[shortfin][python] Skip tokenization in sd.server if request holds `i…

…nput_ids` (#842)

`sd.server` currently performs tokenization irrespective of whether the
request has `input_ids` attached to it.

This patch adds a check to skip tokenization if the request has
tokenized inputs attached to it, and adds `input_ids` to `gen_inputs` so
that such requests can be created from
`InferenceExecRequests.from_batch()`.

shortfin/python/shortfin_apps/sd/components/messages.py

@@ -114,6 +114,7 @@ def from_batch(gen_req: GenerateReqInput, index: int) -> "InferenceExecRequest":
             "steps",
             "guidance_scale",
             "seed",
+            "input_ids",
         ]
         rec_inputs = {}
         for item in gen_inputs:

shortfin/python/shortfin_apps/sd/components/service.py

@@ -393,12 +393,15 @@ async def _prepare(self, device, requests):
             # Tokenize prompts and negative prompts. We tokenize in bs1 for now and join later.
             input_ids_list = []
             neg_ids_list = []
-            for tokenizer in self.service.tokenizers:
-                input_ids = tokenizer.encode(request.prompt)
-                input_ids_list.append(input_ids)
-                neg_ids = tokenizer.encode(request.neg_prompt)
-                neg_ids_list.append(neg_ids)
-            ids_list = [*input_ids_list, *neg_ids_list]
+            ids_list = request.input_ids
+            # Tokenize the prompts if the request does not hold input_ids.
+            if ids_list is None:
+                for tokenizer in self.service.tokenizers:
+                    input_ids = tokenizer.encode(request.prompt)
+                    input_ids_list.append(input_ids)
+                    neg_ids = tokenizer.encode(request.neg_prompt)
+                    neg_ids_list.append(neg_ids)
+                ids_list = [*input_ids_list, *neg_ids_list]
 
             request.input_ids = ids_list