Upload folder using huggingface_hub
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- checkpoint-100/adapter_config.json +19 -0
- checkpoint-100/adapter_model.bin +3 -0
- checkpoint-100/latest +1 -0
- checkpoint-100/special_tokens_map.json +17 -0
- checkpoint-100/tokenizer.json +0 -0
- checkpoint-100/tokenizer_config.json +7 -0
- checkpoint-100/training_config.yaml +143 -0
- checkpoint-100/zero_to_fp32.py +483 -0
- checkpoint-200/adapter_config.json +19 -0
- checkpoint-200/adapter_model.bin +3 -0
- checkpoint-200/latest +1 -0
- checkpoint-200/special_tokens_map.json +17 -0
- checkpoint-200/tokenizer.json +0 -0
- checkpoint-200/tokenizer_config.json +7 -0
- checkpoint-200/training_config.yaml +143 -0
- checkpoint-200/zero_to_fp32.py +483 -0
- checkpoint-300/adapter_config.json +19 -0
- checkpoint-300/adapter_model.bin +3 -0
- checkpoint-300/latest +1 -0
- checkpoint-300/special_tokens_map.json +17 -0
- checkpoint-300/tokenizer.json +0 -0
- checkpoint-300/tokenizer_config.json +7 -0
- checkpoint-300/training_config.yaml +143 -0
- checkpoint-300/zero_to_fp32.py +483 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy +3 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json +0 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json +1 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy +3 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json +0 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json +1 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy +3 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json +0 -0
- logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/metrics.json +1 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy +3 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json +0 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/metrics.json +1 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy +3 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json +0 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/metrics.json +1 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy +3 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json +0 -0
- logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/metrics.json +1 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy +3 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json +0 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json +1 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy +3 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json +0 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json +1 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy +3 -0
- reclor_test_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json +0 -0
checkpoint-100/adapter_config.json
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"base_model_name_or_path": "/tmp/falcon-40b",
|
3 |
+
"bias": "none",
|
4 |
+
"fan_in_fan_out": false,
|
5 |
+
"inference_mode": true,
|
6 |
+
"init_lora_weights": true,
|
7 |
+
"lora_alpha": 16,
|
8 |
+
"lora_dropout": 0.05,
|
9 |
+
"modules_to_save": null,
|
10 |
+
"peft_type": "LORA",
|
11 |
+
"r": 64,
|
12 |
+
"target_modules": [
|
13 |
+
"dense",
|
14 |
+
"query_key_value",
|
15 |
+
"dense_h_to_4h",
|
16 |
+
"dense_4h_to_h"
|
17 |
+
],
|
18 |
+
"task_type": "CAUSAL_LM"
|
19 |
+
}
|
checkpoint-100/adapter_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:66c10e793fd79eea15522bf32fe6486cfa94ec0c22dcf34b0ade12e48a5aae5d
|
3 |
+
size 888747453
|
checkpoint-100/latest
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
global_step100
|
checkpoint-100/special_tokens_map.json
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"additional_special_tokens": [
|
3 |
+
">>TITLE<<",
|
4 |
+
">>ABSTRACT<<",
|
5 |
+
">>INTRODUCTION<<",
|
6 |
+
">>SUMMARY<<",
|
7 |
+
">>COMMENT<<",
|
8 |
+
">>ANSWER<<",
|
9 |
+
">>QUESTION<<",
|
10 |
+
">>DOMAIN<<",
|
11 |
+
">>PREFIX<<",
|
12 |
+
">>SUFFIX<<",
|
13 |
+
">>MIDDLE<<"
|
14 |
+
],
|
15 |
+
"eos_token": "<|endoftext|>",
|
16 |
+
"pad_token": "<|endoftext|>"
|
17 |
+
}
|
checkpoint-100/tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
checkpoint-100/tokenizer_config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_prefix_space": false,
|
3 |
+
"clean_up_tokenization_spaces": true,
|
4 |
+
"eos_token": "<|endoftext|>",
|
5 |
+
"model_max_length": 2048,
|
6 |
+
"tokenizer_class": "PreTrainedTokenizerFast"
|
7 |
+
}
|
checkpoint-100/training_config.yaml
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
2 |
+
train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
|
3 |
+
test_file: null
|
4 |
+
model:
|
5 |
+
_target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
|
6 |
+
pad_token_id: 11
|
7 |
+
use_peft: true
|
8 |
+
lora_config:
|
9 |
+
_recursive_: false
|
10 |
+
_target_: models.rw.LoraConfig
|
11 |
+
task_type: CAUSAL_LM
|
12 |
+
inference_mode: false
|
13 |
+
target_modules:
|
14 |
+
_target_: models.rw.find_all_linear_names
|
15 |
+
bits: 4
|
16 |
+
r: 64
|
17 |
+
lora_alpha: 16
|
18 |
+
lora_dropout: 0.05
|
19 |
+
gradient_checkpointing: true
|
20 |
+
torch_dtype:
|
21 |
+
_target_: general_util.training_utils.return_torch_dtype
|
22 |
+
dtype: bfloat16
|
23 |
+
quantization_config:
|
24 |
+
_target_: transformers.utils.quantization_config.BitsAndBytesConfig
|
25 |
+
load_in_4bit: true
|
26 |
+
bnb_4bit_compute_dtype:
|
27 |
+
_target_: general_util.training_utils.return_torch_dtype
|
28 |
+
dtype: bfloat16
|
29 |
+
bnb_4bit_use_double_quant: true
|
30 |
+
bnb_4bit_quant_type: nf4
|
31 |
+
device_map:
|
32 |
+
_target_: models.rw.return_single_device_map
|
33 |
+
load_in_4bit: true
|
34 |
+
max_memory: true
|
35 |
+
read_tensor_train:
|
36 |
+
_target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
|
37 |
+
max_neg_num: 3
|
38 |
+
aug_num: 3
|
39 |
+
max_seq_length: 512
|
40 |
+
shuffle_context: true
|
41 |
+
min_rep_num: 5
|
42 |
+
geo_p: 0.4
|
43 |
+
deduct_ratio: 1.0
|
44 |
+
context_ratio: 1.0
|
45 |
+
noise_sent_ratio: 0.0
|
46 |
+
num_workers: 128
|
47 |
+
extended_vocab: null
|
48 |
+
collator:
|
49 |
+
_target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
|
50 |
+
max_seq_length: 512
|
51 |
+
tokenizer: ${model_name_or_path}
|
52 |
+
causal_lm: true
|
53 |
+
causal_lm_add_eos: false
|
54 |
+
generative_mode: true
|
55 |
+
num_workers: 4
|
56 |
+
prefetch_factor: 2
|
57 |
+
do_preprocess: false
|
58 |
+
model_name_or_path: /tmp/falcon-40b
|
59 |
+
pretrain: null
|
60 |
+
exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
61 |
+
exp_notes: null
|
62 |
+
output_dir: /tmp/${exp_name}
|
63 |
+
do_train: true
|
64 |
+
evaluate_during_training: false
|
65 |
+
do_eval: true
|
66 |
+
eval_sub_path: checkpoint-*
|
67 |
+
per_gpu_train_batch_size: 16
|
68 |
+
per_gpu_eval_batch_size: 8
|
69 |
+
learning_rate: 0.0005
|
70 |
+
gradient_accumulation_steps: 16
|
71 |
+
weight_decay: 0.0
|
72 |
+
adam_epsilon: 1.0e-06
|
73 |
+
adam_betas: (0.9, 0.99)
|
74 |
+
max_grad_norm: 0.3
|
75 |
+
num_train_epochs: 1
|
76 |
+
max_steps: -1
|
77 |
+
warmup_proportion: 0
|
78 |
+
warmup_steps: 50
|
79 |
+
optimizer: null
|
80 |
+
use_nvlamb: null
|
81 |
+
bit_training: null
|
82 |
+
logging_steps: 1
|
83 |
+
save_best: false
|
84 |
+
save_steps: 100
|
85 |
+
eval_steps: -1
|
86 |
+
ddp_eval: true
|
87 |
+
no_cuda: false
|
88 |
+
seed: 42
|
89 |
+
local_rank: 0
|
90 |
+
fp16: true
|
91 |
+
fp16_opt_level: O1
|
92 |
+
fp16_bfloat16: true
|
93 |
+
prediction_cfg:
|
94 |
+
metric: acc
|
95 |
+
measure: 1
|
96 |
+
best_checkpoint: null
|
97 |
+
best_result: null
|
98 |
+
eval_forward_fn:
|
99 |
+
_target_: general_util.evaluator.DiscriminatorForwardFn
|
100 |
+
post_process: null
|
101 |
+
compile: false
|
102 |
+
fairscale_config: null
|
103 |
+
fsdp_config: null
|
104 |
+
ds_cfg:
|
105 |
+
train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
|
106 |
+
gradient_accumulation_steps: ${gradient_accumulation_steps}
|
107 |
+
optimizer:
|
108 |
+
type: AdamW
|
109 |
+
params:
|
110 |
+
lr: ${learning_rate}
|
111 |
+
betas:
|
112 |
+
- 0.9
|
113 |
+
- 0.999
|
114 |
+
eps: ${adam_epsilon}
|
115 |
+
weight_decay: ${weight_decay}
|
116 |
+
scheduler:
|
117 |
+
type: WarmupLR
|
118 |
+
params:
|
119 |
+
warmup_max_lr: ${learning_rate}
|
120 |
+
warmup_num_steps: 50
|
121 |
+
warmup_type: linear
|
122 |
+
gradient_clipping: ${max_grad_norm}
|
123 |
+
bf16:
|
124 |
+
enabled: ${fp16}
|
125 |
+
zero_optimization:
|
126 |
+
stage: 1
|
127 |
+
contiguous_gradients: true
|
128 |
+
overlap_comm: true
|
129 |
+
reduce_scatter: true
|
130 |
+
reduce_bucket_size: 500000000.0
|
131 |
+
allgather_bucket_size: 500000000.0
|
132 |
+
steps_per_print: 1024
|
133 |
+
with_lightseq: false
|
134 |
+
summary_helper:
|
135 |
+
_target_: general_util.tensorboard_helper.WandbWriter
|
136 |
+
batch_index_or_keys: null
|
137 |
+
outputs_index_or_keys:
|
138 |
+
train/mlm_loss: mlm_loss
|
139 |
+
n_gpu: 1
|
140 |
+
device: cuda:0
|
141 |
+
train_batch_size: 16
|
142 |
+
eval_batch_size: null
|
143 |
+
world_size: 16
|
checkpoint-100/zero_to_fp32.py
ADDED
@@ -0,0 +1,483 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
'''Copyright The Microsoft DeepSpeed Team'''
