombineSegments_244700" type="CombineSegments" version="extension">
+			<input>
+				<port id="0" precision="I32" />
+				<port id="1" precision="I32" />
+				<port id="2" precision="I32">
+					<dim>1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32" />
+				<port id="7" precision="I32" />
+				<port id="8" precision="I32">
+					<dim>1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>3</dim>
+				</port>
+			</input>
+			<output>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="13" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="14" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="15" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="54" name="Subtract_244701" type="Subtract" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="55" name="Constant_244702" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="56" name="ReduceMax_244703" type="ReduceMax" version="opset1">
+			<data keep_dims="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32" />
+			</input>
+			<output>
+				<port id="2" precision="I32" />
+			</output>
+		</layer>
+		<layer id="57" name="Constant_244704" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="2014425" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="58" name="RaggedToDense_244705" type="RaggedToDense" version="extension">
+			<data pad_right="true" m_pad_max_length="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32" />
+				<port id="4" precision="I32" />
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="BOOL">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="59" name="Convert_244706" type="Convert" version="opset1">
+			<data destination_type="i32" />
+			<input>
+				<port id="0" precision="BOOL">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="60" name="Convert_244706.0" type="Convert" version="opset1">
+			<data destination_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64" names="attention_mask">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="61" name="RaggedToDense_244705.0" type="Convert" version="opset1">
+			<data destination_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64" names="input_ids">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="62" name="Assign_244684" type="Assign" version="opset6">
+			<data variable_id="truncation" />
+			<input>
+				<port id="0" precision="BOOL" />
+			</input>
+			<output>
+				<port id="1" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="63" name="Result_244709" type="Result" version="opset1" output_names="input_ids">
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+		<layer id="64" name="Result_244711" type="Result" version="opset1" output_names="attention_mask">
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+	</layers>
+	<edges>
+		<edge from-layer="0" from-port="0" to-layer="7" to-port="0" />
+		<edge from-layer="1" from-port="0" to-layer="2" to-port="0" />
+		<edge from-layer="2" from-port="1" to-layer="45" to-port="0" />
+		<edge from-layer="2" from-port="1" to-layer="62" to-port="0" />
+		<edge from-layer="3" from-port="0" to-layer="53" to-port="0" />
+		<edge from-layer="4" from-port="0" to-layer="53" to-port="1" />
+		<edge from-layer="5" from-port="0" to-layer="53" to-port="2" />
+		<edge from-layer="6" from-port="0" to-layer="13" to-port="0" />
+		<edge from-layer="7" from-port="1" to-layer="8" to-port="0" />
+		<edge from-layer="7" from-port="3" to-layer="20" to-port="4" />
+		<edge from-layer="7" from-port="2" to-layer="20" to-port="3" />
+		<edge from-layer="7" from-port="1" to-layer="20" to-port="2" />
+		<edge from-layer="8" from-port="1" to-layer="11" to-port="0" />
+		<edge from-layer="9" from-port="0" to-layer="11" to-port="1" />
+		<edge from-layer="10" from-port="0" to-layer="11" to-port="2" />
+		<edge from-layer="11" from-port="3" to-layer="13" to-port="1" />
+		<edge from-layer="11" from-port="3" to-layer="16" to-port="0" />
+		<edge from-layer="12" from-port="0" to-layer="13" to-port="2" />
+		<edge from-layer="13" from-port="3" to-layer="20" to-port="0" />
+		<edge from-layer="14" from-port="0" to-layer="18" to-port="0" />
+		<edge from-layer="15" from-port="0" to-layer="16" to-port="1" />
+		<edge from-layer="16" from-port="2" to-layer="18" to-port="1" />
+		<edge from-layer="17" from-port="0" to-layer="18" to-port="2" />
+		<edge from-layer="18" from-port="3" to-layer="20" to-port="1" />
+		<edge from-layer="19" from-port="0" to-layer="20" to-port="5" />
+		<edge from-layer="20" from-port="7" to-layer="26" to-port="1" />
+		<edge from-layer="20" from-port="6" to-layer="26" to-port="0" />
+		<edge from-layer="20" from-port="11" to-layer="21" to-port="3" />
+		<edge from-layer="20" from-port="10" to-layer="21" to-port="2" />
+		<edge from-layer="20" from-port="9" to-layer="21" to-port="1" />
+		<edge from-layer="20" from-port="8" to-layer="21" to-port="0" />
+		<edge from-layer="21" from-port="7" to-layer="24" to-port="3" />
+		<edge from-layer="21" from-port="4" to-layer="24" to-port="0" />
+		<edge from-layer="21" from-port="5" to-layer="24" to-port="1" />
+		<edge from-layer="21" from-port="6" to-layer="24" to-port="2" />
+		<edge from-layer="22" from-port="0" to-layer="24" to-port="4" />
+		<edge from-layer="23" from-port="0" to-layer="24" to-port="5" />
