mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-23 12:12:39 +08:00
455 lines
20 KiB
Python
455 lines
20 KiB
Python
# 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.
|
|
import configparser
|
|
import time
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
import torch
|
|
|
|
import tensorrt_llm
|
|
from tensorrt_llm.functional import is_gated_activation
|
|
from tensorrt_llm.models import ChatGLM6BHeadModel
|
|
from tensorrt_llm.quantization import QuantMode
|
|
|
|
|
|
def gen_suffix(rank, use_smooth_quant, quant_per_channel):
|
|
suffix = f"{rank}.bin"
|
|
if use_smooth_quant:
|
|
sq_prefix = "int8."
|
|
if quant_per_channel:
|
|
sq_prefix += "col."
|
|
suffix = sq_prefix + suffix
|
|
return suffix
|
|
|
|
|
|
def extract_layer_idx(name):
|
|
ss = name.split('.')
|
|
for s in ss:
|
|
if s.isdigit():
|
|
return s
|
|
return None
|
|
|
|
|
|
def split(v, tp_size, idx, dim=0):
|
|
if tp_size == 1:
|
|
return v
|
|
if len(v.shape) == 1:
|
|
return np.ascontiguousarray(np.split(v, tp_size)[idx])
|
|
elif len(v.shape) == 2:
|
|
return np.ascontiguousarray(np.split(v, tp_size, axis=dim)[idx])
|
|
return None
|
|
|
|
|
|
def parse_ft_config(ini_file):
|
|
chatglm6b_config = configparser.ConfigParser()
|
|
chatglm6b_config.read(ini_file)
|
|
|
|
n_embd = chatglm6b_config.getint('chatglm6b', 'hidden_size')
|
|
n_head = chatglm6b_config.getint('chatglm6b', 'num_attention_heads')
|
|
n_layer = chatglm6b_config.getint('chatglm6b', 'num_layers')
|
|
n_positions = chatglm6b_config.getint('chatglm6b', 'max_sequence_length')
|
|
vocab_size = chatglm6b_config.getint('chatglm6b', 'vocab_size')
|
|
do_layer_norm_before = chatglm6b_config.getboolean('chatglm6b',
|
|
'do_layer_norm_before',
|
|
fallback=True)
|
|
rotary_pct = chatglm6b_config.getfloat('chatglm6b',
|
|
'rotary_pct',
|
|
fallback=0.0)
|
|
hidden_act = 'gelu' #chatglm6b_config.get('chatglm6b', 'activation_function')
|
|
bias = chatglm6b_config.getboolean('chatglm6b', 'bias', fallback=True)
|
|
inter_size = chatglm6b_config.getint('chatglm6b',
|
|
'intermediate_size',
|
|
fallback=None)
|
|
|
|
if inter_size is None:
|
|
inter_size = 4 * n_embd
|
|
|
|
multi_query_mode = chatglm6b_config.getboolean('chatglm6b',
|
|
'multi_query_mode',
|
|
fallback=False)
|
|
return n_embd, n_head, n_layer, n_positions, vocab_size, do_layer_norm_before, hidden_act, rotary_pct, bias, inter_size, multi_query_mode
|
|
|
|
|
|
def load_from_ft(chatglm6bModel: ChatGLM6BHeadModel,
|
|
dir_path,
|
|
rank=0,
|
|
tensor_parallel=1,
|
|
fp16=False):
|
|
tensorrt_llm.logger.info('Loading weights from FT...')
