@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 90979c38f8
Fork 0

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 17, 06:45 UTC

[FIX] Don't initialize parameter by default (#1067)

Zhuohan Li committed Sep 17, 2023 at 17:15 UTC 90979c38f87c17d53a7cd0eb430373ecb0b64b9a
1 file changed +4 -4
vllm/model_executor/parallel_utils/tensor_parallel/layers.py
+4 -4
@@ -83,7 +83,7 @@ class VocabParallelEmbedding(torch.nn.Module):
83 init_method=init.xavier_normal_,
84 params_dtype: torch.dtype=None,
85 use_cpu_initialization: bool=False,
86 - perform_initialization: bool=True):
86 + perform_initialization: bool=False):
87 super(VocabParallelEmbedding, self).__init__()
88 assert not perform_initialization
89 assert not use_cpu_initialization
@@ -113,7 +113,7 @@ class VocabParallelEmbedding(torch.nn.Module):
113 self.weight = Parameter(torch.empty(
114 self.num_embeddings_per_partition, self.embedding_dim,
115 device=torch.cuda.current_device(), dtype=params_dtype))
116 -
116 +
117 def forward(self, input_):
118 if self.tensor_model_parallel_size > 1:
119 # Build the mask.
@@ -172,7 +172,7 @@ class ColumnParallelLinear(torch.nn.Module):
172 skip_bias_add=False,
173 params_dtype=None,
174 use_cpu_initialization=False,
175 - perform_initialization=True,
175 + perform_initialization=False,
176 quant_config=None,
177 ):
178 super(ColumnParallelLinear, self).__init__()
@@ -288,7 +288,7 @@ class RowParallelLinear(torch.nn.Module):
288 skip_bias_add=False,
289 params_dtype=None,
290 use_cpu_initialization=False,
291 - perform_initialization=True,
291 + perform_initialization=False,
292 reduce_results=True,
293 quant_config=None,
294 ):