初始化项目，由ModelHub XC社区提供模型
Model: SidhaarthMurali/llama3.2-1b-gsm8k-full Source: Original Platform
This commit is contained in:
ModelHub XC
40
.gitattributes
vendored
Normal file
40
.gitattributes
vendored
Normal file
@@ -0,0 +1,40 @@
|
||||
*.7z filter=lfs diff=lfs merge=lfs -text
|
||||
*.arrow filter=lfs diff=lfs merge=lfs -text
|
||||
*.bin filter=lfs diff=lfs merge=lfs -text
|
||||
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
||||
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
||||
*.ftz filter=lfs diff=lfs merge=lfs -text
|
||||
*.gz filter=lfs diff=lfs merge=lfs -text
|
||||
*.h5 filter=lfs diff=lfs merge=lfs -text
|
||||
*.joblib filter=lfs diff=lfs merge=lfs -text
|
||||
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
||||
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
||||
*.model filter=lfs diff=lfs merge=lfs -text
|
||||
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
||||
*.npy filter=lfs diff=lfs merge=lfs -text
|
||||
*.npz filter=lfs diff=lfs merge=lfs -text
|
||||
*.onnx filter=lfs diff=lfs merge=lfs -text
|
||||
*.ot filter=lfs diff=lfs merge=lfs -text
|
||||
*.parquet filter=lfs diff=lfs merge=lfs -text
|
||||
*.pb filter=lfs diff=lfs merge=lfs -text
|
||||
*.pickle filter=lfs diff=lfs merge=lfs -text
|
||||
*.pkl filter=lfs diff=lfs merge=lfs -text
|
||||
*.pt filter=lfs diff=lfs merge=lfs -text
|
||||
*.pth filter=lfs diff=lfs merge=lfs -text
|
||||
*.rar filter=lfs diff=lfs merge=lfs -text
|
||||
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
||||
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
||||
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
||||
*.tar filter=lfs diff=lfs merge=lfs -text
|
||||
*.tflite filter=lfs diff=lfs merge=lfs -text
|
||||
*.tgz filter=lfs diff=lfs merge=lfs -text
|
||||
*.wasm filter=lfs diff=lfs merge=lfs -text
|
||||
*.xz filter=lfs diff=lfs merge=lfs -text
|
||||
*.zip filter=lfs diff=lfs merge=lfs -text
|
||||
*.zst filter=lfs diff=lfs merge=lfs -text
|
||||
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
||||
checkpoint-1000/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
||||
checkpoint-1500/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
||||
checkpoint-1632/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
||||
checkpoint-500/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
||||
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
||||
BIN
.ipynb_checkpoints/training_loss-checkpoint.png
Normal file
BIN
.ipynb_checkpoints/training_loss-checkpoint.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 44 KiB |
59
README.md
Normal file
59
README.md
Normal file
@@ -0,0 +1,59 @@
|
||||
---
|
||||
library_name: transformers
|
||||
license: other
|
||||
base_model: meta-llama/Llama-3.2-1B
|
||||
tags:
|
||||
- llama-factory
|
||||
- full
|
||||
- generated_from_trainer
|
||||
model-index:
|
||||
- name: gsm8k
|
||||
results: []
|
||||
---
|
||||
|
||||
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
||||
should probably proofread and complete it, then remove this comment. -->
|
||||
|
||||
# gsm8k
|
||||
|
||||
This model is a fine-tuned version of [meta-llama/Llama-3.2-1B](https://huggingface.co/meta-llama/Llama-3.2-1B) on the identity and the gsm8k datasets.
