[ModelRunnerV2] Support prompt embeds (#42963)

Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Signed-off-by: Canlin Guo <961750412@qq.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Canlin Guo
2026-08-17 21:58:50 +00:00
committed by GitHub
co-authored by Nick Hill
parent e68fb75b25
commit 455edc022b
17 changed files with 551 additions and 54 deletions
@@ -776,7 +776,7 @@ class DiffusionGemmaModelState(ModelState):
) -> None:
super().__init__(vllm_config, model, encoder_cache, device)
# Per-step MM data produced by get_mm_embeddings and consumed by
# Per-step MM data produced by prepare_inputs_embeds and consumed by
# prepare_inputs. Stored as raw (mm_embeds, is_mm_embed) so that
# prepare_inputs can call embed_input_ids directly into the
# persistent _inputs_embeds_buf, avoiding the intermediate copy
@@ -874,7 +874,7 @@ class DiffusionGemmaModelState(ModelState):
if idx is not None:
self.diffusion_states.remove_request(idx)
def get_mm_embeddings(
def prepare_inputs_embeds(
self,
scheduled_encoder_inputs: dict[str, list[int]],
input_batch: InputBatch,