|
|
tik = time.time()
|
|
|
|
quant_mode = getattr(chatglm6bModel, 'quant_mode', QuantMode(0))
|
|
if quant_mode.is_int8_weight_only():
|
|
plugin_weight_only_quant_type = torch.int8
|
|
elif quant_mode.is_int4_weight_only():
|
|
plugin_weight_only_quant_type = torch.quint4x2
|
|
n_embd, n_head, n_layer, n_positions, vocab_size, do_layer_norm_before, hidden_act, rotary_pct, bias, inter_size, multi_query_mode = parse_ft_config(
|
|
Path(dir_path) / 'config.ini')
|
|
np_dtype = np.float16 if fp16 else np.float32
|
|
|
|
def fromfile(dir_path, name, shape=None, dtype=None):
|
|
dtype = np_dtype if dtype is None else dtype
|
|
p = dir_path + '/' + name
|
|
if Path(p).exists():
|
|
t = np.fromfile(p, dtype=dtype)
|
|
if shape is not None:
|
|
t = t.reshape(shape)
|
|
return t
|
|
return None
|
|
|
|
def set_smoothquant_scale_factors(module,
|
|
pre_scale_weight,
|
|
dir_path,
|
|
basename,
|
|
shape,
|
|
per_tok_dyn,
|
|
per_channel,
|
|
is_qkv=False,
|
|
rank=None):
|
|
suffix = "bin"
|
|
if per_channel:
|
|
if rank is not None:
|
|
suffix = f"{rank}." + suffix
|
|
suffix = "col." + suffix
|
|
|
|
col_shape = shape if (per_channel or is_qkv) else [1, 1]
|
|
if per_tok_dyn:
|
|
if pre_scale_weight is not None:
|
|
pre_scale_weight.value = np.array([1.0], dtype=np.float32)
|
|
t = fromfile(dir_path, f"{basename}scale_w_quant_orig.{suffix}",
|
|
col_shape, np.float32)
|
|
module.per_channel_scale.value = t
|
|
else:
|
|
t = fromfile(dir_path, f"{basename}scale_x_orig_quant.bin", [1],
|
|
np.float32)
|
|
pre_scale_weight.value = t
|
|
t = fromfile(dir_path, f"{basename}scale_y_accum_quant.{suffix}",
|
|
col_shape, np.float32)
|
|
module.per_channel_scale.value = t
|
|
t = fromfile(dir_path, f"{basename}scale_y_quant_orig.bin", [1, 1],
|
|
np.float32)
|
|
module.act_scale.value = t
|
|
|
|
# Determine the quantization mode.
|
|
quant_mode = getattr(chatglm6bModel, "quant_mode", QuantMode(0))
|
|
# Do we use SmoothQuant?
|
|
use_smooth_quant = quant_mode.has_act_and_weight_quant()
|
|
# Do we use quantization per token?
|
|
quant_per_token_dyn = quant_mode.has_per_token_dynamic_scaling()
|
|
# Do we use quantization per channel?
|
|
quant_per_channel = quant_mode.has_per_channel_scaling()
|
|
|
|
# Do we use INT4/INT8 weight-only?
|
|
use_weight_only = quant_mode.is_weight_only()
|
|
|
|
# Int8 KV cache
|
|
use_int8_kv_cache = quant_mode.has_int8_kv_cache()
|
|
|
|
def sq_trick(x):
|
|
return x.view(np.float32) if use_smooth_quant else x
|
|
|
|
# Debug
|
|
suffix = gen_suffix(rank, use_smooth_quant, quant_per_channel)
|
|
# The type of weights.
|
|
w_type = np_dtype if not use_smooth_quant else np.int8
|
|
|
|
# add weight of LM and position embedding
|
|
chatglm6bModel.embedding.weight.value = (fromfile(
|
|
dir_path, 'model.word_embeddings.weight.bin', [vocab_size, n_embd]))
|
|
chatglm6bModel.position_embedding_cos.weight.value = (fromfile(
|
|
dir_path, 'model.cosTable.weight.bin',
|
|
[n_positions, n_embd // n_head // 2]))
|
|
chatglm6bModel.position_embedding_sin.weight.value = (fromfile(
|
|
dir_path, 'model.sinTable.weight.bin',
|
|
[n_positions, n_embd // n_head // 2]))
|
|
if do_layer_norm_before:
|
|
chatglm6bModel.ln_f.bias.value = (fromfile(
|
|
dir_path, 'model.final_layernorm.bias.bin'))
|
|
chatglm6bModel.ln_f.weight.value = (fromfile(
|
|
dir_path, 'model.final_layernorm.weight.bin'))
|
|
# share input embedding
|
|
lm_head_weight = fromfile(dir_path, 'model.lm.weight.bin',
|
|
[vocab_size, n_embd])
|
|
if lm_head_weight is None:
|
|
lm_head_weight = fromfile(dir_path, 'model.wte.bin',
|
|
[vocab_size, n_embd])
|
|
if vocab_size % tensor_parallel != 0:
|
|
# padding
|
|
vocab_size_padded = chatglm6bModel.lm_head.out_features * tensor_parallel
