mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2025-12-15 08:35:48 +08:00
339 lines
12 KiB
Python
339 lines
12 KiB
Python
# -*- coding: utf-8 -*-
|
|
from typing import Any, Dict, Iterable, List, Optional, Tuple
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers import PretrainedConfig
|
|
|
|
from vllm.attention import Attention, AttentionMetadata
|
|
from vllm.config import CacheConfig
|
|
from vllm.distributed import get_tensor_model_parallel_world_size
|
|
from vllm.model_executor.layers.activation import SiluAndMul
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
|
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, SamplerOutput
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
|
ParallelLMHead, VocabParallelEmbedding)
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
|
from vllm.sequence import IntermediateTensors
|
|
|
|
|
|
class InternLM2MLP(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
intermediate_size: int,
|
|
hidden_act: str,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.gate_up_proj = MergedColumnParallelLinear(
|
|
hidden_size, [intermediate_size] * 2,
|
|
bias=False,
|
|
quant_config=quant_config)
|
|
self.w2 = RowParallelLinear(intermediate_size,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config)
|
|
if hidden_act != "silu":
|
|
raise ValueError(f"Unsupported activation: {hidden_act}. "
|
|
"Only silu is supported for now.")
|
|
self.act_fn = SiluAndMul()
|
|
|
|
def forward(self, x):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.w2(x)
|
|
return x
|
|
|
|
|
|
class InternLM2Attention(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
rope_theta: float = 10000,
|
|
rope_scaling: Optional[Dict[str, Any]] = None,
|
|
max_position_embeddings: int = 8192,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
self.total_num_heads = num_heads
|
|
assert self.total_num_heads % tp_size == 0
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.total_num_kv_heads = num_kv_heads
|
|
if self.total_num_kv_heads >= tp_size:
|
|
# Number of KV heads is greater than TP size, so we partition
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert self.total_num_kv_heads % tp_size == 0
|
|
else:
|
|
# Number of KV heads is less than TP size, so we replicate
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
|
self.head_dim = hidden_size // self.total_num_heads
|
|
self.q_size = self.num_heads * self.head_dim
|
|
self.kv_size = self.num_kv_heads * self.head_dim
|
|
self.key_value_groups = int(self.num_heads / self.num_kv_heads)
|
|
self.scaling = self.head_dim**-0.5
|
|
self.rope_theta = rope_theta
|
|
self.max_position_embeddings = max_position_embeddings
|
|
|
|
self.wqkv = QKVParallelLinear(
|
|
hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
)
|
|
self.wo = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
)
|
|
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
rotary_dim=self.head_dim,
|
|
max_position=max_position_embeddings,
|
|
base=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
)
|
|
self.attn = Attention(self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config)
|
|
|
|
def split_qkv(self, qkv: torch.Tensor):
|
|
qkv = qkv.view(-1, self.num_kv_heads, self.key_value_groups + 2, 128)
|
|
q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=2)
|
|
q = q.reshape(-1, self.q_size)
|
|
k = k.reshape(-1, self.kv_size)
|
|
v = v.reshape(-1, self.kv_size)
|
|
return q, k, v
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
kv_cache: torch.Tensor,
|
|
attn_metadata: AttentionMetadata,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.wqkv(hidden_states)
|
|
q, k, v = self.split_qkv(qkv)
|
|
q, k = self.rotary_emb(positions, q, k)
|
|
attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
|
|
output, _ = self.wo(attn_output)
|
|
return output
|
|
|
|
|
|
class InternLMDecoderLayer(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
rope_theta = getattr(config, "rope_theta", 10000)
|
|
rope_scaling = getattr(config, "rope_scaling", None)
|
|
max_position_embeddings = getattr(config, "max_position_embeddings",
|
|
8192)
|
|
self.attention = InternLM2Attention(
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
rope_theta=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
max_position_embeddings=max_position_embeddings,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
)
|
|
