fix(xnnpack): preserve PLE dequantization and sensitive projection precision

README.md

@@ -23,7 +23,7 @@ library as ExecuTorch `.pte` programs, ready to run on device.
 | --- | --- | --- | --- |
 | `e2b/mlx/gemma_4_e2b_mlx_int4.pte` | mlx | int4 | 4,096 |
 | `e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte` | vulkan | 8da4w | 2,048 |
-| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w | 4,096 |
+| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w + fp16 | 4,096 |
 
 
 
@@ -38,9 +38,18 @@ e2b/tokenizer_config.json                  21.8 KiB
 e2b/vulkan/config.json                     1.3 KiB
 e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte    2.8 GiB
 e2b/xnnpack/config.json                    1.3 KiB
-e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte  2.6 GiB
+e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte  3.0 GiB
 ```
 
+## XNNPACK precision
+
+The XNNPACK export uses mixed precision: MLP and language-model head linears
+use 8da4w, while attention projections and the PLE residual output projections
+remain fp16. Token embeddings remain int4. The small PLE context projection
+weights are materialized from their quantized values before lowering to
+preserve dequantization semantics. The model filename and runtime API remain
+unchanged; maximum prefill is 512 tokens and context length is 4,096 tokens.
+
 ## Compatibility
 
 These files are published for the **ExecuTorch v1.4.1** runtime. ExecuTorch

e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9f52126bea0f7293a2e45d62ec746a780b9a7fd936f2836b2eee8a22336311b8
-size 2809192448
+oid sha256:d11e8587cd20015e01aed66c0880e3a2cefd464073a2f80b74a666246ab71cc7
+size 3215277312