+		<edge from-layer="24" from-port="7" to-layer="26" to-port="3" />
+		<edge from-layer="24" from-port="8" to-layer="26" to-port="4" />
+		<edge from-layer="24" from-port="9" to-layer="26" to-port="5" />
+		<edge from-layer="24" from-port="6" to-layer="26" to-port="2" />
+		<edge from-layer="25" from-port="0" to-layer="26" to-port="6" />
+		<edge from-layer="26" from-port="11" to-layer="28" to-port="4" />
+		<edge from-layer="26" from-port="12" to-layer="28" to-port="5" />
+		<edge from-layer="26" from-port="10" to-layer="28" to-port="3" />
+		<edge from-layer="26" from-port="9" to-layer="28" to-port="2" />
+		<edge from-layer="26" from-port="8" to-layer="28" to-port="1" />
+		<edge from-layer="26" from-port="7" to-layer="28" to-port="0" />
+		<edge from-layer="27" from-port="0" to-layer="28" to-port="6" />
+		<edge from-layer="28" from-port="11" to-layer="42" to-port="4" />
+		<edge from-layer="28" from-port="7" to-layer="42" to-port="0" />
+		<edge from-layer="28" from-port="8" to-layer="42" to-port="1" />
+		<edge from-layer="28" from-port="9" to-layer="42" to-port="2" />
+		<edge from-layer="28" from-port="10" to-layer="42" to-port="3" />
+		<edge from-layer="29" from-port="0" to-layer="42" to-port="5" />
+		<edge from-layer="30" from-port="0" to-layer="42" to-port="6" />
+		<edge from-layer="31" from-port="0" to-layer="42" to-port="7" />
+		<edge from-layer="32" from-port="0" to-layer="42" to-port="8" />
+		<edge from-layer="33" from-port="0" to-layer="42" to-port="9" />
+		<edge from-layer="34" from-port="0" to-layer="42" to-port="10" />
+		<edge from-layer="35" from-port="0" to-layer="42" to-port="11" />
+		<edge from-layer="36" from-port="0" to-layer="42" to-port="12" />
+		<edge from-layer="37" from-port="0" to-layer="42" to-port="13" />
+		<edge from-layer="38" from-port="0" to-layer="42" to-port="14" />
+		<edge from-layer="39" from-port="0" to-layer="42" to-port="15" />
+		<edge from-layer="40" from-port="0" to-layer="42" to-port="16" />
+		<edge from-layer="41" from-port="0" to-layer="42" to-port="17" />
+		<edge from-layer="42" from-port="18" to-layer="48" to-port="0" />
+		<edge from-layer="42" from-port="19" to-layer="48" to-port="1" />
+		<edge from-layer="42" from-port="20" to-layer="48" to-port="2" />
+		<edge from-layer="43" from-port="0" to-layer="45" to-port="1" />
+		<edge from-layer="44" from-port="0" to-layer="45" to-port="2" />
+		<edge from-layer="45" from-port="3" to-layer="48" to-port="3" />
+		<edge from-layer="46" from-port="0" to-layer="48" to-port="4" />
+		<edge from-layer="47" from-port="0" to-layer="48" to-port="5" />
+		<edge from-layer="48" from-port="8" to-layer="53" to-port="5" />
+		<edge from-layer="48" from-port="6" to-layer="53" to-port="3" />
+		<edge from-layer="48" from-port="7" to-layer="53" to-port="4" />
+		<edge from-layer="49" from-port="0" to-layer="53" to-port="6" />
+		<edge from-layer="50" from-port="0" to-layer="53" to-port="7" />
+		<edge from-layer="51" from-port="0" to-layer="53" to-port="8" />
+		<edge from-layer="52" from-port="0" to-layer="53" to-port="9" />
+		<edge from-layer="53" from-port="11" to-layer="58" to-port="1" />
+		<edge from-layer="53" from-port="12" to-layer="58" to-port="2" />
+		<edge from-layer="53" from-port="10" to-layer="58" to-port="0" />
+		<edge from-layer="53" from-port="10" to-layer="54" to-port="1" />
+		<edge from-layer="53" from-port="11" to-layer="54" to-port="0" />
+		<edge from-layer="54" from-port="2" to-layer="56" to-port="0" />
+		<edge from-layer="55" from-port="0" to-layer="56" to-port="1" />
+		<edge from-layer="56" from-port="2" to-layer="58" to-port="3" />
+		<edge from-layer="57" from-port="0" to-layer="58" to-port="4" />
+		<edge from-layer="58" from-port="6" to-layer="59" to-port="0" />
+		<edge from-layer="58" from-port="5" to-layer="61" to-port="0" />
+		<edge from-layer="59" from-port="1" to-layer="60" to-port="0" />
+		<edge from-layer="60" from-port="1" to-layer="64" to-port="0" />
+		<edge from-layer="61" from-port="1" to-layer="63" to-port="0" />
+	</edges>
+	<rt_info>
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
+		<bos_token_id value="49406" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
+		<eos_token_id value="49407" />
+		<handle_special_tokens_with_re />
+		<max_length />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2026.4.0.0-736-a04accf6282" />
+		<openvino_version value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<original_post_processor_template value="{&quot;type&quot;: &quot;RobertaProcessing&quot;, &quot;sep&quot;: [&quot;&lt;|endoftext|>&quot;, 49407], &quot;cls&quot;: [&quot;&lt;|startoftext|>&quot;, 49406], &quot;trim_offsets&quot;: false, &quot;add_prefix_space&quot;: false}" />
+		<original_tokenizer_class value="&lt;class 'transformers.models.clip.tokenization_clip_fast.CLIPTokenizerFast'>" />
+		<pad_token_id value="49407" />
+		<processed_post_processor_template value="{&quot;single&quot;: {&quot;ids&quot;: [49406, -1, 49407], &quot;type_ids&quot;: [0, 0, 0]}, &quot;pair&quot;: {&quot;ids&quot;: [49406, -1, 49407, 49407, -1, 49407], &quot;type_ids&quot;: [0, 0, 0, 0, 0, 0]}}" />
+		<sentencepiece_version value="0.2.2" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.22.0" />
