636 lines
23 KiB
Python
636 lines
23 KiB
Python
import logging
|
|
from functools import partial
|
|
from typing import Iterable, List, Optional, Tuple, Union
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from sglang.srt.configs import NanbeigeConfig
|
|
from sglang.srt.layers.activation import SiluAndMul
|
|
from sglang.srt.layers.aux_hidden_states import AuxHiddenStateList
|
|
from sglang.srt.layers.dp_attention import is_dp_attention_enabled
|
|
from sglang.srt.layers.layer_boundary import (
|
|
append_stages,
|
|
declare_attn,
|
|
declare_ffn,
|
|
)
|
|
from sglang.srt.layers.layer_boundary.residual import batch as residual_batch
|
|
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 EmbeddingPoolerOutput, 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.utils import PPMissingLayer, get_layer_id
|
|
from sglang.srt.layers.vocab_parallel_embedding import (
|
|
ParallelLMHead,
|
|
VocabParallelEmbedding,
|
|
)
|
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
|
|
from sglang.srt.model_loader.weight_utils import default_weight_loader
|
|
from sglang.srt.runtime_context import get_parallel
|
|
from sglang.srt.utils import add_prefix, make_pp_layers
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class NanbeigeRMSNorm(nn.Module):
|
|
def __init__(self, hidden_size, eps=1e-6):
|
|
"""
|
|
NanbeigeRMSNorm is equivalent to T5LayerNorm
|
|
"""
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
self.variance_epsilon = eps
|
|
|
|
def forward(self, hidden_states):
|
|
input_dtype = hidden_states.dtype
|
|
hidden_states = hidden_states.to(torch.float32)
|
|
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
|
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
|
return self.weight * hidden_states.to(input_dtype)
|
|
|
|
|
|
class NanbeigeMLP(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,
|
|
reduce_results=False,
|
|
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, use_reduce_scatter: bool = False):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.down_proj(
|
|
x,
|
|
skip_all_reduce=use_reduce_scatter,
|
|
)
|
|
return x
|
|
|
|
|
|
class NanbeigeAttention(nn.Module):
|
|
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
|
|
|
def __init__(
|
|
self,
|
|
config: NanbeigeConfig,
|
|
layer_id: Optional[int] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.config = config
|
|
self.layer_id = layer_id
|
|
tp_size = get_parallel().tp_size
|
|
|
|
self.attention_dropout = config.attention_dropout
|
|
self.hidden_size = config.hidden_size
|
|
|
|
self.total_num_heads = config.num_attention_heads
|
|
assert self.total_num_heads % tp_size == 0, (
|
|
"num_attention_heads must be divisible by tp_size."
|
|
)
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.head_dim = getattr(
|
|
config, "head_dim", self.hidden_size // self.total_num_heads
|
|
)
|
|
|
|
self.total_num_kv_heads = config.num_key_value_heads
|
|
assert self.total_num_kv_heads >= tp_size, (
|
|
"num_key_value_heads must be greater than tp_size."
|
|
)
|
|
assert self.total_num_kv_heads % tp_size == 0, (
|
|
"num_key_value_heads must be divisible by tp_size."
|
|
)
|
|
self.num_kv_heads = config.num_key_value_heads // tp_size
|
|
|
|
self.q_size = self.num_heads * self.head_dim
|
|
self.kv_size = self.num_kv_heads * self.head_dim
|
|
|
|
self.max_position_embeddings = config.max_position_embeddings
|
|
self.rope_theta = config.rope_theta
|
|
self.is_causal = True
|
|
self.scaling = self.head_dim**-0.5
|
|
self.total_layers = config.num_hidden_layers
|
|
self.num_loops = config.num_loops
|
|
|
|
self.qkv_proj = QKVParallelLinear(
|
|
self.hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("qkv_proj", prefix),
|
|
)
|
|
|
|
self.o_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
self.hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
reduce_results=False,
|
|
prefix=add_prefix("o_proj", prefix),
|
|
)
|
|
|
|
self.attn = nn.ModuleList()
|
|
base_layer_id = layer_id
|
|
for loop_idx in range(self.num_loops):
|
|
layer_id = base_layer_id + loop_idx * self.total_layers
|
|
self.attn.append(
|
|
RadixAttention(
|
|
self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
layer_id=layer_id,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("attn", prefix),
|
|
)
|
|
)
|
|
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
rotary_dim=self.head_dim,
|
|
max_position=self.max_position_embeddings,
|
|
base=self.rope_theta,
|
|
rope_scaling=self.config.rope_scaling,
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
loop_idx: int,
|
|
) -> 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[loop_idx](q, k, v, forward_batch)
|
|
output, _ = self.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class NanbeigeDecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: NanbeigeConfig,
|
|
layer_id: int = 0,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
alt_stream: Optional[torch.cuda.Stream] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
self.self_attn = NanbeigeAttention(
|
|
config=config,
|
|
layer_id=layer_id,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("self_attn", prefix),
|
|
)
|
|
|
|
self.mlp = NanbeigeMLP(
|
|
hidden_size=config.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
|
|
)
