初始化项目，由ModelHub XC社区提供模型
Model: jondurbin/spicyboros-7b-2.2-checkpoints Source: Original Platform
This commit is contained in:
ModelHub XC
26
checkpoints/checkpoint-100/config.json
Normal file
26
checkpoints/checkpoint-100/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-100/generation_config.json
Normal file
4
checkpoints/checkpoint-100/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-100/latest
Normal file
1
checkpoints/checkpoint-100/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step100
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:5e3e1e87153d5c5f361cf24da62c427cea619e82a7a5b0151468f05eeb53be4c
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:a2f79a892136c06a9d22e18ac4c031310ae7d70e523c32778eca0e0d44a4f18b
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-100/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-100/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-100/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-100/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-100/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-100/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-100/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-100/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
699
checkpoints/checkpoint-100/trainer_state.json
Normal file
699
checkpoints/checkpoint-100/trainer_state.json
Normal file
@@ -0,0 +1,699 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 2.527646129541864,
|
||||
"eval_steps": 10,
|
||||
"global_step": 100,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.03,
|
||||
"learning_rate": 2.2e-05,
|
||||
"loss": 0.9856,
|
||||
"step": 1
|
||||
},
|
||||
{
|
||||
"epoch": 0.05,
|
||||
"learning_rate": 2.1998557720321954e-05,
|
||||
"loss": 0.8657,
|
||||
"step": 2
|
||||
},
|
||||
{
|
||||
"epoch": 0.08,
|
||||
"learning_rate": 2.1994231259500664e-05,
|
||||
"loss": 0.7599,
|
||||
"step": 3
|
||||
},
|
||||
{
|
||||
"epoch": 0.1,
|
||||
"learning_rate": 2.1987021752075498e-05,
|
||||
"loss": 0.6924,
|
||||
"step": 4
|
||||
},
|
||||
{
|
||||
"epoch": 0.13,
|
||||
"learning_rate": 2.1976931088614826e-05,
|
||||
"loss": 0.818,
|
||||
"step": 5
|
||||
},
|
||||
{
|
||||
"epoch": 0.15,
|
||||
"learning_rate": 2.1963961915220263e-05,
|
||||
"loss": 1.2993,
|
||||
"step": 6
|
||||
},
|
||||
{
|
||||
"epoch": 0.18,
|
||||
"learning_rate": 2.1948117632832752e-05,
|
||||
"loss": 1.4107,
|
||||
"step": 7
|
||||
},
|
||||
{
|
||||
"epoch": 0.2,
|
||||
"learning_rate": 2.1929402396340756e-05,
|
||||
"loss": 0.8539,
|
||||
"step": 8
|
||||
},
|
||||
{
|
||||
"epoch": 0.23,
|
||||
"learning_rate": 2.1907821113490676e-05,
|
||||
"loss": 0.7822,
|
||||
"step": 9
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"learning_rate": 2.1883379443599917e-05,
|
||||
"loss": 0.6498,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"eval_loss": 0.777102530002594,
|
||||
"eval_runtime": 26.948,
|
||||
"eval_samples_per_second": 30.726,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.28,
|
||||
"learning_rate": 2.1856083796072795e-05,
|
||||
"loss": 0.7439,
|
||||
"step": 11
|
||||
},
|
||||
{
|
||||
"epoch": 0.3,
|
||||
"learning_rate": 2.182594132871982e-05,
|
||||
"loss": 1.0097,
|
||||
"step": 12
|
||||
},
|
||||
{
|
||||
"epoch": 0.33,
|
||||
"learning_rate": 2.179295994588063e-05,
|
||||
"loss": 1.3189,
|
||||
"step": 13
|
||||
},
|
||||
{
|
||||
"epoch": 0.35,
|
||||
"learning_rate": 2.175714829635128e-05,
|
||||
"loss": 0.7509,
|
||||
"step": 14
|
||||
},
|
||||
{
|
||||
"epoch": 0.38,
|
||||
"learning_rate": 2.1718515771116186e-05,
|
||||
"loss": 0.7273,
|
||||
"step": 15
|
||||
},
|
||||
{
|
||||
"epoch": 0.4,
|
||||
"learning_rate": 2.167707250088555e-05,
|
||||
"loss": 0.6242,
|
||||
"step": 16
|
||||
},
|
||||
{
|
||||
"epoch": 0.43,
|
||||
"learning_rate": 2.1632829353438724e-05,
|
||||
"loss": 0.6135,
|
||||
"step": 17
|
||||
},
|
||||
{
|
||||
"epoch": 0.45,
|
||||
"learning_rate": 2.1585797930774335e-05,
|
||||
"loss": 0.9365,
|
||||
"step": 18
|
||||
},
|
||||
{
|
||||
"epoch": 0.48,
|
||||
"learning_rate": 2.1535990566067858e-05,
|
||||
"loss": 1.4216,
|
||||
"step": 19
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"learning_rate": 2.148342032043747e-05,
|
||||
"loss": 0.7369,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"eval_loss": 0.7404242157936096,
|
||||
"eval_runtime": 26.98,
|
||||
"eval_samples_per_second": 30.689,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.53,
|
||||
"learning_rate": 2.1428100979518976e-05,
|
||||
"loss": 0.7605,
|
||||
"step": 21
|
||||
},
|
||||
{
|
||||
"epoch": 0.56,
|
||||
"learning_rate": 2.1370047049850775e-05,
|
||||
"loss": 0.6095,
|
||||
"step": 22
|
||||
},
|
||||
{
|
||||
"epoch": 0.58,
|
||||
"learning_rate": 2.130927375506978e-05,
|
||||
"loss": 0.5955,
|
||||
"step": 23
|
||||
},
|
||||
{
|
||||
"epoch": 0.61,
|
||||
"learning_rate": 2.1245797031919265e-05,
|
||||
"loss": 0.851,
|
||||
"step": 24
|
||||
},
|
||||
{
|
||||
"epoch": 0.63,
|
||||
"learning_rate": 2.1179633526069748e-05,
|
||||
"loss": 1.324,
|
||||
"step": 25
|
||||
},
|
||||
{
|
||||
"epoch": 0.66,
|
||||
"learning_rate": 2.1110800587753934e-05,
|
||||
"loss": 0.9368,
|
||||
"step": 26
|
||||
},
|
||||
{
|
||||
"epoch": 0.68,
|
||||
"learning_rate": 2.103931626721694e-05,
|
||||
"loss": 0.7307,
|
||||
"step": 27
|
||||
},
|
||||
{
|
||||
"epoch": 0.71,
|
||||
"learning_rate": 2.0965199309982913e-05,
|
||||
"loss": 0.6489,
|
||||
"step": 28
|
||||
},
|
||||
{
|
||||
"epoch": 0.73,
|
||||
"learning_rate": 2.0888469151939347e-05,
|
||||
"loss": 0.536,
|
||||
"step": 29
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"learning_rate": 2.0809145914240356e-05,
|
||||
"loss": 0.732,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"eval_loss": 0.7266919612884521,
|
||||
"eval_runtime": 26.9346,
|
||||
"eval_samples_per_second": 30.741,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.78,
|
||||
"learning_rate": 2.072725039803026e-05,
|
||||
"loss": 1.1248,
|
||||
"step": 31
|
||||
},
|
||||
{
|
||||
"epoch": 0.81,
|
||||
"learning_rate": 2.0642804078988827e-05,
|
||||
"loss": 1.0917,
|
||||
"step": 32
|
||||
},
|
||||
{
|
||||
"epoch": 0.83,
|
||||
"learning_rate": 2.055582910169966e-05,
|
||||
"loss": 0.735,
|
||||
"step": 33
|
||||
},
|
||||
{
|
||||
"epoch": 0.86,
|
||||
"learning_rate": 2.0466348273843176e-05,
|
||||
"loss": 0.6749,
|
||||
"step": 34
|
||||
},
|
||||
{
|
||||
"epoch": 0.88,
|
||||
"learning_rate": 2.0374385060215662e-05,
|
||||
"loss": 0.5519,
|
||||
"step": 35
|
||||
},
|
||||
{
|
||||
"epoch": 0.91,
|
||||
"learning_rate": 2.027996357657605e-05,
|
||||
"loss": 0.6593,
|
||||
"step": 36
|
||||
},
|
||||
{
|
||||
"epoch": 0.94,
|
||||
"learning_rate": 2.018310858332198e-05,
|
||||
"loss": 0.959,
|
||||
"step": 37
|
||||
},
|
||||
{
|
||||
"epoch": 0.96,
|
||||
"learning_rate": 2.008384547899682e-05,
|
||||
"loss": 1.168,
|
||||
"step": 38
|
||||
},
|
||||
{
|
||||
"epoch": 0.99,
|
||||
"learning_rate": 1.9982200293629318e-05,
|
||||
"loss": 0.6186,
|
||||
"step": 39
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"learning_rate": 1.9878199681907702e-05,
|
||||
"loss": 1.1207,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"eval_loss": 0.7203702926635742,
|
||||
"eval_runtime": 26.9926,
|
||||
"eval_samples_per_second": 30.675,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.04,
|
||||
"learning_rate": 1.977187091618993e-05,
|
||||
"loss": 0.7095,
|
||||
"step": 41
|
||||
},
|
||||
{
|
||||
"epoch": 1.06,
|
||||
"learning_rate": 1.9663241879351998e-05,
|
||||
"loss": 0.6781,
|
||||
"step": 42
|
||||
},
|
||||
{
|
||||
"epoch": 1.09,
|
||||
"learning_rate": 1.9552341057476144e-05,
|
||||
"loss": 0.5649,
|
||||
"step": 43
|
||||
},
|
||||
{
|
||||
"epoch": 1.11,
|
||||
"learning_rate": 1.9439197532380845e-05,
|
||||
"loss": 0.5466,
|
||||
"step": 44
|
||||
},
|
||||
{
|
||||
"epoch": 1.14,
|
||||
"learning_rate": 1.9323840973994647e-05,
|
||||
"loss": 0.8721,
|
||||
"step": 45
|
||||
},
|
||||
{
|
||||
"epoch": 1.16,
|
||||
"learning_rate": 1.9206301632575707e-05,
|
||||
"loss": 1.2152,
|
||||
"step": 46
|
||||
},
|
||||
{
|
||||
"epoch": 1.19,
|
||||
"learning_rate": 1.9086610330779205e-05,
|
||||
"loss": 0.6863,
|
||||
"step": 47
|
||||
},
|
||||
{
|
||||
"epoch": 1.21,
|
||||
"learning_rate": 1.8964798455574637e-05,
|
||||
"loss": 0.6792,
|
||||
"step": 48
|
||||
},
|
||||
{
|
||||
"epoch": 1.24,
|
||||
"learning_rate": 1.884089795001512e-05,
|
||||
"loss": 0.5667,
|
||||
"step": 49
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"learning_rate": 1.871494130486089e-05,
|
||||
"loss": 0.5256,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"eval_loss": 0.6949857473373413,
|
||||
"eval_runtime": 26.8774,
|
||||
"eval_samples_per_second": 30.807,
|
||||
"eval_steps_per_second": 0.484,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.29,
|
||||
"learning_rate": 1.8586961550059147e-05,
|
||||
"loss": 0.7481,
|
||||
"step": 51
|
||||
},
|
||||
{
|
||||
"epoch": 1.31,
|
||||
"learning_rate": 1.8456992246082534e-05,
|
||||
"loss": 1.2028,
|
||||
"step": 52
|
||||
},
|
||||
{
|
||||
"epoch": 1.34,
|
||||
"learning_rate": 1.8325067475128493e-05,
|
||||
"loss": 0.8791,
