Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -10,25 +10,30 @@ from transformers import PretrainedConfig
|
||||
from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import (divide, get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.model_executor.layers.activation import (get_act_and_mul_fn,
|
||||
get_act_fn)
|
||||
from vllm.model_executor.layers.fused_moe import (activation_without_mul,
|
||||
fused_topk)
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.distributed import (
|
||||
divide,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce,
|
||||
)
|
||||
from vllm.model_executor.layers.activation import get_act_and_mul_fn, get_act_fn
|
||||
from vllm.model_executor.layers.fused_moe import activation_without_mul, fused_topk
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding)
|
||||
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.models.utils import (AutoWeightsLoader, WeightsMapper,
|
||||
maybe_prefix)
|
||||
from vllm.model_executor.models.utils import (
|
||||
AutoWeightsLoader,
|
||||
WeightsMapper,
|
||||
maybe_prefix,
|
||||
)
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.sequence import IntermediateTensors
|
||||
@@ -40,24 +45,24 @@ from .interfaces_base import default_pooling_type
|
||||
|
||||
|
||||
class BertWithRopeEmbedding(nn.Module):
|
||||
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
|
||||
super().__init__()
|
||||
if config.position_embedding_type not in ["rope", "rotary"]:
|
||||
raise ValueError("Only 'rotary'('rope') position_embedding_type" +
|
||||
" is supported")
|
||||
raise ValueError(
|
||||
"Only 'rotary'('rope') position_embedding_type" + " is supported"
|
||||
)
|
||||
|
||||
self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
|
||||
config.hidden_size)
|
||||
self.word_embeddings = VocabParallelEmbedding(
|
||||
config.vocab_size, config.hidden_size
|
||||
)
|
||||
if config.type_vocab_size > 0:
|
||||
self.token_type_embeddings = VocabParallelEmbedding(
|
||||
config.type_vocab_size, config.hidden_size)
|
||||
config.type_vocab_size, config.hidden_size
|
||||
)
|
||||
else:
|
||||
self.token_type_embeddings = None
|
||||
|
||||
self.LayerNorm = nn.LayerNorm(config.hidden_size,
|
||||
eps=config.layer_norm_eps)
|
||||
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -70,9 +75,9 @@ class BertWithRopeEmbedding(nn.Module):
|
||||
embeddings = inputs_embeds
|
||||
if self.token_type_embeddings is not None:
|
||||
if token_type_ids is None:
|
||||
token_type_ids = torch.zeros(input_shape,
|
||||
dtype=torch.long,
|
||||
device=inputs_embeds.device)
|
||||
token_type_ids = torch.zeros(
|
||||
input_shape, dtype=torch.long, device=inputs_embeds.device
|
||||
)
|
||||
|
||||
token_type_embeddings = self.token_type_embeddings(token_type_ids)
|
||||
embeddings += token_type_embeddings
|
||||
@@ -82,7 +87,6 @@ class BertWithRopeEmbedding(nn.Module):
|
||||
|
||||
|
||||
class BertWithRopeAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
@@ -119,23 +123,28 @@ class BertWithRopeAttention(nn.Module):
|
||||
total_num_kv_heads=self.total_num_kv_heads,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.qkv_proj")
|
||||
prefix=f"{prefix}.qkv_proj",
|
||||
)
|
||||
|
||||
self.rotary_emb = get_rope(**rotary_kwargs)
|
||||
|
||||
self.attn = EncoderOnlyAttention(num_heads=self.num_heads,
|
||||
head_size=self.head_dim,
|
||||
scale=self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn")
|
||||
self.attn = EncoderOnlyAttention(
|
||||
num_heads=self.num_heads,
|
||||
head_size=self.head_dim,
|
||||
scale=self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn",
|
||||
)
|
||||
|
||||
self.out_proj = RowParallelLinear(input_size=hidden_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.dense")
|
||||
self.out_proj = RowParallelLinear(
|
||||
input_size=hidden_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.dense",
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -151,14 +160,15 @@ class BertWithRopeAttention(nn.Module):
