From 776eb4657709f599e6ba3684acbb33805d5765d2 Mon Sep 17 00:00:00 2001
From: Yong Yue <yongyue@microsoft.com>
Date: Thu, 9 Jul 2026 22:32:37 +0800
Subject: [PATCH] recipe(cross-encoder/ms-marco-MiniLM-L4-v2): add
 text-classification fp16 and w8a16 recipes

Add recipe configurations for cross-encoder/ms-marco-MiniLM-L4-v2
(BertForSequenceClassification, 19.2M params, 4 layers) for text-classification
task with fp16 (baseline) and w8a16 (quantized) variants.

- fp16: opset 17, seq_len 512, vocab 30522, eval on GLUE MRPC
- w8a16: QDQ uint8/uint16, minmax calibration, 59% size reduction (73.2MB -> 29.9MB)
- Both recipes include GLUE MRPC eval configuration
---
 examples/recipes/README.md                    |  1 +
 .../text-classification_fp16_config.json      | 66 +++++++++++++++
 .../text-classification_w8a16_config.json     | 83 +++++++++++++++++++
 3 files changed, 150 insertions(+)
 create mode 100644 examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_fp16_config.json
 create mode 100644 examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_w8a16_config.json

diff --git a/examples/recipes/README.md b/examples/recipes/README.md
index 1077e4f74..cf8415ae0 100644
--- a/examples/recipes/README.md
+++ b/examples/recipes/README.md
@@ -38,6 +38,7 @@ Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) bu
 | ahotrod/electra_large_discriminator_squad2_512 | question-answering |
 | apple/mobilevit-small | image-classification |
 | cardiffnlp/twitter-roberta-base-sentiment-latest | text-classification |
+| cross-encoder/ms-marco-MiniLM-L4-v2 | text-classification |
 | dbmdz/bert-large-cased-finetuned-conll03-english | token-classification |
 | deepset/bert-large-uncased-whole-word-masking-squad2 | question-answering |
 | deepset/roberta-base-squad2 | question-answering |
diff --git a/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_fp16_config.json b/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_fp16_config.json
new file mode 100644
index 000000000..b8612e5ac
--- /dev/null
+++ b/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_fp16_config.json
@@ -0,0 +1,66 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          30522
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          2
+        ]
+      },
+      {
+        "name": "token_type_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          2
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      }
+    ]
+  },
+  "optim": {
+    "clamp_constant_values": true
+  },
+  "quant": null,
+  "compile": null,
+  "loader": {
+    "task": "text-classification",
+    "model_class": "AutoModelForSequenceClassification",
+    "model_type": "bert"
+  }
+}
diff --git a/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_w8a16_config.json b/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_w8a16_config.json
new file mode 100644
index 000000000..96a605be5
--- /dev/null
+++ b/examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/text-classification_w8a16_config.json
@@ -0,0 +1,83 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          30522
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          2
+        ]
+      },
+      {
+        "name": "token_type_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          512
+        ],
+        "value_range": [
+          0,
+          2
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      }
+    ]
+  },
+  "optim": {
+    "clamp_constant_values": true
+  },
+  "quant": {
+    "mode": "qdq",
+    "samples": 10,
+    "calibration_method": "minmax",
+    "weight_type": "uint8",
+    "activation_type": "uint16",
+    "per_channel": false,
+    "symmetric": false,
+    "save_calibration": false,
+    "distribution": "uniform",
+    "seed": null,
+    "calibration_load_path": null,
+    "calibration_save_path": null,
+    "op_types_to_quantize": null,
+    "nodes_to_exclude": null,
+    "task": "text-classification",
+    "model_id": "cross-encoder/ms-marco-MiniLM-L4-v2"
+  },
+  "compile": null,
+  "loader": {
+    "task": "text-classification",
+    "model_class": "AutoModelForSequenceClassification",
+    "model_type": "bert"
+  }
+}