|
||||
"step": 53
|
||||
},
|
||||
{
|
||||
"epoch": 1.36,
|
||||
"learning_rate": 1.8191221832181777e-05,
|
||||
"loss": 0.6651,
|
||||
"step": 54
|
||||
},
|
||||
{
|
||||
"epoch": 1.39,
|
||||
"learning_rate": 1.8055490415942538e-05,
|
||||
"loss": 0.5875,
|
||||
"step": 55
|
||||
},
|
||||
{
|
||||
"epoch": 1.42,
|
||||
"learning_rate": 1.791790881962229e-05,
|
||||
"loss": 0.5143,
|
||||
"step": 56
|
||||
},
|
||||
{
|
||||
"epoch": 1.44,
|
||||
"learning_rate": 1.77785131216102e-05,
|
||||
"loss": 0.6447,
|
||||
"step": 57
|
||||
},
|
||||
{
|
||||
"epoch": 1.47,
|
||||
"learning_rate": 1.7637339876012172e-05,
|
||||
"loss": 1.0286,
|
||||
"step": 58
|
||||
},
|
||||
{
|
||||
"epoch": 1.49,
|
||||
"learning_rate": 1.749442610306515e-05,
|
||||
"loss": 1.1227,
|
||||
"step": 59
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"learning_rate": 1.7349809279429216e-05,
|
||||
"loss": 0.6634,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"eval_loss": 0.6882586479187012,
|
||||
"eval_runtime": 26.9816,
|
||||
"eval_samples_per_second": 30.688,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.54,
|
||||
"learning_rate": 1.7203527328359972e-05,
|
||||
"loss": 0.6145,
|
||||
"step": 61
|
||||
},
|
||||
{
|
||||
"epoch": 1.57,
|
||||
"learning_rate": 1.705561860976384e-05,
|
||||
"loss": 0.5139,
|
||||
"step": 62
|
||||
},
|
||||
{
|
||||
"epoch": 1.59,
|
||||
"learning_rate": 1.6906121910138824e-05,
|
||||
"loss": 0.5819,
|
||||
"step": 63
|
||||
},
|
||||
{
|
||||
"epoch": 1.62,
|
||||
"learning_rate": 1.675507643240343e-05,
|
||||
"loss": 0.8716,
|
||||
"step": 64
|
||||
},
|
||||
{
|
||||
"epoch": 1.64,
|
||||
"learning_rate": 1.6602521785616394e-05,
|
||||
"loss": 1.2125,
|
||||
"step": 65
|
||||
},
|
||||
{
|
||||
"epoch": 1.67,
|
||||
"learning_rate": 1.6448497974589863e-05,
|
||||
"loss": 0.6803,
|
||||
"step": 66
|
||||
},
|
||||
{
|
||||
"epoch": 1.69,
|
||||
"learning_rate": 1.6293045389398855e-05,
|
||||
"loss": 0.6762,
|
||||
"step": 67
|
||||
},
|
||||
{
|
||||
"epoch": 1.72,
|
||||
"learning_rate": 1.6136204794789646e-05,
|
||||
"loss": 0.5324,
|
||||
"step": 68
|
||||
},
|
||||
{
|
||||
"epoch": 1.74,
|
||||
"learning_rate": 1.597801731948993e-05,
|
||||
"loss": 0.541,
|
||||
"step": 69
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"learning_rate": 1.5818524445423504e-05,
|
||||
"loss": 0.8724,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"eval_loss": 0.6789196729660034,
|
||||
"eval_runtime": 26.8064,
|
||||
"eval_samples_per_second": 30.888,
|
||||
"eval_steps_per_second": 0.485,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.79,
|
||||
"learning_rate": 1.5657767996832395e-05,
|
||||
"loss": 1.204,
|
||||
"step": 71
|
||||
},
|
||||
{
|
||||
"epoch": 1.82,
|
||||
"learning_rate": 1.549579012930912e-05,
|
||||
"loss": 0.6885,
|
||||
"step": 72
|
||||
},
|
||||
{
|
||||
"epoch": 1.85,
|
||||
"learning_rate": 1.5332633318742157e-05,
|
||||
"loss": 0.6975,
|
||||
"step": 73
|
||||
},
|
||||
{
|
||||
"epoch": 1.87,
|
||||
"learning_rate": 1.5168340350177374e-05,
|
||||
"loss": 0.5451,
|
||||
"step": 74
|
||||
},
|
||||
{
|
||||
"epoch": 1.9,
|
||||
"learning_rate": 1.5002954306598364e-05,
|
||||
"loss": 0.4803,
|
||||
"step": 75
|
||||
},
|
||||
{
|
||||
"epoch": 1.92,
|
||||
"learning_rate": 1.4836518557628703e-05,
|
||||
"loss": 0.6776,
|
||||
"step": 76
|
||||
},
|
||||
{
|
||||
"epoch": 1.95,
|
||||
"learning_rate": 1.4669076748159028e-05,
|
||||
"loss": 1.0698,
|
||||
"step": 77
|
||||
},
|
||||
{
|
||||
"epoch": 1.97,
|
||||
"learning_rate": 1.4500672786901884e-05,
|
||||
"loss": 0.7064,
|
||||
"step": 78
|
||||
},
|
||||
{
|
||||
"epoch": 2.0,
|
||||
"learning_rate": 1.4331350834877457e-05,
|
||||
"loss": 0.7407,
|
||||
"step": 79
|
||||
},
|
||||
{
|
||||
"epoch": 2.02,
|
||||
"learning_rate": 1.4161155293833112e-05,
|
||||
"loss": 0.9098,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 2.02,
|
||||
"eval_loss": 0.6750211715698242,
|
||||
"eval_runtime": 27.1108,
|
||||
"eval_samples_per_second": 30.541,
|
||||
"eval_steps_per_second": 0.48,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 2.05,
|
||||
"learning_rate": 1.3990130794599772e-05,
|
||||
"loss": 0.6304,
|
||||
"step": 81
|
||||
},
|
||||
{
|
||||
"epoch": 2.07,
|
||||
"learning_rate": 1.3818322185388293e-05,
|
||||
"loss": 0.5601,
|
||||
"step": 82
|
||||
},
|
||||
{
|
||||
"epoch": 2.1,
|
||||
"learning_rate": 1.3645774520028791e-05,
|
||||
"loss": 0.4661,
|
||||
"step": 83
|
||||
},
|
||||
{
|
||||
"epoch": 2.12,
|
||||
"learning_rate": 1.3472533046156025e-05,
|
||||
"loss": 0.5971,
|
||||
"step": 84
|
||||
},
|
||||
{
|
||||
"epoch": 2.15,
|
||||
"learning_rate": 1.3298643193344028e-05,
|
||||
"loss": 0.8505,
|
||||
"step": 85
|
||||
},
|
||||
{
|
||||
"epoch": 2.17,
|
||||
"learning_rate": 1.3124150561192969e-05,
|
||||
"loss": 1.1119,
|
||||
"step": 86
|
||||
},
|
||||
{
|
||||
"epoch": 2.2,
|
||||
"learning_rate": 1.2949100907371452e-05,
|
||||
"loss": 0.6298,
|
||||
"step": 87
|
||||
},
|
||||
{
|
||||
"epoch": 2.22,
|
||||
"learning_rate": 1.2773540135617358e-05,
|
||||
"loss": 0.5901,
|
||||
"step": 88
|
||||
},
|
||||
{
|
||||
"epoch": 2.25,
|
||||
"learning_rate": 1.2597514283700393e-05,
|
||||
"loss": 0.483,
|
||||
"step": 89
|
||||
},
|
||||
{
|
||||
"epoch": 2.27,
|
||||
"learning_rate": 1.2421069511349471e-05,
|
||||
"loss": 0.522,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 2.27,
|
||||
"eval_loss": 0.6687783002853394,
|
||||
"eval_runtime": 26.924,
|
||||
"eval_samples_per_second": 30.753,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 2.3,
|
||||
"learning_rate": 1.2244252088148128e-05,
|
||||
"loss": 0.8623,
|
||||
"step": 91
|
||||
},
|
||||
{
|
||||
"epoch": 2.33,
|
||||
"learning_rate": 1.2067108381401128e-05,
|
||||
"loss": 1.1752,
|
||||
"step": 92
|
||||
},
|
||||
{
|
||||
"epoch": 2.35,
|
||||
"learning_rate": 1.1889684843975447e-05,
|
||||
"loss": 0.6268,
|
||||
"step": 93
|
||||
},
|
||||
{
|
||||
"epoch": 2.38,
|
||||
"learning_rate": 1.1712028002118794e-05,
|
||||
"loss": 0.6146,
|
||||
"step": 94
|
||||
},
|
||||
{
|
||||
"epoch": 2.4,
|
||||
"learning_rate": 1.1534184443258931e-05,
|
||||
"loss": 0.4776,
|
||||
"step": 95
|
||||
},
|
||||
{
|
||||
"epoch": 2.43,
|
||||
"learning_rate": 1.1356200803786916e-05,
|
||||
"loss": 0.4863,
|
||||
"step": 96
|
||||
},
|
||||
{
|
||||
"epoch": 2.45,
|
||||
"learning_rate": 1.1178123756827518e-05,
|
||||
"loss": 0.7674,
|
||||
"step": 97
|
||||
},
|
||||
{
|
||||
"epoch": 2.48,
|
||||
"learning_rate": 1.1e-05,
|
||||
"loss": 1.1923,
|
||||
"step": 98
|
||||
},
|
||||
{
|
||||
"epoch": 2.5,
|
||||
"learning_rate": 1.0821876243172483e-05,
|
||||
"loss": 0.6922,
|
||||
"step": 99
|
||||
},
|
||||
{
|
||||
"epoch": 2.53,
|
||||
"learning_rate": 1.0643799196213087e-05,
|
||||
"loss": 0.6416,
|
||||
"step": 100
|
||||
},
|
||||
{
|
||||
"epoch": 2.53,
|
||||
"eval_loss": 0.6664387583732605,
|
||||
"eval_runtime": 27.0713,
|
||||
"eval_samples_per_second": 30.586,
|
||||
"eval_steps_per_second": 0.48,
|
||||
"step": 100
|
||||
}
|
||||
],
|
||||
"logging_steps": 1,
|
||||
"max_steps": 195,
|
||||
"num_train_epochs": 5,
|
||||
"save_steps": 25,
|
||||
"total_flos": 2.5513979468725617e+18,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoints/checkpoint-100/training_args.bin
Normal file
3
checkpoints/checkpoint-100/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-100/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-100/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
26
checkpoints/checkpoint-125/config.json
Normal file
26
checkpoints/checkpoint-125/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-125/generation_config.json
Normal file
4
checkpoints/checkpoint-125/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-125/latest
Normal file
1
checkpoints/checkpoint-125/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step125
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:20a1e5d81d6b254c8d719ab3ffcae9c1b9a75f56ec9a43d378c735417190125c
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:6133f7a1f96f5e5f5bf862dd11804989100df8b47bb9e61600d56d62711a4e17
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-125/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-125/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-125/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-125/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-125/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-125/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-125/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-125/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
865
checkpoints/checkpoint-125/trainer_state.json
Normal file
865
checkpoints/checkpoint-125/trainer_state.json
Normal file
@@ -0,0 +1,865 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 3.1595576619273302,
|
||||
"eval_steps": 10,
|
||||
"global_step": 125,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.03,
|
||||
"learning_rate": 2.2e-05,
|
||||
"loss": 0.9856,
|
||||
"step": 1
|
||||
},
|
||||
{
|
||||
"epoch": 0.05,
|
||||
"learning_rate": 2.1998557720321954e-05,
|
||||
"loss": 0.8657,
|
||||
"step": 2
|
||||
},
|
||||
{
|
||||