|
|
# Each loop runs the physical layers as one stack: the model folds the
|
|
# residual back in and re-enters at layer 0.
|
|
attn, ffn = self.stage_facts(config, layer_id)
|
|
self.attn_boundary, self.ffn_boundary = append_stages(
|
|
(attn, self.input_layernorm),
|
|
(ffn, self.post_attention_layernorm),
|
|
)
|
|
|
|
@staticmethod
|
|
def stage_facts(config: NanbeigeConfig, layer_id: int):
|
|
"""The stages the layer at ``layer_id`` declares, in order: the model's
|
|
shared declaration function, which the layer declares with too (see
|
|
make_layers)."""
|
|
return (
|
|
declare_attn(),
|
|
declare_ffn(sparse=False, next_layer_sparse=False),
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
loop_idx: int,
|
|
capture_output=None,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.attn_boundary.prepare(
|
|
hidden_states, forward_batch, capture=capture_output
|
|
)
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
forward_batch=forward_batch,
|
|
loop_idx=loop_idx,
|
|
)
|
|
|
|
# Fully Connected
|
|
hidden_states = self.attn_boundary.finish(hidden_states, forward_batch)
|
|
hidden_states = self.ffn_boundary.prepare(hidden_states, forward_batch)
|
|
hidden_states = self.mlp(hidden_states)
|
|
return self.ffn_boundary.finish(hidden_states, forward_batch)
|
|
|
|
|
|
class NanbeigeModel(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: NanbeigeConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
decoder_layer_type: type[nn.Module] = NanbeigeDecoderLayer,
|
|
alt_stream: Optional[torch.cuda.Stream] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.vocab_size = config.vocab_size
|
|
self.pp_group = get_parallel().pp_group
|
|
pp_size = self.pp_group.world_size
|
|
assert pp_size == 1, (
|
|
"The NanbeigeModel only supports a pipeline parallelism (PP) value of 1."
|
|
)
|
|
|
|
if self.pp_group.is_first_rank:
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
use_attn_tp_group=is_dp_attention_enabled(),
|
|
prefix=add_prefix("embed_tokens", prefix),
|
|
)
|
|
else:
|
|
self.embed_tokens = PPMissingLayer()
|
|
|
|
# Use the provided decoder layer type or default to NanbeigeDecoderLayer
|
|
decoder_layer_type = decoder_layer_type or NanbeigeDecoderLayer
|
|
self.layers, self.start_layer, self.end_layer = make_pp_layers(
|
|
config.num_hidden_layers,
|
|
lambda idx, prefix: decoder_layer_type(
|
|
layer_id=idx,
|
|
config=config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
alt_stream=alt_stream,
|
|
),
|
|
prefix=add_prefix("layers", prefix),
|
|
stage_facts=partial(decoder_layer_type.stage_facts, config),
|
|
)
|
|
if self.pp_group.is_last_rank:
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
else:
|
|
self.norm = PPMissingLayer(return_tuple=True)