|
3 |
+
|
4 |
+
# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
|
5 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
6 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
7 |
+
# application.
|
8 |
+
#
|
9 |
+
# example: python zero_to_fp32.py . pytorch_model.bin
|
10 |
+
|
11 |
+
import argparse
|
12 |
+
import torch
|
13 |
+
import glob
|
14 |
+
import math
|
15 |
+
import os
|
16 |
+
import re
|
17 |
+
from collections import OrderedDict
|
18 |
+
|
19 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
20 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
21 |
+
from deepspeed.utils import logger
|
22 |
+
from deepspeed.checkpoint.constants import (DS_VERSION,
|
23 |
+
OPTIMIZER_STATE_DICT,
|
24 |
+
SINGLE_PARTITION_OF_FP32_GROUPS,
|
25 |
+
FP32_FLAT_GROUPS,
|
26 |
+
ZERO_STAGE,
|
27 |
+
PARTITION_COUNT,
|
28 |
+
PARAM_SHAPES,
|
29 |
+
BUFFER_NAMES)
|
30 |
+
|
31 |
+
debug = 0
|
32 |
+
|
33 |
+
# load to cpu
|
34 |
+
device = torch.device('cpu')
|
35 |
+
|
36 |
+
|
37 |
+
def atoi(text):
|
38 |
+
return int(text) if text.isdigit() else text
|
39 |
+
|
40 |
+
|
41 |
+
def natural_keys(text):
|
42 |
+
'''
|
43 |
+
alist.sort(key=natural_keys) sorts in human order
|
44 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
45 |
+
(See Toothy's implementation in the comments)
|
46 |
+
'''
|
47 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
48 |
+
|
49 |
+
|
50 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
51 |
+
if not os.path.isdir(checkpoint_dir):
|
52 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
53 |
+
|
54 |
+
# there should be only one file
|
55 |
+
if zero_stage == 2:
|
56 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
57 |
+
elif zero_stage == 3:
|
58 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
59 |
+
|
60 |
+
if not os.path.exists(file):
|
61 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
62 |
+
|
63 |
+
return file
|
64 |
+
|
65 |
+
|
66 |
+
def get_optim_files(checkpoint_dir):
|
67 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
68 |
+
optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
|
69 |
+
"*_optim_states.pt")),
|
70 |
+
key=natural_keys)
|
71 |
+
|
72 |
+
if len(optim_files) == 0:
|
73 |
+
raise FileNotFoundError(
|
74 |
+
f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
|
75 |
+
|
76 |
+
return optim_files
|
77 |
+
|
78 |
+
|
79 |
+
def parse_model_state(file):
|
80 |
+
state_dict = torch.load(file, map_location=device)
|
81 |
+
|
82 |
+
if BUFFER_NAMES not in state_dict:
|
83 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
84 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
85 |
+
if debug:
|
86 |
+
print("Found buffers:", buffer_names)
|
87 |
+
|
88 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
89 |
+
buffers = {
|
90 |
+
k: v.float()
|
91 |
+
for k,
|
92 |
+
v in state_dict["module"].items() if k in buffer_names
|
93 |
+
}
|
94 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
95 |
+
|
96 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
97 |
+
|
98 |
+
return buffers, param_shapes, ds_version
|
99 |
+
|
100 |
+
|
101 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
102 |
+
|
103 |
+
total_files = len(files)
|
104 |
+
state_dicts = []
|
105 |
+
for f in files:
|
106 |
+
state_dicts.append(torch.load(f, map_location=device))
|
107 |
+
|
108 |
+
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
109 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
110 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
111 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
112 |
+
|
113 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
114 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
115 |
+
# use the max of the partition_count to get the dp world_size.
|
116 |
+
|
117 |
+
if type(world_size) is list:
|
118 |
+
world_size = max(world_size)
|
119 |
+
|
120 |
+
if world_size != total_files:
|
121 |
+
raise ValueError(
|
122 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
123 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
124 |
+
)
|
125 |
+
|
126 |
+
# the groups are named differently in each stage
|
127 |
+
if zero_stage == 2:
|
128 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
129 |
+
elif zero_stage == 3:
|
130 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
131 |
+
else:
|
132 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
133 |
+
|
134 |
+
if zero_stage == 2:
|
135 |
+
fp32_flat_groups = [
|
136 |
+
state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
|
137 |
+
for i in range(len(state_dicts))
|
138 |
+
]
|
139 |
+
elif zero_stage == 3:
|
140 |
+
# if there is more than one param group, there will be multiple flattened tensors - one
|
141 |
+
# flattened tensor per group - for simplicity merge them into a single tensor
|
142 |
+
#
|
143 |
+
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
144 |
+
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
145 |
+
|
146 |
+
fp32_flat_groups = [
|
147 |
+
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
|
148 |
+
0) for i in range(len(state_dicts))
|
149 |
+
]
|
150 |
+
|
151 |
+
return zero_stage, world_size, fp32_flat_groups
|
152 |
+
|
153 |
+
|
154 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
155 |
+
"""
|
156 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
157 |
+
|
158 |
+
Args:
|
159 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
160 |
+
|
161 |
+
"""
|
162 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
163 |
+
|
164 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
165 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
166 |
+
print(
|
167 |
+
f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
168 |
+
|
169 |
+
model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
|
170 |
+
buffers, param_shapes, ds_version = parse_model_state(model_file)
|
171 |
+
print(f'Parsing checkpoint created by deepspeed=={ds_version}')
|
172 |
+
|
173 |
+
if zero_stage == 2:
|
174 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
175 |
+
param_shapes,
|
176 |
+
fp32_flat_groups,
|
177 |
+
buffers)
|
178 |
+
elif zero_stage == 3:
|
179 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
180 |
+
param_shapes,
|
181 |
+
fp32_flat_groups,
|
182 |
+
buffers)
|
183 |
+
|
184 |
+
|
185 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
186 |
+
param_shapes,
|
187 |
+
fp32_flat_groups,
|
188 |
+
buffers):
|
189 |
+
|
190 |
+
# Reconstruction protocol:
|
191 |
+
#
|
192 |
+
# XXX: document this
|
193 |
+
|
194 |
+
if debug:
|
195 |
+
for i in range(world_size):
|
196 |
+
for j in range(len(fp32_flat_groups[0])):
|
197 |
+
print(
|
198 |
+
f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
199 |
+
|
200 |
+
# XXX: memory usage doubles here (zero2)
|
201 |
+
num_param_groups = len(fp32_flat_groups[0])
|
202 |
+
merged_single_partition_of_fp32_groups = []
|
203 |
+
for i in range(num_param_groups):
|
204 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
205 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
206 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
207 |
+
avail_numel = sum([
|
208 |
+
full_single_fp32_vector.numel()
|
209 |
+
for full_single_fp32_vector in merged_single_partition_of_fp32_groups
|
210 |
+
])
|
211 |
+
|
212 |
+
if debug:
|
213 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
214 |
+
wanted_numel = sum(
|
215 |
+
[sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
216 |
+
# not asserting if there is a mismatch due to possible padding
|
217 |
+
print(f"Have {avail_numel} numels to process.")
