# SPDX-FileCopyrightText: Copyright (c) 2022-2023 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 copy import deepcopy
from types import SimpleNamespace
import tensorrt as trt
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.quantization import QuantMode
from ..._common import default_net
from ..._utils import pad_vocab_size, str_dtype_to_trt
from ...functional import (PositionEmbeddingType, Tensor, concat,
gather_last_token_logits, shape)
from ...layers import (MLP, Attention, AttentionMaskType, AttentionParams,
ColumnLinear, Embedding, KeyValueCacheParams, LayerNorm,
RmsNorm)
from ...module import Module, ModuleList
from ..generation_mixin import GenerationMixin
class ChatGLMParams:
apply_query_key_layer_scaling: bool = None
apply_residual_connection_post_layernorm: bool = None
dtype: str = None
enable_debug_output: bool = None
ffn_hidden_size: int = None
hidden_act: str = None
hidden_size: int = None
linear_bias: bool = None
logits_dtype: str = None
mapping: Mapping = None
max_input_len: int = None
max_output_len: int = None
max_seq_length: int = None
model_name: str = None
norm_epsilon: float = None
num_heads: int = None
num_kv_heads: int = None
num_layers: int = None
qkv_bias: bool = None
quant_mode: QuantMode = None
rmsnorm: bool = None
rotary_embedding_scaling: float = None
tokens_per_block: int = None
use_cache: bool = None
vocab_size: int = None
# default values
default_config = {}
default_config["chatglm_6b"] = SimpleNamespace(
apply_query_key_layer_scaling=False,
apply_residual_connection_post_layernorm=False,
dtype="float16",
ffn_hidden_size=16384,
hidden_act='gelu',
hidden_size=4096,
linear_bias=True,
logits_dtype="float16",
mapping=Mapping(),
max_input_len=1024,
max_output_len=1024,
max_seq_length=2048,
norm_epsilon=1.0e-5,
num_heads=32,
num_kv_heads=32,
num_layers=28,
qkv_bias=True,
quant_mode=QuantMode(0),
rmsnorm=False,
rotary_embedding_scaling=1.0,
use_cache=True,
vocab_size=130528,
)
default_config["chatglm2_6b"] = SimpleNamespace(
apply_query_key_layer_scaling=False,
apply_residual_connection_post_layernorm=False,
dtype="float16",
ffn_hidden_size=13696,
hidden_act='swiglu',
hidden_size=4096,
linear_bias=False,
logits_dtype="float16",
mapping=Mapping(),
max_input_len=1024,
max_output_len=1024,
max_seq_length=32768,
norm_epsilon=1.0e-5,
num_heads=32,
num_kv_heads=2,
num_layers=28,
qkv_bias=True,
quant_mode=QuantMode(0),
rmsnorm=True,
rotary_embedding_scaling=1.0,
use_cache=True,
vocab_size=65024,
)
default_config["chatglm3_6b"] = default_config["chatglm2_6b"]
default_config["chatglm2_6b_32k"] = deepcopy(default_config["chatglm2_6b"])
default_config["chatglm2_6b_32k"].rotary_embedding_scaling = 50.0
default_config["chatglm3_6b_32k"] = default_config["chatglm2_6b_32k"]
default_config["glm_10b"] = SimpleNamespace(
apply_query_key_layer_scaling=False,
apply_residual_connection_post_layernorm=False,
dtype="float16",
ffn_hidden_size=16384,
hidden_act='gelu',
hidden_size=4096,
linear_bias=True,
logits_dtype="float16",
mapping=Mapping(),
max_input_len=1024,
max_output_len=1024,
max_seq_length=2048,
norm_epsilon=1.0e-5,
num_heads=32,
num_kv_heads=32,
num_layers=28,
qkv_bias=True,
quant_mode=QuantMode(0),
rmsnorm=False,
rotary_embedding_scaling=1.0,
use_cache=True,
vocab_size=50304,
)
def __init__(self, **args):
for key, value in args.items():
assert key in dir(
self), f"{key} is not in configuration of ChatGLMHeadModel"
if value is not None:
self.__setattr__(key, value)