+		<transformers_version value="4.57.6" />
+		<truncation value="False" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode value="replace" />
+		<with_detokenizer value="True" />
+	</rt_info>
+</net>

tokenizer/special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer/tokenizer_config.json

@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "49406": {
+      "content": "<|startoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49407": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|startoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": true,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 77,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": "<|endoftext|>"
+}

tokenizer_2/openvino_detokenizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c2f093137fe95a10c67c37897aa809730314d9a3e9da26be05c3cf4b39d4772
+size 814407

tokenizer_2/openvino_detokenizer.xml

@@ -0,0 +1,378 @@
+<?xml version="1.0"?>
+<net name="detokenizer" version="11">
+	<layers>
+		<layer id="0" name="Parameter_245150" type="Parameter" version="opset1">
+			<data shape="?,?" element_type="i64" />
+			<output>
+				<port id="0" precision="I64" names="Parameter_245150">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1" name="Convert_245200" type="Convert" version="opset1">
+			<data destination_type="i32" />
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2" name="Constant_245152" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="0" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3" name="Constant_245154" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="197632" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="4" name="Constant_245156" type="Const" version="opset1">
+			<data element_type="u8" shape="419070" offset="395264" size="419070" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>419070</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="5" name="Slice_245161" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="814334" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="6" name="VocabDecoder_245163" type="VocabDecoder" version="extension">
+			<data skip_tokens="" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>419070</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>3</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="7" name="FuzeRagged_245164" type="FuzeRagged" version="extension">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="8" name="UTF8Validate_245165" type="UTF8Validate" version="extension">
+			<data replace_mode="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="9" name="Constant_245167" type="Const" version="opset1">
+			<data element_type="u8" shape="4" offset="814346" size="4" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>4</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="10" name="Constant_245169" type="Const" version="opset1">
+			<data element_type="u8" shape="1" offset="814350" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="11" name="RegexNormalization_245170" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>4</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="12" name="Constant_245172" type="Const" version="opset1">
+			<data element_type="u8" shape="2" offset="814351" size="2" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="13" name="Constant_245174" type="Const" version="opset1">
+			<data element_type="u8" shape="0" offset="814353" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>0</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="14" name="RegexNormalization_245175" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>2</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>0</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="15" name="Constant_245177" type="Const" version="opset1">
+			<data element_type="u8" shape="51" offset="814354" size="51" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>51</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="16" name="Constant_245179" type="Const" version="opset1">
+			<data element_type="u8" shape="2" offset="814405" size="2" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>2</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="17" name="RegexNormalization_245180" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>51</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>2</dim>
+				</port>
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="18" name="StringTensorPack_245181" type="StringTensorPack" version="opset15">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="3" precision="STRING" names="Result_245182,string_output">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="19" name="Result_245182" type="Result" version="opset1" output_names="Result_245182,string_output">
+			<input>
+				<port id="0" precision="STRING">
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+	</layers>
+	<edges>
+		<edge from-layer="0" from-port="0" to-layer="1" to-port="0" />
+		<edge from-layer="1" from-port="1" to-layer="6" to-port="0" />
+		<edge from-layer="2" from-port="0" to-layer="6" to-port="1" />
+		<edge from-layer="3" from-port="0" to-layer="6" to-port="2" />
+		<edge from-layer="4" from-port="0" to-layer="6" to-port="3" />
