Add initial support for gpt-oss (#8824)
This commit is contained in:
Ying Sheng
923
python/sglang/srt/models/gpt_oss.py
Normal file
923
python/sglang/srt/models/gpt_oss.py
Normal file
@@ -0,0 +1,923 @@
|
||||
# Copyright 2023-2024 SGLang Team
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ==============================================================================
|
||||
|
||||
|
||||
"""Inference-only GptOss model compatible with HuggingFace weights."""
|
||||
|
||||
import logging
|
||||
from collections.abc import Iterable
|
||||
from functools import partial
|
||||
from typing import Any, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from sglang.srt.distributed import (
|
||||
get_moe_tensor_parallel_rank,
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce,
|
||||
)
|
||||
from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
|
||||
from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
|
||||
from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
|
||||
from sglang.srt.layers.dp_attention import (
|
||||
get_attention_tp_rank,
|
||||
get_attention_tp_size,
|
||||
get_local_attention_dp_size,
|
||||
)
|
||||
from sglang.srt.layers.layernorm import RMSNorm
|
||||
from sglang.srt.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from sglang.srt.layers.logits_processor import LogitsProcessor
|
||||
from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
|
||||
from sglang.srt.layers.moe.topk import TopK
|
||||
from sglang.srt.layers.moe.utils import DeepEPMode
|
||||
from sglang.srt.layers.quantization.base_config import QuantizationConfig
|
||||
from sglang.srt.layers.quantization.fp8_utils import dequant_mxfp4
|
||||
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.managers.schedule_batch import global_server_args_dict
|
||||
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.utils import add_prefix, make_layers
|
||||
|
||||
|
||||
class GptOssConfig(PretrainedConfig):
|
||||
model_type = "gpt_oss"
|
||||
|
||||
def __init__(self, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
|
||||
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 GptOssSparseMoeBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
layer_id: int,
|
||||
config: GptOssConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.tp_size = get_tensor_model_parallel_world_size()
|
||||
self.layer_id = layer_id
|
||||
self.activation = config.hidden_act
|
||||
self.activation_alpha = getattr(config, "hidden_act_alpha", 1.702)
|
||||
self.swiglu_limit = config.swiglu_limit
|
||||
if self.tp_size > config.num_local_experts:
|
||||
raise ValueError(
|
||||
f"Tensor parallel size {self.tp_size} is greater than "
|
||||
f"the number of experts {config.num_local_experts}."
|
||||
)
|
||||
|
||||
self.topk = TopK(
|
||||
top_k=config.num_experts_per_tok,
|
||||
renormalize=True,
|
||||
)
|
||||
|
||||
experts_type = get_moe_impl_class()
|
||||
extra_kwargs = {}
|
||||
if experts_type.__name__ == "FusedMoE":
|
||||
extra_kwargs = {
|
||||
"enable_flashinfer_cutlass_moe": global_server_args_dict[
|
||||
"enable_flashinfer_cutlass_moe"
|
||||
],
|
||||
"use_weight_loader_fused": True, # for moe gate_up_proj and down_proj and their bias loading
|
||||
}
|
||||
self.experts = experts_type(
|
||||
num_experts=config.num_local_experts
|
||||
+ global_server_args_dict["ep_num_redundant_experts"],
|
||||
top_k=config.num_experts_per_tok,
|
||||
layer_id=layer_id,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
quant_config=quant_config,
|
||||
activation=self.activation,
|
||||
activation_alpha=self.activation_alpha,
|
||||
swiglu_limit=self.swiglu_limit,
|
||||
with_bias=True,
|
||||
prefix=add_prefix("experts", prefix),
