[V1] V1 Enablement Oracle (#13726)

Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
Robert Shaw
2025-03-14 22:02:20 -07:00
committed by GitHub
co-authored by rshaw@neuralmagic.com <rshaw@neuralmagic.com> Nicolò Lucchesi Tyler Michael Smith Michael Goin
parent 8c0d15d5c5
commit d4d93db2c5
96 changed files with 1537 additions and 512 deletions
+16 -3
View File
@@ -184,7 +184,7 @@ class Processor:
# Only applicable to multimodal models with legacy input processor.
processed_inputs = self.input_processor(preprocessed_inputs)
self._validate_model_inputs(processed_inputs)
self._validate_model_inputs(processed_inputs, lora_request)
if is_encoder_decoder_inputs(processed_inputs):
decoder_inputs = SingletonInputsAdapter(
@@ -200,8 +200,12 @@ class Processor:
raise NotImplementedError
assert isinstance(params, SamplingParams)
# TODO: can we avoid cloning here in multiproc case
# TODO: can we avoid cloning here in multiproc case?
sampling_params = params.clone()
# If unset max tokens, then generate up to the max_model_len.
if sampling_params.max_tokens is None:
sampling_params.max_tokens = (self.model_config.max_model_len -
len(decoder_inputs.prompt_token_ids))
sampling_params.update_from_generation_config(
self.generation_config_fields, eos_token_id)
sampling_params.update_from_tokenizer(
@@ -296,7 +300,9 @@ class Processor:
lora_request=lora_request,
)
def _validate_model_inputs(self, inputs: ProcessorInputs):
def _validate_model_inputs(self,
inputs: ProcessorInputs,
lora_request: Optional[LoRARequest] = None):
if is_encoder_decoder_inputs(inputs):
# For encoder-decoder multimodal models, the max_prompt_len
# restricts the decoder prompt length
@@ -310,6 +316,13 @@ class Processor:
if prompt_ids is None or len(prompt_ids) == 0:
raise ValueError("Prompt cannot be empty")
max_input_id = max(prompt_ids)
max_allowed = self.tokenizer.get_lora_tokenizer(
lora_request).max_token_id
if max_input_id > max_allowed:
raise ValueError(
"Token id {} is out of vocabulary".format(max_input_id))
if len(prompt_ids) >= self.model_config.max_model_len:
raise ValueError(
f"Prompt length of {len(prompt_ids)} is longer than the "