|
|
|
|
# For EAGLE3 / DFLASH: capture *before* unrolled layer id (ids already +1'd).
|
|
self.layers_to_capture = []
|
|
|
|
def get_input_embedding(self, input_ids):
|
|
return self.embed_tokens(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: torch.Tensor = None,
|
|
pp_proxy_tensors: Optional[PPProxyTensors] = None,
|
|
) -> Union[torch.Tensor, PPProxyTensors]:
|
|
if self.pp_group.is_first_rank:
|
|
if input_embeds is None:
|
|
hidden_states = self.embed_tokens(input_ids)
|
|
else:
|
|
hidden_states = input_embeds
|
|
residual_batch.start(forward_batch)
|
|
else:
|
|
assert pp_proxy_tensors is not None
|
|
hidden_states = self.layers[self.start_layer].attn_boundary.from_pp(
|
|
pp_proxy_tensors, forward_batch
|
|
)
|
|
|
|
aux_hidden_states = AuxHiddenStateList()
|
|
# Draft target_layer_ids are in unrolled depth space:
|
|
# logical_id = loop_idx * num_hidden_layers + physical_i
|
|
# e.g. num_hidden_layers=22, num_loops=2 → ids in [0, 44).
|
|
# DFLASH style: setter stores k+1; we capture before that unrolled step.
|
|
num_physical_layers = self.config.num_hidden_layers
|
|
for loop_idx in range(self.config.num_loops):
|
|
for i in range(self.start_layer, self.end_layer):
|
|
logical_id = loop_idx * num_physical_layers + i
|
|
hidden_states = self.layers[i](
|
|
positions,
|
|
hidden_states,
|
|
forward_batch,
|
|
loop_idx,
|
|
capture_output=aux_hidden_states.capture
|
|
if logical_id in self.layers_to_capture
|
|
else None,
|
|
)
|
|
|
|
# Match the reference HF semantics for Nanbeige "loop models":
|
|
# - At the end of each full loop (except the last), convert the
|
|
# (hidden_states, residual) representation into real hidden_states
|
|
# by applying the missing residual addition.
|
|
# - If skip_loop_final_norm=False, HF applies RMSNorm after each loop
|
|
# (including intermediate loops) before entering the next loop.
|
|
if loop_idx != self.config.num_loops - 1:
|
|
hidden_states = residual_batch.fold(hidden_states, forward_batch)
|
|
if not self.config.skip_loop_final_norm:
|
|
hidden_states = self.norm(hidden_states)
|
|
residual_batch.start(forward_batch)
|
|
|
|
if not self.pp_group.is_last_rank:
|
|
return residual_batch.to_pp(hidden_states, forward_batch)
|
|
hidden_states = residual_batch.final_norm(
|
|
hidden_states, forward_batch, self.norm, skip_empty=True
|
|
)
|
|
|
|
if len(aux_hidden_states) == 0:
|
|
return hidden_states
|
|
|
|
return hidden_states, aux_hidden_states
|
|
|
|
|
|
class NanbeigeForCausalLM(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: NanbeigeConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.pp_group = get_parallel().pp_group
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.model = NanbeigeModel(
|
|
config, quant_config=quant_config, prefix=add_prefix("model", prefix)
|
|
)
|
|
|
|
# handle the lm head on different pp ranks
|
|
if self.pp_group.is_last_rank:
|
|
if self.pp_group.world_size == 1 and 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),
|
|
)
|
|
else:
|
|
# ranks other than the last rank will have a placeholder layer
|
|
self.lm_head = PPMissingLayer()
|
|
|
|
# perform weight tying for PP
|
|
if self.pp_group.world_size > 1 and config.tie_word_embeddings:
|
|
if self.pp_group.is_first_rank:
|
|
self.pp_group.send(
|
|
self.model.embed_tokens.weight, dst=self.pp_group.last_rank
|
|
)
|
|
else:
|
|
emb_token_weight = self.pp_group.recv(
|
|
size=(config.vocab_size, config.hidden_size),
|
|
dtype=next(self.model.parameters()).dtype,
|
|
src=self.pp_group.first_rank,
|
|
)
|
|
self.lm_head.weight.copy_(emb_token_weight)
|
|
|
|
self.logits_processor = LogitsProcessor(config)
|
|
self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
|
|
# For EAGLE3 support
|
|
self.capture_aux_hidden_states = False
|
|
|
|
def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.model.get_input_embedding(input_ids)
|
|
|
|
def get_input_embeddings(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,
|
|
pp_proxy_tensors: Optional[PPProxyTensors] = None,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(
|
|
input_ids,
|
|
positions,
|
|
forward_batch,
|
|
input_embeds,
|
|
pp_proxy_tensors=pp_proxy_tensors,
|
|
)
|
|
aux_hidden_states = None
|
|
if self.capture_aux_hidden_states:
|
|
hidden_states, aux_hidden_states = hidden_states
|
|
|
|
if self.pp_group.is_last_rank:
|
|
if not get_embedding:
|
|
return self.logits_processor(
|
|
input_ids,
|
|
hidden_states,
|
|
self.lm_head,
|
|
forward_batch,
|
|
aux_hidden_states,
|
|
)
|
|
else:
|
|
return self.pooler(hidden_states, forward_batch)
|
|
else:
|
|
return hidden_states
|
|
|
|
@property
|
|
def start_layer(self):
|
|
return self.model.start_layer
|
|
|
|
@property
|
|
def end_layer(self):
|
|
return self.model.end_layer
|
|
|
|
@torch.no_grad()
|
|
def forward_split_prefill(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
split_interval: Tuple[int, int], # [start, end) 0-based
|
|
input_embeds: torch.Tensor = None,
|
|
):
|
|
assert False, "NanbeigeModel does not support split_prefill."
