Add XNNPACK int8 exports

README.md

@@ -14,15 +14,31 @@ Upstream model: [clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-b
 
 ## Variants
 
-| Path | Component | Backend | Precision |
-| --- | --- | --- | --- |
-| `coreml/clip_vit_base_patch32_image_coreml_fp16.pte` | image | coreml | fp16 |
-| `coreml/clip_vit_base_patch32_text_coreml_fp16.pte` | text | coreml | fp16 |
-| `mlx/clip_vit_base_patch32_image_mlx_int8.pte` | - | mlx | int8 |
-| `vulkan/clip_vit_base_patch32_image_vulkan_fp16.pte` | image | vulkan | fp16 |
-| `vulkan/clip_vit_base_patch32_text_vulkan_fp16.pte` | text | vulkan | fp16 |
-| `xnnpack/clip_vit_base_patch32_image_xnnpack_fp32.pte` | image | xnnpack | fp32 |
-| `xnnpack/clip_vit_base_patch32_text_xnnpack_fp32.pte` | text | xnnpack | fp32 |
+| Path | Backend | Precision |
+| --- | --- | --- |
+| `coreml/clip_vit_base_patch32_image_coreml_fp16.pte` | coreml | fp16 |
+| `coreml/clip_vit_base_patch32_text_coreml_fp16.pte` | coreml | fp16 |
+| `coreml/clip_vit_base_patch32_text_coreml_int8.pte` | coreml | int8 |
+| `mlx/clip_vit_base_patch32_image_mlx_int8.pte` | mlx | int8 |
+| `vulkan/clip_vit_base_patch32_image_vulkan_fp16.pte` | vulkan | fp16 |
+| `vulkan/clip_vit_base_patch32_text_vulkan_fp16.pte` | vulkan | fp16 |
+| `xnnpack/clip_vit_base_patch32_image_xnnpack_fp32.pte` | xnnpack | fp32 |
+| `xnnpack/clip_vit_base_patch32_image_xnnpack_int8.pte` | xnnpack | int8 |
+| `xnnpack/clip_vit_base_patch32_text_xnnpack_fp32.pte` | xnnpack | fp32 |
+| `xnnpack/clip_vit_base_patch32_text_xnnpack_int8.pte` | xnnpack | int8 |
+
+## XNNPACK int8
+
+The `xnnpack/*_int8.pte` towers are 8da8w (int8 dynamic activations, int8 per-channel
+weights) on the Linear layers; the text tower also has an int8 token embedding table.
+ImageNetV2 zero-shot top-1 (1000 images, one per class, "a photo of a {class}."):
+
+| image | text | top-1 |
+| --- | --- | --- |
+| fp32 | fp32 | 51.4 |
+| int8 | fp32 | 51.5 |
+| fp32 | int8 | 51.2 |
+| int8 | int8 | 51.3 |
 
 ## Repository structure
 
@@ -30,6 +46,7 @@ Upstream model: [clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-b
 config.json                                           43 B
 coreml/clip_vit_base_patch32_image_coreml_fp16.pte    168 MB
 coreml/clip_vit_base_patch32_text_coreml_fp16.pte     122 MB
+coreml/clip_vit_base_patch32_text_coreml_int8.pte     61.0 MB
 coreml/config.json                                    1.7 kB
 mlx/clip_vit_base_patch32_image_mlx_int8.pte          93.7 MB
 mlx/config.json                                       840 B
@@ -38,8 +55,10 @@ vulkan/clip_vit_base_patch32_image_vulkan_fp16.pte    168 MB
 vulkan/clip_vit_base_patch32_text_vulkan_fp16.pte     121 MB
 vulkan/config.json                                    1.7 kB
 xnnpack/clip_vit_base_patch32_image_xnnpack_fp32.pte  335 MB
+xnnpack/clip_vit_base_patch32_image_xnnpack_int8.pte  91.5 MB
 xnnpack/clip_vit_base_patch32_text_xnnpack_fp32.pte   242 MB
-xnnpack/config.json                                   1.6 kB
+xnnpack/clip_vit_base_patch32_text_xnnpack_int8.pte   61.4 MB
+xnnpack/config.json                                   2.9 kB
 ```
 
 ## Compatibility

xnnpack/clip_vit_base_patch32_image_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bc4c3a737813a065f17e5a271443cafd667898b50c558b82df08cc85275a84f
+size 95947264

xnnpack/clip_vit_base_patch32_text_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d770290b59fda3c653e0b00f56bfc1705dd42c77f44f943b49b4fcb1ae2d0bc4
+size 64353408

xnnpack/config.json

@@ -71,6 +71,69 @@
         "image": "clip_vit_base_patch32_image_xnnpack_fp32.pte",
         "text": "clip_vit_base_patch32_text_xnnpack_fp32.pte"
       }
+    },
+    {
+      "file": null,
+      "precision": "int8",
+      "methods": {
+        "image": {
+          "forward": {
+            "inputs": [
+              {
+                "shape": [
+                  1,
+                  3,
+                  224,
+                  224
+                ],
+                "dtype": "float32"
+              }
+            ],
+            "outputs": [
+              {
+                "shape": [
+                  1,
+                  512
+                ],
+                "dtype": "float32"
+              }
+            ]
+          }
+        },
+        "text": {
+          "forward": {
+            "inputs": [
+              {
+                "shape": [
+                  1,
+                  74
+                ],
+                "dtype": "int64"
+              },
+              {
+                "shape": [
+                  1,
+                  74
+                ],
+                "dtype": "int64"
+              }
+            ],
+            "outputs": [
+              {
+                "shape": [
+                  1,
+                  512
+                ],
+                "dtype": "float32"
+              }
+            ]
+          }
+        }
+      },
+      "components": {
+        "image": "clip_vit_base_patch32_image_xnnpack_int8.pte",
+        "text": "clip_vit_base_patch32_text_xnnpack_int8.pte"
+      }
     }
   ]
 }