Add XNNPACK 8da4w export

.gitattributes

@@ -38,3 +38,4 @@ multi-qa-mpnet-base-dot-v1_xnnpack.pte filter=lfs diff=lfs merge=lfs -text
 xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
 vulkan/multi_qa_mpnet_base_dot_v1_vulkan_fp16.pte filter=lfs diff=lfs merge=lfs -text
 vulkan/multi_qa_mpnet_base_dot_v1_vulkan_int8.pte filter=lfs diff=lfs merge=lfs -text
+xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text

README.md

@@ -18,19 +18,32 @@ Upstream model: [multi-qa-mpnet-base-dot-v1](https://huggingface.co/sentence-tra
 | --- | --- | --- |
 | `vulkan/multi_qa_mpnet_base_dot_v1_vulkan_fp16.pte` | vulkan | fp16 |
 | `vulkan/multi_qa_mpnet_base_dot_v1_vulkan_int8.pte` | vulkan | int8 |
+| `xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte` | xnnpack | 8da4w |
 | `xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_fp32.pte` | xnnpack | fp32 |
 
+## XNNPACK int8
+
+`xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte` is 8da4w (int8 dynamic activations, int4 weights,
+group 32) on the Linear layers, with an int8 word-embedding table. SQuAD v1.1 dev
+passage retrieval (2000 questions against their 315 paragraphs), dot product:
+
+| | top-1 | MRR |
+| --- | --- | --- |
+| fp32 | 64.60 | 0.7645 |
+| 8da4w | 63.80 | 0.7578 |
+
 ## Repository structure
 
 ```
-config.json                                          48 B
-tokenizer.json                                       694 kB
-tokenizer_config.json                                363 B
-vulkan/config.json                                   1.6 kB
-vulkan/multi_qa_mpnet_base_dot_v1_vulkan_fp16.pte    208 MB
-vulkan/multi_qa_mpnet_base_dot_v1_vulkan_int8.pte    127 MB
-xnnpack/config.json                                  951 B
-xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_fp32.pte  416 MB
+config.json                                           48 B
+tokenizer.json                                        694 kB
+tokenizer_config.json                                 363 B
+vulkan/config.json                                    1.6 kB
+vulkan/multi_qa_mpnet_base_dot_v1_vulkan_fp16.pte     208 MB
+vulkan/multi_qa_mpnet_base_dot_v1_vulkan_int8.pte     127 MB
+xnnpack/config.json                                   1.6 kB
+xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte  69.1 MB
+xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_fp32.pte   416 MB
 ```
 
 ## Compatibility

xnnpack/config.json

@@ -40,6 +40,39 @@
           ]
         }
       }
+    },
+    {
+      "file": "multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte",
+      "precision": "8da4w",
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                510
+              ],
+              "dtype": "int64"
+            },
+            {
+              "shape": [
+                1,
+                510
+              ],
+              "dtype": "int64"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                768
+              ],
+              "dtype": "float32"
+            }
+          ]
+        }
+      }
     }
   ]
 }

xnnpack/multi_qa_mpnet_base_dot_v1_xnnpack_8da4w.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4942c08d111e825df300d90ae780882d457721ab39f27d731031db97ee3bdf68
+size 72474880