|
||||
|
||||
|
||||
class BertWithRopeGatedMLP(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
bias: bool = True,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = ""):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
bias: bool = True,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.act_fn = get_act_and_mul_fn(hidden_act)
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
@@ -168,11 +178,13 @@ class BertWithRopeGatedMLP(nn.Module):
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.gate_up_proj",
|
||||
)
|
||||
self.down_proj = RowParallelLinear(input_size=intermediate_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.down_proj")
|
||||
self.down_proj = RowParallelLinear(
|
||||
input_size=intermediate_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.down_proj",
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
gate_up, _ = self.gate_up_proj(hidden_states)
|
||||
@@ -182,26 +194,31 @@ class BertWithRopeGatedMLP(nn.Module):
|
||||
|
||||
|
||||
class BertWithRopeMLP(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
bias: bool = True,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = ""):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
bias: bool = True,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.act_fn = get_act_fn(hidden_act)
|
||||
self.up_proj = ColumnParallelLinear(input_size=hidden_size,
|
||||
output_size=intermediate_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.up_proj")
|
||||
self.down_proj = RowParallelLinear(input_size=intermediate_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.down_proj")
|
||||
self.up_proj = ColumnParallelLinear(
|
||||
input_size=hidden_size,
|
||||
output_size=intermediate_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.up_proj",
|
||||
)
|
||||
self.down_proj = RowParallelLinear(
|
||||
input_size=intermediate_size,
|
||||
output_size=hidden_size,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.down_proj",
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
hidden_states, _ = self.up_proj(hidden_states)
|
||||
@@ -211,7 +228,6 @@ class BertWithRopeMLP(nn.Module):
|
||||
|
||||
|
||||
class NomicMoE(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
num_experts: int,
|
||||
@@ -236,28 +252,40 @@ class NomicMoE(nn.Module):
|
||||
params_dtype = torch.get_default_dtype()
|
||||
self.params_dtype = params_dtype
|
||||
|
||||
self.router = ReplicatedLinear(self.hidden_size,
|
||||
self.num_total_experts,
|
||||
bias=False)
|
||||
self.router = ReplicatedLinear(
|
||||
self.hidden_size, self.num_total_experts, bias=False
|
||||
)
|
||||
self.w1 = nn.Parameter(
|
||||
torch.empty(self.num_total_experts,
|
||||
self.intermediate_size,
|
||||
self.hidden_size,
|
||||
device=current_platform.device_type,
|
||||
dtype=self.params_dtype))
|
||||
torch.empty(
|
||||
self.num_total_experts,
|
||||
self.intermediate_size,
|
||||
self.hidden_size,
|
||||
device=current_platform.device_type,
|
||||
dtype=self.params_dtype,
|
||||
)
|
||||
)
|
||||
self.w2 = nn.Parameter(
|
||||
torch.empty(self.num_total_experts,
|
||||
self.hidden_size,
|
||||
self.intermediate_size,
|
||||
device=current_platform.device_type,
|
||||
dtype=self.params_dtype))
|
||||
torch.empty(
|
||||
self.num_total_experts,
|
||||
self.hidden_size,
|
||||
self.intermediate_size,
|
||||
device=current_platform.device_type,
|
||||
dtype=self.params_dtype,
|
||||
)
|
||||
)
|
||||
self.bias = nn.Parameter(torch.zeros(self.hidden_size))
|
||||
set_weight_attrs(self.w1, {
|
||||
"weight_loader": self.weight_loader,
|
||||
})
|
||||
set_weight_attrs(self.w2, {
|
||||
"weight_loader": self.weight_loader,
|
||||
})
|
||||
set_weight_attrs(
|
||||
self.w1,
|
||||
{
|
||||
"weight_loader": self.weight_loader,
|
||||
},
|
||||
)
|
||||
set_weight_attrs(
|
||||
self.w2,
|
||||
{
|
||||
"weight_loader": self.weight_loader,
|
||||
},
|
||||
)
|
||||
|
||||
def weight_loader(
|
||||
self,
|
||||
@@ -293,10 +321,9 @@ class NomicMoE(nn.Module):
