[diffusion] chore: use native ernie prompt enhancer (#34951)

docs/cookbook/diffusion/Ernie-Image/Ernie-Image.mdx

@@ -77,5 +77,5 @@ with open("ernie_image.png", "wb") as f:
 
 - ERNIE-Image is a text-to-image pipeline; do not pass `--image-path`.
 - `--performance-mode auto` keeps conservative defaults while preserving explicit user flags.
-- If the checkpoint includes a PE component, SGLang loads it automatically from `model_index.json`.
+- If the checkpoint includes a PE component, SGLang loads it automatically with the native Ministral3 runtime. Use `--layerwise-offload-components pe` when the local PE decoder needs to trade latency for lower GPU memory usage.
 - Treat FSDP, SP/Ulysses/Ring, and TP as explicit benchmark knobs. Measure the target resolution, step count, and GPU type before making them production defaults.

docs/docs/sglang-diffusion/models_with_pe.mdx

@@ -5,9 +5,9 @@ description: "Run ERNIE-Image with built-in prompt enhancement or a separate SGL
 
 ## Quick Start
 
-By default, the PE model is loaded by the diffusion model server, which may not provide optimal performance. For higher performance, the PE model can be deployed as a separate SGLang server. This document uses `baidu/ERNIE-Image` as an example.
+By default, the diffusion server loads the PE model in-process with SGLang's native Ministral3 implementation. Deploy the PE model as a separate SGLang server when it needs independent resources or scaling. This document uses `baidu/ERNIE-Image` as an example.
 
-Run the model with the built-in Transformers PE implementation (default):
+Run the model with the built-in native PE implementation (default):
 
 ```bash
 # Terminal 1: launch server
@@ -27,7 +27,7 @@ curl -X POST http://${HOST}:${PORT}/v1/images/generations \
   }'
 ```
 
-Run the model with an SGLang-served PE model (high performance):
+Run the model with an SGLang-served PE model (high performance):
 
 ```bash
 # Terminal 1: launch SGLang PE model server
@@ -51,9 +51,21 @@ curl -X POST http://${HOST}:${PORT}/v1/images/generations \
     "guidance_scale": 4.0,
     "use_pe": true
   }'
-```
-
-## Support matrix
+```
+
+For a memory-constrained in-process deployment, the native PE decoder supports
+layerwise offload:
+
+```bash
+sglang serve --model-path baidu/ERNIE-Image \
+  --layerwise-offload-components pe \
+  --port ${PORT}
+```
+
+This option streams PE decoder layers from CPU and can increase prompt-enhancement
+latency. It does not apply when `--pe-server-url` selects an external PE server.
+
+## Support matrix
 
 | Model | Built-in PE | SGLang PE Server |
 |-------|-------------|------------------|