557 lines
19 KiB
Python
557 lines
19 KiB
Python
import logging
|
|
from typing import Iterable, Optional, Tuple, Union
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from sglang.srt.layers.activation import GeluAndMul
|
|
from sglang.srt.layers.dp_attention import is_dp_attention_enabled
|
|
from sglang.srt.layers.layer_boundary import (
|
|
declare_attn,
|
|
declare_ffn,
|
|
make_stages,
|
|
)
|
|
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 (
|
|
ColumnParallelLinear,
|
|
MergedColumnParallelLinear,
|
|
QKVParallelLinear,
|
|
RowParallelLinear,
|
|
)
|
|
from sglang.srt.layers.logits_processor import LogitsProcessor
|
|
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.platforms import current_platform
|
|
from sglang.srt.runtime_context import get_parallel
|
|
from sglang.srt.utils import add_prefix, make_pp_layers
|
|
|
|
Spark2_5Config = None
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
# Aligned with HF's implementation, using sliding window inclusive with the last token
|
|
# SGLang assumes exclusive
|
|
def _get_attention_sliding_window_size(config):
|
|
return config.sliding_window - 1
|
|
|
|
|
|
class Spark2_5MLP(nn.Module):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
intermediate_size: int,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
reduce_results: bool = True,
|
|
) -> 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),
|
|
reduce_results=reduce_results,
|
|
)
|
|
self.act_fn = GeluAndMul()
|
|
|
|
def forward(
|
|
self,
|
|
x,
|
|
forward_batch=None,
|
|
):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.down_proj(x)
|
|
return x
|
|
|
|
|
|
class Spark2_5Attention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
head_dim: Optional[int] = None,
|
|
layer_id: int = 0,
|
|
rope_theta: float = 10000,
|
|
partial_rotary_factor: float = 1.0,
|
|
max_position_embeddings: int = 8192,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
sliding_window: int = 512,
|
|
layer_type: str = "sliding_attention",
|
|
headwise_attn_output_gate: bool = True,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
self.total_num_heads = num_heads
|
|
attn_tp_rank = get_parallel().attn_tp_rank
|
|
attn_tp_size = get_parallel().attn_tp_size
|
|
|
|
assert self.total_num_heads % attn_tp_size == 0
|
|
self.num_heads = self.total_num_heads // attn_tp_size
|
|
self.total_num_kv_heads = num_kv_heads
|
|
if self.total_num_kv_heads >= attn_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 % attn_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 attn_tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
|
|
if head_dim is not None:
|
|
self.head_dim = head_dim
|
|
else:
|
|
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.partial_rotary_factor = partial_rotary_factor
|
|
self.headwise_attn_output_gate = headwise_attn_output_gate
|
|
|
|
self.q_k_v_proj = QKVParallelLinear(
|
|
hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
tp_rank=attn_tp_rank,
|
|
tp_size=attn_tp_size,
|
|
prefix=add_prefix("q_k_v_proj", prefix),
|
|
)
|
|
if self.headwise_attn_output_gate:
|
|
self.g_proj = ColumnParallelLinear(
|
|
hidden_size,
|
|
self.total_num_heads,
|
|
bias=False,
|
|
quant_config=None, # g_proj keeps bf16.
