450 lines
17 KiB
Python
450 lines
17 KiB
Python
|
#!/usr/bin/env python
|
||
|
|
# -*- coding: utf-8 -*-
|
||
|
|
|
||
|
|
import math
|
||
|
|
from typing import Any
|
||
|
|
from typing import Dict
|
||
|
|
from typing import Iterable
|
||
|
|
from typing import List
|
||
|
|
from typing import Optional
|
||
|
|
from typing import Tuple
|
||
|
|
from typing import Union
|
||
|
|
|
||
|
|
import torch
|
||
|
|
from torch import nn
|
||
|
|
from transformers import PretrainedConfig
|
||
|
|
from vllm.attention import Attention
|
||
|
|
from vllm.attention import AttentionMetadata
|
||
|
|
from vllm.config import CacheConfig
|
||
|
|
from vllm.config import VllmConfig
|
||
|
|
from vllm.distributed import get_pp_group
|
||
|
|
from vllm.distributed import get_tensor_model_parallel_world_size
|
||
|
|
from vllm.model_executor import SamplingMetadata
|
||
|
|
from vllm.model_executor.layers.activation import SiluAndMul
|
||
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
||
|
|
from vllm.model_executor.layers.linear import MergedColumnParallelLinear
|
||
|
|
from vllm.model_executor.layers.linear import QKVParallelLinear
|
||
|
|
from vllm.model_executor.layers.linear import RowParallelLinear
|
||
|
|
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||
|
|
from vllm.model_executor.layers.quantization 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.sampler import SamplerOutput
|
||
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import DEFAULT_VOCAB_PADDING_SIZE
|
||
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
|
||
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
|
||
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||
|
|
from vllm.model_executor.model_loader.weight_utils import maybe_remap_kv_scale_name
|
||
|
|
from vllm.model_executor.models import SupportsLoRA
|
||
|
|
from vllm.model_executor.models import SupportsPP
|
||
|
|
from vllm.model_executor.models.utils import AutoWeightsLoader
|
||
|
|
from vllm.model_executor.models.utils import is_pp_missing_parameter
|
||
|
|
from vllm.model_executor.models.utils import make_empty_intermediate_tensors_factory
|
||
|
|
from vllm.model_executor.models.utils import make_layers
|
||
|
|
from vllm.model_executor.models.utils import maybe_prefix
|
||
|
|
from vllm.sequence import IntermediateTensors
|
||
|
|
|
||
|
|
|
||
|
|
class HairuoMLP(nn.Module):
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
hidden_size,
|
||
|
|
intermediate_size: int,
|
||
|
|
hidden_act: str,
|
||
|
|
quant_config: Optional[QuantizationConfig] = None,
|
||
|
|
prefix: str = "",
|
||
|
|
):
|
||
|
|
super().__init__()
|
||
|
|
|
||
|
|
self.gate_up_proj = MergedColumnParallelLinear(
|
||
|
|
hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config, prefix=f"{prefix}.gate_up_proj"
|
||
|
|
)
|
||
|
|
|
||
|
|
self.down_proj = RowParallelLinear(
|
||
|
|
intermediate_size, hidden_size, bias=False, quant_config=quant_config, prefix=f"{prefix}.down_proj"
|
||
|
|
)
|
||
|
|
|
||
|
|
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 HairuoAttention(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,
|
||
|
|
prefix: str = "",
|
||
|
|
):
|
||
|
|
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:
|
||
|
|
assert self.total_num_kv_heads % tp_size == 0
|
||
|
|
else:
|
||
|
|
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=hidden_size,