|
||||
**(
|
||||
dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
|
||||
if global_server_args_dict["moe_a2a_backend"].is_deepep()
|
||||
else {}
|
||||
),
|
||||
**extra_kwargs,
|
||||
)
|
||||
|
||||
self.router = ReplicatedLinear(
|
||||
config.hidden_size,
|
||||
config.num_local_experts,
|
||||
bias=True,
|
||||
quant_config=None,
|
||||
prefix=add_prefix("gate", prefix),
|
||||
params_dtype=config.torch_dtype,
|
||||
)
|
||||
|
||||
def forward(
|
||||
self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
|
||||
) -> torch.Tensor:
|
||||
if not global_server_args_dict["moe_a2a_backend"].is_deepep():
|
||||
return self.forward_normal(hidden_states)
|
||||
else:
|
||||
raise Exception("forward_deepep branch not implemented yet")
|
||||
|
||||
def get_moe_weights(self):
|
||||
return [
|
||||
x.data
|
||||
for name, x in self.experts.named_parameters()
|
||||
if name not in ["correction_bias"]
|
||||
]
|
||||
|
||||
def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
num_tokens, hidden_dim = hidden_states.shape
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
# router_logits: (num_tokens, n_experts)
|
||||
router_logits, _ = self.router(hidden_states)
|
||||
|
||||
kwargs = {"hidden_states": hidden_states}
|
||||
if self.topk is not None:
|
||||
kwargs["topk_output"] = self.topk(hidden_states, router_logits)
|
||||
else:
|
||||
kwargs["router_logits"] = router_logits
|
||||
final_hidden_states = self.experts(**kwargs)
|
||||
|
||||
if self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
|
||||
|
||||
ans = final_hidden_states.view(num_tokens, hidden_dim)
|
||||
return ans
|
||||
|
||||
|
||||
class GptOssAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
layer_id: int = 0,
|
||||
rope_theta: float = 10000,
|
||||
rope_scaling: Optional[Dict[str, Any]] = None,
|
||||
max_position_embeddings: int = 8192,
|
||||
head_dim: Optional[int] = None,
|
||||
rms_norm_eps: float = 1e-06,
|
||||
attention_bias: bool = False,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
sliding_window_size: int = -1, # if -1, normal attention, else, window attention.
|
||||
layer_type: str = "",
|
||||
params_dtype: torch.dtype = torch.bfloat16,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.sliding_window_size = sliding_window_size
|
||||
|
||||
attn_tp_rank = get_attention_tp_rank()
|
||||
attn_tp_size = get_attention_tp_size()
|
||||
|
||||
self.total_num_heads = num_heads
|
||||
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)
|
||||
self.head_dim = head_dim or 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.tp_rank = get_tensor_model_parallel_rank()
|
||||
|
||||
self.qkv_proj = QKVParallelLinear(
|
||||
hidden_size,
|
||||
self.head_dim,
|
||||
self.total_num_heads,
|
||||
self.total_num_kv_heads,
|
||||
bias=attention_bias,
|
||||
params_dtype=params_dtype,
|
||||
quant_config=quant_config,
|
||||
tp_rank=attn_tp_rank,
|
||||
tp_size=attn_tp_size,
|
||||
prefix=add_prefix("qkv_proj", prefix),
|
||||
)
|
||||
|
||||
self.sinks = nn.Parameter(
|
||||
torch.empty(self.num_heads, dtype=params_dtype), requires_grad=False
|
||||
)
|
||||
|
||||
self.o_proj = RowParallelLinear(
|
||||
self.total_num_heads * self.head_dim,
|
||||
hidden_size,
|
||||
bias=attention_bias,
|
||||
quant_config=quant_config,
|
||||
tp_rank=attn_tp_rank,
|
||||
tp_size=attn_tp_size,
|
||||
reduce_results=False,
|
||||
params_dtype=params_dtype,
|
||||
prefix=add_prefix("o_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=rope_scaling,
|
||||
)
|
||||
|
||||
assert layer_type in {"sliding_attention", "full_attention"}
|
||||
use_sliding_window = layer_type == "sliding_attention"
|
||||
self.attn = RadixAttention(