|
|
pad_width = vocab_size_padded - vocab_size
|
|
lm_head_weight = np.pad(lm_head_weight, ((0, pad_width), (0, 0)),
|
|
'constant',
|
|
constant_values=0)
|
|
chatglm6bModel.lm_head.weight.value = np.ascontiguousarray(
|
|
split(lm_head_weight, tensor_parallel, rank))
|
|
for i in range(n_layer):
|
|
chatglm6bModel.layers[i].input_layernorm.weight.value = (fromfile(
|
|
dir_path, 'model.layers.' + str(i) + '.input_layernorm.weight.bin'))
|
|
chatglm6bModel.layers[i].input_layernorm.bias.value = (fromfile(
|
|
dir_path, 'model.layers.' + str(i) + '.input_layernorm.bias.bin'))
|
|
t = fromfile(
|
|
dir_path, 'model.layers.' + str(i) +
|
|
'.attention.query_key_value.weight.' + suffix,
|
|
[3 * n_embd // tensor_parallel, n_embd], w_type)
|
|
if t is not None:
|
|
dst = chatglm6bModel.layers[i].attention.qkv.weight
|
|
if use_smooth_quant:
|
|
dst.value = sq_trick(
|
|
np.ascontiguousarray(np.transpose(t, [1, 0])))
|
|
set_smoothquant_scale_factors(
|
|
chatglm6bModel.layers[i].attention.qkv,
|
|
chatglm6bModel.layers[i].input_layernorm.scale_to_int,
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.attention.query_key_value.',
|
|
[1, 3 * n_embd // tensor_parallel],
|
|
quant_per_token_dyn,
|
|
quant_per_channel,
|
|
rank=rank,
|
|
is_qkv=True)
|
|
elif use_weight_only:
|
|
processed_torch_weights, torch_weight_scales = torch.ops.fastertransformer.symmetric_quantize_last_axis_of_batched_matrix(
|
|
torch.tensor(t), plugin_weight_only_quant_type)
|
|
# workaround for trt not supporting int8 inputs in plugins currently
|
|
dst.value = processed_torch_weights.view(
|
|
dtype=torch.float32).numpy()
|
|
scales = chatglm6bModel.layers[
|
|
i].attention.qkv.per_channel_scale
|
|
scales.value = torch_weight_scales.numpy()
|
|
else:
|
|
dst.value = np.ascontiguousarray(t)
|
|
if bias:
|
|
t = fromfile(
|
|
dir_path, 'model.layers.' + str(i) +
|
|
'.attention.query_key_value.bias.' + str(rank) + '.bin')
|
|
if t is not None:
|
|
dst = chatglm6bModel.layers[i].attention.qkv.bias
|
|
dst.value = np.ascontiguousarray(t)
|
|
|
|
dst = chatglm6bModel.layers[i].attention.dense.weight
|
|
t = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.attention.dense.weight.' + suffix,
|
|
[n_embd, n_embd // tensor_parallel], w_type)
|
|
if use_smooth_quant:
|
|
dst.value = sq_trick(np.ascontiguousarray(np.transpose(t, [1, 0])))
|
|
dense_scale = getattr(chatglm6bModel.layers[i].attention,
|
|
"quantization_scaling_factor", None)
|
|
set_smoothquant_scale_factors(
|
|
chatglm6bModel.layers[i].attention.dense, dense_scale, dir_path,
|
|
'model.layers.' + str(i) + '.attention.dense.', [1, n_embd],
|
|
quant_per_token_dyn, quant_per_channel)
|
|
elif use_weight_only:
|
|
processed_torch_weights, torch_weight_scales = torch.ops.fastertransformer.symmetric_quantize_last_axis_of_batched_matrix(
|
|
torch.tensor(t), plugin_weight_only_quant_type)
|
|
# workaround for trt not supporting int8 inputs in plugins currently
|
|
dst.value = processed_torch_weights.view(
|
|
dtype=torch.float32).numpy()
|
|
scales = chatglm6bModel.layers[i].attention.dense.per_channel_scale
|
|
scales.value = torch_weight_scales.numpy()
|
|
else:
|
|
dst.value = np.ascontiguousarray(t)
|
|
|
|
if bias:
|
|
dst = chatglm6bModel.layers[i].attention.dense.bias
|
|
dst.value = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.attention.dense.bias.bin')
|
|
|
|
dst = chatglm6bModel.layers[i].post_layernorm.weight
|
|
dst.value = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.post_attention_layernorm.weight.bin')
|
|
|
|
dst = chatglm6bModel.layers[i].post_layernorm.bias
|
|
dst.value = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.post_attention_layernorm.bias.bin')