+		<edge from-layer="5" from-port="0" to-layer="6" to-port="4" />
+		<edge from-layer="6" from-port="5" to-layer="7" to-port="0" />
+		<edge from-layer="6" from-port="6" to-layer="7" to-port="1" />
+		<edge from-layer="6" from-port="7" to-layer="7" to-port="2" />
+		<edge from-layer="6" from-port="8" to-layer="7" to-port="3" />
+		<edge from-layer="6" from-port="9" to-layer="8" to-port="2" />
+		<edge from-layer="7" from-port="4" to-layer="8" to-port="0" />
+		<edge from-layer="7" from-port="5" to-layer="8" to-port="1" />
+		<edge from-layer="8" from-port="3" to-layer="11" to-port="0" />
+		<edge from-layer="8" from-port="4" to-layer="11" to-port="1" />
+		<edge from-layer="8" from-port="5" to-layer="11" to-port="2" />
+		<edge from-layer="9" from-port="0" to-layer="11" to-port="3" />
+		<edge from-layer="10" from-port="0" to-layer="11" to-port="4" />
+		<edge from-layer="11" from-port="5" to-layer="14" to-port="0" />
+		<edge from-layer="11" from-port="6" to-layer="14" to-port="1" />
+		<edge from-layer="11" from-port="7" to-layer="14" to-port="2" />
+		<edge from-layer="12" from-port="0" to-layer="14" to-port="3" />
+		<edge from-layer="13" from-port="0" to-layer="14" to-port="4" />
+		<edge from-layer="14" from-port="5" to-layer="17" to-port="0" />
+		<edge from-layer="14" from-port="6" to-layer="17" to-port="1" />
+		<edge from-layer="14" from-port="7" to-layer="17" to-port="2" />
+		<edge from-layer="15" from-port="0" to-layer="17" to-port="3" />
+		<edge from-layer="16" from-port="0" to-layer="17" to-port="4" />
+		<edge from-layer="17" from-port="5" to-layer="18" to-port="0" />
+		<edge from-layer="17" from-port="6" to-layer="18" to-port="1" />
+		<edge from-layer="17" from-port="7" to-layer="18" to-port="2" />
+		<edge from-layer="18" from-port="3" to-layer="19" to-port="0" />
+	</edges>
+	<rt_info>
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
+		<bos_token_id value="49406" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
+		<eos_token_id value="49407" />
+		<handle_special_tokens_with_re />
+		<max_length />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2026.4.0.0-736-a04accf6282" />
+		<openvino_version value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<original_post_processor_template value="{&quot;type&quot;: &quot;RobertaProcessing&quot;, &quot;sep&quot;: [&quot;&lt;|endoftext|>&quot;, 49407], &quot;cls&quot;: [&quot;&lt;|startoftext|>&quot;, 49406], &quot;trim_offsets&quot;: false, &quot;add_prefix_space&quot;: false}" />
+		<original_tokenizer_class value="&lt;class 'transformers.models.clip.tokenization_clip_fast.CLIPTokenizerFast'>" />
+		<pad_token_id value="0" />
+		<processed_post_processor_template value="{&quot;single&quot;: {&quot;ids&quot;: [49406, -1, 49407], &quot;type_ids&quot;: [0, 0, 0]}, &quot;pair&quot;: {&quot;ids&quot;: [49406, -1, 49407, 49407, -1, 49407], &quot;type_ids&quot;: [0, 0, 0, 0, 0, 0]}}" />
+		<sentencepiece_version value="0.2.2" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.22.0" />
+		<transformers_version value="4.57.6" />
+		<truncation value="False" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode value="replace" />
+		<with_detokenizer value="True" />
+	</rt_info>
+</net>

tokenizer_2/openvino_tokenizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3986625afe9c098fc3bba75ff400c9e6774b0b058202df4b896d62d89009a04
+size 2014457

tokenizer_2/openvino_tokenizer.xml

@@ -0,0 +1,995 @@
+<?xml version="1.0"?>
+<net name="tokenizer" version="11">
+	<layers>
+		<layer id="0" name="Parameter_245003" type="Parameter" version="opset1">
+			<data shape="?" element_type="string" />
+			<output>
+				<port id="0" precision="STRING" names="Parameter_245003">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1" name="Constant_245120" type="Const" version="opset1">
+			<data element_type="boolean" shape="" offset="0" size="1" />
+			<output>
+				<port id="0" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="2" name="ReadValue_245121" type="ReadValue" version="opset6">
+			<data variable_id="truncation" variable_type="dynamic" variable_shape="..." />
+			<input>
+				<port id="0" precision="BOOL" />
+			</input>
+			<output>
+				<port id="1" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="3" name="Constant_245131" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="4" name="Constant_245132" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="5" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="5" name="Constant_245133" type="Const" version="opset1">
+			<data element_type="i32" shape="1" offset="9" size="4" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="6" name="Constant_245009" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="7" name="StringTensorUnpack_245004" type="StringTensorUnpack" version="opset15">
+			<input>
+				<port id="0" precision="STRING">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="U8">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="8" name="ShapeOf_245005" type="ShapeOf" version="opset3">
+			<data output_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="9" name="Constant_245006" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="10" name="Constant_245007" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="13" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="11" name="Gather_245008" type="Gather" version="opset8">
