[Misc][Refactor] Generalize linear_method to be quant_method (#4373)
This commit is contained in:
@@ -27,11 +27,12 @@ from vllm.distributed import get_tensor_model_parallel_world_size
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import GeluAndMul
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
MergedColumnParallelLinear,
|
||||
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
@@ -77,17 +78,17 @@ class GemmaMLP(nn.Module):
|
||||
intermediate_size: int,
|
||||
hidden_act: Optional[str] = None,
|
||||
hidden_activation: Optional[str] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
hidden_size, [intermediate_size] * 2,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
quant_config=quant_config)
|
||||
self.down_proj = RowParallelLinear(intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
quant_config=quant_config)
|
||||
self.act_fn = _get_gemma_act_fn(hidden_act, hidden_activation)
|
||||
|
||||
def forward(self, x):
|
||||
@@ -106,7 +107,7 @@ class GemmaAttention(nn.Module):
|
||||
head_dim: int,
|
||||
max_position_embeddings: int = 8192,
|
||||
rope_theta: float = 10000,
|
||||
linear_method: Optional[LinearMethodBase] = None) -> None:
|
||||
quant_config: Optional[QuantizationConfig] = None) -> None:
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
@@ -135,13 +136,13 @@ class GemmaAttention(nn.Module):
|
||||
self.total_num_heads,
|
||||
self.total_num_kv_heads,
|
||||
bias=False,
|
||||
linear_method=linear_method,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
self.o_proj = RowParallelLinear(
|
||||
self.total_num_heads * self.head_dim,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
linear_method=linear_method,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
|
||||
self.rotary_emb = get_rope(
|
||||
@@ -176,7 +177,7 @@ class GemmaDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: GemmaConfig,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.hidden_size = config.hidden_size
|
||||
@@ -187,14 +188,14 @@ class GemmaDecoderLayer(nn.Module):
|
||||
head_dim=config.head_dim,
|
||||
max_position_embeddings=config.max_position_embeddings,
|
||||
rope_theta=config.rope_theta,
|
||||
linear_method=linear_method,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
self.mlp = GemmaMLP(
|
||||
hidden_size=self.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
hidden_activation=getattr(config, "hidden_activation", None),
|
||||
linear_method=linear_method,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
self.input_layernorm = RMSNorm(config.hidden_size,
|
||||
eps=config.rms_norm_eps)
|
||||
@@ -235,7 +236,7 @@ class GemmaModel(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: GemmaConfig,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.config = config
|
||||
@@ -245,7 +246,7 @@ class GemmaModel(nn.Module):
|
||||
config.hidden_size,
|
||||
)
|
||||
self.layers = nn.ModuleList([
|
||||
GemmaDecoderLayer(config, linear_method)
|
||||
GemmaDecoderLayer(config, quant_config)
|
||||
for _ in range(config.num_hidden_layers)
|
||||
])
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
@@ -308,14 +309,14 @@ class GemmaForCausalLM(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: GemmaConfig,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
del lora_config # Unused.
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.linear_method = linear_method
|
||||
self.model = GemmaModel(config, linear_method)
|
||||
self.quant_config = quant_config
|
||||
self.model = GemmaModel(config, quant_config)
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.sampler = Sampler()
|
||||
|
||||
|
||||
Reference in New Issue
Block a user