|
218 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
219 |
+
|
220 |
+
state_dict = OrderedDict()
|
221 |
+
|
222 |
+
# buffers
|
223 |
+
state_dict.update(buffers)
|
224 |
+
if debug:
|
225 |
+
print(f"added {len(buffers)} buffers")
|
226 |
+
|
227 |
+
# params
|
228 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
229 |
+
# out-of-core computing solution
|
230 |
+
total_numel = 0
|
231 |
+
total_params = 0
|
232 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
233 |
+
offset = 0
|
234 |
+
avail_numel = full_single_fp32_vector.numel()
|
235 |
+
for name, shape in shapes.items():
|
236 |
+
|
237 |
+
unpartitioned_numel = shape.numel()
|
238 |
+
total_numel += unpartitioned_numel
|
239 |
+
total_params += 1
|
240 |
+
|
241 |
+
if debug:
|
242 |
+
print(
|
243 |
+
f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
|
244 |
+
)
|
245 |
+
state_dict[name] = full_single_fp32_vector.narrow(
|
246 |
+
0,
|
247 |
+
offset,
|
248 |
+
unpartitioned_numel).view(shape)
|
249 |
+
offset += unpartitioned_numel
|
250 |
+
|
251 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
252 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
253 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
254 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
255 |
+
align_to = 2 * world_size
|
256 |
+
|
257 |
+
def zero2_align(x):
|
258 |
+
return align_to * math.ceil(x / align_to)
|
259 |
+
|
260 |
+
if debug:
|
261 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
262 |
+
|
263 |
+
offset = zero2_align(offset)
|
264 |
+
avail_numel = zero2_align(avail_numel)
|
265 |
+
|
266 |
+
if debug:
|
267 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
268 |
+
|
269 |
+
# Sanity check
|
270 |
+
if offset != avail_numel:
|
271 |
+
raise ValueError(
|
272 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
273 |
+
|
274 |
+
print(
|
275 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
276 |
+
)
|
277 |
+
|
278 |
+
return state_dict
|
279 |
+
|
280 |
+
|
281 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
282 |
+
remainder = unpartitioned_numel % world_size
|
283 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
284 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
285 |
+
return partitioned_numel, padding_numel
|
286 |
+
|
287 |
+
|
288 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
289 |
+
param_shapes,
|
290 |
+
fp32_flat_groups,
|
291 |
+
buffers):
|
292 |
+
|
293 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
294 |
+
# param, re-consolidating each param, while dealing with padding if any
|
295 |
+
|
296 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
297 |
+
# merge list of dicts, preserving order
|
298 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
299 |
+
|
300 |
+
if debug:
|
301 |
+
for i in range(world_size):
|
302 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
303 |
+
|
304 |
+
wanted_params = len(param_shapes)
|
305 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
306 |
+
# not asserting if there is a mismatch due to possible padding
|
307 |
+
print(f"Have {avail_numel} numels to process.")
|
308 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
309 |
+
|
310 |
+
state_dict = OrderedDict()
|
311 |
+
|
312 |
+
# buffers
|
313 |
+
state_dict.update(buffers)
|
314 |
+
if debug:
|
315 |
+
print(f"added {len(buffers)} buffers")
|
316 |
+
|
317 |
+
# params
|
318 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
319 |
+
# out-of-core computing solution
|
320 |
+
offset = 0
|
321 |
+
total_numel = 0
|
322 |
+
total_params = 0
|
323 |
+
for name, shape in param_shapes.items():
|
324 |
+
|
325 |
+
unpartitioned_numel = shape.numel()
|
326 |
+
total_numel += unpartitioned_numel
|
327 |
+
total_params += 1
|
328 |
+
|
329 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
330 |
+
|
331 |
+
if debug:
|
332 |
+
print(
|
333 |
+
f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
334 |
+
)
|
335 |
+
|
336 |
+
# XXX: memory usage doubles here
|
337 |
+
state_dict[name] = torch.cat(
|
338 |
+
tuple(fp32_flat_groups[i].narrow(0,
|
339 |
+
offset,
|
340 |
+
partitioned_numel)
|
341 |
+
for i in range(world_size)),
|
342 |
+
0).narrow(0,
|
343 |
+
0,
|
344 |
+
unpartitioned_numel).view(shape)
|
345 |
+
offset += partitioned_numel
|
346 |
+
|
347 |
+
offset *= world_size
|
348 |
+
|
349 |
+
# Sanity check
|
350 |
+
if offset != avail_numel:
|
351 |
+
raise ValueError(
|
352 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
353 |
+
|
354 |
+
print(
|
355 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
356 |
+
)
|
357 |
+
|
358 |
+
return state_dict
|
359 |
+
|
360 |
+
|
361 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
362 |
+
"""
|
363 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
364 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
365 |
+
via a model hub.
|
366 |
+
|
367 |
+
Args:
|
368 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
369 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
370 |
+
|
371 |
+
Returns:
|
372 |
+
- pytorch ``state_dict``
|
373 |
+
|
374 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
375 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
376 |
+
the checkpoint.
|
377 |
+
|
378 |
+
A typical usage might be ::
|
379 |
+
|
380 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
381 |
+
# do the training and checkpoint saving
|
382 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
383 |
+
model = model.cpu() # move to cpu
|
384 |
+
model.load_state_dict(state_dict)
|
385 |
+
# submit to model hub or save the model to share with others
|
386 |
+
|
387 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
388 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
389 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
390 |
+
|
391 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
392 |
+
|
393 |
+
"""
|
394 |
+
if tag is None:
|
395 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
396 |
+
if os.path.isfile(latest_path):
|
397 |
+
with open(latest_path, 'r') as fd:
|
398 |
+
tag = fd.read().strip()
|
399 |
+
else:
|
400 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
401 |
+
|
402 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
403 |
+
|
404 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
405 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
406 |
+
|
407 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
408 |
+
|
409 |
+
|
410 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
411 |
+
"""
|
412 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
413 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
414 |
+
|
415 |
+
Args:
|
416 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
417 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
418 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
419 |
+
"""
|
420 |
+
|
421 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
422 |
+
print(f"Saving fp32 state dict to {output_file}")
|
423 |
+
torch.save(state_dict, output_file)
|
424 |
+
|
425 |
+
|
426 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
427 |
+
"""
|
428 |
+
1. Put the provided model to cpu
|
429 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
430 |
+
3. Load it into the provided model
|
431 |
+
|
432 |
+
Args:
|
433 |
+
- ``model``: the model object to update
|
434 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
435 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
436 |
+
|
437 |
+
Returns:
|
438 |
+
- ``model`: modified model
|
439 |
+
|
440 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
441 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
442 |
+
conveniently placed for you in the checkpoint folder.
|
443 |
+
|
444 |
+
A typical usage might be ::
|
445 |
+
|
446 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
447 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
448 |
+
# submit to model hub or save the model to share with others
|
449 |
+
|
450 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
451 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
452 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
453 |
+
|
454 |
+
"""
|
455 |
+
logger.info(f"Extracting fp32 weights")
|
456 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
457 |
+
|
458 |
+
logger.info(f"Overwriting model with fp32 weights")
|
459 |
+
model = model.cpu()
|
460 |
+
model.load_state_dict(state_dict, strict=False)
|
461 |
+
|
462 |
+
return model
|
463 |
+
|
464 |
+
|
465 |
+
if __name__ == "__main__":
|
466 |
+
|
467 |
+
parser = argparse.ArgumentParser()
|
468 |
+
parser.add_argument(
|
469 |
+
"checkpoint_dir",
|
470 |
+
type=str,
|
471 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
472 |
+
parser.add_argument(
|
473 |
+
"output_file",
|
474 |
+
type=str,
|
475 |
+
help=
|
476 |
+
"path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
|
477 |
+
)
|
478 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
479 |
+
args = parser.parse_args()
|
480 |
+
|
481 |
+
debug = args.debug
|
482 |
+
|
483 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)
|
checkpoint-200/adapter_config.json
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"base_model_name_or_path": "/tmp/falcon-40b",
|
3 |
+
"bias": "none",
|
4 |
+
"fan_in_fan_out": false,
|
5 |
+
"inference_mode": true,
|
6 |
+
"init_lora_weights": true,
|
7 |
+
"lora_alpha": 16,
|
8 |
+
"lora_dropout": 0.05,
|
9 |
+