"epoch": 0.08,
|
||||
"learning_rate": 2.1994231259500664e-05,
|
||||
"loss": 0.7599,
|
||||
"step": 3
|
||||
},
|
||||
{
|
||||
"epoch": 0.1,
|
||||
"learning_rate": 2.1987021752075498e-05,
|
||||
"loss": 0.6924,
|
||||
"step": 4
|
||||
},
|
||||
{
|
||||
"epoch": 0.13,
|
||||
"learning_rate": 2.1976931088614826e-05,
|
||||
"loss": 0.818,
|
||||
"step": 5
|
||||
},
|
||||
{
|
||||
"epoch": 0.15,
|
||||
"learning_rate": 2.1963961915220263e-05,
|
||||
"loss": 1.2993,
|
||||
"step": 6
|
||||
},
|
||||
{
|
||||
"epoch": 0.18,
|
||||
"learning_rate": 2.1948117632832752e-05,
|
||||
"loss": 1.4107,
|
||||
"step": 7
|
||||
},
|
||||
{
|
||||
"epoch": 0.2,
|
||||
"learning_rate": 2.1929402396340756e-05,
|
||||
"loss": 0.8539,
|
||||
"step": 8
|
||||
},
|
||||
{
|
||||
"epoch": 0.23,
|
||||
"learning_rate": 2.1907821113490676e-05,
|
||||
"loss": 0.7822,
|
||||
"step": 9
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"learning_rate": 2.1883379443599917e-05,
|
||||
"loss": 0.6498,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"eval_loss": 0.777102530002594,
|
||||
"eval_runtime": 26.948,
|
||||
"eval_samples_per_second": 30.726,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.28,
|
||||
"learning_rate": 2.1856083796072795e-05,
|
||||
"loss": 0.7439,
|
||||
"step": 11
|
||||
},
|
||||
{
|
||||
"epoch": 0.3,
|
||||
"learning_rate": 2.182594132871982e-05,
|
||||
"loss": 1.0097,
|
||||
"step": 12
|
||||
},
|
||||
{
|
||||
"epoch": 0.33,
|
||||
"learning_rate": 2.179295994588063e-05,
|
||||
"loss": 1.3189,
|
||||
"step": 13
|
||||
},
|
||||
{
|
||||
"epoch": 0.35,
|
||||
"learning_rate": 2.175714829635128e-05,
|
||||
"loss": 0.7509,
|
||||
"step": 14
|
||||
},
|
||||
{
|
||||
"epoch": 0.38,
|
||||
"learning_rate": 2.1718515771116186e-05,
|
||||
"loss": 0.7273,
|
||||
"step": 15
|
||||
},
|
||||
{
|
||||
"epoch": 0.4,
|
||||
"learning_rate": 2.167707250088555e-05,
|
||||
"loss": 0.6242,
|
||||
"step": 16
|
||||
},
|
||||
{
|
||||
"epoch": 0.43,
|
||||
"learning_rate": 2.1632829353438724e-05,
|
||||
"loss": 0.6135,
|
||||
"step": 17
|
||||
},
|
||||
{
|
||||
"epoch": 0.45,
|
||||
"learning_rate": 2.1585797930774335e-05,
|
||||
"loss": 0.9365,
|
||||
"step": 18
|
||||
},
|
||||
{
|
||||
"epoch": 0.48,
|
||||
"learning_rate": 2.1535990566067858e-05,
|
||||
"loss": 1.4216,
|
||||
"step": 19
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"learning_rate": 2.148342032043747e-05,
|
||||
"loss": 0.7369,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"eval_loss": 0.7404242157936096,
|
||||
"eval_runtime": 26.98,
|
||||
"eval_samples_per_second": 30.689,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.53,
|
||||
"learning_rate": 2.1428100979518976e-05,
|
||||
"loss": 0.7605,
|
||||
"step": 21
|
||||
},
|
||||
{
|
||||
"epoch": 0.56,
|
||||
"learning_rate": 2.1370047049850775e-05,
|
||||
"loss": 0.6095,
|
||||
"step": 22
|
||||
},
|
||||
{
|
||||
"epoch": 0.58,
|
||||
"learning_rate": 2.130927375506978e-05,
|
||||
"loss": 0.5955,
|
||||
"step": 23
|
||||
},
|
||||
{
|
||||
"epoch": 0.61,
|
||||
"learning_rate": 2.1245797031919265e-05,
|
||||
"loss": 0.851,
|
||||
"step": 24
|
||||
},
|
||||
{
|
||||
"epoch": 0.63,
|
||||
"learning_rate": 2.1179633526069748e-05,
|
||||
"loss": 1.324,
|
||||
"step": 25
|
||||
},
|
||||
{
|
||||
"epoch": 0.66,
|
||||
"learning_rate": 2.1110800587753934e-05,
|
||||
"loss": 0.9368,
|
||||
"step": 26
|
||||
},
|
||||
{
|
||||
"epoch": 0.68,
|
||||
"learning_rate": 2.103931626721694e-05,
|
||||
"loss": 0.7307,
|
||||
"step": 27
|
||||
},
|
||||
{
|
||||
"epoch": 0.71,
|
||||
"learning_rate": 2.0965199309982913e-05,
|
||||
"loss": 0.6489,
|
||||
"step": 28
|
||||
},
|
||||
{
|
||||
"epoch": 0.73,
|
||||
"learning_rate": 2.0888469151939347e-05,
|
||||
"loss": 0.536,
|
||||
"step": 29
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"learning_rate": 2.0809145914240356e-05,
|
||||
"loss": 0.732,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"eval_loss": 0.7266919612884521,
|
||||
"eval_runtime": 26.9346,
|
||||
"eval_samples_per_second": 30.741,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.78,
|
||||
"learning_rate": 2.072725039803026e-05,
|
||||
"loss": 1.1248,
|
||||
"step": 31
|
||||
},
|
||||
{
|
||||
"epoch": 0.81,
|
||||
"learning_rate": 2.0642804078988827e-05,
|
||||
"loss": 1.0917,
|
||||
"step": 32
|
||||
},
|
||||
{
|
||||
"epoch": 0.83,
|
||||
"learning_rate": 2.055582910169966e-05,
|
||||
"loss": 0.735,
|
||||
"step": 33
|
||||
},
|
||||
{
|
||||
"epoch": 0.86,
|
||||
"learning_rate": 2.0466348273843176e-05,
|
||||
"loss": 0.6749,
|
||||
"step": 34
|
||||
},
|
||||
{
|
||||
"epoch": 0.88,
|
||||
"learning_rate": 2.0374385060215662e-05,
|
||||
"loss": 0.5519,
|
||||
"step": 35
|
||||
},
|
||||
{
|
||||
"epoch": 0.91,
|
||||
"learning_rate": 2.027996357657605e-05,
|
||||
"loss": 0.6593,
|
||||
"step": 36
|
||||
},
|
||||
{
|
||||
"epoch": 0.94,
|
||||
"learning_rate": 2.018310858332198e-05,
|
||||
"loss": 0.959,
|
||||
"step": 37
|
||||
},
|
||||
{
|
||||
"epoch": 0.96,
|
||||
"learning_rate": 2.008384547899682e-05,
|
||||
"loss": 1.168,
|
||||
"step": 38
|
||||
},
|
||||
{
|
||||
"epoch": 0.99,
|
||||
"learning_rate": 1.9982200293629318e-05,
|
||||
"loss": 0.6186,
|
||||
"step": 39
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"learning_rate": 1.9878199681907702e-05,
|
||||
"loss": 1.1207,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"eval_loss": 0.7203702926635742,
|
||||
"eval_runtime": 26.9926,
|
||||
"eval_samples_per_second": 30.675,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.04,
|
||||
"learning_rate": 1.977187091618993e-05,
|
||||
"loss": 0.7095,
|
||||
"step": 41
|
||||
},
|
||||
{
|
||||
"epoch": 1.06,
|
||||
"learning_rate": 1.9663241879351998e-05,
|
||||
"loss": 0.6781,
|
||||
"step": 42
|
||||
},
|
||||
{
|
||||
"epoch": 1.09,
|
||||
"learning_rate": 1.9552341057476144e-05,
|
||||
"loss": 0.5649,
|
||||
"step": 43
|
||||
},
|
||||
{
|
||||
"epoch": 1.11,
|
||||
"learning_rate": 1.9439197532380845e-05,
|
||||
"loss": 0.5466,
|
||||
"step": 44
|
||||
},
|
||||
{
|
||||
"epoch": 1.14,
|
||||
"learning_rate": 1.9323840973994647e-05,
|
||||
"loss": 0.8721,
|
||||
"step": 45
|
||||
},
|
||||
{
|
||||
"epoch": 1.16,
|
||||
"learning_rate": 1.9206301632575707e-05,
|
||||
"loss": 1.2152,
|
||||
"step": 46
|
||||
},
|
||||
{
|
||||
"epoch": 1.19,
|
||||
"learning_rate": 1.9086610330779205e-05,
|
||||
"loss": 0.6863,
|
||||
"step": 47
|
||||
},
|
||||
{
|
||||
"epoch": 1.21,
|
||||
"learning_rate": 1.8964798455574637e-05,
|
||||
"loss": 0.6792,
|
||||
"step": 48
|
||||
},
|
||||
{
|
||||
"epoch": 1.24,
|
||||
"learning_rate": 1.884089795001512e-05,
|
||||
"loss": 0.5667,
|
||||
"step": 49
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"learning_rate": 1.871494130486089e-05,
|
||||
"loss": 0.5256,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"eval_loss": 0.6949857473373413,
|
||||
"eval_runtime": 26.8774,
|
||||
"eval_samples_per_second": 30.807,
|
||||
"eval_steps_per_second": 0.484,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.29,
|
||||
"learning_rate": 1.8586961550059147e-05,
|
||||
"loss": 0.7481,
|
||||
"step": 51
|
||||
},
|
||||
{
|
||||
"epoch": 1.31,
|
||||
"learning_rate": 1.8456992246082534e-05,
|
||||
"loss": 1.2028,
|
||||
"step": 52
|
||||
},
|
||||
{
|
||||
"epoch": 1.34,
|
||||
"learning_rate": 1.8325067475128493e-05,
|
||||
"loss": 0.8791,
|
||||
"step": 53
|
||||
},
|
||||
{
|
||||
"epoch": 1.36,
|
||||
"learning_rate": 1.8191221832181777e-05,
|
||||
"loss": 0.6651,
|
||||
"step": 54
|
||||
},
|
||||
{
|
||||
"epoch": 1.39,
|
||||
"learning_rate": 1.8055490415942538e-05,
|
||||
"loss": 0.5875,
|
||||
"step": 55
|
||||
},
|
||||
{
|
||||
"epoch": 1.42,
|
||||
"learning_rate": 1.791790881962229e-05,
|
||||
"loss": 0.5143,
|
||||
"step": 56
|
||||
},
|
||||
{
|
||||
"epoch": 1.44,
|
||||
"learning_rate": 1.77785131216102e-05,
|
||||
"loss": 0.6447,
|
||||
"step": 57
|
||||
},
|
||||
{
|
||||
"epoch": 1.47,
|
||||
"learning_rate": 1.7637339876012172e-05,
|
||||
"loss": 1.0286,
|
||||
"step": 58
|
||||
},
|
||||
{
|
||||
"epoch": 1.49,
|
||||
"learning_rate": 1.749442610306515e-05,
|
||||
"loss": 1.1227,
|
||||
"step": 59
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"learning_rate": 1.7349809279429216e-05,
|
||||
"loss": 0.6634,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"eval_loss": 0.6882586479187012,
|
||||
"eval_runtime": 26.9816,
|
||||
"eval_samples_per_second": 30.688,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.54,
|
||||
"learning_rate": 1.7203527328359972e-05,
|
||||
"loss": 0.6145,
|
||||
"step": 61
|
||||
},
|
||||
{
|
||||
"epoch": 1.57,
|
||||
"learning_rate": 1.705561860976384e-05,
|
||||
"loss": 0.5139,
|
||||
"step": 62
|
||||
},
|
||||
{
|
||||
"epoch": 1.59,
|
||||
"learning_rate": 1.6906121910138824e-05,
|
||||
"loss": 0.5819,
|
||||
"step": 63
|
||||
},
|
||||
{
|
||||
"epoch": 1.62,
|
||||
"learning_rate": 1.675507643240343e-05,
|
||||
"loss": 0.8716,
|
||||
"step": 64
|
||||
},
|
||||
{
|
||||
"epoch": 1.64,
|
||||
"learning_rate": 1.6602521785616394e-05,
|
||||
"loss": 1.2125,
|
||||
"step": 65
|
||||
},
|
||||
{
|
||||
"epoch": 1.67,
|
||||
"learning_rate": 1.6448497974589863e-05,
