Upload folder using huggingface_hub

granite-97m-multilingual-r2/CHECKSUMS.sha256

@@ -0,0 +1,9 @@
+ae37c06948edcc0f030369f8563d63b80a5bcd349ecb6d4219dcc7d3d3525fe9  ./granite_97m_512.mlmodelc/analytics/coremldata.bin
+e8e470b2d49b73cf350eaa2c2f97fb39c99355c4bc507501675a8e53282cc337  ./granite_97m_512.mlmodelc/coremldata.bin
+635299df02dfde6115bbcdb7a8a2cdbe26ecef6be35a393276c5381a32a8f893  ./granite_97m_512.mlmodelc/metadata.json
+b00d8da3bd408b23aa00b6935d35376f88d7d82c7c3f02c19b13375cbea42610  ./granite_97m_512.mlmodelc/model.mil
+276bc93c49a4f37ffefdfb2e10f7d7e1ef57db9027c7ad0d3f2e4160f81a79be  ./granite_97m_512.mlmodelc/weights/weight.bin
+e72839d2e0fc97a9fc9464ffa5c3b869179a5a66f575d848c0efacfd2477f9d3  ./granite_97m_512.mlpackage/Data/com.apple.CoreML/model.mlmodel
+276bc93c49a4f37ffefdfb2e10f7d7e1ef57db9027c7ad0d3f2e4160f81a79be  ./granite_97m_512.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+3c33dc56181d64a539db3f0b365e0eb142dae760413eda48e84ebb8bf19153e7  ./granite_97m_512.mlpackage/Manifest.json
+a9bacaf49d780b5a6de07043557805a183f3eb4a191600bd7f89785ad3d90796  ./README.md

granite-97m-multilingual-r2/README.md

@@ -0,0 +1,119 @@
+---
+license: apache-2.0
+base_model: ibm-granite/granite-embedding-97m-multilingual-r2
+library_name: coreml
+pipeline_tag: sentence-similarity
+tags:
+  - coreml
+  - modernbert
+  - granite
+  - text-embeddings
+  - multilingual
+  - embedkit
+---
+
+# granite-embedding-97m-multilingual-r2 — CoreML (embedkit)
+
+CoreML conversion of **[ibm-granite/granite-embedding-97m-multilingual-r2](https://huggingface.co/ibm-granite/granite-embedding-97m-multilingual-r2)**
+(a ModernBERT encoder, 97M params, 384-dim embeddings, multilingual) for on-device
+inference via [`embedkit`](https://github.com/…) on Apple platforms.
+
+This is a **format conversion** of the upstream model (Apache-2.0). No weights were
+retrained or altered in value; the PyTorch graph was lowered to Apple CoreML
+(`.mlpackage` / `.mlmodelc`) at a fixed sequence length.
+
+## What this is
+
+The sentence-transformers inference graph **in-graph**, except the final L2
+normalization (done by the caller in Rust):
+
+```
+input_ids, attention_mask  ─▶  ModernBERT encoder (12 layers, hidden 384)
+                                 · dual RoPE (global θ=150000 / local θ=160000)
+                                 · local sliding window 128 (±64) on 8 layers,
+                                   full attention on layers 0/3/6/9
+                            ─▶  CLS pooling (hidden[:, 0])
+                            ─▶  embedding [1, 384]   (PRE-L2-norm)
+```
+
+### I/O contract
+
+| | name | dtype | shape |
+|---|---|---|---|
+| input | `input_ids` | int32 | `[1, 512]` |
+| input | `attention_mask` | int32 | `[1, 512]` |
+| output | `embedding` | float32 | `[1, 384]` (pre-L2-norm CLS) |
+
+- **Fixed sequence length 512.** RoPE ⇒ any fixed length is sound; the sliding-window
+  local attention (±64) is genuinely windowed at 512 and is faithfully present in the
+  exported graph (verified: the CoreML output tracks the local-masked reference at
+  cosine 1.0 and diverges from an all-global reference at cosine 0.934). Longer inputs
+  are the caller's windowing layer's job.
+- **L2 normalization is NOT in the graph** — the caller normalizes the 384-d output.
+- **CLS pooling** — the embedding is `hidden_states[:, 0]` (the `<|startoftext|>`
+  BOS/CLS token at position 0), taken after the final LayerNorm.
+
+### Prompts: NONE (prompt-free)
+
+granite-embedding r2 retrieval is **prompt-free**: `config_sentence_transformers.json`
+carries empty query/document prompts. Feed **raw strings** — no task prefixes. (This is
+the model's documented retrieval contract; it differs from instruction-tuned embedders.)
+
+## Files
+
+| file | precision | notes |
+|---|---|---|
+| `granite_97m_512.mlpackage` / `.mlmodelc` | **fp16** | the shipped model — clean on CPU / GPU / ANE |
+| `CHECKSUMS.sha256` | — | sha256 of every artifact file |
+
+**fp16 is shipped.** fp16 was measured NaN-free and accurate on every compute unit
+(worst cosine ≥ 0.99996 vs fp32), and the graph is ANE-capable (see below) — so fp32
+fallback was not triggered. An fp32 build exists as a verification reference (cosine
+1.0 vs the transformers-fp32 oracle) but is not shipped.
+
+## Precision & placement (measured, NaN-aware, 16 multilingual inputs)
+
+| precision · placement | worst cosine vs reference | NaN |
+|---|---|---|
+| **fp32 · CPU** vs transformers-fp32 | **1.00000000** ✅ | 0/16 |
+| **fp16 · ALL** vs fp32 | **0.99999926** ✅ | 0/16 |
+| **fp16 · CPU + GPU** vs fp32 | **0.99999934** ✅ | 0/16 |
+| **fp16 · CPU** vs fp32 | **0.99997885** ✅ | 0/16 |
+| **fp16 · ANE** (`CPU_AND_NE`) vs fp32 | **0.99996360** ✅ | 0/16 |
+
+- **CoreML fp32 vs transformers-fp32** (the sentence-transformers granite pipeline,
+  a lossless fp32 upcast of the bf16 checkpoint): **worst cosine 1.00000000**, 0 NaN,
+  over EN / zh / ja / ko / ar (RTL) / ru / es / de / emoji / mixed-script / code /
+  url-number / a 512-token input / short / special-character strings.
+- **fp16 is viable on this model** (unlike some larger encoders whose residual-stream
+  activations overflow fp16): no NaN on any placement, worst cosine ≥ 0.99996.
+- **ANE:** the graph **compiles for the Apple Neural Engine** (ANECCompile accepts it;
+  `CPU_AND_NE` inference runs clean with finite output). Under a CPU-vs-ANE plan,
+  **482/493 ops (97.8%) prefer the ANE** (11 fall back to CPU, 0 to GPU). Under the
+  default `ALL` plan the planner prefers the GPU (this graph is small and cheapest on
+  GPU there), but the graph is fully ANE-capable when the ANE is targeted.
+
+### fp16 vanishing-guard audit — CLEAN
+
+MIL guard-site audit: **25 sites, all CLEAN**, threshold 2⁻²⁴. Every site is a
+`layer_norm` with epsilon `0x1.5p-17` = 1.0013580322265625e-05 (the fp16 rounding of
+the model's `norm_eps` 1e-5), ≫ 2⁻²⁴. There is **no** `rsqrt`/`sqrt`/`real_div`/`log`
+guard class in the graph (the L2-norm rsqrt is kept in Rust; attention `softmax` has
+no decomposed log tail). RoPE cos/sin are precomputed fp32 constants.
+
+## Provenance
+
+- Upstream: `ibm-granite/granite-embedding-97m-multilingual-r2` (Apache-2.0),
+  revision `835ad14087e140460703cf0fae09f97d469d65c2`. Weights ship bf16.
+- The CoreML graph drives the stock `transformers` ModernBERT encoder layers with
+  static per-layer local/global attention masks (lowered at fixed 512) + precomputed
+  fp32 dual-RoPE, then CLS-slices — validated bit-faithful (cosine ≥ 0.99999998) to the
+  sentence-transformers fp32 pipeline before conversion.
+- Toolchain: coremltools 9.0 · torch 2.6.0 · transformers 5.14.0 ·
+  sentence-transformers 5.6.0 · numpy 1.26.4 · python 3.11.
+
+## License
+
+This artifact is a format conversion of `ibm-granite/granite-embedding-97m-multilingual-r2`
+and is distributed under the **Apache License 2.0**, inheriting the upstream license.
+Attribution: IBM Granite team. See the upstream model card for the full license text.