+			<data batch_dims="0" />
+			<input>
+				<port id="0" precision="I64">
+					<dim>1</dim>
+				</port>
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I64" />
+			</output>
+		</layer>
+		<layer id="12" name="Constant_245010" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="13" name="Range_245011" type="Range" version="opset4">
+			<data output_type="i32" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="14" name="Constant_245012" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="15" name="Constant_245013" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="16" name="Add_245014" type="Add" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+			</input>
+			<output>
+				<port id="2" precision="I64" />
+			</output>
+		</layer>
+		<layer id="17" name="Constant_245015" type="Const" version="opset1">
+			<data element_type="i64" shape="" offset="21" size="8" />
+			<output>
+				<port id="0" precision="I64" />
+			</output>
+		</layer>
+		<layer id="18" name="Range_245016" type="Range" version="opset4">
+			<data output_type="i32" />
+			<input>
+				<port id="0" precision="I64" />
+				<port id="1" precision="I64" />
+				<port id="2" precision="I64" />
+			</input>
+			<output>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="19" name="Constant_245080" type="Const" version="opset1">
+			<data element_type="u8" shape="42" offset="29" size="42" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>42</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="20" name="SpecialTokensSplit_245081" type="SpecialTokensSplit" version="extension">
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>42</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="21" name="CharsMapNormalization_245082" type="CharsMapNormalization" version="extension">
+			<data add_dummy_prefix="false" remove_extra_whitespaces="false" escape_whitespaces="false" normalization_form="nfc" case_fold="true" nmt="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="22" name="Constant_245084" type="Const" version="opset1">
+			<data element_type="u8" shape="3" offset="71" size="3" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="23" name="Constant_245086" type="Const" version="opset1">
+			<data element_type="u8" shape="1" offset="74" size="1" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="24" name="RegexNormalization_245087" type="RegexNormalization" version="extension">
+			<data global_replace="true" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>3</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="25" name="Constant_245089" type="Const" version="opset1">
+			<data element_type="u8" shape="57" offset="75" size="57" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>57</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="26" name="RegexSplit_245090" type="RegexSplit" version="extension">
+			<data behaviour="remove" invert="true" max_splits="-1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>57</dim>
+				</port>
+			</input>
+			<output>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="27" name="Constant_245092" type="Const" version="opset1">
+			<data element_type="u8" shape="74" offset="132" size="74" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>74</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="28" name="RegexSplit_245093" type="RegexSplit" version="extension">
+			<data behaviour="isolate" invert="false" max_splits="-1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="U8">
+					<dim>74</dim>
+				</port>
+			</input>
+			<output>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="BOOL">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="29" name="Constant_245095" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="206" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="30" name="Constant_245097" type="Const" version="opset1">
+			<data element_type="i32" shape="49408" offset="197838" size="197632" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>49408</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="31" name="Constant_245099" type="Const" version="opset1">
+			<data element_type="u8" shape="419070" offset="395470" size="419070" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>419070</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="32" name="Constant_245107" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="814540" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="33" name="Constant_245109" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1010116" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="34" name="Constant_245111" type="Const" version="opset1">
+			<data element_type="u8" shape="145154" offset="1205692" size="145154" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>145154</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="35" name="Constant_245113" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1350846" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="36" name="Constant_245115" type="Const" version="opset1">
+			<data element_type="i32" shape="48894" offset="1546422" size="195576" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>48894</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="37" name="Constant_245117" type="Const" version="opset1">
