chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Update model_loader deps and qqq quantization deps (#2220) (#2318)

Browse Source 
Co-authored-by: HandH1998 <1335248067@qq.com>
This commit is contained in:
Yineng Zhang
2024-12-02 23:22:13 +08:00
committed by GitHub
parent 33deca81b5
commit 85e1a6f3aa
58 changed files with 2363 additions and 366 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/models/deepseek.py
View File

@@ -27,7 +27,6 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.fused_moe_triton import fused_moe
@@ -46,6 +45,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
class DeepseekMLP(nn.Module):
@@ -184,7 +184,6 @@ class DeepseekAttention(nn.Module):
rope_theta: float = 10000,
rope_scaling: Optional[Dict[str, Any]] = None,
max_position_embeddings: int = 8192,
cache_config=None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -261,7 +260,6 @@ class DeepseekDecoderLayer(nn.Module):
self,
config: PretrainedConfig,
layer_id: int,
cache_config=None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -277,7 +275,6 @@ class DeepseekDecoderLayer(nn.Module):
rope_theta=rope_theta,
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,
)
if (
@@ -330,7 +327,6 @@ class DeepseekModel(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config=None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
@@ -343,9 +339,7 @@ class DeepseekModel(nn.Module):
)
self.layers = nn.ModuleList(
[
DeepseekDecoderLayer(
config, layer_id, cache_config, quant_config=quant_config
)
DeepseekDecoderLayer(config, layer_id, quant_config=quant_config)
for layer_id in range(config.num_hidden_layers)
]
)
@@ -373,13 +367,12 @@ class DeepseekForCausalLM(nn.Module):
def __init__(
self,
config: PretrainedConfig,
cache_config=None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.config = config
self.quant_config = quant_config
self.model = DeepseekModel(config, cache_config, quant_config)
self.model = DeepseekModel(config, quant_config)
self.lm_head = ParallelLMHead(
config.vocab_size, config.hidden_size, quant_config=quant_config
)