assert self.model_name is not None, "model_name must be set for ChatGLMHeadModel"
# fill other parameters as default values
for key, value in self.default_config[self.model_name].__dict__.items():
if self.__getattribute__(key) is None:
self.__setattr__(key, value)
def report(self):
for key, value in self.__dict__.items():
print(f"{key} = {value}")
class ChatGLMDecoderLayer(Module):
def __init__(self, layer_id, config):
super().__init__()
self.model_name = config.model_name
self.use_cache = config.use_cache
self.rotary_embedding_base = 10000.0
rotary_embedding_scaling = None
if self.model_name in ["chatglm_6b"]:
self.alpha = (2 * config.num_layers)**0.5
self.norm = LayerNorm
attention_mask_type = AttentionMaskType.bidirectional
position_embedding_type = PositionEmbeddingType.chatglm
elif config.model_name in [
"chatglm2_6b", "chatglm2_6b_32k", "chatglm3_6b",
"chatglm3_6b_base", "chatglm3_6b_32k"
]:
self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
self.norm = RmsNorm if config.rmsnorm else LayerNorm
attention_mask_type = AttentionMaskType.causal
position_embedding_type = PositionEmbeddingType.rope_gptj
if config.model_name in ["chatglm2_6b_32k"]:
rotary_embedding_scaling = {
"type": "linear",
"factor": config.rotary_embedding_scaling
}
elif config.model_name in ["chatglm3_6b_32k"]:
self.rotary_embedding_base *= config.rotary_embedding_scaling
elif config.model_name in ["glm_10b"]:
self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
self.norm = LayerNorm
attention_mask_type = AttentionMaskType.bidirectionalglm
position_embedding_type = PositionEmbeddingType.learned_absolute
self.pre_norm = self.norm(
normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
elementwise_affine=True,
dtype=config.dtype,
)
self.attention = Attention(
hidden_size=config.hidden_size,
num_attention_heads=config.num_heads,
num_kv_heads=config.num_kv_heads,
max_position_embeddings=config.max_seq_length,
num_layers=config.num_layers,
apply_query_key_layer_scaling=config.apply_query_key_layer_scaling,
attention_mask_type=attention_mask_type,
bias=config.qkv_bias,
dtype=config.dtype,
position_embedding_type=position_embedding_type,
rotary_embedding_base=self.rotary_embedding_base,
rotary_embedding_scaling=rotary_embedding_scaling,
use_int8_kv_cache=config.quant_mode.has_int8_kv_cache(),
rotary_embedding_percentage=0.5,
tp_group=config.mapping.tp_group,
tp_size=config.mapping.tp_size,
tp_rank=config.mapping.rank,
quant_mode=config.quant_mode,
q_scaling=1.0,
cross_attention=False,
relative_attention=False,
max_distance=0,
num_buckets=0,
instance_id=layer_id * 2,
dense_bias=config.linear_bias,
)
self.mlp = MLP(
hidden_size=config.hidden_size,
ffn_hidden_size=config.ffn_hidden_size,
hidden_act=config.hidden_act,
bias=config.linear_bias,
dtype=config.dtype,
tp_group=config.mapping.tp_group,
tp_size=config.mapping.tp_size,
quant_mode=config.quant_mode,
instance_id=layer_id * 2 + 1,
)
self.post_norm = self.norm(
normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
elementwise_affine=True,
dtype=config.dtype,
)
def forward(
self,
hidden_states: Tensor,
position_ids: Tensor = None, # only used in ChatGLM-6B
kv_cache_params: KeyValueCacheParams = None,
attention_params: AttentionParams = None,
):
norm_output = self.pre_norm(hidden_states)
attention_output = self.attention(
hidden_states=norm_output,
attention_mask=None,
use_cache=self.use_cache,
kv_cache_params=kv_cache_params,
attention_params=attention_params,
encoder_output=None,