"modules_to_save": null,
|
10 |
+
"peft_type": "LORA",
|
11 |
+
"r": 64,
|
12 |
+
"target_modules": [
|
13 |
+
"dense",
|
14 |
+
"query_key_value",
|
15 |
+
"dense_h_to_4h",
|
16 |
+
"dense_4h_to_h"
|
17 |
+
],
|
18 |
+
"task_type": "CAUSAL_LM"
|
19 |
+
}
|
checkpoint-200/adapter_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aa1c94a8e5e5bf4b41ac24e627215502d3592c241d511b1106844eda5fef35c1
|
3 |
+
size 888747453
|
checkpoint-200/latest
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
global_step200
|
checkpoint-200/special_tokens_map.json
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"additional_special_tokens": [
|
3 |
+
">>TITLE<<",
|
4 |
+
">>ABSTRACT<<",
|
5 |
+
">>INTRODUCTION<<",
|
6 |
+
">>SUMMARY<<",
|
7 |
+
">>COMMENT<<",
|
8 |
+
">>ANSWER<<",
|
9 |
+
">>QUESTION<<",
|
10 |
+
">>DOMAIN<<",
|
11 |
+
">>PREFIX<<",
|
12 |
+
">>SUFFIX<<",
|
13 |
+
">>MIDDLE<<"
|
14 |
+
],
|
15 |
+
"eos_token": "<|endoftext|>",
|
16 |
+
"pad_token": "<|endoftext|>"
|
17 |
+
}
|
checkpoint-200/tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
checkpoint-200/tokenizer_config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_prefix_space": false,
|
3 |
+
"clean_up_tokenization_spaces": true,
|
4 |
+
"eos_token": "<|endoftext|>",
|
5 |
+
"model_max_length": 2048,
|
6 |
+
"tokenizer_class": "PreTrainedTokenizerFast"
|
7 |
+
}
|
checkpoint-200/training_config.yaml
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
2 |
+
train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
|
3 |
+
test_file: null
|
4 |
+
model:
|
5 |
+
_target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
|
6 |
+
pad_token_id: 11
|
7 |
+
use_peft: true
|
8 |
+
lora_config:
|
9 |
+
_recursive_: false
|
10 |
+
_target_: models.rw.LoraConfig
|
11 |
+
task_type: CAUSAL_LM
|
12 |
+
inference_mode: false
|
13 |
+
target_modules:
|
14 |
+
_target_: models.rw.find_all_linear_names
|
15 |
+
bits: 4
|
16 |
+
r: 64
|
17 |
+
lora_alpha: 16
|
18 |
+
lora_dropout: 0.05
|
19 |
+
gradient_checkpointing: true
|
20 |
+
torch_dtype:
|
21 |
+
_target_: general_util.training_utils.return_torch_dtype
|
22 |
+
dtype: bfloat16
|
23 |
+
quantization_config:
|
24 |
+
_target_: transformers.utils.quantization_config.BitsAndBytesConfig
|
25 |
+
load_in_4bit: true
|
26 |
+
bnb_4bit_compute_dtype:
|
27 |
+
_target_: general_util.training_utils.return_torch_dtype
|
28 |
+
dtype: bfloat16
|
29 |
+
bnb_4bit_use_double_quant: true
|
30 |
+
bnb_4bit_quant_type: nf4
|
31 |
+
device_map:
|
32 |
+
_target_: models.rw.return_single_device_map
|
33 |
+
load_in_4bit: true
|
34 |
+
max_memory: true
|
35 |
+
read_tensor_train:
|
36 |
+
_target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
|
37 |
+
max_neg_num: 3
|
38 |
+
aug_num: 3
|
39 |
+
max_seq_length: 512
|
40 |
+
shuffle_context: true
|
41 |
+
min_rep_num: 5
|
42 |
+
geo_p: 0.4
|
43 |
+
deduct_ratio: 1.0
|
44 |
+
context_ratio: 1.0
|
45 |
+
noise_sent_ratio: 0.0
|
46 |
+
num_workers: 128
|
47 |
+
extended_vocab: null
|
48 |
+
collator:
|
49 |
+
_target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
|
50 |
+
max_seq_length: 512
|
51 |
+
tokenizer: ${model_name_or_path}
|
52 |
+
causal_lm: true
|
53 |
+
causal_lm_add_eos: false
|
54 |
+
generative_mode: true
|
55 |
+
num_workers: 4
|
56 |
+
prefetch_factor: 2
|
57 |
+
do_preprocess: false
|
58 |
+
model_name_or_path: /tmp/falcon-40b
|
59 |
+
pretrain: null
|
60 |
+
exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
61 |
+
exp_notes: null
|
62 |
+
output_dir: /tmp/${exp_name}
|
63 |
+
do_train: true
|
64 |
+
evaluate_during_training: false
|
65 |
+
do_eval: true
|
66 |
+
eval_sub_path: checkpoint-*
|
67 |
+
per_gpu_train_batch_size: 16
|
68 |
+
per_gpu_eval_batch_size: 8
|
69 |
+
learning_rate: 0.0005
|
70 |
+
gradient_accumulation_steps: 16
|
71 |
+
weight_decay: 0.0
|
72 |
+
adam_epsilon: 1.0e-06
|
73 |
+
adam_betas: (0.9, 0.99)
|
74 |
+
max_grad_norm: 0.3
|
75 |
+
num_train_epochs: 1
|
76 |
+
max_steps: -1
|
77 |
+
warmup_proportion: 0
|
78 |
+
warmup_steps: 50
|
79 |
+
optimizer: null
|
80 |
+
use_nvlamb: null
|
81 |
+
bit_training: null
|
82 |
+
logging_steps: 1
|
83 |
+
save_best: false
|
84 |
+
save_steps: 100
|
85 |
+
eval_steps: -1
|
86 |
+
ddp_eval: true
|
87 |
+
no_cuda: false
|
88 |
+
seed: 42
|
89 |
+
local_rank: 0
|
90 |
+
fp16: true
|
91 |
+
fp16_opt_level: O1
|
92 |
+
fp16_bfloat16: true
|
93 |
+
prediction_cfg:
|
94 |
+
metric: acc
|
95 |
+
measure: 1
|
96 |
+
best_checkpoint: null
|
97 |
+
best_result: null
|
98 |
+
eval_forward_fn:
|
99 |
+
_target_: general_util.evaluator.DiscriminatorForwardFn
|
100 |
+
post_process: null
|
101 |
+
compile: false
|
102 |
+
fairscale_config: null
|
103 |
+
fsdp_config: null
|
104 |
+
ds_cfg:
|
105 |
+
train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
|
106 |
+
gradient_accumulation_steps: ${gradient_accumulation_steps}
|
107 |
+
optimizer:
|
108 |
+
type: AdamW
|
109 |
+
params:
|
110 |
+
lr: ${learning_rate}
|
111 |
+
betas:
|
112 |
+
- 0.9
|
113 |
+
- 0.999
|
114 |
+
eps: ${adam_epsilon}
|
115 |
+
weight_decay: ${weight_decay}
|
116 |
+
scheduler:
|
117 |
+
type: WarmupLR
|
118 |
+
params:
|
119 |
+
warmup_max_lr: ${learning_rate}
|
120 |
+
warmup_num_steps: 50
|
121 |
+
warmup_type: linear
|
122 |
+
gradient_clipping: ${max_grad_norm}
|
123 |
+
bf16:
|
124 |
+
enabled: ${fp16}
|
125 |
+
zero_optimization:
|
126 |
+
stage: 1
|
127 |
+
contiguous_gradients: true
|
128 |
+
overlap_comm: true
|
129 |
+
reduce_scatter: true
|
130 |
+
reduce_bucket_size: 500000000.0
|
131 |
+
allgather_bucket_size: 500000000.0
|
132 |
+
steps_per_print: 1024
|
133 |
+
with_lightseq: false
|
134 |
+
summary_helper:
|
135 |
+
_target_: general_util.tensorboard_helper.WandbWriter
|
136 |
+
batch_index_or_keys: null
|
137 |
+
outputs_index_or_keys:
|
138 |
+
train/mlm_loss: mlm_loss
|
139 |
+
n_gpu: 1
|
140 |
+
device: cuda:0
|
141 |
+
train_batch_size: 16
|
142 |
+
eval_batch_size: null
|
143 |
+
world_size: 16
|
checkpoint-200/zero_to_fp32.py
ADDED
@@ -0,0 +1,483 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
'''Copyright The Microsoft DeepSpeed Team'''
|
3 |
+
|
4 |
+
# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
|
5 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
6 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
7 |
+
# application.
|
8 |
+
#
|
9 |
+
# example: python zero_to_fp32.py . pytorch_model.bin
|
10 |
+
|
11 |
+
import argparse
|
12 |
+
import torch
|
13 |
+
import glob
|
14 |
+
import math
|
15 |
+
import os
|
16 |
+
import re
|
17 |
+
from collections import OrderedDict
|
18 |
+
|
19 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
20 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
21 |
+
from deepspeed.utils import logger
|
22 |
+
from deepspeed.checkpoint.constants import (DS_VERSION,
|
23 |
+
OPTIMIZER_STATE_DICT,
|
24 |
+
SINGLE_PARTITION_OF_FP32_GROUPS,
|
25 |
+
FP32_FLAT_GROUPS,
|
26 |
+
ZERO_STAGE,
|
27 |
+
PARTITION_COUNT,
|
28 |
+
PARAM_SHAPES,
|
29 |
+
BUFFER_NAMES)
|
30 |
+
|
31 |
+
debug = 0
|
32 |
+
|
33 |
+
# load to cpu
|
34 |
+
device = torch.device('cpu')
|
35 |
+
|
36 |
+
|
37 |
+
def atoi(text):
|
38 |
+
return int(text) if text.isdigit() else text
|
39 |
+
|
40 |
+
|
41 |
+
def natural_keys(text):
|
42 |
+
'''
|
43 |
+
alist.sort(key=natural_keys) sorts in human order
|
44 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
45 |
+
(See Toothy's implementation in the comments)
|
46 |
+
'''
|
47 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
48 |
+
|
49 |
+
|
50 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
51 |
+
if not os.path.isdir(checkpoint_dir):
|
52 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
53 |
+
|
54 |
+
# there should be only one file
|
55 |
+
if zero_stage == 2:
|
56 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
57 |
+
elif zero_stage == 3:
|
58 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
59 |
+
|
60 |
+
if not os.path.exists(file):
|
61 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
62 |
+
|
63 |
+
return file
|
64 |
+
|
65 |
+
|
66 |
+
def get_optim_files(checkpoint_dir):
|
67 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
68 |
+
optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
|
69 |
+
"*_optim_states.pt")),
|
70 |
+
key=natural_keys)
|
71 |
+
|
72 |
+
if len(optim_files) == 0:
|
73 |
+
raise FileNotFoundError(
|
74 |
+
f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
|
75 |
+
|
76 |
+
return optim_files
|
77 |
+
|
78 |
+
|
79 |
+
def parse_model_state(file):
|
80 |
+
state_dict = torch.load(file, map_location=device)
|
81 |
+
|
82 |
+
if BUFFER_NAMES not in state_dict:
|
83 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
84 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
85 |
+
if debug:
|
86 |
+
print("Found buffers:", buffer_names)
|
87 |
+
|
88 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
89 |
+
buffers = {
|
90 |
+
k: v.float()
|
91 |
+
for k,
|
92 |
+
v in state_dict["module"].items() if k in buffer_names
|
93 |
+
}
|
94 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
95 |
+
|
96 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
97 |
+
|
98 |
+
return buffers, param_shapes, ds_version
|
99 |
+
|
100 |
+
|
101 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
102 |
+
|
103 |
+
total_files = len(files)
|
104 |
+
state_dicts = []
|
105 |
+
for f in files:
|
106 |
+
state_dicts.append(torch.load(f, map_location=device))
