This reverts commit 5c976a7e1a.
This commit is contained in:
Philipp Moritz
@@ -18,19 +18,305 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Inference-only BaiChuan model compatible with HuggingFace weights."""
|
||||
from typing import Optional
|
||||
import math
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from vllm.config import LoRAConfig
|
||||
from torch import nn
|
||||
|
||||
from vllm.model_executor.layers.linear import LinearMethodBase
|
||||
from vllm.model_executor.models.llama import LlamaForCausalLM
|
||||
from vllm.model_executor.input_metadata import InputMetadata
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import PagedAttention
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
RowParallelLinear)
|
||||
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 (
|
||||
VocabParallelEmbedding, ParallelLMHead)
|
||||
from vllm.model_executor.parallel_utils.parallel_state import (
|
||||
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.model_executor.weight_utils import (default_weight_loader,
|
||||
hf_model_weights_iterator)
|
||||
from vllm.sequence import SamplerOutput
|
||||
from vllm.transformers_utils.configs.baichuan import BaiChuanConfig
|
||||
|
||||
KVCache = Tuple[torch.Tensor, torch.Tensor]
|
||||
|
||||
|
||||
class BaiChuanBaseForCausalLM(LlamaForCausalLM):
|
||||
def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
|
||||
closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
|
||||
base = torch.tensor(
|
||||
2**(-(2**-(math.log2(closest_power_of_2) - 3))),
|
||||
dtype=torch.float32,
|
||||
)
|
||||
powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
|
||||
slopes = torch.pow(base, powers)
|
||||
|
||||
if closest_power_of_2 != total_num_heads:
|
||||
extra_base = torch.tensor(
|
||||
2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
|
||||
dtype=torch.float32,
|
||||
)
|
||||
num_remaining_heads = min(closest_power_of_2,
|
||||
total_num_heads - closest_power_of_2)
|
||||
extra_powers = torch.arange(start=1,
|
||||
end=1 + 2 * num_remaining_heads,
|
||||
step=2,
|
||||
dtype=torch.int32)
|
||||
slopes = torch.cat(
|
||||
[slopes, torch.pow(extra_base, extra_powers)], dim=0)
|
||||
return slopes
|
||||
|
||||
|
||||
class BaiChuanMLP(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
hidden_size, [intermediate_size] * 2,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
self.down_proj = RowParallelLinear(intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
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 BaiChuanAttention(nn.Module):
|
||||
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
position_embedding: str,
|
||||
rope_theta: float = 10000,
|
||||
max_position_embeddings: int = 8192,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
|
||||
)
|
||||
self.total_num_heads = num_heads
|
||||
assert self.total_num_heads % tensor_model_parallel_world_size == 0
|
||||
self.num_heads = (self.total_num_heads //
|
||||
tensor_model_parallel_world_size)
|
||||
self.head_dim = hidden_size // self.total_num_heads
|
||||
self.postion_embedding = position_embedding
|
||||
self.rope_theta = rope_theta
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
|
||||
# pylint: disable=invalid-name
|
||||
self.W_pack = QKVParallelLinear(
|
||||
hidden_size,
|
||||
self.head_dim,
|
||||
self.total_num_heads,
|
||||
self.total_num_heads,
|
||||
bias=False,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.o_proj = RowParallelLinear(
|
||||
self.total_num_heads * self.head_dim,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
# Create the alibi slopes and slice them.
|
||||
if self.postion_embedding == "ALIBI":
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
head_start = tp_rank * self.num_heads
|
||||
head_end = (tp_rank + 1) * self.num_heads
|
||||
alibi_slopes = _get_alibi_slopes(self.total_num_heads)
|
||||
alibi_slopes = alibi_slopes[head_start:head_end].tolist()
|
||||
|
||||
scaling = self.head_dim**-0.5
|
||||
self.attn = PagedAttention(self.num_heads,
|
||||
self.head_dim,
|
||||
scaling,
|
||||
alibi_slopes=alibi_slopes)
|
||||
else:
|
||||
self.rotary_emb = get_rope(
|
||||
self.head_dim,
|
||||
rotary_dim=self.head_dim,
|
||||
max_position=self.max_position_embeddings,
|
||||
base=self.rope_theta,
|
||||
)
|
||||
self.scaling = self.head_dim**-0.5
|
||||
self.attn = PagedAttention(self.num_heads, self.head_dim,
|
||||
self.scaling)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
kv_cache: KVCache,
|
||||
input_metadata: InputMetadata,
|
||||
) -> torch.Tensor:
|
||||
qkv, _ = self.W_pack(hidden_states)
|
||||
q, k, v = qkv.chunk(chunks=3, dim=-1)
|
||||
if self.postion_embedding != "ALIBI":
|
||||
q, k = self.rotary_emb(positions, q, k)
|
||||
k_cache, v_cache = kv_cache
|
||||
attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
|
||||
output, _ = self.o_proj(attn_output)
|
||||
return output
|
||||
|
||||
|
||||