|
||||
self.num_heads,
|
||||
self.head_dim,
|
||||
self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
layer_id=layer_id,
|
||||
prefix=add_prefix("attn", prefix),
|
||||
sliding_window_size=(sliding_window_size if use_sliding_window else -1),
|
||||
)
|
||||
self.layer_id = layer_id
|
||||
|
||||
def forward_prepare(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
forward_batch: ForwardBatch,
|
||||
):
|
||||
if hidden_states.shape[0] == 0:
|
||||
return hidden_states, forward_batch, None
|
||||
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)
|
||||
inner_state = q, k, v, forward_batch
|
||||
return None, forward_batch, inner_state
|
||||
|
||||
def forward_core(self, intermediate_state):
|
||||
hidden_states, forward_batch, inner_state = intermediate_state
|
||||
if inner_state is None:
|
||||
return hidden_states
|
||||
attn_output = self.attn(*inner_state, sk=self.sinks)
|
||||
output, _ = self.o_proj(attn_output)
|
||||
return output
|
||||
|
||||
def forward(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
forward_batch: ForwardBatch,
|
||||
) -> torch.Tensor:
|
||||
s = self.forward_prepare(
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
forward_batch=forward_batch,
|
||||
)
|
||||
return self.forward_core(s)
|
||||
|
||||
|
||||
class GptOssDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: GptOssConfig,
|
||||
layer_id: int,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
sliding_window_size: int | None = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.hidden_size = config.hidden_size
|
||||
rope_theta = getattr(config, "rope_theta", 10000)
|
||||
rope_scaling = getattr(config, "rope_scaling", None)
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
||||
head_dim = getattr(
|
||||
config, "head_dim", config.hidden_size // config.num_attention_heads
|
||||
)
|
||||
rms_norm_eps = config.rms_norm_eps
|
||||
attention_bias = config.attention_bias
|
||||
|
||||
if sliding_window_size is None:
|
||||
self.sliding_window_size = get_attention_sliding_window_size(self.config)
|
||||
else:
|
||||
self.sliding_window_size = sliding_window_size
|
||||
|
||||
self.self_attn = GptOssAttention(
|
||||
hidden_size=self.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
num_kv_heads=config.num_key_value_heads,
|
||||
layer_id=layer_id,
|
||||
rope_theta=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
head_dim=head_dim,
|
||||
rms_norm_eps=rms_norm_eps,
|
||||
attention_bias=attention_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("self_attn", prefix),
|
||||
sliding_window_size=self.sliding_window_size,
|
||||
layer_type=config.layer_types[layer_id],
|
||||
params_dtype=config.torch_dtype,
|
||||
)
|
||||
|
||||
self.layer_id = layer_id
|
||||
|
||||
self.attn_tp_size = get_attention_tp_size()
|
||||
self.attn_tp_rank = get_attention_tp_rank()
|
||||
self.local_dp_size = get_local_attention_dp_size()
|
||||
|
||||
# GptOss all layers are sparse and have no nextn now
|
||||
self.is_layer_sparse = True
|
||||
is_previous_layer_sparse = True
|
||||
|
||||
self.layer_scatter_modes = LayerScatterModes.init_new(
|
||||
layer_id=layer_id,
|
||||
num_layers=config.num_hidden_layers,
|
||||
is_layer_sparse=self.is_layer_sparse,
|
||||
is_previous_layer_sparse=is_previous_layer_sparse,
|
||||
)
|
||||
|
||||
if self.is_layer_sparse:
|
||||
self.mlp = GptOssSparseMoeBlock(
|
||||
layer_id=self.layer_id,
|
||||
config=config,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("mlp", prefix),
|
||||
)
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
"Dense MLP is not implemented for GptOssDecoderLayer. "
|
||||
"Please use GptOssSparseMoeBlock instead."