|
107 |
+
|
108 |
+
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
109 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
110 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
111 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
112 |
+
|
113 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
114 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
115 |
+
# use the max of the partition_count to get the dp world_size.
|
116 |
+
|
117 |
+
if type(world_size) is list:
|
118 |
+
world_size = max(world_size)
|
119 |
+
|
120 |
+
if world_size != total_files:
|
121 |
+
raise ValueError(
|
122 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
123 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
124 |
+
)
|
125 |
+
|
126 |
+
# the groups are named differently in each stage
|
127 |
+
if zero_stage == 2:
|
128 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
129 |
+
elif zero_stage == 3:
|
130 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
131 |
+
else:
|
132 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
133 |
+
|
134 |
+
if zero_stage == 2:
|
135 |
+
fp32_flat_groups = [
|
136 |
+
state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
|
137 |
+
for i in range(len(state_dicts))
|
138 |
+
]
|
139 |
+
elif zero_stage == 3:
|
140 |
+
# if there is more than one param group, there will be multiple flattened tensors - one
|
141 |
+
# flattened tensor per group - for simplicity merge them into a single tensor
|
142 |
+
#
|
143 |
+
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
144 |
+
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
145 |
+
|
146 |
+
fp32_flat_groups = [
|
147 |
+
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
|
148 |
+
0) for i in range(len(state_dicts))
|
149 |
+
]
|
150 |
+
|
151 |
+
return zero_stage, world_size, fp32_flat_groups
|
152 |
+
|
153 |
+
|
154 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
155 |
+
"""
|
156 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
157 |
+
|
158 |
+
Args:
|
159 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
160 |
+
|
161 |
+
"""
|
162 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
163 |
+
|
164 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
165 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
166 |
+
print(
|
167 |
+
f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
168 |
+
|
169 |
+
model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
|
170 |
+
buffers, param_shapes, ds_version = parse_model_state(model_file)
|
171 |
+
print(f'Parsing checkpoint created by deepspeed=={ds_version}')
|
172 |
+
|
173 |
+
if zero_stage == 2:
|
174 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
175 |
+
param_shapes,
|
176 |
+
fp32_flat_groups,
|
177 |
+
buffers)
|
178 |
+
elif zero_stage == 3:
|
179 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
180 |
+
param_shapes,
|
181 |
+
fp32_flat_groups,
|
182 |
+
buffers)
|
183 |
+
|
184 |
+
|
185 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
186 |
+
param_shapes,
|
187 |
+
fp32_flat_groups,
|
188 |
+
buffers):
|
189 |
+
|
190 |
+
# Reconstruction protocol:
|
191 |
+
#
|
192 |
+
# XXX: document this
|
193 |
+
|
194 |
+
if debug:
|
195 |
+
for i in range(world_size):
|
196 |
+
for j in range(len(fp32_flat_groups[0])):
|
197 |
+
print(
|
198 |
+
f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
199 |
+
|
200 |
+
# XXX: memory usage doubles here (zero2)
|
201 |
+
num_param_groups = len(fp32_flat_groups[0])
|
202 |
+
merged_single_partition_of_fp32_groups = []
|
203 |
+
for i in range(num_param_groups):
|
204 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
205 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
206 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
207 |
+
avail_numel = sum([
|
208 |
+
full_single_fp32_vector.numel()
|
209 |
+
for full_single_fp32_vector in merged_single_partition_of_fp32_groups
|
210 |
+
])
|
211 |
+
|
212 |
+
if debug:
|
213 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
214 |
+
wanted_numel = sum(
|
215 |
+
[sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
216 |
+
# not asserting if there is a mismatch due to possible padding
|
217 |
+
print(f"Have {avail_numel} numels to process.")
|
218 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
219 |
+
|
220 |
+
state_dict = OrderedDict()
|
221 |
+
|
222 |
+
# buffers
|
223 |
+
state_dict.update(buffers)
|
224 |
+
if debug:
|
225 |
+
print(f"added {len(buffers)} buffers")
|
226 |
+
|
227 |
+
# params
|
228 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
229 |
+
# out-of-core computing solution
|
230 |
+
total_numel = 0
|
231 |
+
total_params = 0
|
232 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
233 |
+
offset = 0
|
234 |
+
avail_numel = full_single_fp32_vector.numel()
|
235 |
+
for name, shape in shapes.items():
|
236 |
+
|
237 |
+
unpartitioned_numel = shape.numel()
|
238 |
+
total_numel += unpartitioned_numel
|
239 |
+
total_params += 1
|
240 |
+
|
241 |
+
if debug:
|
242 |
+
print(
|
243 |
+
f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
|
244 |
+
)
|
245 |
+
state_dict[name] = full_single_fp32_vector.narrow(
|
246 |
+
0,
|
247 |
+
offset,
|
248 |
+
unpartitioned_numel).view(shape)
|
249 |
+
offset += unpartitioned_numel
|
250 |
+
|
251 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
252 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
253 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
254 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
255 |
+
align_to = 2 * world_size
|
256 |
+
|
257 |
+
def zero2_align(x):
|
258 |
+
return align_to * math.ceil(x / align_to)
|
259 |
+
|
260 |
+
if debug:
|
261 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
262 |
+
|
263 |
+
offset = zero2_align(offset)
|
264 |
+
avail_numel = zero2_align(avail_numel)
|
265 |
+
|
266 |
+
if debug:
|
267 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
268 |
+
|
269 |
+
# Sanity check
|
270 |
+
if offset != avail_numel:
|
271 |
+
raise ValueError(
|
272 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
273 |
+
|
274 |
+
print(
|
275 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
276 |
+
)
|
277 |
+
|
278 |
+
return state_dict
|
279 |
+
|
280 |
+
|
281 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
282 |
+
remainder = unpartitioned_numel % world_size
|
283 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
284 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
285 |
+
return partitioned_numel, padding_numel
|
286 |
+
|
287 |
+
|
288 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
289 |
+
param_shapes,
|
290 |
+
fp32_flat_groups,
|
291 |
+
buffers):
|
292 |
+
|
293 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
294 |
+
# param, re-consolidating each param, while dealing with padding if any
|
295 |
+
|
296 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
297 |
+
# merge list of dicts, preserving order
|
298 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
299 |
+
|
300 |
+
if debug:
|
301 |
+
for i in range(world_size):
|
302 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
303 |
+
|
304 |
+
wanted_params = len(param_shapes)
|
305 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
306 |
+
# not asserting if there is a mismatch due to possible padding
|
307 |
+
print(f"Have {avail_numel} numels to process.")
|
308 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
309 |
+
|
310 |
+
state_dict = OrderedDict()
|
311 |
+
|
312 |
+
# buffers
|
313 |
+
state_dict.update(buffers)
|
314 |
+
if debug:
|
315 |
+
print(f"added {len(buffers)} buffers")
|
316 |
+
|
317 |
+
# params
|
318 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
319 |
+
# out-of-core computing solution
|
320 |
+
offset = 0
|
321 |
+
total_numel = 0
|
322 |
+
total_params = 0
|
323 |
+
for name, shape in param_shapes.items():
|
324 |
+
|
325 |
+
unpartitioned_numel = shape.numel()
|
326 |
+
total_numel += unpartitioned_numel
|
327 |
+
total_params += 1
|
328 |
+
|
329 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
330 |
+
|
331 |
+
if debug:
|
332 |
+
print(
|
333 |
+
f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
334 |
+
)
|
335 |
+
|
336 |
+
# XXX: memory usage doubles here
|
337 |
+
state_dict[name] = torch.cat(
|
338 |
+
tuple(fp32_flat_groups[i].narrow(0,
|
339 |
+
offset,
|
340 |
+
partitioned_numel)
|
341 |
+
for i in range(world_size)),
|
342 |
+
0).narrow(0,
|
343 |
+
0,
|
344 |
+
unpartitioned_numel).view(shape)
|
345 |
+
offset += partitioned_numel
|
346 |
+
|
347 |
+
offset *= world_size
|
348 |
+
|
349 |
+
# Sanity check
|
350 |
+
if offset != avail_numel:
|
351 |
+
raise ValueError(
|
352 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
353 |
+
|
354 |
+
print(
|
355 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
356 |
+
)
|
357 |
+
|
358 |
+
return state_dict
|
359 |
+
|
360 |
+
|
361 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
362 |
+
"""
|
363 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
364 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
365 |
+
via a model hub.
|
366 |
+
|
367 |
+
Args:
|
368 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
369 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
370 |
+
|
371 |
+
Returns:
|
372 |
+
- pytorch ``state_dict``
|
373 |
+
|
374 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
375 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
376 |
+
the checkpoint.
