From b34869c359784f1511e2cec9b7e6e40298c62742 Mon Sep 17 00:00:00 2001
From: "github-actions[bot]" <github-actions[bot]@users.noreply.github.com>
Date: Thu, 16 Jul 2026 10:39:23 +0800
Subject: [PATCH] recipe(deberta): add mixedbread-ai/mxbai-rerank-base-v1
 (text-classification, reranker)

Curated CPU float recipe for the mxbai-rerank-base-v1 cross-encoder reranker
(config model_type deberta-v2, DebertaV2ForSequenceClassification, num_labels=1 ->
single relevance logit). Ships fp32 + fp16 under cpu/cpu/ (both quant:null; no CPU
quantized variant per repo convention). Recipe-only: Optimum covers deberta-v2
text-classification natively, so zero source changes. Inputs are input_ids +
attention_mask only (deberta-v2 has type_vocab_size=0, no token_type_ids).

The recipe is byte-identical to cross-encoder/nli-deberta-v3-base's (model_id is
CLI-passed; output width is config-driven), confirming one deberta-v2
text-classification recipe serves the task family.

Validated on CPU:
- L0 build: recipe-driven winml build exit 0 (633.6s, no quantize); ONNX opset 17,
  output logits[1,1] (trained reranker head preserved); 786 MB fp32.
- L1 perf: CPU/fp32 avg 7027.85 ms, 0.14 samples/s, RAM +624.7 MB (perf also
  auto-ran on QNN/NPU at 523.99 ms / 1.91 samples/s).
- L2 numeric parity vs PyTorch on real (query, document) reranking pairs:
  cosine 1.000000, max-abs 1.097e-5, ranking order 4/4 identical, top-1 doc agrees.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
---
 .../cpu/text-classification_fp16_config.json  | 54 +++++++++++++++++++
 .../cpu/text-classification_fp32_config.json  | 54 +++++++++++++++++++
 2 files changed, 108 insertions(+)
 create mode 100644 examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp16_config.json
 create mode 100644 examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp32_config.json

diff --git a/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp16_config.json b/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp16_config.json
new file mode 100644
index 000000000..7b36e6952
--- /dev/null
+++ b/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp16_config.json
@@ -0,0 +1,54 @@
+{
+  "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,
+          128100
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "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": "deberta-v2"
+  }
+}
diff --git a/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp32_config.json b/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp32_config.json
new file mode 100644
index 000000000..7b36e6952
--- /dev/null
+++ b/examples/recipes/mixedbread-ai_mxbai-rerank-base-v1/cpu/cpu/text-classification_fp32_config.json
@@ -0,0 +1,54 @@
+{
+  "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,
+          128100
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "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": "deberta-v2"
+  }
+}