# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.
from typing import Optional
import torch
from ..._utils import pad_vocab_size, torch_dtype_to_str
from ...functional import Tensor, non_gated_version, recv, send
from ...layers import (Attention, AttentionMaskType, ColumnLinear, Embedding,
GatedMLP, MoeConfig, PositionEmbeddingType, RmsNorm,
SharedMoE)
from ...mapping import Mapping
from ...module import Module
from ...plugin import init_all_reduce_helper
from ..modeling_utils import (DecoderLayerList, DecoderModelForCausalLM,
PretrainedConfig)
from .convert import convert_deepseek, create_trt_config_from_hf
class DeepseekDecoderLayer(Module):
def __init__(self, config: PretrainedConfig, layer_idx: int):
super().__init__()
self.layer_idx = layer_idx
self.config = config
### Input layernorm in Deepseek v1 is same as Llama
self.input_layernorm = RmsNorm(normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
dtype=config.dtype)
layers_range = config.mapping.pp_layers(config.num_hidden_layers)
local_layer_idx = layer_idx - layers_range[0]
### Deepseek v1 model with standard attention
self.attention = Attention(
local_layer_idx=local_layer_idx,
hidden_size=config.hidden_size,
attention_head_size=config.head_size,
num_attention_heads=config.num_attention_heads,
num_kv_heads=config.num_key_value_heads,
max_position_embeddings=config.max_position_embeddings,
dtype=config.dtype,
attention_mask_type=AttentionMaskType.causal,
bias=False,
position_embedding_type=PositionEmbeddingType.rope_gpt_neox,
rotary_embedding_base=config.rotary_base,
rotary_embedding_scaling=config.rotary_scaling,
tp_group=config.mapping.tp_group,
tp_size=config.mapping.tp_size,
tp_rank=config.mapping.tp_rank,
quant_mode=config.quant_mode)
ClsMLP = GatedMLP
moe_config = MoeConfig.from_dict(config.moe)
mlp_kwargs = {}
if moe_config.num_experts > 0 and layer_idx > 0:
mlp_hidden_size = moe_config.num_shared_experts * moe_config.moe_intermediate_size
hidden_act = config.hidden_act
ClsMLP = SharedMoE
mlp_kwargs = {"moe_config": moe_config, "mapping": config.mapping}
else:
ClsMLP = GatedMLP
mlp_hidden_size = config.intermediate_size
hidden_act = non_gated_version(
config.hidden_act) # back to non gated for dense layers
self.mlp = ClsMLP(hidden_size=config.hidden_size,
ffn_hidden_size=mlp_hidden_size,
hidden_act=hidden_act,
dtype=config.dtype,
bias=False,
tp_group=config.mapping.tp_group,
tp_size=config.mapping.tp_size,
quant_mode=config.quant_mode,
**mlp_kwargs)
### Pose layernorm in Deepseek v1 is same as Llama )
self.post_layernorm = RmsNorm(normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
dtype=config.dtype)
def forward(self,
hidden_states,
attention_mask=None,
use_cache=False,
spec_decoding_params=None,
kv_cache_params=None,
attention_params=None):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attention_output = self.attention(
hidden_states,
attention_mask=attention_mask,
use_cache=use_cache,
spec_decoding_params=spec_decoding_params,
kv_cache_params=kv_cache_params,
attention_params=attention_params)
if use_cache:
attention_output, presents = attention_output
hidden_states = residual + attention_output
residual = hidden_states
hidden_states = self.post_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
if use_cache:
return (hidden_states, presents)
return hidden_states
class DeepseekModel(Module):
def __init__(self, config: PretrainedConfig) -> None:
super().__init__()
init_all_reduce_helper() # enable use_customer_all_reduce
self.mapping = config.mapping
if self.mapping.is_first_pp_rank():
self.vocab_embedding = Embedding(config.vocab_size,
config.hidden_size,
dtype=config.dtype)
self.layers = DecoderLayerList(DeepseekDecoderLayer, config)
if self.mapping.is_last_pp_rank():
self.ln_f = RmsNorm(normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
dtype=config.dtype)
def forward(self,
input_ids,
position_ids=None,
use_cache=False,
attention_mask=None,
spec_decoding_params=None,
kv_cache_params=None,
attention_params=None,
hidden_states=None,
prompt_embedding_table: Optional[Tensor] = None,
prompt_tasks: Optional[Tensor] = None,
prompt_vocab_size: Optional[Tensor] = None):
ptuning_args = [
prompt_embedding_table, prompt_tasks, prompt_vocab_size
] if prompt_embedding_table is not None else []
if self.mapping.is_first_pp_rank():
hidden_states = self.vocab_embedding(input_ids, *ptuning_args)
else:
hidden_states = recv(hidden_states, self.mapping.prev_pp_rank())
hidden_states = self.layers.forward(
hidden_states,
use_cache=use_cache,
attention_mask=attention_mask,
kv_cache_params=kv_cache_params,
attention_params=attention_params,
spec_decoding_params=spec_decoding_params)
if use_cache:
hidden_states, presents = hidden_states
if self.mapping.is_last_pp_rank():
hidden_states = self.ln_f(hidden_states)
else:
hidden_states = send(hidden_states, self.mapping.next_pp_rank())
if use_cache:
return (hidden_states, tuple(presents))
return hidden_states
[docs]
class DeepseekForCausalLM(DecoderModelForCausalLM):
def __init__(self, config: PretrainedConfig):
transformer = DeepseekModel(config)
vocab_size_padded = pad_vocab_size(config.vocab_size,
config.mapping.tp_size)
if config.mapping.is_last_pp_rank():
lm_head = ColumnLinear(config.hidden_size,
vocab_size_padded,
bias=False,
dtype=config.dtype,
tp_group=config.mapping.tp_group,
tp_size=config.mapping.tp_size,
gather_output=True)
else:
lm_head = None
self.mapping = config.mapping
super().__init__(config, transformer, lm_head)
[docs]
@classmethod
def from_hugging_face(cls,
hf_model,
model_dir,
dtype: str = 'auto',
mapping: Optional[Mapping] = None,
override_fields={},
**kwargs):
assert hf_model is not None
if mapping is None:
mapping = Mapping()
config = create_trt_config_from_hf(model_dir,
dtype,
mapping=mapping,
override_fields=override_fields)
print(config)
pretrained_config = PretrainedConfig.from_dict(config)
pretrained_config.set_rank(mapping.rank) # TODO:remove this hack
if dtype == 'auto':
dtype = getattr(config, 'torch_dtype', None)
if dtype is None:
dtype = 'float16'
if isinstance(dtype, torch.dtype):
dtype = torch_dtype_to_str(dtype)
if dtype == 'float32': # should remove "float32"
dtype = 'float16'
if dtype == 'bfloat16' and torch.cuda.get_device_properties(
0).major < 8:
logger.warning(
"Pre SM 80 GPUs do not support bfloat16, fallback to float16")
dtype = 'float16'
deepseek = cls.from_config(pretrained_config)
weights = convert_deepseek(
hf_model,
config,
mapping,
dtype=dtype,
use_parallel_embedding=config.get('use_parallel_embedding', False),
sharding_dim=config.get('embedding_sharding_dim', 0),
share_embedding_table=config.get('share_embedding_table', False))
#check_share_embedding(weights, config)
deepseek.load(weights)
return deepseek