|
377 |
+
|
378 |
+
A typical usage might be ::
|
379 |
+
|
380 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
381 |
+
# do the training and checkpoint saving
|
382 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
383 |
+
model = model.cpu() # move to cpu
|
384 |
+
model.load_state_dict(state_dict)
|
385 |
+
# submit to model hub or save the model to share with others
|
386 |
+
|
387 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
388 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
389 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
390 |
+
|
391 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
392 |
+
|
393 |
+
"""
|
394 |
+
if tag is None:
|
395 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
396 |
+
if os.path.isfile(latest_path):
|
397 |
+
with open(latest_path, 'r') as fd:
|
398 |
+
tag = fd.read().strip()
|
399 |
+
else:
|
400 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
401 |
+
|
402 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
403 |
+
|
404 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
405 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
406 |
+
|
407 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
408 |
+
|
409 |
+
|
410 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
411 |
+
"""
|
412 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
413 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
414 |
+
|
415 |
+
Args:
|
416 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
417 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
418 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
419 |
+
"""
|
420 |
+
|
421 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
422 |
+
print(f"Saving fp32 state dict to {output_file}")
|
423 |
+
torch.save(state_dict, output_file)
|
424 |
+
|
425 |
+
|
426 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
427 |
+
"""
|
428 |
+
1. Put the provided model to cpu
|
429 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
430 |
+
3. Load it into the provided model
|
431 |
+
|
432 |
+
Args:
|
433 |
+
- ``model``: the model object to update
|
434 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
435 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
436 |
+
|
437 |
+
Returns:
|
438 |
+
- ``model`: modified model
|
439 |
+
|
440 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
441 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
442 |
+
conveniently placed for you in the checkpoint folder.
|
443 |
+
|
444 |
+
A typical usage might be ::
|
445 |
+
|
446 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
447 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
448 |
+
# submit to model hub or save the model to share with others
|
449 |
+
|
450 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
451 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
452 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
453 |
+
|
454 |
+
"""
|
455 |
+
logger.info(f"Extracting fp32 weights")
|
456 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
457 |
+
|
458 |
+
logger.info(f"Overwriting model with fp32 weights")
|
459 |
+
model = model.cpu()
|
460 |
+
model.load_state_dict(state_dict, strict=False)
|
461 |
+
|
462 |
+
return model
|
463 |
+
|
464 |
+
|
465 |
+
if __name__ == "__main__":
|
466 |
+
|
467 |
+
parser = argparse.ArgumentParser()
|
468 |
+
parser.add_argument(
|
469 |
+
"checkpoint_dir",
|
470 |
+
type=str,
|
471 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
472 |
+
parser.add_argument(
|
473 |
+
"output_file",
|
474 |
+
type=str,
|
475 |
+
help=
|
476 |
+
"path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
|
477 |
+
)
|
478 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
479 |
+
args = parser.parse_args()
|
480 |
+
|
481 |
+
debug = args.debug
|
482 |
+
|
483 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)
|
checkpoint-300/adapter_config.json
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"base_model_name_or_path": "/tmp/falcon-40b",
|
3 |
+
"bias": "none",
|
4 |
+
"fan_in_fan_out": false,
|
5 |
+
"inference_mode": true,
|
6 |
+
"init_lora_weights": true,
|
7 |
+
"lora_alpha": 16,
|
8 |
+
"lora_dropout": 0.05,
|
9 |
+
"modules_to_save": null,
|
10 |
+
"peft_type": "LORA",
|
11 |
+
"r": 64,
|
12 |
+
"target_modules": [
|
13 |
+
"dense",
|
14 |
+
"query_key_value",
|
15 |
+
"dense_h_to_4h",
|
16 |
+
"dense_4h_to_h"
|
17 |
+
],
|
18 |
+
"task_type": "CAUSAL_LM"
|
19 |
+
}
|
checkpoint-300/adapter_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:128586f354a558aadb064f6441acdc59655fde2f22a204c4ad3d1765b7b455b8
|
3 |
+
size 888747453
|
checkpoint-300/latest
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
global_step300
|
checkpoint-300/special_tokens_map.json
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"additional_special_tokens": [
|
3 |
+
">>TITLE<<",
|
4 |
+
">>ABSTRACT<<",
|
5 |
+
">>INTRODUCTION<<",
|
6 |
+
">>SUMMARY<<",
|
7 |
+
">>COMMENT<<",
|
8 |
+
">>ANSWER<<",
|
9 |
+
">>QUESTION<<",
|
10 |
+
">>DOMAIN<<",
|
11 |
+
">>PREFIX<<",
|
12 |
+
">>SUFFIX<<",
|
13 |
+
">>MIDDLE<<"
|
14 |
+
],
|
15 |
+
"eos_token": "<|endoftext|>",
|
16 |
+
"pad_token": "<|endoftext|>"
|
17 |
+
}
|
checkpoint-300/tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
checkpoint-300/tokenizer_config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_prefix_space": false,
|
3 |
+
"clean_up_tokenization_spaces": true,
|
4 |
+
"eos_token": "<|endoftext|>",
|
5 |
+
"model_max_length": 2048,
|
6 |
+
"tokenizer_class": "PreTrainedTokenizerFast"
|
7 |
+
}
|
checkpoint-300/training_config.yaml
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
2 |
+
train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
|
3 |
+
test_file: null
|
4 |
+
model:
|
5 |
+
_target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
|
6 |
+
pad_token_id: 11
|
7 |
+
use_peft: true
|
8 |
+
lora_config:
|
9 |
+
_recursive_: false
|
10 |
+
_target_: models.rw.LoraConfig
|
11 |
+
task_type: CAUSAL_LM
|
12 |
+
inference_mode: false
|
13 |
+
target_modules:
|
14 |
+
_target_: models.rw.find_all_linear_names
|
15 |
+
bits: 4
|
16 |
+
r: 64
|
17 |
+
lora_alpha: 16
|
18 |
+
lora_dropout: 0.05
|
19 |
+
gradient_checkpointing: true
|
20 |
+
torch_dtype:
|
21 |
+
_target_: general_util.training_utils.return_torch_dtype
|
22 |
+
dtype: bfloat16
|
23 |
+
quantization_config:
|
24 |
+
_target_: transformers.utils.quantization_config.BitsAndBytesConfig
|
25 |
+
load_in_4bit: true
|
26 |
+
bnb_4bit_compute_dtype:
|
27 |
+
_target_: general_util.training_utils.return_torch_dtype
|
28 |
+
dtype: bfloat16
|
29 |
+
bnb_4bit_use_double_quant: true
|
30 |
+
bnb_4bit_quant_type: nf4
|
31 |
+
device_map:
|
32 |
+
_target_: models.rw.return_single_device_map
|
33 |
+
load_in_4bit: true
|
34 |
+
max_memory: true
|
35 |
+
read_tensor_train:
|
36 |
+
_target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
|
37 |
+
max_neg_num: 3
|
38 |
+
aug_num: 3
|
39 |
+
max_seq_length: 512
|
40 |
+
shuffle_context: true
|
41 |
+
min_rep_num: 5
|
42 |
+
geo_p: 0.4
|
43 |
+
deduct_ratio: 1.0
|
44 |
+
context_ratio: 1.0
|
45 |
+
noise_sent_ratio: 0.0
|
46 |
+
num_workers: 128
|
47 |
+
extended_vocab: null
|
48 |
+
collator:
|
49 |
+
_target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
|
50 |
+
max_seq_length: 512
|
51 |
+
tokenizer: ${model_name_or_path}
|
52 |
+
causal_lm: true
|
53 |
+
causal_lm_add_eos: false
|
54 |
+
generative_mode: true
|
55 |
+
num_workers: 4
|
56 |
+
prefetch_factor: 2
|
57 |
+
do_preprocess: false
|
58 |
+
model_name_or_path: /tmp/falcon-40b
|
59 |
+
pretrain: null
|
60 |
+
exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
|
61 |
+
exp_notes: null
|
62 |
+
output_dir: /tmp/${exp_name}
|
63 |
+
do_train: true
|
64 |
+
evaluate_during_training: false
|
65 |
+
do_eval: true
|
66 |
+
eval_sub_path: checkpoint-*
|
67 |
+
per_gpu_train_batch_size: 16
|
68 |
+
per_gpu_eval_batch_size: 8
|
69 |
+
learning_rate: 0.0005
|
70 |
+
gradient_accumulation_steps: 16
|
71 |
+
weight_decay: 0.0
|
72 |
+
adam_epsilon: 1.0e-06
|
73 |
+
adam_betas: (0.9, 0.99)
|
74 |
+
max_grad_norm: 0.3
|
75 |
+
num_train_epochs: 1
|
76 |
+
max_steps: -1
|
77 |
+
warmup_proportion: 0
|
78 |
+
warmup_steps: 50
|
79 |
+
optimizer: null
|
80 |
+
use_nvlamb: null
|
81 |
+
bit_training: null
|
82 |
+
logging_steps: 1
|
83 |
+
save_best: false
|
84 |
+
save_steps: 100
|
85 |
+
eval_steps: -1
|
86 |
+
ddp_eval: true
|
87 |
+
no_cuda: false
|
88 |
+
seed: 42
|
89 |
+
local_rank: 0
|
90 |
+
fp16: true
|
91 |
+
fp16_opt_level: O1
|
92 |
+
fp16_bfloat16: true
|
93 |
+
prediction_cfg:
|
94 |
+
metric: acc
|
95 |
+
measure: 1
|
96 |
+
best_checkpoint: null
|
97 |
+
best_result: null
|
98 |
+
eval_forward_fn:
|
99 |
+
_target_: general_util.evaluator.DiscriminatorForwardFn
|
100 |
+
post_process: null
|
101 |
+
compile: false
|
102 |
+
fairscale_config: null
|
103 |
+
fsdp_config: null
|
104 |
+
ds_cfg:
|
105 |
+
train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
|
106 |
+
gradient_accumulation_steps: ${gradient_accumulation_steps}
|
107 |
+
optimizer:
|
108 |
+
type: AdamW
|
109 |
+
params:
|
110 |
+
lr: ${learning_rate}
|
111 |
+
betas:
|
112 |
+
- 0.9
|
113 |
+
- 0.999
|
114 |
+
eps: ${adam_epsilon}
|
115 |
+
weight_decay: ${weight_decay}
|
116 |
+
scheduler:
|
117 |
+
type: WarmupLR
|
118 |
+
params:
|
119 |
+
warmup_max_lr: ${learning_rate}
|
120 |
+
warmup_num_steps: 50
|
121 |
+
warmup_type: linear
|
122 |
+
gradient_clipping: ${max_grad_norm}
|
123 |
+
bf16:
|
124 |
+
enabled: ${fp16}
|
125 |
+
zero_optimization:
|
126 |
+
stage: 1
|
127 |
+
contiguous_gradients: true
|
128 |
+
overlap_comm: true
|
129 |
+
reduce_scatter: true
|
130 |
+
reduce_bucket_size: 500000000.0
|
131 |
+
allgather_bucket_size: 500000000.0
|
132 |
+
steps_per_print: 1024
|
133 |
+
with_lightseq: false
|
134 |
+
summary_helper:
|
135 |
+
_target_: general_util.tensorboard_helper.WandbWriter
|
136 |
+
batch_index_or_keys: null
|
137 |
+
outputs_index_or_keys:
|
138 |
+
train/mlm_loss: mlm_loss
|
139 |
+
n_gpu: 1
|
140 |
+
device: cuda:0
|
141 |
+
train_batch_size: 16
|
142 |
+
eval_batch_size: null
|
143 |
+
world_size: 16
|
checkpoint-300/zero_to_fp32.py
ADDED
@@ -0,0 +1,483 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
'''Copyright The Microsoft DeepSpeed Team'''