|
||||
"loss": 0.6803,
|
||||
"step": 66
|
||||
},
|
||||
{
|
||||
"epoch": 1.69,
|
||||
"learning_rate": 1.6293045389398855e-05,
|
||||
"loss": 0.6762,
|
||||
"step": 67
|
||||
},
|
||||
{
|
||||
"epoch": 1.72,
|
||||
"learning_rate": 1.6136204794789646e-05,
|
||||
"loss": 0.5324,
|
||||
"step": 68
|
||||
},
|
||||
{
|
||||
"epoch": 1.74,
|
||||
"learning_rate": 1.597801731948993e-05,
|
||||
"loss": 0.541,
|
||||
"step": 69
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"learning_rate": 1.5818524445423504e-05,
|
||||
"loss": 0.8724,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"eval_loss": 0.6789196729660034,
|
||||
"eval_runtime": 26.8064,
|
||||
"eval_samples_per_second": 30.888,
|
||||
"eval_steps_per_second": 0.485,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.79,
|
||||
"learning_rate": 1.5657767996832395e-05,
|
||||
"loss": 1.204,
|
||||
"step": 71
|
||||
},
|
||||
{
|
||||
"epoch": 1.82,
|
||||
"learning_rate": 1.549579012930912e-05,
|
||||
"loss": 0.6885,
|
||||
"step": 72
|
||||
},
|
||||
{
|
||||
"epoch": 1.85,
|
||||
"learning_rate": 1.5332633318742157e-05,
|
||||
"loss": 0.6975,
|
||||
"step": 73
|
||||
},
|
||||
{
|
||||
"epoch": 1.87,
|
||||
"learning_rate": 1.5168340350177374e-05,
|
||||
"loss": 0.5451,
|
||||
"step": 74
|
||||
},
|
||||
{
|
||||
"epoch": 1.9,
|
||||
"learning_rate": 1.5002954306598364e-05,
|
||||
"loss": 0.4803,
|
||||
"step": 75
|
||||
},
|
||||
{
|
||||
"epoch": 1.92,
|
||||
"learning_rate": 1.4836518557628703e-05,
|
||||
"loss": 0.6776,
|
||||
"step": 76
|
||||
},
|
||||
{
|
||||
"epoch": 1.95,
|
||||
"learning_rate": 1.4669076748159028e-05,
|
||||
"loss": 1.0698,
|
||||
"step": 77
|
||||
},
|
||||
{
|
||||
"epoch": 1.97,
|
||||
"learning_rate": 1.4500672786901884e-05,
|
||||
"loss": 0.7064,
|
||||
"step": 78
|
||||
},
|
||||
{
|
||||
"epoch": 2.0,
|
||||
"learning_rate": 1.4331350834877457e-05,
|
||||
"loss": 0.7407,
|
||||
"step": 79
|
||||
},
|
||||
{
|
||||
"epoch": 2.02,
|
||||
"learning_rate": 1.4161155293833112e-05,
|
||||
"loss": 0.9098,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 2.02,
|
||||
"eval_loss": 0.6750211715698242,
|
||||
"eval_runtime": 27.1108,
|
||||
"eval_samples_per_second": 30.541,
|
||||
"eval_steps_per_second": 0.48,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 2.05,
|
||||
"learning_rate": 1.3990130794599772e-05,
|
||||
"loss": 0.6304,
|
||||
"step": 81
|
||||
},
|
||||
{
|
||||
"epoch": 2.07,
|
||||
"learning_rate": 1.3818322185388293e-05,
|
||||
"loss": 0.5601,
|
||||
"step": 82
|
||||
},
|
||||
{
|
||||
"epoch": 2.1,
|
||||
"learning_rate": 1.3645774520028791e-05,
|
||||
"loss": 0.4661,
|
||||
"step": 83
|
||||
},
|
||||
{
|
||||
"epoch": 2.12,
|
||||
"learning_rate": 1.3472533046156025e-05,
|
||||
"loss": 0.5971,
|
||||
"step": 84
|
||||
},
|
||||
{
|
||||
"epoch": 2.15,
|
||||
"learning_rate": 1.3298643193344028e-05,
|
||||
"loss": 0.8505,
|
||||
"step": 85
|
||||
},
|
||||
{
|
||||
"epoch": 2.17,
|
||||
"learning_rate": 1.3124150561192969e-05,
|
||||
"loss": 1.1119,
|
||||
"step": 86
|
||||
},
|
||||
{
|
||||
"epoch": 2.2,
|
||||
"learning_rate": 1.2949100907371452e-05,
|
||||
"loss": 0.6298,
|
||||
"step": 87
|
||||
},
|
||||
{
|
||||
"epoch": 2.22,
|
||||
"learning_rate": 1.2773540135617358e-05,
|
||||
"loss": 0.5901,
|
||||
"step": 88
|
||||
},
|
||||
{
|
||||
"epoch": 2.25,
|
||||
"learning_rate": 1.2597514283700393e-05,
|
||||
"loss": 0.483,
|
||||
"step": 89
|
||||
},
|
||||
{
|
||||
"epoch": 2.27,
|
||||
"learning_rate": 1.2421069511349471e-05,
|
||||
"loss": 0.522,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 2.27,
|
||||
"eval_loss": 0.6687783002853394,
|
||||
"eval_runtime": 26.924,
|
||||
"eval_samples_per_second": 30.753,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 2.3,
|
||||
"learning_rate": 1.2244252088148128e-05,
|
||||
"loss": 0.8623,
|
||||
"step": 91
|
||||
},
|
||||
{
|
||||
"epoch": 2.33,
|
||||
"learning_rate": 1.2067108381401128e-05,
|
||||
"loss": 1.1752,
|
||||
"step": 92
|
||||
},
|
||||
{
|
||||
"epoch": 2.35,
|
||||
"learning_rate": 1.1889684843975447e-05,
|
||||
"loss": 0.6268,
|
||||
"step": 93
|
||||
},
|
||||
{
|
||||
"epoch": 2.38,
|
||||
"learning_rate": 1.1712028002118794e-05,
|
||||
"loss": 0.6146,
|
||||
"step": 94
|
||||
},
|
||||
{
|
||||
"epoch": 2.4,
|
||||
"learning_rate": 1.1534184443258931e-05,
|
||||
"loss": 0.4776,
|
||||
"step": 95
|
||||
},
|
||||
{
|
||||
"epoch": 2.43,
|
||||
"learning_rate": 1.1356200803786916e-05,
|
||||
"loss": 0.4863,
|
||||
"step": 96
|
||||
},
|
||||
{
|
||||
"epoch": 2.45,
|
||||
"learning_rate": 1.1178123756827518e-05,
|
||||
"loss": 0.7674,
|
||||
"step": 97
|
||||
},
|
||||
{
|
||||
"epoch": 2.48,
|
||||
"learning_rate": 1.1e-05,
|
||||
"loss": 1.1923,
|
||||
"step": 98
|
||||
},
|
||||
{
|
||||
"epoch": 2.5,
|
||||
"learning_rate": 1.0821876243172483e-05,
|
||||
"loss": 0.6922,
|
||||
"step": 99
|
||||
},
|
||||
{
|
||||
"epoch": 2.53,
|
||||
"learning_rate": 1.0643799196213087e-05,
|
||||
"loss": 0.6416,
|
||||
"step": 100
|
||||
},
|
||||
{
|
||||
"epoch": 2.53,
|
||||
"eval_loss": 0.6664387583732605,
|
||||
"eval_runtime": 27.0713,
|
||||
"eval_samples_per_second": 30.586,
|
||||
"eval_steps_per_second": 0.48,
|
||||
"step": 100
|
||||
},
|
||||
{
|
||||
"epoch": 2.55,
|
||||
"learning_rate": 1.0465815556741072e-05,
|
||||
"loss": 0.5438,
|
||||
"step": 101
|
||||
},
|
||||
{
|
||||
"epoch": 2.58,
|
||||
"learning_rate": 1.0287971997881207e-05,
|
||||
"loss": 0.4733,
|
||||
"step": 102
|
||||
},
|
||||
{
|
||||
"epoch": 2.6,
|
||||
"learning_rate": 1.0110315156024555e-05,
|
||||
"loss": 0.6894,
|
||||
"step": 103
|
||||
},
|
||||
{
|
||||
"epoch": 2.63,
|
||||
"learning_rate": 9.932891618598871e-06,
|
||||
"loss": 1.0594,
|
||||
"step": 104
|
||||
},
|
||||
{
|
||||
"epoch": 2.65,
|
||||
"learning_rate": 9.755747911851873e-06,
|
||||
"loss": 0.8682,
|
||||
"step": 105
|
||||
},
|
||||
{
|
||||
"epoch": 2.68,
|
||||
"learning_rate": 9.578930488650532e-06,
|
||||
"loss": 0.6275,
|
||||
"step": 106
|
||||
},
|
||||
{
|
||||
"epoch": 2.7,
|
||||
"learning_rate": 9.40248571629961e-06,
|
||||
"loss": 0.5676,
|
||||
"step": 107
|
||||
},
|
||||
{
|
||||
"epoch": 2.73,
|
||||
"learning_rate": 9.226459864382642e-06,
|
||||
"loss": 0.4443,
|
||||
"step": 108
|
||||
},
|
||||
{
|
||||
"epoch": 2.76,
|
||||
"learning_rate": 9.05089909262855e-06,
|
||||
"loss": 0.5907,
|
||||
"step": 109
|
||||
},
|
||||
{
|
||||
"epoch": 2.78,
|
||||
"learning_rate": 8.875849438807032e-06,
|
||||
"loss": 0.9166,
|
||||
"step": 110
|
||||
},
|
||||
{
|
||||
"epoch": 2.78,
|
||||
"eval_loss": 0.6656625270843506,
|
||||
"eval_runtime": 26.8682,
|
||||
"eval_samples_per_second": 30.817,
|
||||
"eval_steps_per_second": 0.484,
|
||||
"step": 110
|
||||
},
|
||||
{
|
||||
"epoch": 2.81,
|
||||
"learning_rate": 8.701356806655973e-06,
|
||||
"loss": 1.0669,
|
||||
"step": 111
|
||||
},
|
||||
{
|
||||
"epoch": 2.83,
|
||||
"learning_rate": 8.52746695384398e-06,
|
||||
"loss": 0.6286,
|
||||
"step": 112
|
||||
},
|
||||
{
|
||||
"epoch": 2.86,
|
||||
"learning_rate": 8.354225479971213e-06,
|
||||
"loss": 0.5951,
|
||||
"step": 113
|
||||
},
|
||||
{
|
||||
"epoch": 2.88,
|
||||
"learning_rate": 8.181677814611703e-06,
|
||||
"loss": 0.4757,
|
||||
"step": 114
|
||||
},
|
||||
{
|
||||
"epoch": 2.91,
|
||||
"learning_rate": 8.009869205400227e-06,
|
||||
"loss": 0.5694,
|
||||
"step": 115
|
||||
},
|
||||
{
|
||||
"epoch": 2.93,
|
||||
"learning_rate": 7.83884470616689e-06,
|
||||
"loss": 0.8569,
|
||||
"step": 116
|
||||
},
|
||||
{
|
||||
"epoch": 2.96,
|
||||
"learning_rate": 7.668649165122545e-06,
|
||||
"loss": 1.0254,
|
||||
"step": 117
|
||||
},
|
||||
{
|
||||
"epoch": 2.98,
|
||||
"learning_rate": 7.49932721309812e-06,
|
||||
"loss": 0.5457,
|
||||
"step": 118
|
||||
},
|
||||
{
|
||||
"epoch": 3.01,
|
||||
"learning_rate": 7.330923251840976e-06,
|
||||
"loss": 0.9681,
|
||||
"step": 119
|
||||
},
|
||||
{
|
||||
"epoch": 3.03,
|
||||
"learning_rate": 7.163481442371295e-06,
|
||||
"loss": 0.5708,
|
||||
"step": 120
|
||||
},
|
||||
{
|
||||
"epoch": 3.03,
|
||||
"eval_loss": 0.6612405776977539,
|
||||
"eval_runtime": 27.1115,
|
||||
"eval_samples_per_second": 30.541,
|
||||
"eval_steps_per_second": 0.48,
|
||||
"step": 120
|
||||
},
|
||||
{
|
||||
"epoch": 3.06,
|
||||
"learning_rate": 6.997045693401637e-06,
|
||||
"loss": 0.5819,
|
||||
"step": 121
|
||||
},
|
||||
{
|
||||
"epoch": 3.08,
|
||||
"learning_rate": 6.831659649822625e-06,
|
||||
"loss": 0.4961,
|
||||
"step": 122
|
||||
},
|
||||
{
|
||||
"epoch": 3.11,
|
||||
"learning_rate": 6.667366681257843e-06,
|
||||
"loss": 0.4893,
|
||||
"step": 123
|
||||
},
|
||||
{
|
||||
"epoch": 3.13,
|
||||
"learning_rate": 6.504209870690885e-06,
|
||||
"loss": 0.7579,
|
||||
"step": 124
|
||||
},
|
||||
{
|
||||
"epoch": 3.16,
|
||||
"learning_rate": 6.342232003167612e-06,
|
||||
"loss": 1.0994,
|
||||
"step": 125
|
||||
}
|
||||
],
|
||||
"logging_steps": 1,
|
||||
"max_steps": 195,
|
||||
"num_train_epochs": 5,
|
||||
"save_steps": 25,
|
||||
"total_flos": 3.1244125843845284e+18,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoints/checkpoint-125/training_args.bin
Normal file
3