|
|
tp_rank=attn_tp_rank,
|
|
tp_size=attn_tp_size,
|
|
prefix=add_prefix("g_proj", prefix),
|
|
)
|
|
|
|
self.out_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
tp_rank=attn_tp_rank,
|
|
tp_size=attn_tp_size,
|
|
reduce_results=False,
|
|
prefix=add_prefix("out_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=None,
|
|
partial_rotary_factor=partial_rotary_factor,
|
|
is_neox_style=True,
|
|
)
|
|
if layer_type not in ("sliding_attention", "full_attention"):
|
|
raise ValueError(f"Unsupported Spark2_5 layer_type: {layer_type}")
|
|
sliding_window_size = (
|
|
sliding_window if layer_type == "sliding_attention" else -1
|
|
)
|
|
self.attn = RadixAttention(
|
|
self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
sliding_window_size=sliding_window_size,
|
|
layer_id=layer_id,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("attn", prefix),
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.q_k_v_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)
|
|
|
|
if self.headwise_attn_output_gate:
|
|
g, _ = self.g_proj(hidden_states)
|
|
g = torch.sigmoid(g.float()).to(attn_output.dtype)
|
|
gate_output = attn_output.view(
|
|
attn_output.shape[0],
|
|
self.num_heads,
|
|
self.head_dim,
|
|
) * g.unsqueeze(-1)
|
|
attn_output = gate_output.view(*attn_output.shape)
|
|
|
|
output, _ = self.out_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class Spark2_5DecoderLayer(nn.Module):
|
|
"""A single transformer layer.
|
|
|
|
Transformer layer takes input with size [s, b, h] and returns an
|
|
output of the same size.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
config: Spark2_5Config,
|
|
layer_id: int = 0,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
|
head_dim = getattr(config, "head_dim", None)
|
|
layer_type = config.layer_types[layer_id]
|
|
rope_params = config.rope_parameters[layer_type]
|
|
rope_theta = rope_params.get("rope_theta")
|
|
partial_rotary_factor = rope_params.get("partial_rotary_factor")
|
|
if partial_rotary_factor is None:
|
|
partial_rotary_factor = 1.0
|
|
|
|
self.self_attn = Spark2_5Attention(
|
|
hidden_size=config.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
head_dim=head_dim,
|
|
layer_id=layer_id,
|
|
rope_theta=rope_theta,
|
|
partial_rotary_factor=partial_rotary_factor,
|
|
max_position_embeddings=max_position_embeddings,
|
|
quant_config=quant_config,
|
|
sliding_window=_get_attention_sliding_window_size(config),
|
|
layer_type=layer_type,
|
|
headwise_attn_output_gate=getattr(
|
|
config, "headwise_attn_output_gate", True
|
|
),
|
|
prefix=add_prefix("self_attn", prefix),
|
|
)
|
|
|
|
# MLP
|
|
self.mlp = Spark2_5MLP(
|
|
config.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("mlp", prefix),
|
|
reduce_results=False,
|
|
)
|
|
|
|
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.attn_boundary, self.ffn_boundary = make_stages(
|
|
(declare_attn(), self.input_layernorm),
|
|
(
|
|
declare_ffn(sparse=False, next_layer_sparse=False),
|
|
self.post_attention_layernorm,
|
|
),
|
|
previous=declare_ffn() if layer_id != 0 else None,
|
|
terminal=layer_id == config.num_hidden_layers - 1,
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
) -> torch.Tensor:
|
|
# Self Attention
|
|
hidden_states = self.attn_boundary.prepare(hidden_states, forward_batch)
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
forward_batch=forward_batch,
|
|
)
|
|
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 Spark2_5Model(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Spark2_5Config,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.vocab_size = config.vocab_size
|
|
self.pp_group = get_parallel().pp_group
|
|
|
|
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()
|
|
|
|
self.layers, self.start_layer, self.end_layer = make_pp_layers(
|
|
config.num_hidden_layers,
|
|
lambda idx, prefix: Spark2_5DecoderLayer(
|
|
layer_id=idx,
|
|
config=config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
),
|
|