|
||
|
|
head_size=self.head_dim,
|
||
|
|
total_num_heads=self.total_num_heads,
|
||
|
|
total_num_kv_heads=self.total_num_kv_heads,
|
||
|
|
bias=False,
|
||
|
|
quant_config=quant_config,
|
||
|
|
prefix=f"{prefix}.qkv_proj",
|
||
|
|
)
|
||
|
|
|
||
|
|
self.o_proj = RowParallelLinear(
|
||
|
|
input_size=self.total_num_heads * self.head_dim,
|
||
|
|
output_size=hidden_size,
|
||
|
|
bias=False,
|
||
|
|
quant_config=quant_config,
|
||
|
|
prefix=f"{prefix}.o_proj",
|
||
|
|
)
|
||
|
|
|
||
|
|
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,
|
||
|
|
prefix=f"{prefix}.attn",
|
||
|
|
)
|
||
|
|
|
||
|
|
def forward(
|
||
|
|
self,
|
||
|
|
positions: torch.Tensor,
|
||
|
|
hidden_states: torch.Tensor,
|
||
|
|
kv_cache: torch.Tensor,
|
||
|
|
attn_metadata: AttentionMetadata,
|
||
|
|
) -> torch.Tensor:
|
||
|
|
qkv, _ = self.qkv_proj(hidden_states)
|
||
|
|
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
||
|
|
|
||
|
|
o_dtype = q.dtype
|
||
|
|
q, k = self.rotary_emb(positions, q.float(), k.float())
|
||
|
|
q, k = q.to(o_dtype), k.to(o_dtype)
|
||
|
|
|
||
|
|
attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
|
||
|
|
output, _ = self.o_proj(attn_output)
|
||
|
|
return output
|
||
|
|
|
||
|
|
|
||
|
|
class HairuoDecoderLayer(nn.Module):
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
config: PretrainedConfig,
|
||
|
|
cache_config: Optional[CacheConfig] = None,
|
||
|
|
quant_config: Optional[QuantizationConfig] = None,
|
||
|
|
prefix: str = "",
|
||
|
|
):
|
||
|
|
super().__init__()
|
||
|
|
|
||
|
|
self.config = config
|
||
|
|
self.cache_config = cache_config
|
||
|
|
self.quant_config = quant_config
|
||
|
|
self.hidden_size = config.hidden_size
|
||
|
|
self.rope_theta = getattr(config, "rope_theta", 10000.0)
|
||
|
|
self.rope_scaling = getattr(config, "rope_scaling", None)
|
||
|
|
if self.rope_scaling is not None and getattr(config, "original_max_position_embeddings", None):
|
||
|
|
self.rope_scaling["original_max_position_embeddings"] = config.original_max_position_embeddings
|
||
|
|
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
||
|
|
|
||
|
|
self.self_attn = HairuoAttention(
|
||
|
|
hidden_size=self.hidden_size,
|
||
|
|
num_heads=config.num_attention_heads,
|
||
|
|
num_kv_heads=getattr(config, "num_key_value_heads", config.num_attention_heads),
|
||
|
|
rope_theta=self.rope_theta,
|
||
|
|
rope_scaling=self.rope_scaling,
|
||
|
|
max_position_embeddings=max_position_embeddings,
|
||
|
|
cache_config=cache_config,
|
||
|
|
quant_config=quant_config,
|
||
|
|
prefix=f"{prefix}.self_attn",
|
||
|
|
)
|
||
|
|
|
||
|
|
self.mlp = HairuoMLP(
|
||
|
|
hidden_size=self.hidden_size,
|
||
|
|
intermediate_size=config.intermediate_size,
|
||
|
|
hidden_act=config.hidden_act,
|
||
|
|
quant_config=quant_config,
|
||
|
|
prefix=f"{prefix}.mlp",
|
||
|
|
)
|
||
|
|
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)
|
||
|
|
|
||
|
|
self.mup_scale_hidden_states = config.mup_scale_depth / math.sqrt(config.num_hidden_layers)
|
||
|
|
|
||
|
|
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]:
|
||
|
|
residual = hidden_states
|
||
|
|
hidden_states = self.input_layernorm(hidden_states)
|
||
|
|
hidden_states = self.self_attn(
|
||
|
|
positions=positions, hidden_states=hidden_states, kv_cache=kv_cache, attn_metadata=attn_metadata
|
||
|
|
)
|
||
|
|