checkpoints/checkpoint-125/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-125/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-125/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
26
checkpoints/checkpoint-150/config.json
Normal file
26
checkpoints/checkpoint-150/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-150/generation_config.json
Normal file
4
checkpoints/checkpoint-150/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-150/latest
Normal file
1
checkpoints/checkpoint-150/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step150
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:78e7fe4ea159cd51597d6c5842f79edaa2a19f65e52fe37a3c3935f04e87e2c0
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:49df5c5317e167cf33e254e2ba47940f768f32e7f0b97f9a19a033fb224436e2
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-150/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-150/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-150/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-150/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-150/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-150/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-150/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-150/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
1039
checkpoints/checkpoint-150/trainer_state.json
Normal file
1039
checkpoints/checkpoint-150/trainer_state.json
Normal file
File diff suppressed because it is too large
Load Diff
3
checkpoints/checkpoint-150/training_args.bin
Normal file
3
checkpoints/checkpoint-150/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-150/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-150/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
26
checkpoints/checkpoint-175/config.json
Normal file
26
checkpoints/checkpoint-175/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-175/generation_config.json
Normal file
4
checkpoints/checkpoint-175/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-175/latest
Normal file
1
checkpoints/checkpoint-175/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step175
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:97d36a2330f57a20959559988f8b3e538c51e59a3bbb98430b97bbdc639cab64
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:508a28d0ccf3b9625b40b6dfdc229ea46d350bc72adddd1c11cf897f0889c92f
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-175/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-175/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-175/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-175/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-175/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-175/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-175/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-175/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
1205
checkpoints/checkpoint-175/trainer_state.json
Normal file
1205
checkpoints/checkpoint-175/trainer_state.json
Normal file
File diff suppressed because it is too large
Load Diff
3
checkpoints/checkpoint-175/training_args.bin
Normal file
3
checkpoints/checkpoint-175/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-175/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-175/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
26
checkpoints/checkpoint-50/config.json
Normal file
26
checkpoints/checkpoint-50/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-50/generation_config.json
Normal file
4
checkpoints/checkpoint-50/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-50/latest
Normal file
1
checkpoints/checkpoint-50/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step50
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:a53d6fadf9dcb62c9d8e1300f97f793978f4213578d592e3d19c71bbfc955c8c
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:953b3c54b89e1eb6dcf2a2385a9128b68ebe453f401130ffa50c4c0379a15f6e
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-50/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-50/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-50/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-50/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-50/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-50/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-50/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-50/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
359
checkpoints/checkpoint-50/trainer_state.json
Normal file
359
checkpoints/checkpoint-50/trainer_state.json
Normal file
@@ -0,0 +1,359 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 1.263823064770932,
|
||||
"eval_steps": 10,
|
||||
"global_step": 50,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.03,
|
||||
"learning_rate": 2.2e-05,
|
||||
"loss": 0.9856,
|
||||
"step": 1
|
||||
},
|
||||
{
|
||||
"epoch": 0.05,
|
||||
"learning_rate": 2.1998557720321954e-05,
|
||||
"loss": 0.8657,
|
||||
"step": 2
|
||||
},
|
||||
{
|
||||
"epoch": 0.08,
|
||||
"learning_rate": 2.1994231259500664e-05,
|
||||
"loss": 0.7599,
|
||||
"step": 3
|
||||
},
|
||||
{
|
||||
"epoch": 0.1,
|
||||
"learning_rate": 2.1987021752075498e-05,
|
||||
"loss": 0.6924,
|
||||
"step": 4
|
||||
},
|
||||
{
|
||||
"epoch": 0.13,
|
||||
"learning_rate": 2.1976931088614826e-05,
|
||||
"loss": 0.818,
|
||||
"step": 5
|
||||
},
|
||||
{
|
||||
"epoch": 0.15,
|
||||
"learning_rate": 2.1963961915220263e-05,
|
||||
"loss": 1.2993,
|
||||
"step": 6
|
||||
},
|
||||
{
|
||||
"epoch": 0.18,
|
||||
"learning_rate": 2.1948117632832752e-05,
|
||||
"loss": 1.4107,
|
||||
"step": 7
|
||||
},
|
||||
{
|
||||
"epoch": 0.2,
|
||||
"learning_rate": 2.1929402396340756e-05,
|
||||
"loss": 0.8539,
|
||||
"step": 8
|
||||
},
|
||||
{
|
||||
"epoch": 0.23,
|
||||
"learning_rate": 2.1907821113490676e-05,
|
||||
"loss": 0.7822,
|
||||
"step": 9
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"learning_rate": 2.1883379443599917e-05,
|
||||
"loss": 0.6498,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"eval_loss": 0.777102530002594,
|
||||
"eval_runtime": 26.948,
|
||||
"eval_samples_per_second": 30.726,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.28,
|
||||
"learning_rate": 2.1856083796072795e-05,
|
||||
"loss": 0.7439,
|
||||
"step": 11
|
||||
},
|
||||
{
|
||||
"epoch": 0.3,
|
||||
"learning_rate": 2.182594132871982e-05,
|
||||
"loss": 1.0097,
|
||||
"step": 12
|
||||
},
|
||||
{
|
||||
"epoch": 0.33,
|
||||
"learning_rate": 2.179295994588063e-05,
|
||||
"loss": 1.3189,
|
||||
"step": 13
|
||||
},
|
||||
{
|
||||
"epoch": 0.35,
|
||||
"learning_rate": 2.175714829635128e-05,
|
||||
"loss": 0.7509,
|
||||
"step": 14
|
||||
},
|
||||
{
|
||||
"epoch": 0.38,
|
||||
"learning_rate": 2.1718515771116186e-05,
|
||||
"loss": 0.7273,
|
||||
"step": 15
|
||||
},
|
||||
{
|
||||
"epoch": 0.4,
|
||||
"learning_rate": 2.167707250088555e-05,
|
||||
"loss": 0.6242,
|
||||
"step": 16
|
||||
},
|
||||
{
|
||||
"epoch": 0.43,
|
||||
"learning_rate": 2.1632829353438724e-05,
|
||||
"loss": 0.6135,
|
||||
"step": 17
|
||||
},
|
||||
{
|
||||
"epoch": 0.45,
|
||||
"learning_rate": 2.1585797930774335e-05,
|
||||
"loss": 0.9365,
|
||||
"step": 18
|
||||
},
|
||||
{
|
||||
"epoch": 0.48,
|
||||
"learning_rate": 2.1535990566067858e-05,
|
||||
"loss": 1.4216,
|
||||
"step": 19
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"learning_rate": 2.148342032043747e-05,
|
||||
"loss": 0.7369,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"eval_loss": 0.7404242157936096,
|
||||
"eval_runtime": 26.98,
|
||||
"eval_samples_per_second": 30.689,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.53,
|
||||
"learning_rate": 2.1428100979518976e-05,
|
||||
"loss": 0.7605,
|
||||
"step": 21
|
||||
},
|
||||
{
|
||||
"epoch": 0.56,
|
||||
"learning_rate": 2.1370047049850775e-05,
|
||||
"loss": 0.6095,
|
||||
"step": 22
|
||||
},
|
||||
{
|
||||
"epoch": 0.58,
|
||||
"learning_rate": 2.130927375506978e-05,
|
||||
"loss": 0.5955,
|
||||
"step": 23
|
||||
},
|
||||
{
|
||||
"epoch": 0.61,
|
||||
"learning_rate": 2.1245797031919265e-05,
|
||||
"loss": 0.851,
|
||||
"step": 24
|
||||
},
|
||||
{
|
||||
"epoch": 0.63,
|
||||
"learning_rate": 2.1179633526069748e-05,
|
||||
"loss": 1.324,
|
||||
"step": 25
|
||||
},
|
||||
{
|
||||
"epoch": 0.66,
|
||||
"learning_rate": 2.1110800587753934e-05,
|
||||
"loss": 0.9368,
|
||||
"step": 26
|
||||
},
|
||||
{
|
||||
"epoch": 0.68,
|
||||
"learning_rate": 2.103931626721694e-05,
|
||||
"loss": 0.7307,
|
||||
"step": 27
|
||||
},
|
||||
{
|
||||
"epoch": 0.71,
|
||||
"learning_rate": 2.0965199309982913e-05,
|
||||
"loss": 0.6489,
|
||||
"step": 28
|
||||
},
|
||||
{
|
||||
"epoch": 0.73,
|
||||
"learning_rate": 2.0888469151939347e-05,
|
||||
"loss": 0.536,
|
||||
"step": 29
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"learning_rate": 2.0809145914240356e-05,
|
||||
"loss": 0.732,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"eval_loss": 0.7266919612884521,
|
||||
"eval_runtime": 26.9346,
|
||||
"eval_samples_per_second": 30.741,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.78,
|
||||
"learning_rate": 2.072725039803026e-05,
|
||||
"loss": 1.1248,
|
||||
"step": 31
|
||||
},
|
||||
{
|
||||
"epoch": 0.81,
|
||||