|
||||
)
|
||||
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.layer_communicator = LayerCommunicator(
|
||||
layer_scatter_modes=self.layer_scatter_modes,
|
||||
input_layernorm=self.input_layernorm,
|
||||
post_attention_layernorm=self.post_attention_layernorm,
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
forward_batch: ForwardBatch,
|
||||
residual: Optional[torch.Tensor],
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
hidden_states, residual = self.layer_communicator.prepare_attn(
|
||||
hidden_states, residual, forward_batch
|
||||
)
|
||||
|
||||
if hidden_states.shape[0] != 0:
|
||||
hidden_states = self.self_attn(
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
forward_batch=forward_batch,
|
||||
)
|
||||
|
||||
hidden_states, residual = self.layer_communicator.prepare_mlp(
|
||||
hidden_states, residual, forward_batch
|
||||
)
|
||||
|
||||
hidden_states = self.mlp(hidden_states, forward_batch)
|
||||
|
||||
hidden_states, residual = self.layer_communicator.postprocess_layer(
|
||||
hidden_states, residual, forward_batch
|
||||
)
|
||||
|
||||
return hidden_states, residual
|
||||
|
||||
|
||||
class GptOssModel(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
decoder_layer_type: type[nn.Module] = GptOssDecoderLayer,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.padding_idx = config.pad_token_id
|
||||
self.vocab_size = config.vocab_size
|
||||
self.pp_group = get_pp_group()
|
||||
|
||||
if self.pp_group.is_first_rank:
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
enable_tp=not global_server_args_dict["enable_dp_attention"],
|
||||
prefix=add_prefix("embed_tokens", prefix),
|
||||
)
|
||||
else:
|
||||
self.embed_tokens = PPMissingLayer()
|
||||
|
||||
# Use the provided decoder layer type or default to GptOssDecoderLayer
|
||||
decoder_layer_type = decoder_layer_type or GptOssDecoderLayer
|
||||
self.layers, self.start_layer, self.end_layer = make_layers(
|
||||
config.num_hidden_layers,
|
||||
lambda idx, prefix: decoder_layer_type(
|
||||
layer_id=idx,
|
||||
config=config,
|
||||
quant_config=quant_config,
|
||||
prefix=prefix,
|
||||
),
|
||||
pp_rank=self.pp_group.rank_in_group,
|
||||
pp_size=self.pp_group.world_size,
|
||||
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)
|
||||
|
||||
self.layers_to_capture = []
|
||||
|
||||
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 = None
|
||||
else:
|
||||
assert pp_proxy_tensors is not None
|
||||
hidden_states = pp_proxy_tensors["hidden_states"]
|
||||
residual = pp_proxy_tensors["residual"]
|
||||
|
||||
aux_hidden_states = []
|
||||
for i in range(self.start_layer, self.end_layer):
|
||||
with get_global_expert_distribution_recorder().with_current_layer(i):
|
||||
if i in self.layers_to_capture:
|
||||
aux_hidden_states.append(hidden_states + residual)
|
||||
layer = self.layers[i]
|
||||
hidden_states, residual = layer(
|
||||
positions, hidden_states, forward_batch, residual
|
||||
)
|
||||
if not self.pp_group.is_last_rank:
|
||||
return PPProxyTensors(
|
||||
{
|
||||
"hidden_states": hidden_states,
|
||||
"residual": residual,
|
||||
}
|
||||
)
|
||||
else:
|
||||
if hidden_states.shape[0] != 0:
|
||||
if residual is None:
|
||||
hidden_states = self.norm(hidden_states)
|
||||
else:
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
if len(aux_hidden_states) == 0:
|
||||
return hidden_states
|
||||
|
||||
return hidden_states, aux_hidden_states
|
||||
|
||||
|
||||
class GptOssForCausalLM(nn.Module):
|
||||
fall_back_to_pt_during_load = False
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: GptOssConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.pp_group = get_pp_group()
|
||||
self.config = config
|
||||
self.quant_config = quant_config
|
||||
self.model = GptOssModel(
|
||||
config, quant_config, prefix=add_prefix("model", prefix)
|
||||
)
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("lm_head", prefix),
|
||||
use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(config)
|
||||
self.capture_aux_hidden_states = False
|
||||
|
||||
@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,
|
||||
)
|
||||
|
||||
aux_hidden_states = None
|
||||
if self.capture_aux_hidden_states:
|
||||
hidden_states, aux_hidden_states = hidden_states
|
||||
|
||||
if self.pp_group.is_last_rank:
|
||||
return self.logits_processor(
|
||||
input_ids,
|
||||
hidden_states,
|
||||
self.lm_head,
|
||||
forward_batch,
|
||||
aux_hidden_states,
|
||||
)
|
||||
else:
|
||||
return hidden_states
|
||||
|
||||
@property
|
||||
def start_layer(self):
|
||||
return self.model.start_layer
|
||||
|
||||
@property
|
||||
def end_layer(self):
|
||||
return self.model.end_layer
|
||||
|
||||
def _get_default_weight_mapping(self):
|
||||
"""Generate default weight name mapping for GptOss safetensors."""
