feat: restructure repo with fp16/w4a16 variants, benchmarks, and metadata

.gitattributes

@@ -33,5 +33,5 @@ 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
-siglip-runner-2.0.0-Linux.sh filter=lfs diff=lfs merge=lfs -text
-siglip-w4a16-runner-2.1.0-Linux.sh filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+*.sh filter=lfs diff=lfs merge=lfs -text

.metadata.yaml

@@ -0,0 +1,20 @@
+model:
+  name: siglip-so400m-patch14-384
+  display_name: SigLIP-SO400M-patch14-384
+  upstream: google/siglip-so400m-patch14-384
+
+architecture:
+  family: siglip
+  modality:
+    - vision
+    - language
+  parameters: 0.9B
+  type: dual-encoder
+
+format:
+  type: gguf
+  version: 3
+
+tasks:
+  - zero-shot-image-classification
+  - image-text-similarity

README.md

@@ -0,0 +1,105 @@
+---
+license: apache-2.0
+base_model:
+- google/siglip-so400m-patch14-384
+pipeline_tag: zero-shot-image-classification
+tags:
+- siglip
+- vision-language
+- dual-encoder
+- renesas
+- x5h
+---
+
+# SigLIP-SO400M-patch14-384 - Renesas X5H
+
+## Introduction
+
+This repository contains the **SigLIP-SO400M-patch14-384** dual-encoder model, optimized for the **Renesas X5H** platform for **zero-shot image-text similarity** inference.
+
+SigLIP is a dual-encoder model consisting of a vision encoder and a text encoder. It is **not** a generative model -- it does not produce text output. Instead, it computes similarity scores between images and text labels, enabling zero-shot image classification and image-text matching.
+
+- **Model Architecture:** SigLIP uses a Vision Transformer (ViT) for image encoding and a Transformer for text encoding. Both encoders produce L2-normalized embeddings that are compared via dot-product similarity with a learned scale and bias, followed by sigmoid activation.
+- **Model Summary:**
+  |   Parameter    |    SigLIP-SO400M-patch14-384     |
+  |:---:|:---:|
+  | HIDDEN_SIZE | 1152 |
+  | INTERMEDIATE_SIZE | 4304 |
+  | NUM_HEADS | 16 |
+  | HEAD_DIM | 72 |
+  | NUM_LAYERS | 27 (both encoders) |
+  | PATCH_SIZE | 14 |
+  | IMAGE_SIZE | 384 |
+  | NUM_PATCHES | 729 |
+  | VOCAB_SIZE | 32000 |
+  | MAX_TEXT_SEQ_LEN | 64 |
+- **Source Model:** [google/siglip-so400m-patch14-384](https://huggingface.co/google/siglip-so400m-patch14-384)
+
+
+
+## Performance
+The following performance metrics were measured on the Renesas X5H board.
+| Precision | Device | VE Latency (ms) | TE Latency (ms) | NPU DDR (MB) |
+|:---:|:---:|:---:|:---:|:---:|
+| FP16 | NPX6 | 1236.5 | 46.3 | 1714.88 |
+| W4A16 | NPX6 | 1008.9 | 79.1 | 1155.76 |
+
+
+
+## Prerequisites
+To run the model, you need:
+
+1.  **Renesas X5H Board with SDK v4.32.0**
+2.  **Hugging Face CLI**: For downloading the model and installer.
+
+## Deployment
+
+### SigLIP-SO400M-patch14-384 (FP16)
+1. Download the installer **siglip-runner-2.1.0-Linux.sh** from **Files and versions** tab under **fp16/binaries/rcar-x5hv1/xOS-v4.32/** folder.
+2. Copy the installer to the X5H board and run the installer.
+   ```bash
+   bash ./siglip-runner-2.1.0-Linux.sh --prefix=./ --exclude-subdir --skip-license
+   ```
+3. Download GGUF **mmproj-siglip-f16.gguf** from **Files and versions** tab under **fp16** folder and copy to the installed directory on the X5H board.
+4. Expected directory structure on the X5H board.
+   ```bash
+    siglip-runner
+    ├── mmproj-siglip-f16.gguf
+    ├── firmwares
+    ├── kernel_modules
+    ├── scripts
+    ├── siglip-runner
+    └── setup_npu.sh
+   ```
+
+### Inference - SigLIP-SO400M-patch14-384 (FP16)
+```bash
+bash ./setup_npu.sh npu0
+./siglip-runner -m mmproj-siglip-f16.gguf -i image.ppm -t <comma_separated_token_ids> -s
+```
+
+
+
+### SigLIP-SO400M-patch14-384 (W4A16)
+1. Download the installer **siglip-w4a16-runner-2.2.0-Linux.sh** from **Files and versions** tab under **w4a16/binaries/rcar-x5hv1/xOS-v4.32/** folder.
+2. Copy the installer to the X5H board and run the installer.
+   ```bash
+   bash ./siglip-w4a16-runner-2.2.0-Linux.sh --prefix=./ --exclude-subdir --skip-license
+   ```
+3. Download GGUF **mmproj-siglip-f16.gguf** from **Files and versions** tab under **fp16** folder and copy to the installed directory on the X5H board.
+4. Expected directory structure on the X5H board.
+   ```bash
+    siglip-w4a16-runner
+    ├── mmproj-siglip-f16.gguf
+    ├── firmwares
+    ├── kernel_modules
+    ├── scripts
+    ├── siglip-w4a16-runner
+    └── setup_npu.sh
+   ```
+
+### Inference - SigLIP-SO400M-patch14-384 (W4A16)
+```bash
+bash ./setup_npu.sh npu0
+./siglip-w4a16-runner -m mmproj-siglip-f16.gguf -i image.ppm -t <comma_separated_token_ids> -g . -s
+```