|
||||
# FIXME(Isotr0py): This implementation is too tricky,
|
||||
# we should use FusedMoE instead in the future
|
||||
# after supporting ungated activation for it.
|
||||
topk_weights, topk_ids, _ = fused_topk(hidden_states,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=False)
|
||||
topk_weights, topk_ids, _ = fused_topk(
|
||||
hidden_states, router_logits, self.top_k, renormalize=False
|
||||
)
|
||||
|
||||
final_hidden_states = torch.ops.vllm.outplace_fused_experts(
|
||||
hidden_states=hidden_states,
|
||||
@@ -308,22 +335,22 @@ class NomicMoE(nn.Module):
|
||||
)
|
||||
|
||||
if self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(
|
||||
final_hidden_states)
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
|
||||
|
||||
return final_hidden_states.view(num_tokens, hidden_size) + self.bias
|
||||
|
||||
|
||||
class BertWithRopeBlock(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
config: PretrainedConfig,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
moe: bool = False,
|
||||
bias: bool = True,
|
||||
rotary_kwargs: Optional[dict] = None,
|
||||
prefix: str = ""):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
moe: bool = False,
|
||||
bias: bool = True,
|
||||
rotary_kwargs: Optional[dict] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.attn = BertWithRopeAttention(
|
||||
hidden_size=config.hidden_size,
|
||||
@@ -332,14 +359,17 @@ class BertWithRopeBlock(nn.Module):
|
||||
quant_config=quant_config,
|
||||
bias=bias,
|
||||
rotary_kwargs=rotary_kwargs,
|
||||
prefix=f"{prefix}.attention")
|
||||
prefix=f"{prefix}.attention",
|
||||
)
|
||||
|
||||
if moe:
|
||||
self.mlp = NomicMoE(num_experts=config.num_experts,
|
||||
top_k=config.moe_top_k,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act)
|
||||
self.mlp = NomicMoE(
|
||||
num_experts=config.num_experts,
|
||||
top_k=config.moe_top_k,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
)
|
||||
else:
|
||||
if config.hidden_act in ["silu", "geglu"]:
|
||||
self.mlp = BertWithRopeGatedMLP(
|
||||
@@ -348,7 +378,8 @@ class BertWithRopeBlock(nn.Module):
|
||||
hidden_act=config.hidden_act,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp")
|
||||
prefix=f"{prefix}.mlp",
|
||||
)
|
||||
else:
|
||||
self.mlp = BertWithRopeMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
@@ -356,12 +387,11 @@ class BertWithRopeBlock(nn.Module):
|
||||
hidden_act=config.hidden_act,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp")
|
||||
prefix=f"{prefix}.mlp",
|
||||
)
|
||||
|
||||
self.attn_ln = nn.LayerNorm(config.hidden_size,
|
||||
eps=config.layer_norm_eps)
|
||||
self.mlp_ln = nn.LayerNorm(config.hidden_size,
|
||||
eps=config.layer_norm_eps)
|
||||
self.attn_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
||||
self.mlp_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
||||
|
||||
def forward(self, positions: torch.Tensor, hidden_states: torch.Tensor):
|
||||
attn_output = self.attn(positions, hidden_states)
|
||||
@@ -372,27 +402,32 @@ class BertWithRopeBlock(nn.Module):
|
||||
|
||||
|
||||
class BertWithRopeEncoder(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
vllm_config: VllmConfig,
|
||||
bias: bool = True,
|
||||
rotary_kwargs: Optional[dict] = None,
|
||||
prefix: str = ""):
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
bias: bool = True,
|
||||
rotary_kwargs: Optional[dict] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
config = vllm_config.model_config.hf_config
|
||||
cache_config = vllm_config.cache_config
|
||||
quant_config = vllm_config.quant_config
|
||||
every_n = getattr(config, "moe_every_n_layers", 0)
|
||||
self.layers = nn.ModuleList([
|
||||
BertWithRopeBlock(config=config,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
bias=bias,
|
||||
moe=every_n > 0 and (layer_idx % every_n == 1),
|
||||
rotary_kwargs=rotary_kwargs,
|
||||
prefix=f"{prefix}.layer.{layer_idx}")
|
||||
for layer_idx in range(config.num_hidden_layers)
|
||||
])
|
||||
self.layers = nn.ModuleList(
|
||||
[
|
||||
BertWithRopeBlock(
|
||||
config=config,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
bias=bias,
|
||||
moe=every_n > 0 and (layer_idx % every_n == 1),
