A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
optimize: eliminate duplicate split_enc_dec_inputs calls (#25573)
Signed-off-by: nicole-lihui <nicole.li@daocloud.io>
Nicole LiHui 🥜 committed
Sep 25, 2025 at 13:03 UTC
c85be1f6dd3e20d9b42cd68ff54b328ffeb6cb4b
1 file changed
+4
-5
vllm/v1/engine/processor.py
+4
-5
@@ -388,9 +388,9 @@ class Processor:
388
389
eos_token_id = self.input_preprocessor.get_eos_token_id()
390
391
- self._validate_model_inputs(processed_inputs)
392
-
391
encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
392
+ self._validate_model_inputs(encoder_inputs, decoder_inputs)
393
+
394
# Mypy does not always properly infer the types of some elements of
395
# discriminated unions of TypedDicts, because of how it handles
396
# inheritance of TypedDict. If we explicitly extract the items we want
@@ -458,9 +458,8 @@ class Processor:
458
trace_headers=trace_headers,
459
)
460
461
- def _validate_model_inputs(self, inputs: ProcessorInputs):
462
- encoder_inputs, decoder_inputs = split_enc_dec_inputs(inputs)
463
-
461
+ def _validate_model_inputs(self, encoder_inputs: Optional[SingletonInputs],
462
+ decoder_inputs: SingletonInputs):
463
if encoder_inputs is not None:
464
self._validate_model_input(encoder_inputs, prompt_type="encoder")
465