|
||||
|
||||
## Model description
|
||||
|
||||
More information needed
|
||||
|
||||
## Intended uses & limitations
|
||||
|
||||
More information needed
|
||||
|
||||
## Training and evaluation data
|
||||
|
||||
More information needed
|
||||
|
||||
## Training procedure
|
||||
|
||||
### Training hyperparameters
|
||||
|
||||
The following hyperparameters were used during training:
|
||||
- learning_rate: 1e-05
|
||||
- train_batch_size: 1
|
||||
- eval_batch_size: 8
|
||||
- seed: 42
|
||||
- distributed_type: multi-GPU
|
||||
- gradient_accumulation_steps: 2
|
||||
- total_train_batch_size: 2
|
||||
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
|
||||
- lr_scheduler_type: cosine
|
||||
- lr_scheduler_warmup_ratio: 0.1
|
||||
- num_epochs: 3.0
|
||||
|
||||
### Training results
|
||||
|
||||
|
||||
|
||||
### Framework versions
|
||||
|
||||
- Transformers 4.46.1
|
||||
- Pytorch 2.4.0
|
||||
- Datasets 3.1.0
|
||||
- Tokenizers 0.20.3
|
||||
8
all_results.json
Normal file
8
all_results.json
Normal file
@@ -0,0 +1,8 @@
|
||||
{
|
||||
"epoch": 3.0,
|
||||
"total_flos": 246796222464.0,
|
||||
"train_loss": 0.3023774076490572,
|
||||
"train_runtime": 943.8917,
|
||||
"train_samples_per_second": 3.458,
|
||||
"train_steps_per_second": 1.729
|
||||
}
|
||||
36
checkpoint-1000/config.json
Normal file
36
checkpoint-1000/config.json
Normal file
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"_name_or_path": "meta-llama/Llama-3.2-1B",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"attention_bias": false,
|
||||
"attention_dropout": 0.0,
|
||||
"bos_token_id": 128000,
|
||||
"eos_token_id": 128001,
|
||||
"head_dim": 64,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 2048,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 8192,
|
||||
"max_position_embeddings": 131072,
|
||||
"mlp_bias": false,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 16,
|
||||
"num_key_value_heads": 8,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": {
|
||||
"factor": 32.0,
|
||||
"high_freq_factor": 4.0,
|
||||
"low_freq_factor": 1.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"rope_type": "llama3"
|
||||
},
|
||||
"rope_theta": 500000.0,
|
||||
"tie_word_embeddings": true,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.46.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 128256
|
||||
}
|
||||
9
checkpoint-1000/generation_config.json
Normal file
9
checkpoint-1000/generation_config.json
Normal file
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"bos_token_id": 128000,
|
||||
"do_sample": true,
|
||||
"eos_token_id": 128001,
|
||||
"temperature": 0.6,
|
||||
"top_p": 0.9,
|
||||
"transformers_version": "4.46.1"
|
||||
}
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:892c2d03db456ff21170b647f052b0cf6e1352777d6d579d4beddc9a855b4675
|
||||
size 14829779262
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:5a04be0563c5f386492666542da6a8d1e4c6a1a9682f2b7a829686ba673404cb
|
||||
size 76664
|
||||
1
checkpoint-1000/latest
Normal file
1
checkpoint-1000/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step1000
|
||||
3
checkpoint-1000/model.safetensors
Normal file
3
checkpoint-1000/model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:58f6ff8a12bf515c19d967824c18ff17da812fd6da6052cac5f0ecfb939257fc
|
||||
size 2471645608
|
||||
3
checkpoint-1000/rng_state.pth
Normal file
3
checkpoint-1000/rng_state.pth
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:7c733f603eddfbf8aea2db46dfb96d2d44052ea3c8d772ba82a9011002700581
|
||||
size 14244
|
||||
3
checkpoint-1000/scheduler.pt
Normal file
3
checkpoint-1000/scheduler.pt
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:37c48f28ae63964e8f5659d7f3fce20932fb92cf5ae691f29ab80ac40fc0fdab
|
||||
size 1064
|
||||
33
checkpoint-1000/special_tokens_map.json
Normal file
33
checkpoint-1000/special_tokens_map.json
Normal file
@@ -0,0 +1,33 @@
|
||||
{
|
||||
"additional_special_tokens": [
|
||||
{
|
||||
"content": "<|eot_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
{
|
||||
"content": "<|eom_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
],
|
||||
"bos_token": {
|
||||
"content": "<|begin_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "<|end_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<|end_of_text|>"
|
||||
}
|
||||
BIN
checkpoint-1000/tokenizer.json
(Stored with Git LFS)
Normal file
BIN
checkpoint-1000/tokenizer.json
(Stored with Git LFS)
Normal file
Binary file not shown.
