biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source
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
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/llama.py
View File

@@ -39,8 +39,7 @@ from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.quantized_linear import ParallelLinear
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)
from vllm.model_executor.parallel_utils.layers import VocabParallelEmbedding
from vllm.model_executor.quantization_utils import QuantizationConfig
from vllm.model_executor.weight_utils import (
convert_pyslice_to_tensor, hf_model_weights_iterator,
@@ -64,13 +63,11 @@ class LlamaMLP(nn.Module):
2 * intermediate_size,
bias=False,
gather_output=False,
perform_initialization=False,
quant_config=quant_config)
self.down_proj = ParallelLinear.row(intermediate_size,
hidden_size,
bias=False,
input_is_parallel=True,
perform_initialization=False,
quant_config=quant_config)
if hidden_act != "silu":
raise ValueError(f"Unsupported activation: {hidden_act}. "
@@ -127,7 +124,6 @@ class LlamaAttention(nn.Module):
self.head_dim,
bias=False,
gather_output=False,
perform_initialization=False,
quant_config=quant_config,
)
self.o_proj = ParallelLinear.row(
@@ -135,7 +131,6 @@ class LlamaAttention(nn.Module):
hidden_size,
bias=False,
input_is_parallel=True,
perform_initialization=False,
quant_config=quant_config,
)
self.attn = PagedAttentionWithRoPE(
@@ -241,7 +236,9 @@ class LlamaModel(nn.Module):
vocab_size = ((config.vocab_size + 63) // 64) * 64
self.embed_tokens = VocabParallelEmbedding(
vocab_size, config.hidden_size, perform_initialization=False)
vocab_size,
config.hidden_size,
)
self.layers = nn.ModuleList([
LlamaDecoderLayer(config, quant_config)
for _ in range(config.num_hidden_layers)
@@ -291,7 +288,6 @@ class LlamaForCausalLM(nn.Module):
vocab_size,
bias=False,
gather_output=False,
perform_initialization=False,
quant_config=None)
self.sampler = Sampler(config.vocab_size)