[Fix] Reset output_ids for requests with input_embeds during retraction (#14110)

python/sglang/srt/managers/schedule_batch.py

@@ -1121,6 +1121,14 @@ class Req(ReqDllmMixin):
         self.extend_batch_idx = 0
         self.decode_batch_idx = 0
 
+        # When using input_embeds, we cannot easily mix the original input embeddings
+        # with the newly generated output token IDs during re-prefill of retracted request.
+        # output_ids will have no use, but will lead to wrong size cache indexes.
+        # Therefore, we discard the generated output_ids and restart prefill and generation
+        # to ensure shape consistency in KV cache.
+        if self.input_embeds is not None:
+            self.output_ids = []
+
     def offload_kv_cache(self, req_to_token_pool, token_to_kv_pool_allocator):
         token_indices = req_to_token_pool.req_to_token[
             self.req_pool_idx, : self.seqlen - 1