[Bugfix] support tie_word_embeddings for all models (#5724)
This commit is contained in:
@@ -37,7 +37,7 @@ from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding)
|
||||
ParallelLMHead, VocabParallelEmbedding)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.sequence import IntermediateTensors, SamplerOutput
|
||||
@@ -291,7 +291,11 @@ class JAISLMHeadModel(nn.Module):
|
||||
self.config = config
|
||||
self.quant_config = quant_config
|
||||
self.transformer = JAISModel(config, cache_config, quant_config)
|
||||
self.lm_head = self.transformer.wte
|
||||
if self.config.tie_word_embeddings:
|
||||
self.lm_head = self.transformer.wte
|
||||
else:
|
||||
self.lm_head = ParallelLMHead(self.config.vocab_size,
|
||||
self.config.hidden_size)
|
||||
if hasattr(config, "width_scale"):
|
||||
self.output_logits_scale = config.width_scale
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user