|
3 |
+
|
4 |
+
# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
|
5 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
6 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
7 |
+
# application.
|
8 |
+
#
|
9 |
+
# example: python zero_to_fp32.py . pytorch_model.bin
|
10 |
+
|
11 |
+
import argparse
|
12 |
+
import torch
|
13 |
+
import glob
|
14 |
+
import math
|
15 |
+
import os
|
16 |
+
import re
|
17 |
+
from collections import OrderedDict
|
18 |
+
|
19 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
20 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
21 |
+
from deepspeed.utils import logger
|
22 |
+
from deepspeed.checkpoint.constants import (DS_VERSION,
|
23 |
+
OPTIMIZER_STATE_DICT,
|
24 |
+
SINGLE_PARTITION_OF_FP32_GROUPS,
|
25 |
+
FP32_FLAT_GROUPS,
|
26 |
+
ZERO_STAGE,
|
27 |
+
PARTITION_COUNT,
|
28 |
+
PARAM_SHAPES,
|
29 |
+
BUFFER_NAMES)
|
30 |
+
|
31 |
+
debug = 0
|
32 |
+
|
33 |
+
# load to cpu
|
34 |
+
device = torch.device('cpu')
|
35 |
+
|
36 |
+
|
37 |
+
def atoi(text):
|
38 |
+
return int(text) if text.isdigit() else text
|
39 |
+
|
40 |
+
|
41 |
+
def natural_keys(text):
|
42 |
+
'''
|
43 |
+
alist.sort(key=natural_keys) sorts in human order
|
44 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
45 |
+
(See Toothy's implementation in the comments)
|
46 |
+
'''
|
47 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
48 |
+
|
49 |
+
|
50 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
51 |
+
if not os.path.isdir(checkpoint_dir):
|
52 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
53 |
+
|
54 |
+
# there should be only one file
|
55 |
+
if zero_stage == 2:
|
56 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
57 |
+
elif zero_stage == 3:
|
58 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
59 |
+
|
60 |
+
if not os.path.exists(file):
|
61 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
62 |
+
|
63 |
+
return file
|
64 |
+
|
65 |
+
|
66 |
+
def get_optim_files(checkpoint_dir):
|
67 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
68 |
+
optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
|
69 |
+
"*_optim_states.pt")),
|
70 |
+
key=natural_keys)
|
71 |
+
|
72 |
+
if len(optim_files) == 0:
|
73 |
+
raise FileNotFoundError(
|
74 |
+
f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
|
75 |
+
|
76 |
+
return optim_files
|
77 |
+
|
78 |
+
|
79 |
+
def parse_model_state(file):
|
80 |
+
state_dict = torch.load(file, map_location=device)
|
81 |
+
|
82 |
+
if BUFFER_NAMES not in state_dict:
|
83 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
84 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
85 |
+
if debug:
|
86 |
+
print("Found buffers:", buffer_names)
|
87 |
+
|
88 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
89 |
+
buffers = {
|
90 |
+
k: v.float()
|
91 |
+
for k,
|
92 |
+
v in state_dict["module"].items() if k in buffer_names
|
93 |
+
}
|
94 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
95 |
+
|
96 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
97 |
+
|
98 |
+
return buffers, param_shapes, ds_version
|
99 |
+
|
100 |
+
|
101 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
102 |
+
|
103 |
+
total_files = len(files)
|
104 |
+
state_dicts = []
|
105 |
+
for f in files:
|
106 |
+
state_dicts.append(torch.load(f, map_location=device))
|
107 |
+
|
108 |
+
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
109 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
110 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
111 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
112 |
+
|
113 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
114 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
115 |
+
# use the max of the partition_count to get the dp world_size.
|
116 |
+
|
117 |
+
if type(world_size) is list:
|
118 |
+
world_size = max(world_size)
|
119 |
+
|
120 |
+
if world_size != total_files:
|
121 |
+
raise ValueError(
|
122 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
123 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
124 |
+
)
|
125 |
+
|
126 |
+
# the groups are named differently in each stage
|
127 |
+
if zero_stage == 2:
|
128 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
129 |
+
elif zero_stage == 3:
|
130 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
131 |
+
else:
|
132 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
133 |
+
|
134 |
+
if zero_stage == 2:
|
135 |
+
fp32_flat_groups = [
|
136 |
+
state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
|
137 |
+
for i in range(len(state_dicts))
|
138 |
+
]
|
139 |
+
elif zero_stage == 3:
|
140 |
+
# if there is more than one param group, there will be multiple flattened tensors - one
|
141 |
+
# flattened tensor per group - for simplicity merge them into a single tensor
|
142 |
+
#
|
143 |
+
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
144 |
+
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
145 |
+
|
146 |
+
fp32_flat_groups = [
|
147 |
+
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
|
148 |
+
0) for i in range(len(state_dicts))
|
149 |
+
]
|
150 |
+
|
151 |
+
return zero_stage, world_size, fp32_flat_groups
|
152 |
+
|
153 |
+
|
154 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
155 |
+
"""
|
156 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
157 |
+
|
158 |
+
Args:
|
159 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
160 |
+
|
161 |
+
"""
|
162 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
163 |
+
|
164 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
165 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
166 |
+
print(
|
167 |
+
f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
168 |
+
|
169 |
+
model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
|
170 |
+
buffers, param_shapes, ds_version = parse_model_state(model_file)
|
171 |
+
print(f'Parsing checkpoint created by deepspeed=={ds_version}')
|
172 |
+
|
173 |
+
if zero_stage == 2:
|
174 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
175 |
+
param_shapes,
|
176 |
+
fp32_flat_groups,
|
177 |
+
buffers)
|
178 |
+
elif zero_stage == 3:
|
179 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
180 |
+
param_shapes,
|
181 |
+
fp32_flat_groups,
|
182 |
+
buffers)
|
183 |
+
|
184 |
+
|
185 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
|
186 |
+
param_shapes,
|
187 |
+
fp32_flat_groups,
|
188 |
+
buffers):
|
189 |
+
|
190 |
+
# Reconstruction protocol:
|
191 |
+
#
|
192 |
+
# XXX: document this
|
193 |
+
|
194 |
+
if debug:
|
195 |
+
for i in range(world_size):
|
196 |
+
for j in range(len(fp32_flat_groups[0])):
|
197 |
+
print(
|
198 |
+
f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
199 |
+
|
200 |
+
# XXX: memory usage doubles here (zero2)
|
201 |
+
num_param_groups = len(fp32_flat_groups[0])
|
202 |
+
merged_single_partition_of_fp32_groups = []
|
203 |
+
for i in range(num_param_groups):
|
204 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
205 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
206 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
207 |
+
avail_numel = sum([
|
208 |
+
full_single_fp32_vector.numel()
|
209 |
+
for full_single_fp32_vector in merged_single_partition_of_fp32_groups
|
210 |
+
])
|
211 |
+
|
212 |
+
if debug:
|
213 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
214 |
+
wanted_numel = sum(
|
215 |
+
[sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
216 |
+
# not asserting if there is a mismatch due to possible padding
|
217 |
+
print(f"Have {avail_numel} numels to process.")