"learning_rate": 2.0642804078988827e-05,
|
||||
"loss": 1.0917,
|
||||
"step": 32
|
||||
},
|
||||
{
|
||||
"epoch": 0.83,
|
||||
"learning_rate": 2.055582910169966e-05,
|
||||
"loss": 0.735,
|
||||
"step": 33
|
||||
},
|
||||
{
|
||||
"epoch": 0.86,
|
||||
"learning_rate": 2.0466348273843176e-05,
|
||||
"loss": 0.6749,
|
||||
"step": 34
|
||||
},
|
||||
{
|
||||
"epoch": 0.88,
|
||||
"learning_rate": 2.0374385060215662e-05,
|
||||
"loss": 0.5519,
|
||||
"step": 35
|
||||
},
|
||||
{
|
||||
"epoch": 0.91,
|
||||
"learning_rate": 2.027996357657605e-05,
|
||||
"loss": 0.6593,
|
||||
"step": 36
|
||||
},
|
||||
{
|
||||
"epoch": 0.94,
|
||||
"learning_rate": 2.018310858332198e-05,
|
||||
"loss": 0.959,
|
||||
"step": 37
|
||||
},
|
||||
{
|
||||
"epoch": 0.96,
|
||||
"learning_rate": 2.008384547899682e-05,
|
||||
"loss": 1.168,
|
||||
"step": 38
|
||||
},
|
||||
{
|
||||
"epoch": 0.99,
|
||||
"learning_rate": 1.9982200293629318e-05,
|
||||
"loss": 0.6186,
|
||||
"step": 39
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"learning_rate": 1.9878199681907702e-05,
|
||||
"loss": 1.1207,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"eval_loss": 0.7203702926635742,
|
||||
"eval_runtime": 26.9926,
|
||||
"eval_samples_per_second": 30.675,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.04,
|
||||
"learning_rate": 1.977187091618993e-05,
|
||||
"loss": 0.7095,
|
||||
"step": 41
|
||||
},
|
||||
{
|
||||
"epoch": 1.06,
|
||||
"learning_rate": 1.9663241879351998e-05,
|
||||
"loss": 0.6781,
|
||||
"step": 42
|
||||
},
|
||||
{
|
||||
"epoch": 1.09,
|
||||
"learning_rate": 1.9552341057476144e-05,
|
||||
"loss": 0.5649,
|
||||
"step": 43
|
||||
},
|
||||
{
|
||||
"epoch": 1.11,
|
||||
"learning_rate": 1.9439197532380845e-05,
|
||||
"loss": 0.5466,
|
||||
"step": 44
|
||||
},
|
||||
{
|
||||
"epoch": 1.14,
|
||||
"learning_rate": 1.9323840973994647e-05,
|
||||
"loss": 0.8721,
|
||||
"step": 45
|
||||
},
|
||||
{
|
||||
"epoch": 1.16,
|
||||
"learning_rate": 1.9206301632575707e-05,
|
||||
"loss": 1.2152,
|
||||
"step": 46
|
||||
},
|
||||
{
|
||||
"epoch": 1.19,
|
||||
"learning_rate": 1.9086610330779205e-05,
|
||||
"loss": 0.6863,
|
||||
"step": 47
|
||||
},
|
||||
{
|
||||
"epoch": 1.21,
|
||||
"learning_rate": 1.8964798455574637e-05,
|
||||
"loss": 0.6792,
|
||||
"step": 48
|
||||
},
|
||||
{
|
||||
"epoch": 1.24,
|
||||
"learning_rate": 1.884089795001512e-05,
|
||||
"loss": 0.5667,
|
||||
"step": 49
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"learning_rate": 1.871494130486089e-05,
|
||||
"loss": 0.5256,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"eval_loss": 0.6949857473373413,
|
||||
"eval_runtime": 26.8774,
|
||||
"eval_samples_per_second": 30.807,
|
||||
"eval_steps_per_second": 0.484,
|
||||
"step": 50
|
||||
}
|
||||
],
|
||||
"logging_steps": 1,
|
||||
"max_steps": 195,
|
||||
"num_train_epochs": 5,
|
||||
"save_steps": 25,
|
||||
"total_flos": 1.2823585474593423e+18,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoints/checkpoint-50/training_args.bin
Normal file
3
checkpoints/checkpoint-50/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-50/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-50/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
26
checkpoints/checkpoint-75/config.json
Normal file
26
checkpoints/checkpoint-75/config.json
Normal file
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"_name_or_path": "/workspace/llama-2-7b-hf",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"bos_token_id": 1,
|
||||
"eos_token_id": 2,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 4096,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 11008,
|
||||
"max_position_embeddings": 4096,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 32,
|
||||
"num_key_value_heads": 32,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": null,
|
||||
"rope_theta": 10000.0,
|
||||
"tie_word_embeddings": false,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.33.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 32000
|
||||
}
|
||||
4
checkpoints/checkpoint-75/generation_config.json
Normal file
4
checkpoints/checkpoint-75/generation_config.json
Normal file
@@ -0,0 +1,4 @@
|
||||
{
|
||||
"max_new_tokens": 4096,
|
||||
"transformers_version": "4.33.1"
|
||||
}
|
||||
1
checkpoints/checkpoint-75/latest
Normal file
1
checkpoints/checkpoint-75/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step75
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:81be02943d0fa188b849710aa3efa3a1630e74ca7a0eeb25104527940db855cc
|
||||
size 9976620122
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:118512f545f0793ec20b5993afc63b1ccecbc597ba450a22e89fd70c6aa2752d
|
||||
size 3500310787
|
||||
3
checkpoints/checkpoint-75/pytorch_model.bin.index.json
Normal file
3
checkpoints/checkpoint-75/pytorch_model.bin.index.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e572e08c4d4e81c7916197f6fcd2956a2f05e5919f28d72c9ba4f351efae1e29
|
||||
size 23950
|
||||
24
checkpoints/checkpoint-75/special_tokens_map.json
Normal file
24
checkpoints/checkpoint-75/special_tokens_map.json
Normal file
@@ -0,0 +1,24 @@
|
||||
{
|
||||
"bos_token": {
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<unk>",
|
||||
"unk_token": {
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
}
|
||||
BIN
checkpoints/checkpoint-75/tokenizer.model
(Stored with Git LFS)
Normal file
BIN
checkpoints/checkpoint-75/tokenizer.model
(Stored with Git LFS)
Normal file
Binary file not shown.
37
checkpoints/checkpoint-75/tokenizer_config.json
Normal file
37
checkpoints/checkpoint-75/tokenizer_config.json
Normal file
@@ -0,0 +1,37 @@
|
||||
{
|
||||
"add_bos_token": true,
|
||||
"add_eos_token": false,
|
||||
"bos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"clean_up_tokenization_spaces": false,
|
||||
"eos_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "</s>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"legacy": false,
|
||||
"model_max_length": 1000000000000000019884624838656,
|
||||
"pad_token": null,
|
||||
"padding_side": "right",
|
||||
"sp_model_kwargs": {},
|
||||
"spaces_between_special_tokens": false,
|
||||
"tokenizer_class": "LlamaTokenizer",
|
||||
"unk_token": {
|
||||
"__type": "AddedToken",
|
||||
"content": "<unk>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"use_default_system_prompt": true
|
||||
}
|
||||
525
checkpoints/checkpoint-75/trainer_state.json
Normal file
525
checkpoints/checkpoint-75/trainer_state.json
Normal file
@@ -0,0 +1,525 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 1.8957345971563981,
|
||||
"eval_steps": 10,
|
||||
"global_step": 75,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.03,
|
||||
"learning_rate": 2.2e-05,
|
||||
"loss": 0.9856,
|
||||
"step": 1
|
||||
},
|
||||
{
|
||||
"epoch": 0.05,
|
||||
"learning_rate": 2.1998557720321954e-05,
|
||||
"loss": 0.8657,
|
||||
"step": 2
|
||||
},
|
||||
{
|
||||
"epoch": 0.08,
|
||||
"learning_rate": 2.1994231259500664e-05,
|
||||
"loss": 0.7599,
|
||||
"step": 3
|
||||
},
|
||||
{
|
||||
"epoch": 0.1,
|
||||
"learning_rate": 2.1987021752075498e-05,
|
||||
"loss": 0.6924,
|
||||
"step": 4
|
||||
},
|
||||
{
|
||||
"epoch": 0.13,
|
||||
"learning_rate": 2.1976931088614826e-05,
|
||||
"loss": 0.818,
|
||||
"step": 5
|
||||
},
|
||||
{
|
||||
"epoch": 0.15,
|
||||
"learning_rate": 2.1963961915220263e-05,
|
||||
"loss": 1.2993,
|
||||
"step": 6
|
||||
},
|
||||
{
|
||||
"epoch": 0.18,
|
||||
"learning_rate": 2.1948117632832752e-05,
|
||||
"loss": 1.4107,
|
||||
"step": 7
|
||||
},
|
||||
{
|
||||
"epoch": 0.2,
|
||||
"learning_rate": 2.1929402396340756e-05,
|
||||
"loss": 0.8539,
|
||||
"step": 8
|
||||
},
|
||||
{
|
||||
"epoch": 0.23,
|
||||
"learning_rate": 2.1907821113490676e-05,
|
||||
"loss": 0.7822,
|
||||
"step": 9
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"learning_rate": 2.1883379443599917e-05,
|
||||
"loss": 0.6498,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.25,
|
||||
"eval_loss": 0.777102530002594,
|
||||
"eval_runtime": 26.948,
|
||||
"eval_samples_per_second": 30.726,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.28,
|
||||
"learning_rate": 2.1856083796072795e-05,
|
||||
"loss": 0.7439,
|
||||
"step": 11
|
||||
},
|
||||
{
|
||||
"epoch": 0.3,
|
||||
"learning_rate": 2.182594132871982e-05,
|
||||
"loss": 1.0097,
|
||||
"step": 12
|
||||
},
|
||||
{
|
||||
"epoch": 0.33,
|
||||
"learning_rate": 2.179295994588063e-05,
|
||||
"loss": 1.3189,
|
||||
"step": 13
|
||||
},
|
||||
{
|
||||
"epoch": 0.35,
|
||||
"learning_rate": 2.175714829635128e-05,
|
||||
"loss": 0.7509,
|
||||
"step": 14
|
||||
},
|
||||
{
|
||||
"epoch": 0.38,
|
||||
"learning_rate": 2.1718515771116186e-05,
|
||||
"loss": 0.7273,
|
||||
"step": 15
|
||||
},
|
||||
{
|
||||
"epoch": 0.4,
|
||||
"learning_rate": 2.167707250088555e-05,
|
||||
"loss": 0.6242,
|
||||
"step": 16
|
||||
},
|
||||
{
|
||||
"epoch": 0.43,
|
||||
"learning_rate": 2.1632829353438724e-05,
|
||||
"loss": 0.6135,
|
||||
"step": 17
|
||||
},
|
||||
{
|
||||
"epoch": 0.45,
|
||||
"learning_rate": 2.1585797930774335e-05,
|
||||
"loss": 0.9365,
|
||||
"step": 18
|
||||
},
|
||||
{
|
||||
"epoch": 0.48,
|
||||
"learning_rate": 2.1535990566067858e-05,