workspace=None,
position_embedding=position_ids,
)
if self.use_cache:
attention_output, presents = attention_output
if self.model_name in ["chatglm_6b"]:
residual = norm_output
norm_input = residual * self.alpha + attention_output
norm_output = self.post_norm(norm_input)
mlp_output = self.mlp(norm_output)
residual = norm_output
output = residual * self.alpha + mlp_output
elif self.model_name in [
"chatglm2_6b", "chatglm2_6b_32k", "chatglm3_6b",
"chatglm3_6b_base", "chatglm3_6b_32k", "glm_10b"
]:
residual = norm_output if self.apply_residual_connection_post_layernorm else hidden_states
norm_input = residual + attention_output
norm_output = self.post_norm(norm_input)
mlp_output = self.mlp(norm_output)
residual = norm_output if self.apply_residual_connection_post_layernorm else norm_input
output = residual + mlp_output
return (output, presents) if self.use_cache else output
[docs]
class ChatGLMModel(Module):
def __init__(self, config):
super().__init__()
self.model_name = config.model_name
if config.model_name in ["chatglm_6b", "glm_10b"]:
self.norm = LayerNorm
elif config.model_name in [
"chatglm2_6b", "chatglm2_6b_32k", "chatglm3_6b",
"chatglm3_6b_base", "chatglm3_6b_32k"
]:
self.norm = RmsNorm
self.use_cache = config.use_cache
self.embedding = Embedding(
num_embeddings=config.vocab_size,
embedding_dim=config.hidden_size,
dtype=config.dtype,
tp_size=1, #config.mapping.tp_size,
tp_group=None, #config.mapping.tp_group,
sharding_dim=0,
tp_rank=0, #config.mapping.rank,
instance_id=config.num_layers * 2,
)
if config.model_name in ["glm_10b"]:
self.position_embeddings = Embedding(
config.max_seq_length + 1,
config.hidden_size,
dtype=config.dtype,
tp_size=1, #config.mapping.tp_size,
tp_group=None, #config.mapping.tp_group,
sharding_dim=0,
tp_rank=0, #config.mapping.rank,
instance_id=config.num_layers * 2,
)
self.block_embeddings = Embedding(
config.max_seq_length + 1,
config.hidden_size,
dtype=config.dtype,
tp_size=1, #config.mapping.tp_size,
tp_group=None, #config.mapping.tp_group,
sharding_dim=0,
tp_rank=0, #config.mapping.rank,
instance_id=config.num_layers * 2,
)
self.layers = ModuleList(
ChatGLMDecoderLayer(i, config) for i in range(config.num_layers))
self.final_norm = self.norm(
normalized_shape=config.hidden_size,
eps=config.norm_epsilon,
elementwise_affine=True,
dtype=config.dtype,
)
[docs]
def forward(
self,
input_ids: Tensor = None,
position_ids: Tensor = None, # only used in ChatGLM-6B
kv_cache_params: KeyValueCacheParams = None,
attention_params: AttentionParams = None,
):
hidden_states = self.embedding(input_ids)
if self.model_name in ["glm_10b"]:
position_ids_list = position_ids.split(1, dim=1)
position_embedding = self.position_embeddings(position_ids_list[0])
block_embedding = self.block_embeddings(position_ids_list[1])
position_embedding = position_embedding + block_embedding
position_embedding = position_embedding.view(
concat([
shape(input_ids, 0),
shape(input_ids, 1),
4096,
]))
hidden_states = hidden_states + position_embedding
kv_cache_params.fill_none_tensor_list(len(self.layers))
if self.use_cache:
presents = []
for layer, past, pointer, max_kv_cache_length in zip(
self.layers, kv_cache_params.past_key_value,
kv_cache_params.kv_cache_block_pointers,
kv_cache_params.host_max_kv_cache_lengths):
layer_output = layer(
hidden_states,
position_ids,
kv_cache_params=KeyValueCacheParams(
past_key_value=[past],
kv_cache_block_pointers=[pointer],
host_past_key_value_lengths=kv_cache_params.