2069
checkpoint-1000/tokenizer_config.json
Normal file
2069
checkpoint-1000/tokenizer_config.json
Normal file
File diff suppressed because it is too large
Load Diff
733
checkpoint-1000/trainer_state.json
Normal file
733
checkpoint-1000/trainer_state.json
Normal file
@@ -0,0 +1,733 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 1.8382352941176472,
|
||||
"eval_steps": 500,
|
||||
"global_step": 1000,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.01838235294117647,
|
||||
"grad_norm": 24.461347652326168,
|
||||
"learning_rate": 6.097560975609757e-07,
|
||||
"loss": 1.2036,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.03676470588235294,
|
||||
"grad_norm": 17.308880580961944,
|
||||
"learning_rate": 1.2195121951219514e-06,
|
||||
"loss": 1.2064,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.05514705882352941,
|
||||
"grad_norm": 29.601720546719697,
|
||||
"learning_rate": 1.8292682926829268e-06,
|
||||
"loss": 1.2741,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.07352941176470588,
|
||||
"grad_norm": 14.115818236245202,
|
||||
"learning_rate": 2.4390243902439027e-06,
|
||||
"loss": 0.7596,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 0.09191176470588236,
|
||||
"grad_norm": 21.73939509164672,
|
||||
"learning_rate": 3.0487804878048782e-06,
|
||||
"loss": 0.6171,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 0.11029411764705882,
|
||||
"grad_norm": 16.96116387460838,
|
||||
"learning_rate": 3.6585365853658537e-06,
|
||||
"loss": 0.794,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 0.12867647058823528,
|
||||
"grad_norm": 15.372652400517632,
|
||||
"learning_rate": 4.268292682926829e-06,
|
||||
"loss": 0.6233,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 0.14705882352941177,
|
||||
"grad_norm": 15.062285237489078,
|
||||
"learning_rate": 4.8780487804878055e-06,
|
||||
"loss": 0.5441,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 0.16544117647058823,
|
||||
"grad_norm": 18.632304280702545,
|
||||
"learning_rate": 5.487804878048781e-06,
|
||||
"loss": 0.5784,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 0.18382352941176472,
|
||||
"grad_norm": 13.945519044585177,
|
||||
"learning_rate": 6.0975609756097564e-06,
|
||||
"loss": 0.4844,
|
||||
"step": 100
|
||||
},
|
||||
{
|
||||
"epoch": 0.20220588235294118,
|
||||
"grad_norm": 6.9751627817275,
|
||||
"learning_rate": 6.707317073170733e-06,
|
||||
"loss": 0.6535,
|
||||
"step": 110
|
||||
},
|
||||
{
|
||||
"epoch": 0.22058823529411764,
|
||||
"grad_norm": 25.188175477401735,
|
||||
"learning_rate": 7.317073170731707e-06,
|
||||
"loss": 0.6897,
|
||||
"step": 120
|
||||
},
|
||||
{
|
||||
"epoch": 0.23897058823529413,
|
||||
"grad_norm": 14.572936912172068,