|
||||
"loss": 1.4216,
|
||||
"step": 19
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"learning_rate": 2.148342032043747e-05,
|
||||
"loss": 0.7369,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.51,
|
||||
"eval_loss": 0.7404242157936096,
|
||||
"eval_runtime": 26.98,
|
||||
"eval_samples_per_second": 30.689,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.53,
|
||||
"learning_rate": 2.1428100979518976e-05,
|
||||
"loss": 0.7605,
|
||||
"step": 21
|
||||
},
|
||||
{
|
||||
"epoch": 0.56,
|
||||
"learning_rate": 2.1370047049850775e-05,
|
||||
"loss": 0.6095,
|
||||
"step": 22
|
||||
},
|
||||
{
|
||||
"epoch": 0.58,
|
||||
"learning_rate": 2.130927375506978e-05,
|
||||
"loss": 0.5955,
|
||||
"step": 23
|
||||
},
|
||||
{
|
||||
"epoch": 0.61,
|
||||
"learning_rate": 2.1245797031919265e-05,
|
||||
"loss": 0.851,
|
||||
"step": 24
|
||||
},
|
||||
{
|
||||
"epoch": 0.63,
|
||||
"learning_rate": 2.1179633526069748e-05,
|
||||
"loss": 1.324,
|
||||
"step": 25
|
||||
},
|
||||
{
|
||||
"epoch": 0.66,
|
||||
"learning_rate": 2.1110800587753934e-05,
|
||||
"loss": 0.9368,
|
||||
"step": 26
|
||||
},
|
||||
{
|
||||
"epoch": 0.68,
|
||||
"learning_rate": 2.103931626721694e-05,
|
||||
"loss": 0.7307,
|
||||
"step": 27
|
||||
},
|
||||
{
|
||||
"epoch": 0.71,
|
||||
"learning_rate": 2.0965199309982913e-05,
|
||||
"loss": 0.6489,
|
||||
"step": 28
|
||||
},
|
||||
{
|
||||
"epoch": 0.73,
|
||||
"learning_rate": 2.0888469151939347e-05,
|
||||
"loss": 0.536,
|
||||
"step": 29
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"learning_rate": 2.0809145914240356e-05,
|
||||
"loss": 0.732,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.76,
|
||||
"eval_loss": 0.7266919612884521,
|
||||
"eval_runtime": 26.9346,
|
||||
"eval_samples_per_second": 30.741,
|
||||
"eval_steps_per_second": 0.483,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.78,
|
||||
"learning_rate": 2.072725039803026e-05,
|
||||
"loss": 1.1248,
|
||||
"step": 31
|
||||
},
|
||||
{
|
||||
"epoch": 0.81,
|
||||
"learning_rate": 2.0642804078988827e-05,
|
||||
"loss": 1.0917,
|
||||
"step": 32
|
||||
},
|
||||
{
|
||||
"epoch": 0.83,
|
||||
"learning_rate": 2.055582910169966e-05,
|
||||
"loss": 0.735,
|
||||
"step": 33
|
||||
},
|
||||
{
|
||||
"epoch": 0.86,
|
||||
"learning_rate": 2.0466348273843176e-05,
|
||||
"loss": 0.6749,
|
||||
"step": 34
|
||||
},
|
||||
{
|
||||
"epoch": 0.88,
|
||||
"learning_rate": 2.0374385060215662e-05,
|
||||
"loss": 0.5519,
|
||||
"step": 35
|
||||
},
|
||||
{
|
||||
"epoch": 0.91,
|
||||
"learning_rate": 2.027996357657605e-05,
|
||||
"loss": 0.6593,
|
||||
"step": 36
|
||||
},
|
||||
{
|
||||
"epoch": 0.94,
|
||||
"learning_rate": 2.018310858332198e-05,
|
||||
"loss": 0.959,
|
||||
"step": 37
|
||||
},
|
||||
{
|
||||
"epoch": 0.96,
|
||||
"learning_rate": 2.008384547899682e-05,
|
||||
"loss": 1.168,
|
||||
"step": 38
|
||||
},
|
||||
{
|
||||
"epoch": 0.99,
|
||||
"learning_rate": 1.9982200293629318e-05,
|
||||
"loss": 0.6186,
|
||||
"step": 39
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"learning_rate": 1.9878199681907702e-05,
|
||||
"loss": 1.1207,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.01,
|
||||
"eval_loss": 0.7203702926635742,
|
||||
"eval_runtime": 26.9926,
|
||||
"eval_samples_per_second": 30.675,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 1.04,
|
||||
"learning_rate": 1.977187091618993e-05,
|
||||
"loss": 0.7095,
|
||||
"step": 41
|
||||
},
|
||||
{
|
||||
"epoch": 1.06,
|
||||
"learning_rate": 1.9663241879351998e-05,
|
||||
"loss": 0.6781,
|
||||
"step": 42
|
||||
},
|
||||
{
|
||||
"epoch": 1.09,
|
||||
"learning_rate": 1.9552341057476144e-05,
|
||||
"loss": 0.5649,
|
||||
"step": 43
|
||||
},
|
||||
{
|
||||
"epoch": 1.11,
|
||||
"learning_rate": 1.9439197532380845e-05,
|
||||
"loss": 0.5466,
|
||||
"step": 44
|
||||
},
|
||||
{
|
||||
"epoch": 1.14,
|
||||
"learning_rate": 1.9323840973994647e-05,
|
||||
"loss": 0.8721,
|
||||
"step": 45
|
||||
},
|
||||
{
|
||||
"epoch": 1.16,
|
||||
"learning_rate": 1.9206301632575707e-05,
|
||||
"loss": 1.2152,
|
||||
"step": 46
|
||||
},
|
||||
{
|
||||
"epoch": 1.19,
|
||||
"learning_rate": 1.9086610330779205e-05,
|
||||
"loss": 0.6863,
|
||||
"step": 47
|
||||
},
|
||||
{
|
||||
"epoch": 1.21,
|
||||
"learning_rate": 1.8964798455574637e-05,
|
||||
"loss": 0.6792,
|
||||
"step": 48
|
||||
},
|
||||
{
|
||||
"epoch": 1.24,
|
||||
"learning_rate": 1.884089795001512e-05,
|
||||
"loss": 0.5667,
|
||||
"step": 49
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"learning_rate": 1.871494130486089e-05,
|
||||
"loss": 0.5256,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.26,
|
||||
"eval_loss": 0.6949857473373413,
|
||||
"eval_runtime": 26.8774,
|
||||
"eval_samples_per_second": 30.807,
|
||||
"eval_steps_per_second": 0.484,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 1.29,
|
||||
"learning_rate": 1.8586961550059147e-05,
|
||||
"loss": 0.7481,
|
||||
"step": 51
|
||||
},
|
||||
{
|
||||
"epoch": 1.31,
|
||||
"learning_rate": 1.8456992246082534e-05,
|
||||
"loss": 1.2028,
|
||||
"step": 52
|
||||
},
|
||||
{
|
||||
"epoch": 1.34,
|
||||
"learning_rate": 1.8325067475128493e-05,
|
||||
"loss": 0.8791,
|
||||
"step": 53
|
||||
},
|
||||
{
|
||||
"epoch": 1.36,
|
||||
"learning_rate": 1.8191221832181777e-05,
|
||||
"loss": 0.6651,
|
||||
"step": 54
|
||||
},
|
||||
{
|
||||
"epoch": 1.39,
|
||||
"learning_rate": 1.8055490415942538e-05,
|
||||
"loss": 0.5875,
|
||||
"step": 55
|
||||
},
|
||||
{
|
||||
"epoch": 1.42,
|
||||
"learning_rate": 1.791790881962229e-05,
|
||||
"loss": 0.5143,
|
||||
"step": 56
|
||||
},
|
||||
{
|
||||
"epoch": 1.44,
|
||||
"learning_rate": 1.77785131216102e-05,
|
||||
"loss": 0.6447,
|
||||
"step": 57
|
||||
},
|
||||
{
|
||||
"epoch": 1.47,
|
||||
"learning_rate": 1.7637339876012172e-05,
|
||||
"loss": 1.0286,
|
||||
"step": 58
|
||||
},
|
||||
{
|
||||
"epoch": 1.49,
|
||||
"learning_rate": 1.749442610306515e-05,
|
||||
"loss": 1.1227,
|
||||
"step": 59
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"learning_rate": 1.7349809279429216e-05,
|
||||
"loss": 0.6634,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.52,
|
||||
"eval_loss": 0.6882586479187012,
|
||||
"eval_runtime": 26.9816,
|
||||
"eval_samples_per_second": 30.688,
|
||||
"eval_steps_per_second": 0.482,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 1.54,
|
||||
"learning_rate": 1.7203527328359972e-05,
|
||||
"loss": 0.6145,
|
||||
"step": 61
|
||||
},
|
||||
{
|
||||
"epoch": 1.57,
|
||||
"learning_rate": 1.705561860976384e-05,
|
||||
"loss": 0.5139,
|
||||
"step": 62
|
||||
},
|
||||
{
|
||||
"epoch": 1.59,
|
||||
"learning_rate": 1.6906121910138824e-05,
|
||||
"loss": 0.5819,
|
||||
"step": 63
|
||||
},
|
||||
{
|
||||
"epoch": 1.62,
|
||||
"learning_rate": 1.675507643240343e-05,
|
||||
"loss": 0.8716,
|
||||
"step": 64
|
||||
},
|
||||
{
|
||||
"epoch": 1.64,
|
||||
"learning_rate": 1.6602521785616394e-05,
|
||||
"loss": 1.2125,
|
||||
"step": 65
|
||||
},
|
||||
{
|
||||
"epoch": 1.67,
|
||||
"learning_rate": 1.6448497974589863e-05,
|
||||
"loss": 0.6803,
|
||||
"step": 66
|
||||
},
|
||||
{
|
||||
"epoch": 1.69,
|
||||
"learning_rate": 1.6293045389398855e-05,
|
||||
"loss": 0.6762,
|
||||
"step": 67
|
||||
},
|
||||
{
|
||||
"epoch": 1.72,
|
||||
"learning_rate": 1.6136204794789646e-05,
|
||||
"loss": 0.5324,
|
||||
"step": 68
|
||||
},
|
||||
{
|
||||
"epoch": 1.74,
|
||||
"learning_rate": 1.597801731948993e-05,
|
||||
"loss": 0.541,
|
||||
"step": 69
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"learning_rate": 1.5818524445423504e-05,
|
||||
"loss": 0.8724,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.77,
|
||||
"eval_loss": 0.6789196729660034,
|
||||
"eval_runtime": 26.8064,
|
||||
"eval_samples_per_second": 30.888,
|
||||
"eval_steps_per_second": 0.485,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 1.79,
|
||||
"learning_rate": 1.5657767996832395e-05,
|
||||
"loss": 1.204,
|
||||
"step": 71
|
||||
},
|
||||
{
|
||||
"epoch": 1.82,
|
||||
"learning_rate": 1.549579012930912e-05,
|
||||
"loss": 0.6885,
|
||||
"step": 72
|
||||
},
|
||||
{
|
||||
"epoch": 1.85,
|
||||
"learning_rate": 1.5332633318742157e-05,
|
||||
"loss": 0.6975,
|
||||
"step": 73
|
||||
},
|
||||
{
|
||||
"epoch": 1.87,
|
||||
"learning_rate": 1.5168340350177374e-05,
|
||||
"loss": 0.5451,
|
||||
"step": 74
|
||||
},
|
||||
{
|
||||
"epoch": 1.9,
|
||||
"learning_rate": 1.5002954306598364e-05,
|
||||
"loss": 0.4803,
|
||||
"step": 75
|
||||
}
|
||||
],
|
||||
"logging_steps": 1,
|
||||
"max_steps": 195,
|
||||
"num_train_epochs": 5,
|
||||
"save_steps": 25,
|
||||
"total_flos": 1.905812798785454e+18,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoints/checkpoint-75/training_args.bin
Normal file
3
checkpoints/checkpoint-75/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:0fff9d08f1bfe26e9db148360b79291f364e4e7b3de0ef25e1877fe4e00d98a6
|
||||
size 6907
|
||||
587
checkpoints/checkpoint-75/zero_to_fp32.py
Normal file
587
checkpoints/checkpoint-75/zero_to_fp32.py
Normal file
@@ -0,0 +1,587 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright (c) Microsoft Corporation.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