+			<data element_type="u8" shape="272352" offset="1741998" size="272352" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>272352</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="38" name="Constant_245101" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="2014350" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="39" name="Constant_245103" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="2014362" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="40" name="Constant_245105" type="Const" version="opset1">
+			<data element_type="u8" shape="29" offset="2014374" size="29" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>29</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="41" name="Constant_245118" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="2014403" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="42" name="BPETokenizer_245119" type="BPETokenizer" version="extension">
+			<data unk_token="&lt;|endoftext|>" fuse_unk="false" suffix_indicator="" end_suffix="&lt;/w>" byte_fallback="false" cache_capacity="20000" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="U8">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="6" precision="I32">
+					<dim>49408</dim>
+				</port>
+				<port id="7" precision="U8">
+					<dim>419070</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="10" precision="U8">
+					<dim>145154</dim>
+				</port>
+				<port id="11" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="12" precision="I32">
+					<dim>48894</dim>
+				</port>
+				<port id="13" precision="U8">
+					<dim>272352</dim>
+				</port>
+				<port id="14" precision="I32">
+					<dim>3</dim>
+				</port>
+				<port id="15" precision="I32">
+					<dim>3</dim>
+				</port>
+				<port id="16" precision="U8">
+					<dim>29</dim>
+				</port>
+				<port id="17" precision="I32">
+					<dim>3</dim>
+				</port>
+			</input>
+			<output>
+				<port id="18" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="19" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="20" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="43" name="Constant_245123" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="2014415" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="44" name="Constant_245124" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="2014419" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="45" name="Select_245125" type="Select" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="BOOL" />
+				<port id="1" precision="I32" />
+				<port id="2" precision="I32" />
+			</input>
+			<output>
+				<port id="3" precision="I32" />
+			</output>
+		</layer>
+		<layer id="46" name="Constant_245127" type="Const" version="opset1">
+			<data element_type="u8" shape="5" offset="2014423" size="5" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>5</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="47" name="Constant_245129" type="Const" version="opset1">
+			<data element_type="u8" shape="13" offset="2014428" size="13" />
+			<output>
+				<port id="0" precision="U8">
+					<dim>13</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="48" name="Truncate_245130" type="Truncate" version="extension">
+			<data m_num_inputs="1" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32" />
+				<port id="4" precision="U8">
+					<dim>5</dim>
+				</port>
+				<port id="5" precision="U8">
+					<dim>13</dim>
+				</port>
+			</input>
+			<output>
+				<port id="6" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="7" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="8" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="49" name="Constant_245134" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="50" name="Constant_245135" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="5" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="51" name="Constant_245136" type="Const" version="opset1">
+			<data element_type="i32" shape="1" offset="2014441" size="4" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="52" name="Constant_245137" type="Const" version="opset1">
+			<data element_type="i32" shape="3" offset="2014445" size="12" />
+			<output>
+				<port id="0" precision="I32">
+					<dim>3</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="53" name="CombineSegments_245138" type="CombineSegments" version="extension">
+			<input>
+				<port id="0" precision="I32" />
+				<port id="1" precision="I32" />
+				<port id="2" precision="I32">
+					<dim>1</dim>
+				</port>
+				<port id="3" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="4" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="I32" />
+				<port id="7" precision="I32" />
+				<port id="8" precision="I32">
+					<dim>1</dim>
+				</port>
+				<port id="9" precision="I32">
+					<dim>3</dim>
+				</port>
+			</input>
+			<output>
+				<port id="10" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="11" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="12" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="13" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="14" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="15" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="54" name="Subtract_245139" type="Subtract" version="opset1">