|
|
t = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.mlp.dense_h_to_4h.weight.' + suffix,
|
|
[inter_size // tensor_parallel, n_embd], w_type)
|
|
if use_smooth_quant:
|
|
chatglm6bModel.layers[i].mlp.fc.weight.value = sq_trick(
|
|
np.ascontiguousarray(np.transpose(t, [1, 0])))
|
|
set_smoothquant_scale_factors(
|
|
chatglm6bModel.layers[i].mlp.fc,
|
|
chatglm6bModel.layers[i].post_layernorm.scale_to_int,
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.mlp.dense_h_to_4h.',
|
|
[1, inter_size // tensor_parallel],
|
|
quant_per_token_dyn,
|
|
quant_per_channel,
|
|
rank=rank)
|
|
elif use_weight_only:
|
|
dst = chatglm6bModel.layers[i].mlp.fc.weight
|
|
processed_torch_weights, torch_weight_scales = torch.ops.fastertransformer.symmetric_quantize_last_axis_of_batched_matrix(
|
|
torch.tensor(t), plugin_weight_only_quant_type)
|
|
# workaround for trt not supporting int8 inputs in plugins currently
|
|
dst.value = processed_torch_weights.view(
|
|
dtype=torch.float32).numpy()
|
|
scales = chatglm6bModel.layers[i].mlp.fc.per_channel_scale
|
|
scales.value = torch_weight_scales.numpy()
|
|
else:
|
|
chatglm6bModel.layers[i].mlp.fc.weight.value = np.ascontiguousarray(
|
|
t)
|
|
if bias:
|
|
chatglm6bModel.layers[i].mlp.fc.bias.value = fromfile(
|
|
dir_path, 'model.layers.' + str(i) +
|
|
'.mlp.dense_h_to_4h.bias.' + str(rank) + '.bin')
|
|
if is_gated_activation(hidden_act):
|
|
t = fromfile(
|
|
dir_path, 'model.layers.' + str(i) +
|
|
'.mlp.dense_h_to_4h.gate.weight.' + suffix,
|
|
[inter_size // tensor_parallel, n_embd], w_type)
|
|
chatglm6bModel.layers[
|
|
i].mlp.gate.weight.value = np.ascontiguousarray(t)
|
|
|
|
t = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.mlp.dense_4h_to_h.weight.' + suffix,
|
|
[n_embd, inter_size // tensor_parallel], w_type)
|
|
if use_smooth_quant:
|
|
chatglm6bModel.layers[i].mlp.proj.weight.value = sq_trick(
|
|
np.ascontiguousarray(np.transpose(t, [1, 0])))
|
|
proj_scale = getattr(chatglm6bModel.layers[i].mlp,
|
|
"quantization_scaling_factor", None)
|
|
set_smoothquant_scale_factors(
|
|
chatglm6bModel.layers[i].mlp.proj, proj_scale, dir_path,
|
|
'model.layers.' + str(i) + '.mlp.dense_4h_to_h.', [1, n_embd],
|
|
quant_per_token_dyn, quant_per_channel)
|
|
elif use_weight_only:
|
|
dst = chatglm6bModel.layers[i].mlp.proj.weight
|
|
processed_torch_weights, torch_weight_scales = torch.ops.fastertransformer.symmetric_quantize_last_axis_of_batched_matrix(
|
|
torch.tensor(t), plugin_weight_only_quant_type)
|
|
# workaround for trt not supporting int8 inputs in plugins currently
|
|
dst.value = processed_torch_weights.view(
|
|
dtype=torch.float32).numpy()
|
|
scales = chatglm6bModel.layers[i].mlp.proj.per_channel_scale
|
|
scales.value = torch_weight_scales.numpy()
|
|
else:
|
|
chatglm6bModel.layers[
|
|
i].mlp.proj.weight.value = np.ascontiguousarray(t)
|
|
if bias:
|
|
chatglm6bModel.layers[i].mlp.proj.bias.value = fromfile(
|
|
dir_path,
|
|
'model.layers.' + str(i) + '.mlp.dense_4h_to_h.bias.bin')
|
|
|
|
if use_int8_kv_cache:
|
|
t = fromfile(
|
|
dir_path, 'model.layers.' + str(i) +
|
|
'.attention.query_key_value.scale_y_quant_orig.bin', [1],
|
|
np.float32)
|
|
chatglm6bModel.layers[
|
|
i].attention.kv_orig_quant_scale.value = 1.0 / t
|
|
chatglm6bModel.layers[i].attention.kv_quant_orig_scale.value = t
|
|
|
|
tok = time.time()
|
|
t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
|
|
tensorrt_llm.logger.info(f'Weights loaded. Total time: {t}')
|
|
|
|
|
|
def load_from_hf_chatglm6b(chatglm6bModel: ChatGLM6BHeadModel,
|
|
hf_chatglm6b,
|
|
rank=0,
|
|
tensor_parallel=1,
|
|
fp16=False):
|
|
tensorrt_llm.logger.info('Loading weights from HF GPT...')
