sglang0.4.5.post1/python/sglang/srt/models/internlm2.py

354 lines
13 KiB
Python

# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/7f62077af5159c625fe3ad1c812e6c1a2b93ba3b/vllm/model_executor/models/internlm2.py
from typing import Any, Dict, Iterable, Optional, Tuple
import torch
from torch import nn
from transformers import PretrainedConfig
from sglang.srt.distributed import get_tensor_model_parallel_world_size
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.linear import (
MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.rotary_embedding import get_rope
from sglang.srt.layers.vocab_parallel_embedding import (
ParallelLMHead,
VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.utils import add_prefix
class InternLM2MLP(nn.Module):
def __init__(
self,
hidden_size: int,
intermediate_size: int,
hidden_act: str,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size,
[intermediate_size] * 2,
bias=False,
quant_config=quant_config,
prefix=add_prefix("gate_up_proj", prefix),
)
self.w2 = RowParallelLinear(
intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
prefix=add_prefix("w2", prefix),
)
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,
layer_id: int = 0,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> 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.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,
prefix=add_prefix("wqkv", prefix),
)
self.wo = RowParallelLinear(
self.total_num_heads * self.head_dim,
hidden_size,
bias=False,
quant_config=quant_config,
prefix=add_prefix("wo", prefix),
)
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 = RadixAttention(
self.num_heads,
self.head_dim,
self.scaling,
self.num_kv_heads,
layer_id,
prefix=add_prefix("attn", prefix),
)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
forward_batch: ForwardBatch,
) -> torch.Tensor:
qkv, _ = self.wqkv(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
q, k = self.rotary_emb(positions, q, k)
attn_output = self.attn(q, k, v, forward_batch)
output, _ = self.wo(attn_output)
return output
class InternLMDecoderLayer(nn.Module):
def __init__(
self,
config: PretrainedConfig,
layer_id: int = 0,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> 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,
layer_id=layer_id,
quant_config=quant_config,
prefix=add_prefix("attention", prefix),
)
self.feed_forward = InternLM2MLP(
hidden_size=self.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
prefix=add_prefix("feed_forward", prefix),
)
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,
forward_batch: ForwardBatch,
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,
forward_batch=forward_batch,
)
# 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,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> 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,
prefix=add_prefix("tok_embeddings", prefix),
)
self.layers = nn.ModuleList(
[
InternLMDecoderLayer(
config, i, quant_config, prefix=add_prefix(f"layers.{i}", prefix)
)
for i in range(config.num_hidden_layers)
]
)
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
forward_batch: ForwardBatch,
input_embeds: torch.Tensor = None,
) -> torch.Tensor:
if input_embeds is None:
hidden_states = self.tok_embeddings(input_ids)
else:
hidden_states = input_embeds
residual = None
for i in range(len(self.layers)):
layer = self.layers[i]
hidden_states, residual = layer(
positions,
hidden_states,
forward_batch,
residual,
)
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
class InternLM2ForCausalLM(nn.Module):
def __init__(
self,
config: PretrainedConfig,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.config = config
self.quant_config = quant_config
self.model = InternLM2Model(
config, quant_config, prefix=add_prefix("model", prefix)
)
self.output = ParallelLMHead(
config.vocab_size, config.hidden_size, prefix=add_prefix("output", prefix)
)
self.logits_processor = LogitsProcessor(config)
@torch.no_grad()
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
forward_batch: ForwardBatch,
input_embeds: torch.Tensor = None,
) -> torch.Tensor:
hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
return self.logits_processor(
input_ids, hidden_states, self.output, forward_batch
)
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]
if "wqkv" in name:
config = self.config
kv_groups = config.num_attention_heads // config.num_key_value_heads
head_dim = config.hidden_size // config.num_attention_heads
loaded_weight = loaded_weight.view(
-1, 2 + kv_groups, head_dim, loaded_weight.shape[-1]
)
wq, wk, wv = torch.split(loaded_weight, [kv_groups, 1, 1], dim=1)
wq = wq.reshape(-1, wq.shape[-1])
wk = wk.reshape(-1, wk.shape[-1])
wv = wv.reshape(-1, wv.shape[-1])
weight_loader = param.weight_loader
weight_loader(param, wq, "q")
weight_loader(param, wk, "k")
weight_loader(param, wv, "v")
else:
weight_loader = getattr(
param, "weight_loader", default_weight_loader
)
weight_loader(param, loaded_weight)
EntryClass = InternLM2ForCausalLM