[feat] Support prompt embedding caching for diffusion model (#2962)

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>
Co-authored-by: Samit <285365963@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

vllm_omni/diffusion/models/magi_human/pipeline_magi_human.py

@@ -2130,6 +2130,28 @@ class MagiHumanPipeline(nn.Module, ProgressBarMixin, DiffusionPipelineProfilerMi
 
         return resample(audio_np, target_len)
 
+    def encode_prompt(
+        self,
+        prompt: str,
+        target_length: int | None = None,
+    ) -> tuple[torch.Tensor, int]:
+        """Encode *prompt* with the T5-Gemma text encoder and pad to fixed length.
+
+        This is the single text-encoder entrypoint so the runner-level
+        prompt-embedding cache (see
+        ``vllm_omni/diffusion/cache/prompt_embed_cache.py``) can transparently
+        memoize results when the same prompt is submitted repeatedly.
+
+        Returns:
+            ``(context, original_context_len)`` matching
+            :func:`_get_padded_t5_gemma_embedding`.
+        """
+        return _get_padded_t5_gemma_embedding(
+            prompt,
+            self.text_encoder,
+            target_length if target_length is not None else self.t5_gemma_target_length,
+        )
+
     @torch.inference_mode()
     def forward(
         self,
@@ -2200,11 +2222,7 @@ class MagiHumanPipeline(nn.Module, ProgressBarMixin, DiffusionPipelineProfilerMi
             device=device,
         )
 
-        context, original_context_len = _get_padded_t5_gemma_embedding(
-            prompt,
-            self.text_encoder,
-            self.t5_gemma_target_length,
-        )
+        context, original_context_len = self.encode_prompt(prompt)
 
         if image_path is not None:
             br_image = self._encode_image(load_image(image_path), br_height, br_width)