|
|
tik = time.time()
|
|
|
|
valid_lm_head_weight = False
|
|
for k, v in hf_chatglm6b.state_dict().items():
|
|
torch_dtype = torch.float16 if fp16 else torch.float32
|
|
v = v.to(torch_dtype).cpu().numpy()
|
|
if 'wte.weight' in k:
|
|
chatglm6bModel.embedding.vocab_embedding.weight.value = v
|
|
elif 'wpe.weight' in k:
|
|
chatglm6bModel.embedding.position_embedding.weight.value = v
|
|
elif 'ln_f.weight' in k:
|
|
chatglm6bModel.ln_f.weight.value = v
|
|
elif 'ln_f.bias' in k:
|
|
chatglm6bModel.ln_f.bias.value = v
|
|
elif 'lm_head.weight' in k:
|
|
chatglm6bModel.lm_head.weight.value = np.ascontiguousarray(
|
|
split(v, tensor_parallel, rank))
|
|
valid_lm_head_weight = True
|
|
else:
|
|
layer_idx = extract_layer_idx(k)
|
|
if layer_idx is None:
|
|
continue
|
|
idx = int(layer_idx)
|
|
if 'ln_1.weight' in k:
|
|
chatglm6bModel.layers[idx].input_layernorm.weight.value = v
|
|
elif 'ln_1.bias' in k:
|
|
chatglm6bModel.layers[idx].input_layernorm.bias.value = v
|
|
elif 'attn.c_attn.weight' in k:
|
|
# HF-GPT uses Conv1D instead of Linear
|
|
v = v.transpose()
|
|
dst = chatglm6bModel.layers[idx].attention.qkv.weight
|
|
dst.value = np.ascontiguousarray(split(v, tensor_parallel,
|
|
rank))
|
|
elif 'attn.c_attn.bias' in k:
|
|
dst = chatglm6bModel.layers[idx].attention.qkv.bias
|
|
dst.value = np.ascontiguousarray(split(v, tensor_parallel,
|
|
rank))
|
|
elif 'attn.c_proj.weight' in k:
|
|
v = v.transpose()
|
|
dst = chatglm6bModel.layers[idx].attention.dense.weight
|
|
dst.value = np.ascontiguousarray(
|
|
split(v, tensor_parallel, rank, dim=1))
|
|
elif 'attn.c_proj.bias' in k:
|
|
dst = chatglm6bModel.layers[idx].attention.dense.bias
|
|
dst.value = v
|
|
elif 'ln_2.weight' in k:
|
|
dst = chatglm6bModel.layers[idx].post_layernorm.weight
|
|
dst.value = v
|
|
elif 'ln_2.bias' in k:
|
|
dst = chatglm6bModel.layers[idx].post_layernorm.bias
|
|
dst.value = v
|
|
elif 'mlp.c_fc.weight' in k:
|
|
v = v.transpose()
|
|
chatglm6bModel.layers[
|
|
idx].mlp.fc.weight.value = np.ascontiguousarray(
|
|
split(v, tensor_parallel, rank))
|
|
elif 'mlp.c_fc.bias' in k:
|
|
chatglm6bModel.layers[
|
|
idx].mlp.fc.bias.value = np.ascontiguousarray(
|
|
split(v, tensor_parallel, rank))
|
|
elif 'mlp.c_proj.weight' in k:
|
|
v = v.transpose()
|
|
chatglm6bModel.layers[
|
|
idx].mlp.proj.weight.value = np.ascontiguousarray(
|
|
split(v, tensor_parallel, rank, dim=1))
|
|
elif 'mlp.c_proj.bias' in k:
|
|
chatglm6bModel.layers[idx].mlp.proj.bias.value = v
|
|
|
|
if not valid_lm_head_weight:
|
|
# Use wte as lm_head weight to match the load_from_ft implementation.
|
|
lm_head_weight = chatglm6bModel.embedding.vocab_embedding.weight._value
|
|
vocab_size = hf_chatglm6b.config.vocab_size
|
|
if vocab_size % tensor_parallel != 0:
|
|
# padding
|
|
vocab_size_padded = chatglm6bModel.lm_head.out_features * tensor_parallel
|
|
pad_width = vocab_size_padded - vocab_size
|
|
lm_head_weight = np.pad(lm_head_weight, ((0, pad_width), (0, 0)),
|
|
'constant',
|
|
constant_values=0)
|
|
chatglm6bModel.lm_head.weight.value = np.ascontiguousarray(
|
|
split(lm_head_weight, tensor_parallel, rank))
|
|
|
|
tok = time.time()
|
|
t = time.strftime('%H:%M:%S', time.gmtime(tok - tik))
|
|
tensorrt_llm.logger.info(f'Weights loaded. Total time: {t}')
|