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 16, 07:00 UTC
Add default value to avoid Falcon crash (#5363) (#10347)
Signed-off-by: wchen61 <wchen61@foxmail.com>
wchen61 committed
Nov 15, 2024 at 16:52 UTC
3d158cdc8dad62dfed45d5d808ae9f14f16e4dae
1 file changed
+3
vllm/model_executor/models/falcon.py
+3
@@ -250,6 +250,9 @@ class FalconDecoderLayer(nn.Module):
250
self.mlp = FalconMLP(config, quant_config)
251
self.config = config
252
253
+ if (not hasattr(config, "num_ln_in_parallel_attn")):
254
+ config.num_ln_in_parallel_attn = None
255
+
256
if (config.num_ln_in_parallel_attn is None
257
and config.new_decoder_architecture):
258
config.num_ln_in_parallel_attn = 2