|
||||
rotary_kwargs=rotary_kwargs,
|
||||
prefix=f"{prefix}.layer.{layer_idx}",
|
||||
)
|
||||
for layer_idx in range(config.num_hidden_layers)
|
||||
]
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -409,11 +444,13 @@ class BertWithRopeEncoder(nn.Module):
|
||||
class BertWithRope(nn.Module, SupportsQuant):
|
||||
hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
|
||||
|
||||
def __init__(self,
|
||||
*,
|
||||
vllm_config: VllmConfig,
|
||||
prefix: str = "",
|
||||
add_pooling_layer: bool = False):
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
vllm_config: VllmConfig,
|
||||
prefix: str = "",
|
||||
add_pooling_layer: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
self.vllm_config = vllm_config
|
||||
self.add_pooling_layer = add_pooling_layer
|
||||
@@ -423,7 +460,8 @@ class BertWithRope(nn.Module, SupportsQuant):
|
||||
vllm_config=vllm_config,
|
||||
bias=getattr(self.config, "bias", True),
|
||||
rotary_kwargs=self.config.rotary_kwargs,
|
||||
prefix=f"{prefix}.encoder")
|
||||
prefix=f"{prefix}.encoder",
|
||||
)
|
||||
self.pooler = BertPooler(self.config) if add_pooling_layer else None
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
@@ -440,12 +478,12 @@ class BertWithRope(nn.Module, SupportsQuant):
|
||||
if inputs_embeds is not None:
|
||||
hidden_states = inputs_embeds
|
||||
else:
|
||||
hidden_states = self.embeddings(input_ids=input_ids,
|
||||
token_type_ids=token_type_ids)
|
||||
hidden_states = self.embeddings(
|
||||
input_ids=input_ids, token_type_ids=token_type_ids
|
||||
)
|
||||
return self.encoder(positions, hidden_states)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = self.hf_to_vllm_mapper.apply(weights)
|
||||
|
||||
if self.config.hidden_act in ["silu", "geglu"]:
|
||||
@@ -462,7 +500,7 @@ class BertWithRope(nn.Module, SupportsQuant):
|
||||
for name, loaded_weight in weights:
|
||||
if not self.add_pooling_layer and "pooler" in name:
|
||||
continue
|
||||
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
||||
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)
|
||||
@@ -478,8 +516,7 @@ class BertWithRope(nn.Module, SupportsQuant):
|
||||
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 = getattr(param, "weight_loader", default_weight_loader)
|
||||
if name.endswith((".w1", ".w2")):
|
||||
# Nomic-MoE has fused experts weights
|
||||
weight_loader(param, loaded_weight, name)
|
||||
@@ -506,7 +543,8 @@ class NomicBertModel(BertWithRope):
|
||||
"experts.mlp.": "",
|
||||
"experts.": "",
|
||||
"router.layer": "router",
|
||||
})
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
class GteNewModel(BertWithRope):
|
||||
@@ -518,7 +556,8 @@ class GteNewModel(BertWithRope):
|
||||
"layer": "layers",
|
||||
"attention.qkv_proj": "attn.qkv_proj",
|
||||
"attention.o_proj": "attn.out_proj",
|
||||
})
|
||||
}
|
||||
)
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs):
|
||||
super().__init__(vllm_config=vllm_config, prefix=prefix, **kwargs)
|
||||
@@ -539,15 +578,13 @@ class GteNewModel(BertWithRope):
|
||||
else:
|
||||
yield name, weight
|
||||
|
||||
def ignore_unnecessary_layers(self,
|
||||
weights: Iterable[tuple[str, torch.Tensor]]):
|
||||
def ignore_unnecessary_layers(self, weights: Iterable[tuple[str, torch.Tensor]]):
|
||||
for name, weight in weights:
|
||||
if name.startswith("classifier"):
|
||||
continue
|
||||
yield name, weight
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = self.ignore_unnecessary_layers(weights)
|
||||
weights = self.split_up_gate_proj(weights)
|
||||
return super().load_weights(weights)
|
||||
@@ -561,7 +598,8 @@ class SnowflakeGteNewModel(GteNewModel):
|
||||
"layer": "layers",
|
||||
"attention.qkv_proj": "attn.qkv_proj",
|
||||
"attention.o_proj": "attn.out_proj",
|
||||
})
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
class JinaRobertaModel(BertWithRope):
|
||||
@@ -576,11 +614,11 @@ class JinaRobertaModel(BertWithRope):
|
||||
"mlp.fc1.": "mlp.up_proj.",
|
||||
"mlp.fc2": "mlp.down_proj",
|
||||
"norm2": "mlp_ln",
|
||||
})
|
||||
}
|
||||
)
|
||||
|
||||
@torch.inference_mode()
|
||||
def jina_merge_lora_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]):