hidden_states = residual + hidden_states * self.mup_scale_hidden_states
|
||
|
|
|
||
|
|
residual = hidden_states
|
||
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
||
|
|
hidden_states = self.mlp(hidden_states)
|
||
|
|
hidden_states = residual + hidden_states * self.mup_scale_hidden_states
|
||
|
|
|
||
|
|
return hidden_states, None
|
||
|
|
|
||
|
|
|
||
|
|
class HairuoModel(nn.Module):
|
||
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
|
||
|
|
super().__init__()
|
||
|
|
|
||
|
|
config = vllm_config.model_config.hf_config
|
||
|
|
cache_config = vllm_config.cache_config
|
||
|
|
quant_config = vllm_config.quant_config
|
||
|
|
lora_config = vllm_config.lora_config
|
||
|
|
|
||
|
|
self.config = config
|
||
|
|
|
||
|
|
self.padding_idx = config.pad_token_id
|
||
|
|
lora_vocab = (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1)) if lora_config else 0
|
||
|
|
self.vocab_size = config.vocab_size + lora_vocab
|
||
|
|
self.org_vocab_size = config.vocab_size
|
||
|
|
self.embed_tokens = VocabParallelEmbedding(
|
||
|
|
self.vocab_size,
|
||
|
|
config.hidden_size,
|
||
|
|
org_num_embeddings=config.vocab_size,
|
||
|
|
prefix=f"{prefix}.embed_tokens",
|
||
|
|
)
|
||
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
||
|
|
config.num_hidden_layers,
|
||
|
|
lambda prefix: HairuoDecoderLayer(
|
||
|
|
config=config, cache_config=cache_config, quant_config=quant_config, prefix=f"{prefix}.layers"
|
||
|
|
),
|
||
|
|
prefix=f"{prefix}.layers",
|
||
|
|
)
|
||
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||
|
|
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||
|
|
["hidden_states", "residual"], config.hidden_size
|
||
|
|
)
|
||
|
|
|
||
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||
|
|
embedding = self.embed_tokens(input_ids)
|
||
|
|
return embedding * self.config.mup_scale_emb
|
||
|
|
|
||
|
|
def forward(
|
||
|
|
self,
|
||
|
|
input_ids: torch.Tensor,
|
||
|
|
positions: torch.Tensor,
|
||
|
|
kv_caches: List[torch.Tensor],
|
||
|
|
attn_metadata: AttentionMetadata,
|
||
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
||
|
|
) -> Union[torch.Tensor, IntermediateTensors]:
|
||
|
|
if get_pp_group().is_first_rank:
|
||
|
|
if inputs_embeds is not None:
|
||
|
|
hidden_states = inputs_embeds
|
||
|
|
else:
|
||
|
|
hidden_states = self.get_input_embeddings(input_ids)
|
||
|
|
residual = None
|
||
|
|
else:
|
||
|
|
hidden_states = intermediate_tensors["hidden_states"]
|
||
|
|
residual = intermediate_tensors["residual"]
|
||
|
|
|
||
|
|
for i in range(self.start_layer, self.end_layer):
|
||
|
|
layer = self.layers[i]
|
||
|
|
hidden_states, residual = layer(
|
||
|
|
positions,
|
||
|
|
hidden_states,
|
||
|
|
kv_caches[i - self.start_layer],
|
||
|
|
attn_metadata,
|
||
|
|
residual,
|
||
|
|
)
|
||
|
|
if not get_pp_group().is_last_rank:
|
||
|
|
return IntermediateTensors({"hidden_states": hidden_states, "residual": residual})
|
||
|
|
hidden_states = self.norm(hidden_states)
|
||
|
|
return hidden_states
|
||
|
|
|
||
|
|
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(remove_duplicate=False))
|
||
|
|
for name, loaded_weight in weights:
|
||
|
|
if "rotary_emb.inv_freq" 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
|
||
|
|
|
||
|
|
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
|
||
|
|
if is_pp_missing_parameter(name, self):