class BaiChuanDecoderLayer(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
config: BaiChuanConfig,
|
||||
position_embedding: str,
|
||||
linear_method: Optional[LinearMethodBase] = None):
|
||||
super().__init__()
|
||||
self.hidden_size = config.hidden_size
|
||||
rope_theta = getattr(config, "rope_theta", 10000)
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings",
|
||||
8192)
|
||||
self.self_attn = BaiChuanAttention(
|
||||
hidden_size=self.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
position_embedding=position_embedding,
|
||||
rope_theta=rope_theta,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.mlp = BaiChuanMLP(
|
||||
hidden_size=self.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
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,
|
||||
kv_cache: KVCache,
|
||||
input_metadata: InputMetadata,
|
||||
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,
|
||||
kv_cache=kv_cache,
|
||||
input_metadata=input_metadata,
|
||||
)
|
||||
|
||||
# Fully Connected
|
||||
hidden_states, residual = self.post_attention_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states = self.mlp(hidden_states)
|
||||
return hidden_states, residual
|
||||
|
||||
|
||||
class BaiChuanModel(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
config: BaiChuanConfig,
|
||||
position_embedding: str,
|
||||
linear_method: Optional[LinearMethodBase] = 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,
|
||||
)
|
||||
self.layers = nn.ModuleList([
|
||||
BaiChuanDecoderLayer(config, position_embedding, linear_method)
|
||||
for _ 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,
|
||||
kv_caches: List[KVCache],
|
||||
input_metadata: InputMetadata,
|
||||
) -> torch.Tensor:
|
||||
hidden_states = self.embed_tokens(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],
|
||||
input_metadata,
|
||||
residual,
|
||||
)
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
return hidden_states
|
||||
|
||||
|
||||
class BaiChuanBaseForCausalLM(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
config,
|
||||
position_embedding: str,
|
||||
linear_method: Optional[LinearMethodBase] = None):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.linear_method = linear_method
|
||||
self.model = BaiChuanModel(config, position_embedding, linear_method)
|
||||
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
|
||||
self.sampler = Sampler(config.vocab_size)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[KVCache],
|
||||
input_metadata: InputMetadata,
|
||||
) -> torch.Tensor:
|
||||
hidden_states = self.model(input_ids, positions, kv_caches,
|
||||
input_metadata)
|
||||
return hidden_states
|
||||
|
||||
def sample(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[SamplerOutput]:
|
||||
next_tokens = self.sampler(self.lm_head.weight, hidden_states,
|
||||
sampling_metadata)
|
||||
return next_tokens
|
||||
|
||||
def load_weights(self,
|
||||
model_name_or_path: str,
|
||||
@@ -42,15 +328,9 @@ class BaiChuanBaseForCausalLM(LlamaForCausalLM):
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
]
|
||||
param_weight_map = [
|
||||
("qkv_proj", "W_pack"),
|
||||
]
|
||||
params_dict = dict(self.named_parameters())
|
||||
for name, loaded_weight in hf_model_weights_iterator(
|
||||
model_name_or_path, cache_dir, load_format, revision):
|
||||
for (param_name, weight_name) in param_weight_map:
|
||||
name = name.replace(weight_name, param_name)
|
||||
|
||||
if "rotary_emb.inv_freq" in name:
|
||||
continue
|
||||
if name == "lm_head.weight":
|
||||
@@ -88,28 +368,19 @@ class BaiChuanBaseForCausalLM(LlamaForCausalLM):
|
||||
class BaichuanForCausalLM(BaiChuanBaseForCausalLM):
|
||||
"""Baichuan 13B and Baichuan2 7B/13B."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: Optional[PretrainedConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
if config.hidden_size != 4096: # baichuan 13b, baichuan2 13b
|
||||
config.postion_embedding = "ALIBI"
|
||||
super().__init__(config=config,
|
||||
linear_method=linear_method,
|
||||
lora_config=lora_config)
|
||||
def __init__(self,
|
||||
config,
|
||||
linear_method: Optional[LinearMethodBase] = None):
|
||||
if config.hidden_size == 4096: # baichuan2 7b
|
||||
super().__init__(config, "ROPE", linear_method)
|
||||
else: # baichuan 13b, baichuan2 13b
|
||||
super().__init__(config, "ALIBI", linear_method)
|
||||
|
||||
|
||||
class BaiChuanForCausalLM(BaiChuanBaseForCausalLM):
|
||||
"""Baichuan 7B."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: Optional[PretrainedConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
) -> None:
|
||||
super().__init__(config=config,
|
||||
linear_method=linear_method,
|
||||
lora_config=lora_config)
|
||||
def __init__(self,
|
||||
config,
|
||||
linear_method: Optional[LinearMethodBase] = None):
|
||||
super().__init__(config, "ROPE", linear_method)
|
||||
|
||||
Reference in New Issue
Block a user