[V1] Make v1 more testable (#9888)
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
This commit is contained in:
@@ -44,7 +44,7 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
|
||||
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
@@ -1141,7 +1141,7 @@ class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(config.output_hidden_states,
|
||||
config.text_config.vocab_size)
|
||||
self.sampler = Sampler()
|
||||
self.sampler = get_sampler()
|
||||
|
||||
def compute_logits(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user