Initial upload: QAT-derived int8 ONNX export of EmbeddingGemma-300M

.gitattributes

@@ -33,3 +33,4 @@ 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.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,110 @@
+---
+license: gemma
+base_model: google/embeddinggemma-300m-qat-q8_0-unquantized
+tags:
+  - sentence-transformers
+  - sentence-similarity
+  - feature-extraction
+  - onnx
+  - quantized
+  - int8
+  - transformers.js
+library_name: transformers.js
+pipeline_tag: feature-extraction
+---
+
+# EmbeddingGemma-300M QAT → ONNX (int8)
+
+ONNX export of [`google/embeddinggemma-300m-qat-q8_0-unquantized`](https://huggingface.co/google/embeddinggemma-300m-qat-q8_0-unquantized), quantized to int8 via ONNX Runtime dynamic quantization. Designed as a drop-in replacement for [`onnx-community/embeddinggemma-300m-ONNX`](https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX) (which is PTQ-derived).
+
+The headline difference: this export starts from Google's **QAT-trained** checkpoint, where weights were trained to tolerate int8 quantization noise during training. PTQ-derived exports apply quantization to a model that was never trained for it. The expected quality recovery is ~1.36 MTEB points.
+
+## Provenance
+
+| | |
+|---|---|
+| Source checkpoint | `google/embeddinggemma-300m-qat-q8_0-unquantized` (safetensors fp32) |
+| Export route | `optimum-cli export onnx --library-name sentence_transformers --opset 18` |
+| Quantization | `onnxruntime.quantization.quantize_dynamic`, `QInt8`, `per_channel=True`, `reduce_range=False` |
+| Toolchain | optimum-onnx 2.1.0, sentence-transformers 4.1.0, onnxruntime 1.26.0, torch 2.12.0, Python 3.14 |
+
+The `per_channel=True` flag is load-bearing: QAT weights were trained with per-channel scales, and collapsing them to per-tensor would discard exactly the calibration QAT performed.
+
+## Files
+
+```
+onnx/
+├── model.onnx              # fp32 (~1.2 GB) — fallback / reference
+├── model_q8.onnx           # int8 (~310 MB) — transformers.js v3 dtype="q8"
+└── model_quantized.onnx    # int8 (~310 MB) — identical to model_q8.onnx, transformers.js v2 alias
+config.json
+tokenizer.json
+tokenizer_config.json
+tokenizer.model
+special_tokens_map.json
+added_tokens.json
+```
+
+## Graph signature
+
+```
+Inputs:
+  input_ids:        [batch_size, sequence_length] int64
+  attention_mask:   [batch_size, sequence_length] int64
+
+Outputs:
+  token_embeddings:    [batch_size, sequence_length, 768] float32
+  sentence_embedding:  [batch_size, 768] float32  ← unit-normalized
+```
+
+The `sentence_embedding` output is the full pipeline: mean-pool over tokens → Dense projection → LayerNorm → L2-normalize. MRL truncation (768 → 512 / 256 / 128) is the consumer's choice and should be applied after fetching this output.
+
+## Validation
+
+Numerical agreement between fp32-QAT and int8-QAT on 5 diverse queries (cosine similarity, higher = more agreement; 1.0 = identical vector):
+
+| Query | Cosine |
+|---|---:|
+| "how do I quantize an ONNX model" | 0.997491 |
+| "difference between QAT and PTQ" | 0.997789 |
+| "sentence embedding similarity" | 0.997953 |
+| "WebGPU on iOS" | 0.997189 |
+| "Adobe Photoshop tutorial" | 0.997584 |
+| **Mean** | **0.9976** |
+
+For comparison, PTQ-int8 against the same fp32 reference typically lands in the 0.985–0.995 range — meaning PTQ produces ~2–4× more drift than what QAT produces here.
+
+## Usage (transformers.js)
+
+```js
+import { pipeline } from '@huggingface/transformers';
+
+const embedder = await pipeline(
+  'feature-extraction',
+  'tooape/embeddinggemma-300m-qat-q8-ONNX',
+  { dtype: 'q8' }
+);
+
+const out = await embedder('your query text here', {
+  pooling: 'none',       // sentence_embedding is already pooled+normalized
+  normalize: false,      // already unit-normalized in the ONNX graph
+});
+```
+
+## Usage (Python / onnxruntime)
+
+```python
+import numpy as np
+import onnxruntime as ort
+from transformers import AutoTokenizer
+
+sess = ort.InferenceSession("onnx/model_q8.onnx")
+tok = AutoTokenizer.from_pretrained("tooape/embeddinggemma-300m-qat-q8-ONNX")
+enc = tok("task: search result | query: your query", return_tensors="np")
+feeds = {i.name: enc[i.name] for i in sess.get_inputs() if i.name in enc}
+sentence_embedding = sess.run(["sentence_embedding"], feeds)[0]  # (1, 768) unit-norm
+```
+
+## License
+
+Inherits the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) from the source checkpoint. Same access conditions apply.

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "<image_soft_token>": 262144
+}

config.json

@@ -0,0 +1,61 @@
+{
+  "_sliding_window_pattern": 6,
+  "architectures": [
+    "Gemma3TextModel"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attn_logit_softcapping": null,
+  "bos_token_id": 2,
+  "dtype": "float32",
+  "eos_token_id": 1,
+  "export_model_type": "transformer",
+  "final_logit_softcapping": null,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 1152,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 2048,
+  "model_type": "gemma3_text",
+  "num_attention_heads": 3,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 1,
+  "pad_token_id": 0,
+  "query_pre_attn_scalar": 256,
+  "rms_norm_eps": 1e-06,
+  "rope_local_base_freq": 10000.0,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 257,
+  "transformers_version": "4.57.6",
+  "use_bidirectional_attention": true,
+  "use_cache": true,
+  "vocab_size": 262144
+}

onnx/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:985793983822dc05d7eeba8d7bd13e88980d4ee9bed18a0694ef0ce7de363375
+size 1231716334

onnx/model_q8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:483fd617174767efc1083950a98a0a0cba2b131920a303811c6882a5591a7dea
+size 310651109

onnx/model_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:483fd617174767efc1083950a98a0a0cba2b131920a303811c6882a5591a7dea
+size 310651109

special_tokens_map.json

@@ -0,0 +1,33 @@
+{
+  "boi_token": "<start_of_image>",
+  "bos_token": {
+    "content": "<bos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoi_token": "<end_of_image>",
+  "eos_token": {
+    "content": "<eos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<image_soft_token>",
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6852f8d561078cc0cebe70ca03c5bfdd0d60a45f9d2e0e1e4cc05b68e9ec329e
+size 33385008

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c
+size 4689074