prefix=add_prefix("layers", prefix),
|
|
)
|
|
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)
|
|
|
|
def get_input_embeddings(self) -> nn.Embedding:
|
|
return self.embed_tokens
|
|
|
|
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
|
|
)
|
|
|
|
for i in range(self.start_layer, self.end_layer):
|
|
hidden_states = self.layers[i](positions, hidden_states, forward_batch)
|
|
if not self.pp_group.is_last_rank:
|
|
return residual_batch.to_pp(hidden_states, forward_batch)
|
|
return residual_batch.final_norm(
|
|
hidden_states, forward_batch, self.norm, skip_empty=True
|
|
)
|
|
|
|
|
|
class Spark2_5ForCausalLM(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Spark2_5Config,
|
|
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 = Spark2_5Model(
|
|
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()
|
|
|
|
self.logits_processor = LogitsProcessor(config)
|
|
|
|
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,
|
|
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,
|
|
)
|
|
|
|
if self.pp_group.is_last_rank:
|
|
return self.logits_processor(
|
|
input_ids,
|
|
hidden_states,
|
|
self.lm_head,
|
|
forward_batch,
|
|
)
|
|
else:
|
|
return hidden_states
|
|
|
|
@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,
|
|
):
|
|
start, end = split_interval
|
|
# embed
|
|
if start == 0:
|
|
residual_batch.start(forward_batch)
|
|
if input_embeds is None:
|
|
forward_batch.hidden_states = self.model.embed_tokens(input_ids)
|
|
else:
|
|
forward_batch.hidden_states = input_embeds
|
|
# decoder layer
|
|
for i in range(start, end):
|
|
layer = self.model.layers[i]
|
|
forward_batch.hidden_states = layer(
|
|
positions,
|
|
forward_batch.hidden_states,
|
|
forward_batch,
|
|
)
|
|
|
|
if end == self.model.config.num_hidden_layers:
|
|
# norm
|
|
forward_batch.hidden_states = residual_batch.final_norm(
|
|
forward_batch.hidden_states, forward_batch, self.model.norm
|
|
)
|
|
# logits process
|
|
result = self.logits_processor(
|
|
input_ids, forward_batch.hidden_states, self.lm_head, forward_batch
|
|
)
|
|
else:
|
|
result = None
|
|
|
|
return result
|
|
|
|
@property
|
|
def start_layer(self):
|
|
return self.model.start_layer
|
|
|
|
@property
|
|
def end_layer(self):
|
|
return self.model.end_layer
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
|
stacked_params_mapping = [
|
|
("gate_up_proj", "gate_proj", 0),
|
|
("gate_up_proj", "up_proj", 1),
|
|
]
|
|
params_dict = dict(self.named_parameters())
|
|
for name, loaded_weight in weights:
|
|
original_name = name
|
|
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 self.config.tie_word_embeddings and "lm_head.weight" in name:
|
|
continue
|
|
|
|
if name in ("model.embedding.weight",):
|
|
name = "model.embed_tokens.weight"
|
|
if (
|
|
name == "model.embed_tokens.weight"
|
|
and self.config.tie_word_embeddings
|
|
and self.pp_group.world_size > 1
|
|
):
|
|
if self.pp_group.is_last_rank:
|
|
name = "lm_head.weight"
|
|
elif not self.pp_group.is_first_rank:
|
|
continue
|
|
|
|
loaded = False
|
|
for param_name, weight_name, shard_id in stacked_params_mapping:
|
|
if weight_name not in name:
|
|
continue
|
|
mapped_name = name.replace(weight_name, param_name)
|
|
if mapped_name not in params_dict:
|
|
continue
|
|
param = params_dict[mapped_name]
|
|
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
|
weight_loader(param, loaded_weight, shard_id)
|
|
loaded = True
|
|
break
|
|
if loaded:
|
|
continue
|
|
|
|
if name in params_dict:
|
|
param = params_dict[name]
|
|
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
|
weight_loader(param, loaded_weight)
|
|
elif original_name in ("model.embedding.weight",):
|
|
continue
|
|
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
|
|
current_platform.empty_cache()
|
|
current_platform.synchronize()
|
|
|
|
def get_attention_sliding_window_size(self) -> int:
|
|
return _get_attention_sliding_window_size(self.config)
|
|
|
|
|
|
EntryClass = [Spark2_5ForCausalLM]
|