|
||||
"learning_rate": 7.926829268292685e-06,
|
||||
"loss": 0.4217,
|
||||
"step": 130
|
||||
},
|
||||
{
|
||||
"epoch": 0.25735294117647056,
|
||||
"grad_norm": 14.934836192557746,
|
||||
"learning_rate": 8.536585365853658e-06,
|
||||
"loss": 0.5868,
|
||||
"step": 140
|
||||
},
|
||||
{
|
||||
"epoch": 0.2757352941176471,
|
||||
"grad_norm": 11.450309910668999,
|
||||
"learning_rate": 9.146341463414635e-06,
|
||||
"loss": 0.53,
|
||||
"step": 150
|
||||
},
|
||||
{
|
||||
"epoch": 0.29411764705882354,
|
||||
"grad_norm": 21.677329690439954,
|
||||
"learning_rate": 9.756097560975611e-06,
|
||||
"loss": 0.5538,
|
||||
"step": 160
|
||||
},
|
||||
{
|
||||
"epoch": 0.3125,
|
||||
"grad_norm": 13.221017712308962,
|
||||
"learning_rate": 9.999587822598768e-06,
|
||||
"loss": 0.4616,
|
||||
"step": 170
|
||||
},
|
||||
{
|
||||
"epoch": 0.33088235294117646,
|
||||
"grad_norm": 17.10580370679984,
|
||||
"learning_rate": 9.997069206794246e-06,
|
||||
"loss": 0.5633,
|
||||
"step": 180
|
||||
},
|
||||
{
|
||||
"epoch": 0.3492647058823529,
|
||||
"grad_norm": 12.4225156252439,
|
||||
"learning_rate": 9.992262114666653e-06,
|
||||
"loss": 0.6035,
|
||||
"step": 190
|
||||
},
|
||||
{
|
||||
"epoch": 0.36764705882352944,
|
||||
"grad_norm": 9.012606555821149,
|
||||
"learning_rate": 9.985168747689706e-06,
|
||||
"loss": 0.5822,
|
||||
"step": 200
|
||||
},
|
||||
{
|
||||
"epoch": 0.3860294117647059,
|
||||
"grad_norm": 7.876601539287651,
|
||||
"learning_rate": 9.975792354368019e-06,
|
||||
"loss": 0.3755,
|
||||
"step": 210
|
||||
},
|
||||
{
|
||||
"epoch": 0.40441176470588236,
|
||||
"grad_norm": 6.10882571982772,
|
||||
"learning_rate": 9.964137228749409e-06,
|
||||
"loss": 0.4914,
|
||||
"step": 220
|
||||
},
|
||||
{
|
||||
"epoch": 0.4227941176470588,
|
||||
"grad_norm": 6.537974128609375,
|
||||
"learning_rate": 9.950208708458371e-06,
|
||||
"loss": 0.4604,
|
||||
"step": 230
|
||||
},
|
||||
{
|
||||
"epoch": 0.4411764705882353,
|
||||
"grad_norm": 7.108775542718406,
|
||||
"learning_rate": 9.934013172251654e-06,
|
||||
"loss": 0.5029,
|
||||
"step": 240
|
||||
},
|
||||
{
|
||||
"epoch": 0.45955882352941174,
|
||||
"grad_norm": 13.652621643454376,
|
||||
"learning_rate": 9.915558037097003e-06,
|
||||
"loss": 0.5404,
|
||||
"step": 250
|
||||
},
|
||||
{
|
||||
"epoch": 0.47794117647058826,
|
||||
"grad_norm": 9.685271320615122,
|
||||
"learning_rate": 9.894851754776473e-06,
|
||||
"loss": 0.3555,
|
||||
"step": 260
|
||||
},
|
||||
{
|
||||
"epoch": 0.4963235294117647,
|
||||
"grad_norm": 10.504810722270348,
|
||||
"learning_rate": 9.871903808015812e-06,
|
||||
"loss": 0.4092,
|
||||
"step": 270
|
||||
},
|
||||
{
|
||||
"epoch": 0.5147058823529411,
|
||||
