[Misc][Refactor] Generalize linear_method to be quant_method (#4373)

This commit is contained in:
Cody Yu
2024-04-26 16:41:14 -04:00
committed by GitHub
parent 603ad84815
commit a62aaf1df5
45 changed files with 759 additions and 713 deletions
+21 -20
View File
@@ -34,11 +34,12 @@ from vllm.distributed import (get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (LinearMethodBase,
QKVParallelLinear,
from vllm.model_executor.layers.linear import (QKVParallelLinear,
ReplicatedLinear,
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 (
@@ -55,7 +56,7 @@ class MixtralMLP(nn.Module):
num_experts: int,
hidden_size: int,
intermediate_size: int,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.num_experts = num_experts
@@ -65,15 +66,15 @@ class MixtralMLP(nn.Module):
self.w1 = ReplicatedLinear(self.hidden_dim,
self.ffn_dim,
bias=False,
linear_method=linear_method)
quant_config=quant_config)
self.w2 = ReplicatedLinear(self.ffn_dim,
self.hidden_dim,
bias=False,
linear_method=linear_method)
quant_config=quant_config)
self.w3 = ReplicatedLinear(self.hidden_dim,
self.ffn_dim,
bias=False,
linear_method=linear_method)
quant_config=quant_config)
# TODO: Use vllm's SiluAndMul
self.act_fn = nn.SiLU()
@@ -92,7 +93,7 @@ class MixtralMoE(nn.Module):
def __init__(
self,
config: MixtralConfig,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
):
super().__init__()
self.config = config
@@ -115,14 +116,14 @@ class MixtralMoE(nn.Module):
MixtralMLP(self.num_total_experts,
config.hidden_size,
config.intermediate_size,
linear_method=linear_method)
quant_config=quant_config)
if idx in self.expert_indicies else None
for idx in range(self.num_total_experts)
])
self.gate = ReplicatedLinear(config.hidden_size,
self.num_total_experts,
bias=False,
linear_method=None)
quant_config=None)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
num_tokens, hidden_dim = hidden_states.shape
@@ -162,7 +163,7 @@ class MixtralAttention(nn.Module):
num_kv_heads: int,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
sliding_window: Optional[int] = None) -> None:
super().__init__()
self.hidden_size = hidden_size
@@ -193,13 +194,13 @@ class MixtralAttention(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(
self.head_dim,
@@ -236,7 +237,7 @@ class MixtralDecoderLayer(nn.Module):
def __init__(
self,
config: MixtralConfig,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
@@ -249,9 +250,9 @@ class MixtralDecoderLayer(nn.Module):
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
sliding_window=config.sliding_window,
linear_method=linear_method)
quant_config=quant_config)
self.block_sparse_moe = MixtralMoE(config=config,
linear_method=linear_method)
quant_config=quant_config)
self.input_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(config.hidden_size,
@@ -291,7 +292,7 @@ class MixtralModel(nn.Module):
def __init__(
self,
config: MixtralConfig,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.padding_idx = config.pad_token_id
@@ -302,7 +303,7 @@ class MixtralModel(nn.Module):
config.hidden_size,
)
self.layers = nn.ModuleList([
MixtralDecoderLayer(config, linear_method=linear_method)
MixtralDecoderLayer(config, quant_config=quant_config)
for _ in range(config.num_hidden_layers)
])
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -331,12 +332,12 @@ class MixtralForCausalLM(nn.Module):
def __init__(
self,
config: MixtralConfig,
linear_method: Optional[LinearMethodBase] = None,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
self.config = config
self.linear_method = linear_method
self.model = MixtralModel(config, linear_method)
self.quant_config = quant_config
self.model = MixtralModel(config, quant_config)
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = Sampler()