+			<data auto_broadcast="numpy" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="55" name="Constant_245140" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="56" name="ReduceMax_245141" type="ReduceMax" version="opset1">
+			<data keep_dims="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32" />
+			</input>
+			<output>
+				<port id="2" precision="I32" />
+			</output>
+		</layer>
+		<layer id="57" name="Constant_245142" type="Const" version="opset1">
+			<data element_type="i32" shape="" offset="1" size="4" />
+			<output>
+				<port id="0" precision="I32" />
+			</output>
+		</layer>
+		<layer id="58" name="RaggedToDense_245143" type="RaggedToDense" version="extension">
+			<data pad_right="true" m_pad_max_length="false" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="2" precision="I32">
+					<dim>-1</dim>
+				</port>
+				<port id="3" precision="I32" />
+				<port id="4" precision="I32" />
+			</input>
+			<output>
+				<port id="5" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+				<port id="6" precision="BOOL">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="59" name="Convert_245144" type="Convert" version="opset1">
+			<data destination_type="i32" />
+			<input>
+				<port id="0" precision="BOOL">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="60" name="Convert_245144.0" type="Convert" version="opset1">
+			<data destination_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64" names="attention_mask">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="61" name="RaggedToDense_245143.0" type="Convert" version="opset1">
+			<data destination_type="i64" />
+			<input>
+				<port id="0" precision="I32">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="I64" names="input_ids">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="62" name="Assign_245122" type="Assign" version="opset6">
+			<data variable_id="truncation" />
+			<input>
+				<port id="0" precision="BOOL" />
+			</input>
+			<output>
+				<port id="1" precision="BOOL" />
+			</output>
+		</layer>
+		<layer id="63" name="Result_245147" type="Result" version="opset1" output_names="input_ids">
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+		<layer id="64" name="Result_245149" type="Result" version="opset1" output_names="attention_mask">
+			<input>
+				<port id="0" precision="I64">
+					<dim>-1</dim>
+					<dim>-1</dim>
+				</port>
+			</input>
+		</layer>
+	</layers>
+	<edges>
+		<edge from-layer="0" from-port="0" to-layer="7" to-port="0" />
+		<edge from-layer="1" from-port="0" to-layer="2" to-port="0" />
+		<edge from-layer="2" from-port="1" to-layer="45" to-port="0" />
+		<edge from-layer="2" from-port="1" to-layer="62" to-port="0" />
+		<edge from-layer="3" from-port="0" to-layer="53" to-port="0" />
+		<edge from-layer="4" from-port="0" to-layer="53" to-port="1" />
+		<edge from-layer="5" from-port="0" to-layer="53" to-port="2" />
+		<edge from-layer="6" from-port="0" to-layer="13" to-port="0" />
+		<edge from-layer="7" from-port="1" to-layer="8" to-port="0" />
+		<edge from-layer="7" from-port="3" to-layer="20" to-port="4" />
+		<edge from-layer="7" from-port="2" to-layer="20" to-port="3" />
+		<edge from-layer="7" from-port="1" to-layer="20" to-port="2" />
+		<edge from-layer="8" from-port="1" to-layer="11" to-port="0" />
+		<edge from-layer="9" from-port="0" to-layer="11" to-port="1" />
+		<edge from-layer="10" from-port="0" to-layer="11" to-port="2" />
+		<edge from-layer="11" from-port="3" to-layer="13" to-port="1" />
+		<edge from-layer="11" from-port="3" to-layer="16" to-port="0" />
+		<edge from-layer="12" from-port="0" to-layer="13" to-port="2" />
+		<edge from-layer="13" from-port="3" to-layer="20" to-port="0" />
+		<edge from-layer="14" from-port="0" to-layer="18" to-port="0" />
+		<edge from-layer="15" from-port="0" to-layer="16" to-port="1" />
+		<edge from-layer="16" from-port="2" to-layer="18" to-port="1" />
+		<edge from-layer="17" from-port="0" to-layer="18" to-port="2" />
+		<edge from-layer="18" from-port="3" to-layer="20" to-port="1" />
+		<edge from-layer="19" from-port="0" to-layer="20" to-port="5" />
+		<edge from-layer="20" from-port="7" to-layer="26" to-port="1" />
+		<edge from-layer="20" from-port="6" to-layer="26" to-port="0" />
+		<edge from-layer="20" from-port="11" to-layer="21" to-port="3" />
+		<edge from-layer="20" from-port="10" to-layer="21" to-port="2" />
+		<edge from-layer="20" from-port="9" to-layer="21" to-port="1" />
+		<edge from-layer="20" from-port="8" to-layer="21" to-port="0" />
+		<edge from-layer="21" from-port="7" to-layer="24" to-port="3" />
+		<edge from-layer="21" from-port="4" to-layer="24" to-port="0" />
+		<edge from-layer="21" from-port="5" to-layer="24" to-port="1" />
+		<edge from-layer="21" from-port="6" to-layer="24" to-port="2" />
+		<edge from-layer="22" from-port="0" to-layer="24" to-port="4" />
+		<edge from-layer="23" from-port="0" to-layer="24" to-port="5" />
+		<edge from-layer="24" from-port="7" to-layer="26" to-port="3" />
+		<edge from-layer="24" from-port="8" to-layer="26" to-port="4" />
+		<edge from-layer="24" from-port="9" to-layer="26" to-port="5" />