|
218 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
219 |
+
|
220 |
+
state_dict = OrderedDict()
|
221 |
+
|
222 |
+
# buffers
|
223 |
+
state_dict.update(buffers)
|
224 |
+
if debug:
|
225 |
+
print(f"added {len(buffers)} buffers")
|
226 |
+
|
227 |
+
# params
|
228 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
229 |
+
# out-of-core computing solution
|
230 |
+
total_numel = 0
|
231 |
+
total_params = 0
|
232 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
233 |
+
offset = 0
|
234 |
+
avail_numel = full_single_fp32_vector.numel()
|
235 |
+
for name, shape in shapes.items():
|
236 |
+
|
237 |
+
unpartitioned_numel = shape.numel()
|
238 |
+
total_numel += unpartitioned_numel
|
239 |
+
total_params += 1
|
240 |
+
|
241 |
+
if debug:
|
242 |
+
print(
|
243 |
+
f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
|
244 |
+
)
|
245 |
+
state_dict[name] = full_single_fp32_vector.narrow(
|
246 |
+
0,
|
247 |
+
offset,
|
248 |
+
unpartitioned_numel).view(shape)
|
249 |
+
offset += unpartitioned_numel
|
250 |
+
|
251 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
252 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
253 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
254 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
255 |
+
align_to = 2 * world_size
|
256 |
+
|
257 |
+
def zero2_align(x):
|
258 |
+
return align_to * math.ceil(x / align_to)
|
259 |
+
|
260 |
+
if debug:
|
261 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
262 |
+
|
263 |
+
offset = zero2_align(offset)
|
264 |
+
avail_numel = zero2_align(avail_numel)
|
265 |
+
|
266 |
+
if debug:
|
267 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
268 |
+
|
269 |
+
# Sanity check
|
270 |
+
if offset != avail_numel:
|
271 |
+
raise ValueError(
|
272 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
273 |
+
|
274 |
+
print(
|
275 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
276 |
+
)
|
277 |
+
|
278 |
+
return state_dict
|
279 |
+
|
280 |
+
|
281 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
282 |
+
remainder = unpartitioned_numel % world_size
|
283 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
284 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
285 |
+
return partitioned_numel, padding_numel
|
286 |
+
|
287 |
+
|
288 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
|
289 |
+
param_shapes,
|
290 |
+
fp32_flat_groups,
|
291 |
+
buffers):
|
292 |
+
|
293 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
294 |
+
# param, re-consolidating each param, while dealing with padding if any
|
295 |
+
|
296 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
297 |
+
# merge list of dicts, preserving order
|
298 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
299 |
+
|
300 |
+
if debug:
|
301 |
+
for i in range(world_size):
|
302 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
303 |
+
|
304 |
+
wanted_params = len(param_shapes)
|
305 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
306 |
+
# not asserting if there is a mismatch due to possible padding
|
307 |
+
print(f"Have {avail_numel} numels to process.")
|
308 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
309 |
+
|
310 |
+
state_dict = OrderedDict()
|
311 |
+
|
312 |
+
# buffers
|
313 |
+
state_dict.update(buffers)
|
314 |
+
if debug:
|
315 |
+
print(f"added {len(buffers)} buffers")
|
316 |
+
|
317 |
+
# params
|
318 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
319 |
+
# out-of-core computing solution
|
320 |
+
offset = 0
|
321 |
+
total_numel = 0
|
322 |
+
total_params = 0
|
323 |
+
for name, shape in param_shapes.items():
|
324 |
+
|
325 |
+
unpartitioned_numel = shape.numel()
|
326 |
+
total_numel += unpartitioned_numel
|
327 |
+
total_params += 1
|
328 |
+
|
329 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
330 |
+
|
331 |
+
if debug:
|
332 |
+
print(
|
333 |
+
f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
334 |
+
)
|
335 |
+
|
336 |
+
# XXX: memory usage doubles here
|
337 |
+
state_dict[name] = torch.cat(
|
338 |
+
tuple(fp32_flat_groups[i].narrow(0,
|
339 |
+
offset,
|
340 |
+
partitioned_numel)
|
341 |
+
for i in range(world_size)),
|
342 |
+
0).narrow(0,
|
343 |
+
0,
|
344 |
+
unpartitioned_numel).view(shape)
|
345 |
+
offset += partitioned_numel
|
346 |
+
|
347 |
+
offset *= world_size
|
348 |
+
|
349 |
+
# Sanity check
|
350 |
+
if offset != avail_numel:
|
351 |
+
raise ValueError(
|
352 |
+
f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
353 |
+
|
354 |
+
print(
|
355 |
+
f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
|
356 |
+
)
|
357 |
+
|
358 |
+
return state_dict
|
359 |
+
|
360 |
+
|
361 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
362 |
+
"""
|
363 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
364 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
365 |
+
via a model hub.
|
366 |
+
|
367 |
+
Args:
|
368 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
369 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
370 |
+
|
371 |
+
Returns:
|
372 |
+
- pytorch ``state_dict``
|
373 |
+
|
374 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
375 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
376 |
+
the checkpoint.
|
377 |
+
|
378 |
+
A typical usage might be ::
|
379 |
+
|
380 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
381 |
+
# do the training and checkpoint saving
|
382 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
383 |
+
model = model.cpu() # move to cpu
|
384 |
+
model.load_state_dict(state_dict)
|
385 |
+
# submit to model hub or save the model to share with others
|
386 |
+
|
387 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
388 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
389 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
390 |
+
|
391 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
392 |
+
|
393 |
+
"""
|
394 |
+
if tag is None:
|
395 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
396 |
+
if os.path.isfile(latest_path):
|
397 |
+
with open(latest_path, 'r') as fd:
|
398 |
+
tag = fd.read().strip()
|
399 |
+
else:
|
400 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
401 |
+
|
402 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
403 |
+
|
404 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
405 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
406 |
+
|
407 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
408 |
+
|
409 |
+
|
410 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
411 |
+
"""
|
412 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
413 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
414 |
+
|
415 |
+
Args:
|
416 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
417 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
418 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
419 |
+
"""
|
420 |
+
|
421 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
422 |
+
print(f"Saving fp32 state dict to {output_file}")
|
423 |
+
torch.save(state_dict, output_file)
|
424 |
+
|
425 |
+
|
426 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
427 |
+
"""
|
428 |
+
1. Put the provided model to cpu
|
429 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
430 |
+
3. Load it into the provided model
|
431 |
+
|
432 |
+
Args:
|
433 |
+
- ``model``: the model object to update
|
434 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
435 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
436 |
+
|
437 |
+
Returns:
|
438 |
+
- ``model`: modified model
|
439 |
+
|
440 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
441 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
442 |
+
conveniently placed for you in the checkpoint folder.
|
443 |
+
|
444 |
+
A typical usage might be ::
|
445 |
+
|
446 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
447 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
448 |
+
# submit to model hub or save the model to share with others
|
449 |
+
|
450 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
451 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
452 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
453 |
+
|
454 |
+
"""
|
455 |
+
logger.info(f"Extracting fp32 weights")
|
456 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
457 |
+
|
458 |
+
logger.info(f"Overwriting model with fp32 weights")
|
459 |
+
model = model.cpu()
|
460 |
+
model.load_state_dict(state_dict, strict=False)
|
461 |
+
|
462 |
+
return model
|
463 |
+
|
464 |
+
|
465 |
+
if __name__ == "__main__":
|
466 |
+
|
467 |
+
parser = argparse.ArgumentParser()
|
468 |
+
parser.add_argument(
|
469 |
+
"checkpoint_dir",
|
470 |
+
type=str,
|
471 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
472 |
+
parser.add_argument(
|
473 |
+
"output_file",
|
474 |
+
type=str,
|
475 |
+
help=
|
476 |
+
"path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
|
477 |
+
)
|
478 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
479 |
+
args = parser.parse_args()
|
480 |
+
|
481 |
+
debug = args.debug
|
482 |
+
|
483 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:694b78b68739e7ae2f39c441f3ee284269b7e67cd2a1580373889c7c2ab6cf82
|
3 |
+
size 12704
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3861}
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:337e73e47317aa6d269c6d88cd5637dbb2ae32505de6cc1fcf56bcfa81a72598
|
3 |
+
size 12704
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3696}
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5fcd86d2b06e468632ac7cb064f9f47febb6ff75f798b7675581b3e4f2c751e4
|
3 |
+
size 12704
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_test_mc_0shot.v1.0/test-checkpoint-300/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3715}
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f640f89a04bcaa15d9ee7318aedd944d973dd0a6ff8bc530c6aecdbec53ab222
|
3 |
+
size 12680
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-100/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3926}
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:99d2f4dc6e59ca6038743efb507a2bc1451c919e1e5fbc22785d67fbc4b973fc
|
3 |
+
size 12680
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-200/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3881}
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:57ed8b5971038f010c9ada982042cfb43409adb836ddb8f015471686dff0ec64
|
3 |
+
size 12680
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
logiqav2_val_mc_0shot.v1.0/test-checkpoint-300/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.3862}
|
reclor_test_mc_0shot.v1.0/test-checkpoint-100/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6d624748d81e273100278a09e8d72d41a3b5aea0a6e2c6c0986657661cc7818e
|
3 |
+
size 8128
|
reclor_test_mc_0shot.v1.0/test-checkpoint-100/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
reclor_test_mc_0shot.v1.0/test-checkpoint-100/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.0}
|
reclor_test_mc_0shot.v1.0/test-checkpoint-200/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3f85fb5eb848073e0793f0c2f25d362d609868d2fca08041b99aa5b923df7c40
|
3 |
+
size 8128
|
reclor_test_mc_0shot.v1.0/test-checkpoint-200/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
reclor_test_mc_0shot.v1.0/test-checkpoint-200/metrics.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"acc": 0.0}
|
reclor_test_mc_0shot.v1.0/test-checkpoint-300/decode_results.npy
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:84cd970cf400a19de4ffd32874c7b6068c509713eb1f3656d9767c6cae4e9f2b
|
3 |
+
size 8128
|
reclor_test_mc_0shot.v1.0/test-checkpoint-300/eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|