TP/quantization/weight loading refactor part 1 - Simplify parallel linear logic (#1181)

This commit is contained in:
Zhuohan Li
2023-10-02 15:36:09 -07:00
committed by GitHub
parent 84e4e37d14
commit ba0bfd40e2
42 changed files with 819 additions and 1547 deletions
+27 -22
View File
@@ -36,8 +36,9 @@ from vllm.model_executor.weight_utils import (
load_padded_tensor_parallel_vocab, load_tensor_parallel_weights)
from vllm.model_executor.parallel_utils.parallel_state import (
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
from vllm.model_executor.parallel_utils.tensor_parallel import (
VocabParallelEmbedding, ColumnParallelLinear, RowParallelLinear)
from vllm.model_executor.parallel_utils.layers import (VocabParallelEmbedding,
ColumnParallelLinear,
RowParallelLinear)
from vllm.sequence import SamplerOutput
KVCache = Tuple[torch.Tensor, torch.Tensor]
@@ -56,16 +57,18 @@ class GPT2Attention(nn.Module):
self.head_dim = self.hidden_size // total_num_heads
self.scale = self.head_dim**-0.5
self.c_attn = ColumnParallelLinear(self.hidden_size,
3 * self.hidden_size,
bias=True,
gather_output=False,
perform_initialization=False)
self.c_proj = RowParallelLinear(self.hidden_size,
self.hidden_size,
bias=True,
input_is_parallel=True,
perform_initialization=False)
self.c_attn = ColumnParallelLinear(
self.hidden_size,
3 * self.hidden_size,
bias=True,
gather_output=False,
)
self.c_proj = RowParallelLinear(
self.hidden_size,
self.hidden_size,
bias=True,
input_is_parallel=True,
)
self.attn = PagedAttention(self.num_heads,
self.head_dim,
scale=self.scale)
@@ -95,16 +98,18 @@ class GPT2MLP(nn.Module):
):
super().__init__()
hidden_size = config.hidden_size
self.c_fc = ColumnParallelLinear(hidden_size,
intermediate_size,
bias=True,
gather_output=False,
perform_initialization=False)
self.c_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=True,
input_is_parallel=True,
perform_initialization=False)
self.c_fc = ColumnParallelLinear(
hidden_size,
intermediate_size,
bias=True,
gather_output=False,
)
self.c_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=True,
input_is_parallel=True,
)
self.act = get_act_fn(config.activation_function)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: