Migrate logits computation and gather to model_runner (#3233)
This commit is contained in:
@@ -37,6 +37,7 @@ from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
QKVParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding, ParallelLMHead, DEFAULT_VOCAB_PADDING_SIZE)
|
||||
@@ -325,7 +326,11 @@ class LlamaForCausalLM(nn.Module):
|
||||
# compatibility
|
||||
if not lora_config else lora_config.lora_vocab_padding_size,
|
||||
)
|
||||
self.sampler = Sampler(self.unpadded_vocab_size, config.vocab_size)
|
||||
|
||||
logit_scale = getattr(config, "logit_scale", 1.0)
|
||||
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
|
||||
config.vocab_size, logit_scale)
|
||||
self.sampler = Sampler()
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -338,13 +343,18 @@ class LlamaForCausalLM(nn.Module):
|
||||
input_metadata)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(self, hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata) -> torch.Tensor:
|
||||
logits = self.logits_processor(self.lm_head.weight, hidden_states,
|
||||
sampling_metadata)
|
||||
return logits
|
||||
|
||||
def sample(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
logits: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[SamplerOutput]:
|
||||
next_tokens = self.sampler(self.lm_head.weight, hidden_states,
|
||||
sampling_metadata)
|
||||
next_tokens = self.sampler(logits, sampling_metadata)
|
||||
return next_tokens
|
||||
|
||||
def load_weights(self,
|
||||
|
||||
Reference in New Issue
Block a user