host_past_key_value_lengths,
host_max_kv_cache_lengths=max_kv_cache_length,
cache_indirection=kv_cache_params.cache_indirection,
),
attention_params=attention_params,
)
if self.use_cache:
hidden_states = layer_output[0]
presents.append(layer_output[1])
hidden_states = self.final_norm(hidden_states)
return (hidden_states,
tuple(presents)) if self.use_cache else hidden_states
[docs]
class ChatGLMHeadModel(ChatGLMModel, GenerationMixin):
def __init__(
self,
apply_query_key_layer_scaling: bool = None,
apply_residual_connection_post_layernorm: bool = None,
dtype: str = None,
enable_debug_output: bool = None,
ffn_hidden_size: int = None,
hidden_act: str = None,
hidden_size: int = None,
linear_bias: bool = None,
logits_dtype: str = None,
mapping: Mapping = None,
max_input_len: int = None,
max_output_len: int = None,
max_seq_length: int = None,
model_name: str = None,
norm_epsilon: float = None,
num_heads: int = None,
num_kv_heads: int = None,
num_layers: int = None,
qkv_bias: bool = None,
quant_mode: QuantMode = None,
rmsnorm: bool = None,
rotary_embedding_scaling: float = None,
tokens_per_block: int = None,
use_cache: bool = None,
vocab_size: int = None,
max_position_embeddings: int = None,
):
# for benchmark scripts
if max_seq_length is None and max_position_embeddings is not None:
max_seq_length = max_position_embeddings
config = ChatGLMParams(
apply_query_key_layer_scaling=apply_query_key_layer_scaling,
apply_residual_connection_post_layernorm=
apply_residual_connection_post_layernorm,
dtype=dtype,
enable_debug_output=enable_debug_output,
ffn_hidden_size=ffn_hidden_size,
hidden_act=hidden_act,
hidden_size=hidden_size,
linear_bias=linear_bias,
logits_dtype=logits_dtype,
mapping=mapping,
max_input_len=max_input_len,
max_output_len=max_output_len,
max_seq_length=max_seq_length,
model_name=model_name,
norm_epsilon=norm_epsilon,
num_heads=num_heads,
num_kv_heads=num_kv_heads,
num_layers=num_layers,
qkv_bias=qkv_bias,
quant_mode=quant_mode,
rmsnorm=rmsnorm,
rotary_embedding_scaling=rotary_embedding_scaling,
tokens_per_block=tokens_per_block,
use_cache=use_cache,
vocab_size=vocab_size,
)
super().__init__(config)
if isinstance(config.dtype, str):
self.kv_dtype = str_dtype_to_trt(config.dtype)
else:
assert isinstance(config.dtype, trt.DataType)
self.kv_dtype = config.dtype
self.dtype = self.kv_dtype
if isinstance(config.logits_dtype, str):
self.logits_dtype = str_dtype_to_trt(config.logits_dtype)
else:
assert isinstance(config.logits_dtype, trt.DataType)
self.logits_dtype = config.logits_dtype
if config.quant_mode.has_int8_kv_cache():
self.kv_dtype = str_dtype_to_trt('int8')
elif config.quant_mode.has_fp8_kv_cache():
self.kv_dtype = str_dtype_to_trt('fp8')
self.hidden_size = config.hidden_size
self.mapping = config.mapping
self.max_num_tokens = config.max_output_len + config.max_input_len
self.model_name = config.model_name
self.num_heads = config.num_heads
self.num_kv_heads = config.num_kv_heads
self.num_layers = config.num_layers
self.tokens_per_block = config.tokens_per_block
self.use_cache = config.use_cache
self.lm_head = ColumnLinear(
in_features=self.hidden_size,
out_features=pad_vocab_size(config.vocab_size,
self.mapping.tp_size),
bias=False,
dtype=self.dtype,
tp_group=self.mapping.tp_group,
tp_size=self.mapping.tp_size,
gather_output=True,
)
[docs]
def forward(
self,
input_ids: Tensor = None,
position_ids: Tensor = None, # only used in ChatGLM-6B
last_token_ids: Tensor = None,
kv_cache_params: KeyValueCacheParams = None,
attention_params: AttentionParams = None,
):
hidden_states = super().forward(
input_ids,
position_ids,
kv_cache_params,
attention_params,
)
if self.use_cache:
hidden_states, presents = hidden_states
hidden_states = gather_last_token_logits(
hidden_states, last_token_ids,
default_net().plugin_config.remove_input_padding)
lm_logits = self.lm_head(hidden_states)
lm_logits.mark_output('logits', self.logits_dtype)
if self.use_cache and default_net(
).plugin_config.paged_kv_cache == False:
for i, present in enumerate(presents):
present.mark_output(f'present_key_value_{i}', self.kv_dtype)
return (lm_logits, presents)
return lm_logits