+		<edge from-layer="24" from-port="6" to-layer="26" to-port="2" />
+		<edge from-layer="25" from-port="0" to-layer="26" to-port="6" />
+		<edge from-layer="26" from-port="11" to-layer="28" to-port="4" />
+		<edge from-layer="26" from-port="12" to-layer="28" to-port="5" />
+		<edge from-layer="26" from-port="10" to-layer="28" to-port="3" />
+		<edge from-layer="26" from-port="9" to-layer="28" to-port="2" />
+		<edge from-layer="26" from-port="8" to-layer="28" to-port="1" />
+		<edge from-layer="26" from-port="7" to-layer="28" to-port="0" />
+		<edge from-layer="27" from-port="0" to-layer="28" to-port="6" />
+		<edge from-layer="28" from-port="11" to-layer="42" to-port="4" />
+		<edge from-layer="28" from-port="7" to-layer="42" to-port="0" />
+		<edge from-layer="28" from-port="8" to-layer="42" to-port="1" />
+		<edge from-layer="28" from-port="9" to-layer="42" to-port="2" />
+		<edge from-layer="28" from-port="10" to-layer="42" to-port="3" />
+		<edge from-layer="29" from-port="0" to-layer="42" to-port="5" />
+		<edge from-layer="30" from-port="0" to-layer="42" to-port="6" />
+		<edge from-layer="31" from-port="0" to-layer="42" to-port="7" />
+		<edge from-layer="32" from-port="0" to-layer="42" to-port="8" />
+		<edge from-layer="33" from-port="0" to-layer="42" to-port="9" />
+		<edge from-layer="34" from-port="0" to-layer="42" to-port="10" />
+		<edge from-layer="35" from-port="0" to-layer="42" to-port="11" />
+		<edge from-layer="36" from-port="0" to-layer="42" to-port="12" />
+		<edge from-layer="37" from-port="0" to-layer="42" to-port="13" />
+		<edge from-layer="38" from-port="0" to-layer="42" to-port="14" />
+		<edge from-layer="39" from-port="0" to-layer="42" to-port="15" />
+		<edge from-layer="40" from-port="0" to-layer="42" to-port="16" />
+		<edge from-layer="41" from-port="0" to-layer="42" to-port="17" />
+		<edge from-layer="42" from-port="18" to-layer="48" to-port="0" />
+		<edge from-layer="42" from-port="19" to-layer="48" to-port="1" />
+		<edge from-layer="42" from-port="20" to-layer="48" to-port="2" />
+		<edge from-layer="43" from-port="0" to-layer="45" to-port="1" />
+		<edge from-layer="44" from-port="0" to-layer="45" to-port="2" />
+		<edge from-layer="45" from-port="3" to-layer="48" to-port="3" />
+		<edge from-layer="46" from-port="0" to-layer="48" to-port="4" />
+		<edge from-layer="47" from-port="0" to-layer="48" to-port="5" />
+		<edge from-layer="48" from-port="8" to-layer="53" to-port="5" />
+		<edge from-layer="48" from-port="6" to-layer="53" to-port="3" />
+		<edge from-layer="48" from-port="7" to-layer="53" to-port="4" />
+		<edge from-layer="49" from-port="0" to-layer="53" to-port="6" />
+		<edge from-layer="50" from-port="0" to-layer="53" to-port="7" />
+		<edge from-layer="51" from-port="0" to-layer="53" to-port="8" />
+		<edge from-layer="52" from-port="0" to-layer="53" to-port="9" />
+		<edge from-layer="53" from-port="11" to-layer="58" to-port="1" />
+		<edge from-layer="53" from-port="12" to-layer="58" to-port="2" />
+		<edge from-layer="53" from-port="10" to-layer="58" to-port="0" />
+		<edge from-layer="53" from-port="10" to-layer="54" to-port="1" />
+		<edge from-layer="53" from-port="11" to-layer="54" to-port="0" />
+		<edge from-layer="54" from-port="2" to-layer="56" to-port="0" />
+		<edge from-layer="55" from-port="0" to-layer="56" to-port="1" />
+		<edge from-layer="56" from-port="2" to-layer="58" to-port="3" />
+		<edge from-layer="57" from-port="0" to-layer="58" to-port="4" />
+		<edge from-layer="58" from-port="6" to-layer="59" to-port="0" />
+		<edge from-layer="58" from-port="5" to-layer="61" to-port="0" />
+		<edge from-layer="59" from-port="1" to-layer="60" to-port="0" />
+		<edge from-layer="60" from-port="1" to-layer="64" to-port="0" />
+		<edge from-layer="61" from-port="1" to-layer="63" to-port="0" />
+	</edges>
+	<rt_info>
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
+		<bos_token_id value="49406" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
+		<eos_token_id value="49407" />
+		<handle_special_tokens_with_re />
+		<max_length />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2026.4.0.0-736-a04accf6282" />
+		<openvino_version value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
+		<original_post_processor_template value="{&quot;type&quot;: &quot;RobertaProcessing&quot;, &quot;sep&quot;: [&quot;&lt;|endoftext|>&quot;, 49407], &quot;cls&quot;: [&quot;&lt;|startoftext|>&quot;, 49406], &quot;trim_offsets&quot;: false, &quot;add_prefix_space&quot;: false}" />
+		<original_tokenizer_class value="&lt;class 'transformers.models.clip.tokenization_clip_fast.CLIPTokenizerFast'>" />
+		<pad_token_id value="0" />
+		<processed_post_processor_template value="{&quot;single&quot;: {&quot;ids&quot;: [49406, -1, 49407], &quot;type_ids&quot;: [0, 0, 0]}, &quot;pair&quot;: {&quot;ids&quot;: [49406, -1, 49407, 49407, -1, 49407], &quot;type_ids&quot;: [0, 0, 0, 0, 0, 0]}}" />
+		<sentencepiece_version value="0.2.2" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.22.0" />
+		<transformers_version value="4.57.6" />
+		<truncation value="False" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode value="replace" />
+		<with_detokenizer value="True" />
+	</rt_info>
+</net>