|
|
|
|
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", ".up_proj", 1),
|
|
(".gate_up_proj", ".gate_proj", 0),
|
|
]
|
|
|
|
params_dict = dict(self.named_parameters())
|
|
for name, loaded_weight in weights:
|
|
layer_id = get_layer_id(name)
|
|
if (
|
|
layer_id is not None
|
|
and hasattr(self.model, "start_layer")
|
|
and (
|
|
layer_id < self.model.start_layer
|
|
or layer_id >= self.model.end_layer
|
|
)
|
|
):
|
|
continue
|
|
|
|
if "rotary_emb.inv_freq" in name or "projector" in name:
|
|
continue
|
|
if self.config.tie_word_embeddings and "lm_head.weight" in name:
|
|
if self.pp_group.world_size > 1 and self.pp_group.is_last_rank:
|
|
# Handle pp weight tying here
|
|
# find the embed_tokens.weight in the weights
|
|
embed_token_weights = next(
|
|
filter(lambda x: x[0] == "model.embed_tokens.weight", weights)
|
|
)[1]
|
|
loaded_weight = embed_token_weights
|
|
else:
|
|
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 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
|
|
|
|
if name in params_dict.keys():
|
|
param = params_dict[name]
|
|
weight_loader = getattr(
|
|
param, "weight_loader", default_weight_loader
|
|
)
|
|
weight_loader(param, loaded_weight)
|
|
else:
|
|
logger.warning(f"Parameter {name} not found in params_dict")
|
|
|
|
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 set_dflash_layers_to_capture(self, layer_ids: List[int]):
|
|
if not self.pp_group.is_last_rank:
|
|
return
|
|
|
|
if layer_ids is None:
|
|
raise ValueError(
|
|
"DFLASH requires explicit layer_ids for aux hidden capture."
|
|
)
|
|
|
|
self.capture_aux_hidden_states = True
|
|
# Unrolled ids: capture before step (k+1) == after HF-style layer k.
|
|
self.model.layers_to_capture = [val + 1 for val in layer_ids]
|
|
|
|
|
|
class NanbeigeForSequenceClassification(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: NanbeigeConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.model = NanbeigeModel(
|
|
config, quant_config=quant_config, prefix=add_prefix("model", prefix)
|
|
)
|
|
self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)
|
|
self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=False)
|
|
|
|
self.eos_token_id = config.eos_token_id
|
|
|
|
@torch.no_grad()
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: torch.Tensor = None,
|
|
get_embedding: bool = True,
|
|
) -> EmbeddingPoolerOutput:
|
|
assert get_embedding, (
|
|
"NanbeigeForSequenceClassification is only used for embedding"
|
|
)
|
|
|
|
hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
|
|
logits = self.score(hidden_states)
|
|
pooled_logits = self.pooler(logits, forward_batch).embeddings
|
|
|
|
return EmbeddingPoolerOutput(pooled_logits)
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
|
# Filter out lm_head weights of NanbeigeForCausalLM
|
|
filtered_weights = [
|
|
(name, w) for name, w in weights if not name.startswith("lm_head")
|
|
]
|
|
return NanbeigeForCausalLM.load_weights(self, filtered_weights)
|
|
|
|
|
|
EntryClass = [NanbeigeForCausalLM, NanbeigeForSequenceClassification]
|