|
||||
weight_mapping = {}
|
||||
|
||||
# Map router weights to gate
|
||||
weight_mapping["embedding.weight"] = "model.embed_tokens.weight"
|
||||
weight_mapping["unembedding.weight"] = "lm_head.weight"
|
||||
weight_mapping["norm.scale"] = "model.norm.weight"
|
||||
for layer_id in range(self.config.num_hidden_layers):
|
||||
weight_mapping[f"block.{layer_id}.attn.q_proj.weight"] = (
|
||||
f"model.layers.{layer_id}.self_attn.q_proj.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.q_proj.bias"] = (
|
||||
f"model.layers.{layer_id}.self_attn.q_proj.bias"
|
||||
)
|
||||
|
||||
weight_mapping[f"block.{layer_id}.attn.k_proj.weight"] = (
|
||||
f"model.layers.{layer_id}.self_attn.k_proj.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.k_proj.bias"] = (
|
||||
f"model.layers.{layer_id}.self_attn.k_proj.bias"
|
||||
)
|
||||
|
||||
weight_mapping[f"block.{layer_id}.attn.v_proj.weight"] = (
|
||||
f"model.layers.{layer_id}.self_attn.v_proj.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.v_proj.bias"] = (
|
||||
f"model.layers.{layer_id}.self_attn.v_proj.bias"
|
||||
)
|
||||
|
||||
weight_mapping[f"block.{layer_id}.attn.out.weight"] = (
|
||||
f"model.layers.{layer_id}.self_attn.o_proj.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.out.bias"] = (
|
||||
f"model.layers.{layer_id}.self_attn.o_proj.bias"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.sinks"] = (
|
||||
f"model.layers.{layer_id}.self_attn.sinks"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.attn.norm.scale"] = (
|
||||
f"model.layers.{layer_id}.input_layernorm.weight"
|
||||
)
|
||||
|
||||
weight_mapping[f"block.{layer_id}.mlp.gate.weight"] = (
|
||||
f"model.layers.{layer_id}.mlp.router.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.gate.bias"] = (
|
||||
f"model.layers.{layer_id}.mlp.router.bias"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.norm.scale"] = (
|
||||
f"model.layers.{layer_id}.post_attention_layernorm.weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.experts.gate_up_proj"] = (
|
||||
f"model.layers.{layer_id}.mlp.experts.gate_up_proj"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.gate_up_proj_bias"] = (
|
||||
f"model.layers.{layer_id}.mlp.experts.gate_up_proj_bias"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.down_proj"] = (
|
||||
f"model.layers.{layer_id}.mlp.experts.mlp2_weight"
|
||||
)
|
||||
weight_mapping[f"block.{layer_id}.mlp.down_proj_bias"] = (
|
||||
f"model.layers.{layer_id}.mlp.experts.mlp2_bias"
|
||||
)
|
||||
|
||||
return weight_mapping
|
||||
|
||||
def load_weights(
|
||||
self,
|
||||
weights: Iterable[Tuple[str, torch.Tensor]],
|
||||
is_nextn: bool = False,
|
||||
weight_name_mapping: dict = None,
|
||||
):
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
if is_nextn:
|
||||
logging.warning(
|
||||
"Loading weights for nextn is currently not supported in GptOssForCausalLM. "
|
||||
)
|
||||
return
|
||||
weights = _canonicalize_weights(self.config, weights)
|
||||
weights = sorted(weights, key=lambda x: x[0]) # Sort by name for consistency
|
||||
|
||||
new_weights = []
|
||||
for name, p in weights:
|
||||
if "qkv.weight" in name:
|
||||
q_proj, k_proj, v_proj = p.split(
|
||||
[
|
||||
self.config.num_attention_heads * self.config.head_dim,
|
||||
self.config.num_key_value_heads * self.config.head_dim,
|
||||
self.config.num_key_value_heads * self.config.head_dim,
|
||||
],
|
||||
dim=0,
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.weight', 'q_proj.weight')}", q_proj)
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.weight', 'k_proj.weight')}", k_proj)
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.weight', 'v_proj.weight')}", v_proj)
|
||||
)
|
||||
elif "qkv.bias" in name:
|