granite-97m-multilingual-r2/granite_97m_512.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae37c06948edcc0f030369f8563d63b80a5bcd349ecb6d4219dcc7d3d3525fe9
+size 243

granite-97m-multilingual-r2/granite_97m_512.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8e470b2d49b73cf350eaa2c2f97fb39c99355c4bc507501675a8e53282cc337
+size 746

granite-97m-multilingual-r2/granite_97m_512.mlmodelc/metadata.json

@@ -0,0 +1,91 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "granite-embedding-97m text encoder: input_ids\/attention_mask [1,512] -> 384-d CLS embedding (PRE-L2-norm; caller L2-normalizes). Prompt-free (raw strings).",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 384)",
+        "shortDescription" : "",
+        "shape" : "[1, 384]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "embedkit T1: ibm-granite\/granite-embedding-97m-multilingual-r2 (ModernBERT encoder + CLS pooling), pre-L2-norm; static per-layer local\/global masks + fp32 dual-RoPE at fixed 512",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Split" : 24,
+      "Ios17.mul" : 96,
+      "Ios16.silu" : 12,
+      "Ios17.linear" : 48,
+      "Select" : 3,
+      "Ios17.transpose" : 48,
+      "Ios17.matmul" : 24,
+      "Ios17.gather" : 1,
+      "Ios17.add" : 61,
+      "Ios17.sliceByIndex" : 49,
+      "Ios17.concat" : 24,
+      "Ios16.softmax" : 12,
+      "Ios17.logicalAnd" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.expandDims" : 2,
+      "Ios17.squeeze" : 36,
+      "Ios17.cast" : 2,
+      "Ios17.layerNorm" : 25,
+      "Ios17.reshape" : 24
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "attention_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-18",
+      "com.github.apple.coremltools.source" : "torch==2.6.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "granite_97m_512",
+    "method" : "predict"
+  }
+]

granite-97m-multilingual-r2/granite_97m_512.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:276bc93c49a4f37ffefdfb2e10f7d7e1ef57db9027c7ad0d3f2e4160f81a79be
+size 196078720

granite-97m-multilingual-r2/granite_97m_512.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e72839d2e0fc97a9fc9464ffa5c3b869179a5a66f575d848c0efacfd2477f9d3
+size 427137

granite-97m-multilingual-r2/granite_97m_512.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:276bc93c49a4f37ffefdfb2e10f7d7e1ef57db9027c7ad0d3f2e4160f81a79be
+size 196078720

granite-97m-multilingual-r2/granite_97m_512.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "6A4C538F-4FDB-47FC-BEED-1CEC07CDAE63": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "F10AB0AC-F201-4DDF-B6C2-39CD7B9B2BA2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "6A4C538F-4FDB-47FC-BEED-1CEC07CDAE63"
+}