|
||||
def jina_merge_lora_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
|
||||
# use for jina-embeddings-v3
|
||||
# Merge Lora weights into a single weight tensor.
|
||||
# This is a temporary solution until we have a better way to handle
|
||||
@@ -601,7 +639,7 @@ class JinaRobertaModel(BertWithRope):
|
||||
if o in name:
|
||||
dtype = weights[name].dtype
|
||||
shape = weights[name].shape
|
||||
weight_name = name[:-len(o)]
|
||||
weight_name = name[: -len(o)]
|
||||
|
||||
if "embeddings" in weight_name:
|
||||
B = weights[weight_name + a][i].to(device).float()
|
||||
@@ -610,20 +648,23 @@ class JinaRobertaModel(BertWithRope):
|
||||
B = weights[weight_name + b][i].to(device).float()
|
||||
A = weights[weight_name + a][i].to(device).float()
|
||||
|
||||
weight = (weights[weight_name + o].to(device) +
|
||||
torch.matmul(B, A).view(shape) * scaling)
|
||||
weight = (
|
||||
weights[weight_name + o].to(device)
|
||||
+ torch.matmul(B, A).view(shape) * scaling
|
||||
)
|
||||
weight = weight.cpu().to(dtype)
|
||||
|
||||
weights[weight_name.replace(".parametrizations", "")] = weight
|
||||
|
||||
del weights[weight_name + o], weights[weight_name +
|
||||
a], weights[weight_name +
|
||||
b]
|
||||
del (
|
||||
weights[weight_name + o],
|
||||
weights[weight_name + a],
|
||||
weights[weight_name + b],
|
||||
)
|
||||
|
||||
return [(name, weight) for name, weight in weights.items()]
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = self.jina_merge_lora_weights(weights)
|
||||
return super().load_weights(weights)
|
||||
|
||||
@@ -637,9 +678,9 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
|
||||
config = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
|
||||
self.new = GteNewModel(vllm_config=vllm_config,
|
||||
prefix=prefix,
|
||||
add_pooling_layer=True)
|
||||
self.new = GteNewModel(
|
||||
vllm_config=vllm_config, prefix=prefix, add_pooling_layer=True
|
||||
)
|
||||
self.classifier = ReplicatedLinear(
|
||||
config.hidden_size,
|
||||
config.num_labels,
|
||||
@@ -647,29 +688,31 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
|
||||
quant_config=quant_config,
|
||||
params_dtype=vllm_config.model_config.head_dtype,
|
||||
prefix=maybe_prefix(prefix, "classifier"),
|
||||
return_bias=False)
|
||||
return_bias=False,
|
||||
)
|
||||
|
||||
pooler_config = vllm_config.model_config.pooler_config
|
||||
assert pooler_config is not None
|
||||
|
||||
self.pooler = DispatchPooler({
|
||||
"encode":
|
||||
Pooler.for_encode(pooler_config),
|
||||
"classify":
|
||||
ClassifierPooler(
|
||||
pooling=self.new.pooler,
|
||||
classifier=self.classifier,
|
||||
act_fn=ClassifierPooler.act_fn_for_seq_cls(
|
||||
vllm_config.model_config),
|
||||
),
|
||||
"score":
|
||||
ClassifierPooler(
|
||||
pooling=self.new.pooler,
|
||||
classifier=self.classifier,
|
||||
act_fn=ClassifierPooler.act_fn_for_cross_encoder(
|
||||
vllm_config.model_config),
|
||||
),
|
||||
})
|
||||
self.pooler = DispatchPooler(
|
||||
{
|
||||
"encode": Pooler.for_encode(pooler_config),
|
||||
"classify": ClassifierPooler(
|
||||
pooling=self.new.pooler,
|
||||
classifier=self.classifier,
|
||||
act_fn=ClassifierPooler.act_fn_for_seq_cls(
|
||||
vllm_config.model_config
|
||||
),
|
||||
),
|
||||
"score": ClassifierPooler(
|
||||
pooling=self.new.pooler,
|
||||
classifier=self.classifier,
|
||||
act_fn=ClassifierPooler.act_fn_for_cross_encoder(
|
||||
vllm_config.model_config
|
||||
),
|
||||
),
|
||||
}
|
||||
)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
|
||||
loader = AutoWeightsLoader(self)
|
||||
@@ -686,8 +729,9 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
|
||||
return self.new(input_ids=input_ids,
|
||||
positions=positions,
|
||||
inputs_embeds=inputs_embeds,
|
||||
intermediate_tensors=intermediate_tensors)
|
||||
return self.new(
|
||||
input_ids=input_ids,
|
||||
positions=positions,
|
||||
inputs_embeds=inputs_embeds,
|
||||
intermediate_tensors=intermediate_tensors,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user