||||
q_bias, k_bias, v_bias = p.split(
|
||||
[
|
||||
self.config.num_attention_heads * self.config.head_dim,
|
||||
self.config.num_key_value_heads * self.config.head_dim,
|
||||
self.config.num_key_value_heads * self.config.head_dim,
|
||||
],
|
||||
dim=0,
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.bias', 'q_proj.bias')}", q_bias)
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.bias', 'k_proj.bias')}", k_bias)
|
||||
)
|
||||
new_weights.append(
|
||||
(f"{name.replace('qkv.bias', 'v_proj.bias')}", v_bias)
|
||||
)
|
||||
else:
|
||||
new_weights.append((name, p))
|
||||
weights = new_weights
|
||||
|
||||
# Use provided weight name mapping if available, otherwise use default
|
||||
if weight_name_mapping is None:
|
||||
weight_name_mapping = self._get_default_weight_mapping()
|
||||
else:
|
||||
# Merge with default mapping
|
||||
default_mapping = self._get_default_weight_mapping()
|
||||
default_mapping.update(weight_name_mapping)
|
||||
weight_name_mapping = default_mapping
|
||||
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("qkv_proj", "q_proj", "q"),
|
||||
("qkv_proj", "k_proj", "k"),
|
||||
("qkv_proj", "v_proj", "v"),
|
||||
]
|
||||
|
||||
expert_params_mapping = get_moe_impl_class().make_expert_params_mapping_fused(
|
||||
ckpt_gate_up_proj_name="gate_up_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
ckpt_gate_up_proj_bias_name="gate_up_proj_bias",
|
||||
ckpt_down_proj_bias_name="down_proj_bias",
|
||||
)
|
||||
|
||||
params_dict = dict(self.named_parameters())
|
||||
params_checker = {k: False for k, v in params_dict.items()}
|
||||
for name, loaded_weight in weights:
|
||||
loaded_weight = _WeightCreator.maybe_materialize(loaded_weight)
|
||||
|
||||
# Apply weight name mapping if provided
|
||||
if weight_name_mapping and name in weight_name_mapping:
|
||||
name = weight_name_mapping[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 "rotary_emb.inv_freq" in name:
|
||||
continue
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
if "mlp.experts" in name:
|
||||
continue
|
||||
|
||||
name = name.replace(weight_name, param_name)
|
||||
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)
|
||||
params_checker[name] = True
|
||||
break
|
||||
else:
|
||||
for mapping in expert_params_mapping:
|
||||
param_name, weight_name, shard_id = mapping
|
||||
if weight_name not in name:
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
if name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
if "bias" not in name:
|
||||
loaded_weight = loaded_weight.transpose(-2, -1)
|
||||
if "w2_weight_bias" in name and get_moe_tensor_parallel_rank() != 0:
|
||||
loaded_weight = loaded_weight.zero_()
|
||||
|
||||
weight_loader(
|
||||
param,
|
||||
loaded_weight,
|
||||
name,
|
||||
shard_id=shard_id,
|
||||
)
|
||||
params_checker[name] = True
|
||||
break
|
||||
else:
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
if name not in params_dict:
|
||||
continue
|
||||
if name in params_dict.keys():
|
||||
param = params_dict[name]
|
||||
if "sinks" in name:
|
||||
start = tp_rank * param.numel()
|
||||
param.data.copy_(
|
||||
loaded_weight[start : start + param.numel()]
|
||||
)
|
||||
else:
|
||||
weight_loader = getattr(
|
||||
param, "weight_loader", default_weight_loader
|
||||
)
|
||||
weight_loader(param, loaded_weight)
|
||||
params_checker[name] = True
|
||||
else:
|
||||
logger.warning(f"Parameter {name} not found in params_dict")
|
||||
|
||||
not_loaded_params = [k for k, v in params_checker.items() if not v]
|
||||
if tp_rank == 0:
|
||||
if len(not_loaded_params) > 0:
|
||||
raise Exception(f"Not all parameters loaded: {not_loaded_params}")
|
||||
else:
|
||||
logging.info("All parameters loaded successfully.")
