442 lines
16 KiB
Python
442 lines
16 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 llama2.py
|
|
# Modify details for the adaptation of Qwen2 model.
|
|
"""Inference-only Qwen2 model compatible with HuggingFace weights."""
|
|
from typing import Any, Dict, Iterable, Optional, Tuple
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from sglang.srt.distributed import (
|
|
get_tensor_model_parallel_rank,
|
|
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.pooler import Pooler, PoolingType
|
|
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,
|
|
kv_cache_scales_loader,
|
|
)
|
|
from sglang.srt.utils import add_prefix, make_layers
|
|
|
|
Qwen2Config = None
|
|
|
|
|
|
class Qwen2MLP(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.down_proj = RowParallelLinear(
|
|
intermediate_size,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("down_proj", 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.down_proj(x)
|
|
return x
|
|
|
|
|
|
class Qwen2Attention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
layer_id: int = 0,
|
|
rope_theta: float = 1000000,
|
|
rope_scaling: Optional[Dict[str, Any]] = None,
|
|
max_position_embeddings: int = 32768,
|
|
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.qkv_proj = QKVParallelLinear(
|
|
hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=True,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("qkv_proj", prefix),
|
|
)
|
|
self.o_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("o_proj", 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,
|
|
num_kv_heads=self.num_kv_heads,
|
|
layer_id=layer_id,
|
|
prefix=add_prefix("attn", prefix),
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.qkv_proj(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.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class Qwen2DecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Qwen2Config,
|
|
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", 1000000)
|
|
rope_scaling = getattr(config, "rope_scaling", None)
|
|
max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
|
|
self.self_attn = Qwen2Attention(
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
layer_id=layer_id,
|
|
rope_theta=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
max_position_embeddings=max_position_embeddings,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("self_attn", prefix),
|
|
)
|
|
self.mlp = Qwen2MLP(
|
|
hidden_size=self.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
hidden_act=config.hidden_act,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("mlp", prefix),
|
|
)
|
|
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.post_attention_layernorm = 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.input_layernorm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
forward_batch=forward_batch,
|
|
)
|
|
|
|
# Fully Connected
|
|
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
|
|
hidden_states = self.mlp(hidden_states)
|
|
return hidden_states, residual
|
|
|
|
|
|
class Qwen2Model(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Qwen2Config,
|
|
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.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("embed_tokens", prefix),
|
|
)
|
|
self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda idx, prefix: Qwen2DecoderLayer(
|
|
layer_id=idx,
|
|
config=config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
),
|
|
prefix=add_prefix("layers", prefix),
|
|
)
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
if hasattr(self.config, "scale_emb"):
|
|
return self.embed_tokens(input_ids) * self.config.scale_emb
|
|
else:
|
|
return self.embed_tokens(input_ids)
|
|
|
|
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.embed_tokens(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
|
|
|
|
# If this function is called, it should always initialize KV cache scale
|
|
# factors (or else raise an exception). Thus, handled exceptions should
|
|
# make sure to leave KV cache scale factors in a known good (dummy) state
|
|
def load_kv_cache_scales(self, quantization_param_path: str) -> None:
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
tp_rank = get_tensor_model_parallel_rank()
|
|
for layer_idx, scaling_factor in kv_cache_scales_loader(
|
|
quantization_param_path,
|
|
tp_rank,
|
|
tp_size,
|
|
self.config.num_hidden_layers,
|
|
self.config.__class__.model_type,
|
|
):
|
|
if not isinstance(self.layers[layer_idx], nn.Identity):
|
|
layer_self_attn = self.layers[layer_idx].self_attn
|
|
if hasattr(layer_self_attn.attn, "k_scale"):
|
|
layer_self_attn.attn.k_scale = scaling_factor
|
|
layer_self_attn.attn.v_scale = scaling_factor
|
|
else:
|
|
raise RuntimeError(
|
|
"Self attention has no KV cache scaling " "factor attribute!"
|
|
)
|
|
|
|
|
|
class Qwen2ForCausalLM(nn.Module):
|
|
# BitandBytes specific attributes
|
|
default_bitsandbytes_target_modules = [
|
|
".gate_proj.",
|
|
".down_proj.",
|
|
".up_proj.",
|
|
".q_proj.",
|
|
".k_proj.",
|
|
".v_proj.",
|
|
".o_proj.",
|
|
]
|
|
bitsandbytes_stacked_params_mapping = {
|
|
# shard_name, weight_name, index
|
|
"q_proj": ("qkv_proj", 0),
|
|
"k_proj": ("qkv_proj", 1),
|
|
"v_proj": ("qkv_proj", 2),
|
|
"gate_proj": ("gate_up_proj", 0),
|
|
"up_proj": ("gate_up_proj", 1),
|
|
}
|
|
|
|
def __init__(
|
|
self,
|
|
config: Qwen2Config,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.model = Qwen2Model(
|
|
config, quant_config=quant_config, prefix=add_prefix("model", prefix)
|
|
)
|
|
if config.tie_word_embeddings:
|
|
self.lm_head = self.model.embed_tokens
|
|
else:
|
|
self.lm_head = ParallelLMHead(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("lm_head", prefix),
|
|
)
|
|
self.logits_processor = LogitsProcessor(config)
|
|
self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.model.get_input_embeddings(input_ids)
|
|
|
|
def get_input_embedding(self) -> nn.Embedding:
|
|
return self.model.embed_tokens
|
|
|
|
@torch.no_grad()
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: torch.Tensor = None,
|
|
get_embedding: bool = False,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
|
|
if not get_embedding:
|
|
return self.logits_processor(
|
|
input_ids, hidden_states, self.lm_head, forward_batch
|
|
)
|
|
else:
|
|
return self.pooler(hidden_states, forward_batch)
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
|
stacked_params_mapping = [
|
|
# (param_name, shard_name, shard_id)
|
|
("qkv_proj", "q_proj", "q"),
|
|
("qkv_proj", "k_proj", "k"),
|
|
("qkv_proj", "v_proj", "v"),
|
|
("gate_up_proj", "gate_proj", 0),
|
|
("gate_up_proj", "up_proj", 1),
|
|
]
|
|
|
|
params_dict = dict(self.named_parameters())
|
|
for name, loaded_weight in weights:
|
|
if "rotary_emb.inv_freq" in name or "projector" in name:
|
|
continue
|
|
if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
|
|
# Models trained using ColossalAI may include these tensors in
|
|
# the checkpoint. Skip them.
|
|
continue
|
|
if self.config.tie_word_embeddings and "lm_head.weight" in name:
|
|
continue
|
|
if name.startswith("model.vision_tower") and name not in params_dict:
|
|
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)
|
|
|
|
def get_embed_and_head(self):
|
|
return self.model.embed_tokens.weight, self.lm_head.weight
|
|
|
|
def set_embed_and_head(self, embed, head):
|
|
del self.model.embed_tokens.weight
|
|
del self.lm_head.weight
|
|
self.model.embed_tokens.weight = embed
|
|
self.lm_head.weight = head
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.synchronize()
|
|
|
|
def load_kv_cache_scales(self, quantization_param_path: str) -> None:
|
|
self.model.load_kv_cache_scales(quantization_param_path)
|
|
|
|
|
|
EntryClass = Qwen2ForCausalLM
|