fix(xnnpack): bound ring-cache prefill to 512 tokens

README.md

@@ -23,10 +23,12 @@ library as ExecuTorch `.pte` programs, ready to run on device.
 | --- | --- | --- | --- |
 | `e2b/mlx/gemma_4_e2b_mlx_int4.pte` | mlx | int4 | 2,048 |
 | `e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte` | vulkan | 8da4w | 2,048 |
-| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w | 4,048 |
+| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w | 4,096 |
 
 
 
+XNNPACK accepts prefill chunks of up to **512 tokens** within its 4,096-token context. The chunk bound prevents sliding-window ring-cache eviction during long prefills. The exported binary, embedded model schema, and companion config use the same bound.
+
 ## Repository structure
 
 ```

e2b/xnnpack/config.json

@@ -22,7 +22,7 @@
           50
         ],
         "get_max_context_len": 4096,
-        "get_max_seq_len": 1024,
+        "get_max_seq_len": 512,
         "get_vocab_size": 262144,
         "use_kv_cache": true,
         "has_ple": true,
@@ -43,7 +43,7 @@
                 1,
                 {
                   "min": 1,
-                  "max": 1024
+                  "max": 512
                 }
               ],
               "dtype": "int64"
@@ -84,7 +84,7 @@
                 1,
                 {
                   "min": 1,
-                  "max": 1024
+                  "max": 512
                 },
                 1536
               ],
@@ -96,7 +96,7 @@
                 1,
                 {
                   "min": 1,
-                  "max": 1024
+                  "max": 512
                 },
                 35,
                 256
@@ -108,7 +108,7 @@
               "shape": [
                 {
                   "min": 1,
-                  "max": 1024
+                  "max": 512
                 }
               ],
               "dtype": "int64"

e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2e0381d531c6db6e8d175652dcd88198729129c74e33b54ea6aca2cbca88aa21
-size 3212918656
+oid sha256:a5799ecbed3db94029a5c2ce847b138159c1d2b6c0f09f2f153a790211eb6409
+size 3213090176

e2b/xnnpack/validation.json

@@ -0,0 +1,40 @@
+{
+  "passed": true,
+  "metadata_verified": true,
+  "sha256": "a5799ecbed3db94029a5c2ce847b138159c1d2b6c0f09f2f153a790211eb6409",
+  "date": "2026-10-15",
+  "device": {
+    "model": "Samsung SM-G986B",
+    "soc": "Exynos 990",
+    "gpu": "Mali-G77"
+  },
+  "backend": "XNNPACK",
+  "max_prefill_tokens": 512,
+  "max_context_tokens": 4096,
+  "transfer_sha256_verified": true,
+  "prefill_513_rejected": true,
+  "checks": {
+    "omni_tool": {
+      "finite_logits": true,
+      "output": "The weather in Warsaw is 18 degrees and sunny.<turn|>",
+      "first_token": 818,
+      "termination_token": 106
+    },
+    "omni_long": {
+      "finite_logits": true,
+      "output": "The weather in Warsaw is 8 degrees and sunny.<turn|>",
+      "first_token": 818,
+      "termination_token": 106
+    },
+    "transcription": {
+      "finite_logits": true,
+      "output": "And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.<turn|>",
+      "first_token": 3133,
+      "termination_token": 106
+    }
+  },
+  "limitations": [
+    "The 2,102-token repetitive tool fixture answers 8 degrees instead of the supplied 18 degrees, also observed with the previous published model. This verifies continuation and finite logits, not factual accuracy.",
+    "These fixtures did not reproduce the reported immediate EOS; they do not establish correctness for every prompt."
+  ]
+}