fp16/.metadata.yaml

@@ -0,0 +1,13 @@
+variant:
+  id: gguf_fp16
+  format: gguf
+  precision: fp16
+  method: null
+
+quantization:
+  datatype: fp16
+  scope:
+    - weights
+  granularity: null
+  calibration: none
+  toolchain: custom

fp16/benchmarks/x5h_npu.yaml

@@ -0,0 +1,41 @@
+hardware:
+  vendor: renesas
+  chip: rcar-x5h
+  cpu: arm-cortex-a720
+  npu: npx6-48k
+  npu_count: 2
+  npu_cores: 12
+  npu_default_freq_mhz: 1066
+  accelerator:
+    - npu
+
+runtime:
+  engine: NPX_Toolkit
+  format: gguf
+  execution_provider: npu
+
+configuration:
+  npu_instances: 1
+  npu_cores_per_instance: 12
+  npu_freq_mhz: 1066
+
+benchmark:
+  type: hil
+  parameters:
+    batch_size: 1
+
+performance:
+  ve_encode_ms: 1236.5
+  te_encode_ms: 46.3
+  total_ms: 3849.0
+
+metrics:
+  accuracy: null
+
+memory:
+  ve_npu_ddr_mb: 920.75
+  te_npu_ddr_mb: 794.12
+  total_npu_ddr_mb: 1714.88
+
+power:
+  avg_w: null

siglip-runner-2.0.0-Linux.sh

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0a9cd54e6630c44d6a67fe8933ee7aeb6ae1a66372d8d27f41552c13d4c30737
-size 11071285

siglip-w4a16-runner-2.1.0-Linux.sh

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b180a8db9f77c09952ef68708702754f8084ef1e733ce5896851a2a8a0e30e06
-size 339097260

w4a16/.metadata.yaml

@@ -0,0 +1,20 @@
+variant:
+  id: gguf_w4a16
+  format: gguf
+  precision: w4a16
+  method: symmetric_int4_per_channel
+
+quantization:
+  datatype: w4a16
+  scope:
+    - weights
+  granularity: per_output_channel
+  calibration:
+    method: per_tensor_max_abs
+    activation_bits: 16
+    vision_dataset: detection-datasets/coco[val]
+    vision_samples: 128
+    text_dataset: Salesforce/wikitext[wikitext-2-raw-v1][train]
+    text_samples: 128
+    seed: 42
+  toolchain: custom

w4a16/benchmarks/x5h_npu.yaml

@@ -0,0 +1,41 @@
+hardware:
+  vendor: renesas
+  chip: rcar-x5h
+  cpu: arm-cortex-a720
+  npu: npx6-48k
+  npu_count: 2
+  npu_cores: 12
+  npu_default_freq_mhz: 1066
+  accelerator:
+    - npu
+
+runtime:
+  engine: NPX_Toolkit
+  format: gguf
+  execution_provider: npu
+
+configuration:
+  npu_instances: 1
+  npu_cores_per_instance: 12
+  npu_freq_mhz: 1066
+
+benchmark:
+  type: hil
+  parameters:
+    batch_size: 1
+
+performance:
+  ve_encode_ms: 1008.9
+  te_encode_ms: 79.1
+  total_ms: 6486.5
+
+metrics:
+  accuracy: null
+
+memory:
+  ve_npu_ddr_mb: 677.84
+  te_npu_ddr_mb: 477.91
+  total_npu_ddr_mb: 1155.76
+
+power:
+  avg_w: null