self.feed_forward = InternLM2MLP(
|
|
hidden_size=self.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
hidden_act=config.hidden_act,
|
|
quant_config=quant_config,
|
|
)
|
|
self.attention_norm = RMSNorm(config.hidden_size,
|
|
eps=config.rms_norm_eps)
|
|
self.ffn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
kv_cache: torch.Tensor,
|
|
attn_metadata: AttentionMetadata,
|
|
residual: Optional[torch.Tensor],
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
# Self Attention
|
|
if residual is None:
|
|
residual = hidden_states
|
|
hidden_states = self.attention_norm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.attention_norm(
|
|
hidden_states, residual)
|
|
hidden_states = self.attention(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
kv_cache=kv_cache,
|
|
attn_metadata=attn_metadata,
|
|
)
|
|
|
|
# Fully Connected
|
|
hidden_states, residual = self.ffn_norm(hidden_states, residual)
|
|
hidden_states = self.feed_forward(hidden_states)
|
|
return hidden_states, residual
|
|
|
|
|
|
class InternLM2Model(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.padding_idx = config.pad_token_id
|
|
self.vocab_size = config.vocab_size
|
|
self.tok_embeddings = VocabParallelEmbedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
)
|
|
self.layers = nn.ModuleList([
|
|
InternLMDecoderLayer(config, cache_config, quant_config)
|
|
for _ in range(config.num_hidden_layers)
|
|
])
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.tok_embeddings(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
kv_caches: List[torch.Tensor],
|
|
attn_metadata: AttentionMetadata,
|
|
intermediate_tensors: IntermediateTensors = None,
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
if inputs_embeds is not None:
|
|
hidden_states = inputs_embeds
|
|
else:
|
|
hidden_states = self.tok_embeddings(input_ids)
|
|
residual = None
|
|
for i in range(len(self.layers)):
|
|
layer = self.layers[i]
|
|
hidden_states, residual = layer(
|
|
positions,
|
|
hidden_states,
|
|
kv_caches[i],
|
|
attn_metadata,
|
|
residual,
|
|
)
|
|
hidden_states, _ = self.norm(hidden_states, residual)
|
|
return hidden_states
|
|
|
|
|
|
class InternLM2ForCausalLM(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.model = InternLM2Model(config, cache_config, quant_config)
|
|
self.output = ParallelLMHead(config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config)
|
|
if self.config.tie_word_embeddings:
|
|
self.output.weight = self.model.tok_embeddings.weight
|
|
self.logits_processor = LogitsProcessor(config.vocab_size)
|
|
self.sampler = Sampler()
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
kv_caches: List[torch.Tensor],
|
|
attn_metadata: AttentionMetadata,
|
|
intermediate_tensors: IntermediateTensors,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(input_ids, positions, kv_caches,
|
|
attn_metadata)
|
|
return hidden_states
|
|
|
|
def compute_logits(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
sampling_metadata: SamplingMetadata,
|
|
) -> Optional[torch.Tensor]:
|
|
logits = self.logits_processor(self.output, hidden_states,
|
|
sampling_metadata)
|
|
return logits
|
|
|
|
def sample(
|
|
self,
|
|
logits: torch.Tensor,
|
|
sampling_metadata: SamplingMetadata,
|
|
) -> Optional[SamplerOutput]:
|
|
next_tokens = self.sampler(logits, sampling_metadata)
|
|
return next_tokens
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
|
stacked_params_mapping = [
|
|
# (param_name, shard_name, shard_id)
|
|
("gate_up_proj", "w1", 0),
|
|
("gate_up_proj", "w3", 1),
|
|
]
|
|
params_dict = dict(self.named_parameters())
|
|
for name, loaded_weight in weights:
|
|
if "rotary_emb.inv_freq" in name:
|
|
continue
|
|
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
|
if weight_name not in name:
|
|
continue
|
|
name = name.replace(weight_name, param_name)
|
|
# Skip loading extra bias for GPTQ models.
|
|
if name.endswith(".bias") and name not in params_dict:
|
|
continue
|
|
param = params_dict[name]
|
|
weight_loader = param.weight_loader
|
|
weight_loader(param, loaded_weight, shard_id)
|
|
break
|
|
else:
|
|
# Skip loading extra bias for GPTQ models.
|
|
if name.endswith(".bias") and name not in params_dict:
|
|
continue
|
|
param = params_dict[name]
|
|
weight_loader = getattr(param, "weight_loader",
|
|
default_weight_loader)
|
|
weight_loader(param, loaded_weight)
|