|
||
|
|
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
|
||
|
|
# Remapping the name of FP8 kv-scale.
|
||
|
|
name = maybe_remap_kv_scale_name(name, params_dict)
|
||
|
|
if name is None:
|
||
|
|
continue
|
||
|
|
if is_pp_missing_parameter(name, self):
|
||
|
|
continue
|
||
|
|
param = params_dict[name]
|
||
|
|
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||
|
|
weight_loader(param, loaded_weight)
|
||
|
|
|
||
|
|
|
||
|
|
class HairuoForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||
|
|
packed_modules_mapping = {
|
||
|
|
"qkv_proj": [
|
||
|
|
"q_proj",
|
||
|
|
"k_proj",
|
||
|
|
"v_proj",
|
||
|
|
],
|
||
|
|
"gate_up_proj": [
|
||
|
|
"gate_proj",
|
||
|
|
"up_proj",
|
||
|
|
],
|
||
|
|
}
|
||
|
|
|
||
|
|
# LoRA specific attributes
|
||
|
|
supported_lora_modules = [
|
||
|
|
"qkv_proj",
|
||
|
|
"o_proj",
|
||
|
|
"gate_up_proj",
|
||
|
|
"down_proj",
|
||
|
|
"embed_tokens",
|
||
|
|
"lm_head",
|
||
|
|
]
|
||
|
|
embedding_modules = {
|
||
|
|
"embed_tokens": "input_embeddings",
|
||
|
|
"lm_head": "output_embeddings",
|
||
|
|
}
|
||
|
|
embedding_padding_modules = ["lm_head"]
|
||
|
|
|
||
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||
|
|
super().__init__()
|
||
|
|
|
||
|
|
config = vllm_config.model_config.hf_config
|
||
|
|
quant_config = vllm_config.quant_config
|
||
|
|
lora_config = vllm_config.lora_config
|
||
|
|
|
||
|
|
self.config = config
|
||
|
|
|
||
|
|
self.model = HairuoModel(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model"))
|
||
|
|
|
||
|
|
unpadded_vocab_size = config.vocab_size
|
||
|
|
if lora_config:
|
||
|
|
unpadded_vocab_size += lora_config.lora_extra_vocab_size
|
||
|
|
|
||
|
|
self.lm_head = ParallelLMHead(
|
||
|
|
unpadded_vocab_size,
|
||
|
|
config.hidden_size,
|
||
|
|
org_num_embeddings=config.vocab_size,
|
||
|
|
padding_size=(
|
||
|
|
DEFAULT_VOCAB_PADDING_SIZE
|
||
|
|
# We need bigger padding if using lora for kernel
|
||
|
|
# compatibility
|
||
|
|
if not lora_config
|
||
|
|
else lora_config.lora_vocab_padding_size
|
||
|
|
),
|
||
|
|
quant_config=quant_config,
|
||
|
|
prefix=maybe_prefix(prefix, "lm_head"),
|
||
|
|
)
|
||
|
|
|
||
|
|
self.logits_processor = LogitsProcessor(unpadded_vocab_size, config.vocab_size)
|
||
|
|
self.sampler = Sampler()
|
||
|
|
self.make_empty_intermediate_tensors = self.model.make_empty_intermediate_tensors
|
||
|
|
|
||
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||
|
|
return self.model.get_input_embeddings(input_ids)
|
||
|
|
|
||
|
|
def forward(
|
||
|
|
self,
|
||
|
|
input_ids: torch.Tensor,
|
||
|
|
positions: torch.Tensor,
|
||
|
|
kv_caches: List[torch.Tensor],
|
||
|
|
attn_metadata: AttentionMetadata,
|
||
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||
|
|
) -> Union[torch.Tensor, IntermediateTensors]:
|
||
|
|
hidden_states = self.model(input_ids, positions, kv_caches, attn_metadata, intermediate_tensors)
|
||
|
|
return hidden_states
|
||
|
|
|
||
|
|
def compute_logits(
|
||
|
|
self,
|
||
|
|
hidden_states: torch.Tensor,
|
||
|
|
sampling_metadata: SamplingMetadata,
|
||
|
|
) -> Optional[torch.Tensor]:
|
||
|
|
hidden_states = hidden_states / self.config.mup_scale_width
|
||
|
|
logits = self.logits_processor(self.lm_head, 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]]):
|
||
|
|
loader = AutoWeightsLoader(
|
||
|
|
self,
|
||
|
|
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
|
||
|
|
)
|
||
|
|
|
||
|
|
loader.load_weights(weights)
|