# DeepSpeed Team
|
||||
|
||||
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
||||
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
||||
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
||||
# application.
|
||||
#
|
||||
# example: python zero_to_fp32.py . pytorch_model.bin
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
from collections import OrderedDict
|
||||
from dataclasses import dataclass
|
||||
|
||||
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
||||
# DeepSpeed data structures it has to be available in the current python environment.
|
||||
from deepspeed.utils import logger
|
||||
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
||||
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
||||
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
||||
|
||||
|
||||
@dataclass
|
||||
class zero_model_state:
|
||||
buffers: dict()
|
||||
param_shapes: dict()
|
||||
shared_params: list
|
||||
ds_version: int
|
||||
frozen_param_shapes: dict()
|
||||
frozen_param_fragments: dict()
|
||||
|
||||
|
||||
debug = 0
|
||||
|
||||
# load to cpu
|
||||
device = torch.device('cpu')
|
||||
|
||||
|
||||
def atoi(text):
|
||||
return int(text) if text.isdigit() else text
|
||||
|
||||
|
||||
def natural_keys(text):
|
||||
'''
|
||||
alist.sort(key=natural_keys) sorts in human order
|
||||
http://nedbatchelder.com/blog/200712/human_sorting.html
|
||||
(See Toothy's implementation in the comments)
|
||||
'''
|
||||
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
||||
|
||||
|
||||
def get_model_state_file(checkpoint_dir, zero_stage):
|
||||
if not os.path.isdir(checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
||||
|
||||
# there should be only one file
|
||||
if zero_stage <= 2:
|
||||
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
||||
elif zero_stage == 3:
|
||||
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
||||
|
||||
if not os.path.exists(file):
|
||||
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
||||
|
||||
return file
|
||||
|
||||
|
||||
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
||||
# XXX: need to test that this simple glob rule works for multi-node setup too
|
||||
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
||||
|
||||
if len(ckpt_files) == 0:
|
||||
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
||||
|
||||
return ckpt_files
|
||||
|
||||
|
||||
def get_optim_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
||||
|
||||
|
||||
def get_model_state_files(checkpoint_dir):
|
||||
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
||||
|
||||
|
||||
def parse_model_states(files):
|
||||
zero_model_states = []
|
||||
for file in files:
|
||||
state_dict = torch.load(file, map_location=device)
|
||||
|
||||
if BUFFER_NAMES not in state_dict:
|
||||
raise ValueError(f"{file} is not a model state checkpoint")
|
||||
buffer_names = state_dict[BUFFER_NAMES]
|
||||
if debug:
|
||||
print("Found buffers:", buffer_names)
|
||||
|
||||
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
||||
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
||||
param_shapes = state_dict[PARAM_SHAPES]
|
||||
|
||||
# collect parameters that are included in param_shapes
|
||||
param_names = []
|
||||
for s in param_shapes:
|
||||
for name in s.keys():
|
||||
param_names.append(name)
|
||||
|
||||
# update with frozen parameters
|
||||
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
||||
if frozen_param_shapes is not None:
|
||||
if debug:
|
||||
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
||||
param_names += list(frozen_param_shapes.keys())
|
||||
|
||||
# handle shared params
|
||||
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
||||
|
||||
ds_version = state_dict.get(DS_VERSION, None)
|
||||
|
||||
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
||||
|
||||
z_model_state = zero_model_state(buffers=buffers,
|
||||
param_shapes=param_shapes,
|
||||
shared_params=shared_params,
|
||||
ds_version=ds_version,
|
||||
frozen_param_shapes=frozen_param_shapes,
|
||||
frozen_param_fragments=frozen_param_fragments)
|
||||
zero_model_states.append(z_model_state)
|
||||
|
||||
return zero_model_states
|
||||
|
||||
|
||||
def parse_optim_states(files, ds_checkpoint_dir):
|
||||
|
||||
total_files = len(files)
|
||||
state_dicts = []
|
||||
for f in files:
|
||||
state_dict = torch.load(f, map_location=device)
|
||||
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
||||
# and also handle the case where it was already removed by another helper script
|
||||
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
||||
state_dicts.append(state_dict)
|
||||
|
||||
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
||||
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
||||
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
||||
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
||||
|
||||
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
||||
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
||||
# use the max of the partition_count to get the dp world_size.
|
||||
|
||||
if type(world_size) is list:
|
||||
world_size = max(world_size)
|
||||
|
||||
if world_size != total_files:
|
||||
raise ValueError(
|
||||
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
||||
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
||||
)
|
||||
|
||||
# the groups are named differently in each stage
|
||||
if zero_stage <= 2:
|
||||
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
||||
elif zero_stage == 3:
|
||||
fp32_groups_key = FP32_FLAT_GROUPS
|
||||
else:
|
||||
raise ValueError(f"unknown zero stage {zero_stage}")
|
||||
|
||||
if zero_stage <= 2:
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
||||
elif zero_stage == 3:
|
||||
# if there is more than one param group, there will be multiple flattened tensors - one
|
||||
# flattened tensor per group - for simplicity merge them into a single tensor
|
||||
#
|
||||
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
||||
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
||||
|
||||
fp32_flat_groups = [
|
||||
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
||||
]
|
||||
|
||||
return zero_stage, world_size, fp32_flat_groups
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
||||
"""
|
||||
Returns fp32 state_dict reconstructed from ds checkpoint
|
||||
|
||||
Args:
|
||||
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
||||
|
||||
"""
|
||||
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
||||
|
||||
optim_files = get_optim_files(ds_checkpoint_dir)
|
||||
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
||||
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
||||
|
||||
model_files = get_model_state_files(ds_checkpoint_dir)
|
||||
|
||||
zero_model_states = parse_model_states(model_files)
|
||||
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
||||
|
||||
if zero_stage <= 2:
|
||||
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
|
||||
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
||||
|
||||
if debug:
|
||||
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
state_dict[name] = frozen_param_fragments[name]
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
|
||||
# Reconstruction protocol:
|
||||
#
|
||||
# XXX: document this
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
for j in range(len(fp32_flat_groups[0])):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
||||
|
||||
# XXX: memory usage doubles here (zero2)
|
||||
num_param_groups = len(fp32_flat_groups[0])
|
||||
merged_single_partition_of_fp32_groups = []
|
||||
for i in range(num_param_groups):
|
||||
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
||||
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
||||
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
||||
avail_numel = sum(
|
||||
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
||||
|
||||
if debug:
|
||||
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
||||
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
print(f"Have {avail_numel} numels to process.")
|
||||
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
||||
offset = 0
|
||||
avail_numel = full_single_fp32_vector.numel()
|
||||
for name, shape in shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
if debug:
|
||||
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
||||
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
||||
offset += unpartitioned_numel
|
||||
|
||||
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
||||
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
||||
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
||||
# live optimizer object, so we are checking that the numbers are within the right range
|
||||
align_to = 2 * world_size
|
||||
|
||||
def zero2_align(x):
|
||||
return align_to * math.ceil(x / align_to)
|
||||
|
||||
if debug:
|
||||
print(f"original offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
offset = zero2_align(offset)
|
||||
avail_numel = zero2_align(avail_numel)
|
||||
|
||||
if debug:
|
||||
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
||||
|
||||
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
||||
remainder = unpartitioned_numel % world_size
|
||||
padding_numel = (world_size - remainder) if remainder else 0
|
||||
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
||||
return partitioned_numel, padding_numel
|
||||
|
||||
|
||||
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
||||
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
||||
return
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
||||
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
||||
|
||||
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
||||
wanted_params = len(frozen_param_shapes)
|
||||
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
||||
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
||||
print(f'Frozen params: Have {avail_numel} numels to process.')
|
||||
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
||||
|
||||
total_params = 0
|
||||
total_numel = 0
|
||||
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
||||
total_params += 1
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
|
||||
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
||||
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
||||
param_shapes = zero_model_states[0].param_shapes
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
||||
# param, re-consolidating each param, while dealing with padding if any
|
||||
|
||||
# merge list of dicts, preserving order
|
||||
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
||||
|
||||
if debug:
|
||||
for i in range(world_size):
|
||||
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
||||
|
||||
wanted_params = len(param_shapes)
|
||||
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
||||
# not asserting if there is a mismatch due to possible padding
|
||||
avail_numel = fp32_flat_groups[0].numel() * world_size
|
||||
print(f"Trainable params: Have {avail_numel} numels to process.")
|
||||
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
||||
|
||||
# params
|
||||
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
||||
# out-of-core computing solution
|
||||
offset = 0
|
||||
total_numel = 0
|
||||
total_params = 0
|
||||
for name, shape in param_shapes.items():
|
||||
|
||||
unpartitioned_numel = shape.numel()
|
||||
total_numel += unpartitioned_numel
|
||||
total_params += 1
|
||||
|
||||
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
||||
)
|
||||
|
||||
# XXX: memory usage doubles here
|
||||
state_dict[name] = torch.cat(
|
||||
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
||||
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
||||
offset += partitioned_numel
|
||||
|
||||
offset *= world_size
|
||||
|
||||
# Sanity check
|
||||
if offset != avail_numel:
|
||||
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
||||
|
||||
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
||||
|
||||
|
||||
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
||||
|
||||
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
||||
|
||||
# recover shared parameters
|
||||
for pair in zero_model_states[0].shared_params:
|
||||
if pair[1] in state_dict:
|
||||
state_dict[pair[0]] = state_dict[pair[1]]
|
||||
|
||||
return state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
||||
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
||||
via a model hub.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
||||
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
# do the training and checkpoint saving
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
||||
model = model.cpu() # move to cpu
|
||||
model.load_state_dict(state_dict)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
||||
application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
||||
|
||||
"""
|
||||
if tag is None:
|
||||
latest_path = os.path.join(checkpoint_dir, 'latest')
|
||||
if os.path.isfile(latest_path):
|
||||
with open(latest_path, 'r') as fd:
|
||||
tag = fd.read().strip()
|
||||
else:
|
||||
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
||||
|
||||
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
||||
|
||||
if not os.path.isdir(ds_checkpoint_dir):
|
||||
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
||||
|
||||
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
||||
"""
|
||||
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
||||
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
||||
|
||||
Args:
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
||||
- ``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``
|
||||
"""
|
||||
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
print(f"Saving fp32 state dict to {output_file}")
|
||||
torch.save(state_dict, output_file)
|
||||
|
||||
|
||||
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
||||
"""
|
||||
1. Put the provided model to cpu
|
||||
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
||||
3. Load it into the provided model
|
||||
|
||||
Args:
|
||||
- ``model``: the model object to update
|
||||
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
||||
- ``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``
|
||||
|
||||
Returns:
|
||||
- ``model`: modified model
|
||||
|
||||
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
||||
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
||||
conveniently placed for you in the checkpoint folder.
|
||||
|
||||
A typical usage might be ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
||||
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
||||
# submit to model hub or save the model to share with others
|
||||
|
||||
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
||||
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
||||
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
||||
|
||||
"""
|
||||
logger.info(f"Extracting fp32 weights")
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
||||
|
||||
logger.info(f"Overwriting model with fp32 weights")
|
||||
model = model.cpu()
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
return model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("checkpoint_dir",
|
||||
type=str,
|
||||
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
||||
parser.add_argument(
|
||||
"output_file",
|
||||
type=str,
|
||||
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
||||
parser.add_argument("-t",
|
||||
"--tag",
|
||||
type=str,
|
||||
default=None,
|
||||
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
||||
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
||||
args = parser.parse_args()
|
||||
|
||||
debug = args.debug
|
||||
|
||||
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
||||
Reference in New Issue
Block a user