@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 462ae5220a
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, 08:00 UTC

[Fix] unwantted bias in InternLM Model (#740)

WRH committed Aug 12, 2023 at 02:40 UTC 462ae5220aeeb2135b841107d2c841f85fc348bd
1 file changed +6 -6
vllm/model_executor/models/internlm.py
+6 -6
@@ -7,15 +7,15 @@ from transformers import LlamaConfig
7
8 from vllm.model_executor.input_metadata import InputMetadata
9 from vllm.model_executor.layers.activation import SiluAndMul
10 -from vllm.model_executor.layers.layernorm import RMSNorm
10 from vllm.model_executor.layers.attention import PagedAttentionWithRoPE
11 +from vllm.model_executor.layers.layernorm import RMSNorm
12 from vllm.model_executor.layers.sampler import Sampler
13 -from vllm.model_executor.weight_utils import (hf_model_weights_iterator,
14 - load_tensor_parallel_weights)
13 from vllm.model_executor.parallel_utils.parallel_state import (
14 get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
15 from vllm.model_executor.parallel_utils.tensor_parallel import (
18 - VocabParallelEmbedding, ColumnParallelLinear, RowParallelLinear)
16 + ColumnParallelLinear, RowParallelLinear, VocabParallelEmbedding)
17 +from vllm.model_executor.weight_utils import (hf_model_weights_iterator,
18 + load_tensor_parallel_weights)
19 from vllm.sequence import SequenceOutputs
20
21 KVCache = Tuple[torch.Tensor, torch.Tensor]
@@ -32,12 +32,12 @@ class InternLMMLP(nn.Module):
32 super().__init__()
33 self.gate_up_proj = ColumnParallelLinear(hidden_size,
34 2 * intermediate_size,
35 - bias=True,
35 + bias=False,
36 gather_output=False,
37 perform_initialization=False)
38 self.down_proj = RowParallelLinear(intermediate_size,
39 hidden_size,
40 - bias=True,
40 + bias=False,
41 input_is_parallel=True,
42 perform_initialization=False)
43 if hidden_act != "silu":