|
||||
|
||||
self.routed_experts_weights_of_layer = {
|
||||
layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
|
||||
for layer_id in range(self.start_layer, self.end_layer)
|
||||
if isinstance(self.model.layers[layer_id].mlp, GptOssSparseMoeBlock)
|
||||
}
|
||||
|
||||
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_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
|
||||
if not self.pp_group.is_last_rank:
|
||||
return
|
||||
|
||||
if layer_ids is None:
|
||||
self.capture_aux_hidden_states = True
|
||||
num_layers = self.config.num_hidden_layers
|
||||
self.model.layers_to_capture = [2, num_layers // 2, num_layers - 3]
|
||||
else:
|
||||
self.capture_aux_hidden_states = True
|
||||
# we plus 1 here because in sglang, for the ith layer, it takes the output
|
||||
# of the (i-1)th layer as aux hidden state
|
||||
self.model.layers_to_capture = [val + 1 for val in layer_ids]
|
||||
|
||||
@classmethod
|
||||
def get_model_config_for_expert_location(cls, config):
|
||||
return ModelConfigForExpertLocation(
|
||||
num_layers=config.num_hidden_layers,
|
||||
num_logical_experts=config.num_local_experts,
|
||||
num_groups=None,
|
||||
)
|
||||
|
||||
def get_attention_sliding_window_size(self):
|
||||
return get_attention_sliding_window_size(self.config)
|
||||
|
||||
|
||||
def _canonicalize_weights(config, weights_in: Iterable[Tuple[str, torch.Tensor]]):
|
||||
weights_out_dict = dict(weights_in)
|
||||
|
||||
for layer_id in range(config.num_hidden_layers):
|
||||
for name_chunk in ["mlp1_weight", "mlp2_weight"]:
|
||||
name_prefix = f"block.{layer_id}.mlp.{name_chunk}"
|
||||
w_blocks = weights_out_dict.pop(f"{name_prefix}.blocks", None)
|
||||
w_scales = weights_out_dict.pop(f"{name_prefix}.scales", None)
|
||||
if w_blocks is not None:
|
||||
weights_out_dict[name_prefix] = _WeightCreator(
|
||||
partial(
|
||||
_dequant_mlp_weight,
|
||||
debug_name=name_prefix,
|
||||
w_blocks=w_blocks,
|
||||
w_scales=w_scales,
|
||||
)
|
||||
)
|
||||
|
||||
return list(weights_out_dict.items())
|
||||
|
||||
|
||||
def _dequant_mlp_weight(debug_name, w_blocks, w_scales):
|
||||
if get_tensor_model_parallel_rank() == 0:
|
||||
logger.info(f"Dequantize {debug_name} start")
|
||||
|
||||
original_device = w_blocks.device
|
||||
|
||||
w_blocks = w_blocks.cuda()
|
||||
w_scales = w_scales.cuda()
|
||||
|
||||
w_bf16 = dequant_mxfp4(w_block=w_blocks, w_scale=w_scales, out_dtype=torch.bfloat16)
|
||||
w_bf16 = w_bf16.transpose(-2, -1).contiguous()
|
||||
|
||||
if get_tensor_model_parallel_rank() == 0:
|
||||
logger.info(
|
||||
f"Dequantize {debug_name} end {w_blocks.shape=} {w_scales.shape=} {w_bf16.shape=}"
|
||||
)
|
||||
|
||||
return w_bf16.to(original_device)
|
||||
|
||||
|
||||
class _WeightCreator:
|
||||
def __init__(self, fn):
|
||||
self._fn = fn
|
||||
|
||||
@staticmethod
|
||||
def maybe_materialize(obj):
|
||||
if isinstance(obj, _WeightCreator):
|
||||
output = obj._fn()
|
||||
obj._fn = None
|
||||
return output
|
||||
|
||||
return obj
|
||||
|
||||
|
||||
EntryClass = GptOssForCausalLM
|
||||
Reference in New Issue
Block a user