"grad_norm": 12.557835687927717,
|
||||
"learning_rate": 9.846724706141718e-06,
|
||||
"loss": 0.4489,
|
||||
"step": 280
|
||||
},
|
||||
{
|
||||
"epoch": 0.5330882352941176,
|
||||
"grad_norm": 19.32420837447236,
|
||||
"learning_rate": 9.819325980268946e-06,
|
||||
"loss": 0.5297,
|
||||
"step": 290
|
||||
},
|
||||
{
|
||||
"epoch": 0.5514705882352942,
|
||||
"grad_norm": 12.56688753149935,
|
||||
"learning_rate": 9.789720178019483e-06,
|
||||
"loss": 0.4301,
|
||||
"step": 300
|
||||
},
|
||||
{
|
||||
"epoch": 0.5698529411764706,
|
||||
"grad_norm": 8.769304725602016,
|
||||
"learning_rate": 9.757920857776188e-06,
|
||||
"loss": 0.5499,
|
||||
"step": 310
|
||||
},
|
||||
{
|
||||
"epoch": 0.5882352941176471,
|
||||
"grad_norm": 6.742725850222398,
|
||||
"learning_rate": 9.723942582473545e-06,
|
||||
"loss": 0.3723,
|
||||
"step": 320
|
||||
},
|
||||
{
|
||||
"epoch": 0.6066176470588235,
|
||||
"grad_norm": 32.81061857477797,
|
||||
"learning_rate": 9.687800912928362e-06,
|
||||
"loss": 0.6103,
|
||||
"step": 330
|
||||
},
|
||||
{
|
||||
"epoch": 0.625,
|
||||
"grad_norm": 7.250155807815409,
|
||||
"learning_rate": 9.649512400713497e-06,
|
||||
"loss": 0.5561,
|
||||
"step": 340
|
||||
},
|
||||
{
|
||||
"epoch": 0.6433823529411765,
|
||||
"grad_norm": 8.449295772111478,
|
||||
"learning_rate": 9.609094580577825e-06,
|
||||
"loss": 0.4384,
|
||||
"step": 350
|
||||
},
|
||||
{
|
||||
"epoch": 0.6617647058823529,
|
||||
"grad_norm": 8.965901929446623,
|
||||
"learning_rate": 9.566565962415958e-06,
|
||||
"loss": 0.492,
|
||||
"step": 360
|
||||
},
|
||||
{
|
||||
"epoch": 0.6801470588235294,
|
||||
"grad_norm": 10.783592668291883,
|
||||
"learning_rate": 9.521946022791402e-06,
|
||||
"loss": 0.5596,
|
||||
"step": 370
|
||||
},
|
||||
{
|
||||
"epoch": 0.6985294117647058,
|
||||
"grad_norm": 6.524203722153494,
|
||||
"learning_rate": 9.475255196016972e-06,
|
||||
"loss": 0.4657,
|
||||
"step": 380
|
||||
},
|
||||
{
|
||||
"epoch": 0.7169117647058824,
|
||||
"grad_norm": 6.844495651307719,
|
||||
"learning_rate": 9.426514864796648e-06,
|
||||
"loss": 0.547,
|
||||
"step": 390
|
||||
},
|
||||
{
|
||||
"epoch": 0.7352941176470589,
|
||||
"grad_norm": 7.093995062440304,
|
||||
"learning_rate": 9.375747350433044e-06,
|
||||
"loss": 0.4821,
|
||||
"step": 400
|
||||
},
|
||||
{
|
||||
"epoch": 0.7536764705882353,
|
||||
"grad_norm": 13.008619062718342,
|
||||
"learning_rate": 9.322975902605082e-06,
|
||||
"loss": 0.3627,
|
||||
"step": 410
|
||||
},
|
||||
{
|
||||
"epoch": 0.7720588235294118,
|
||||
"grad_norm": 7.658314017103231,
|
||||
"learning_rate": 9.268224688720475e-06,
|
||||
"loss": 0.2779,
|
||||
"step": 420
|
||||
},
|
||||
{
|
||||
"epoch": 0.7904411764705882,
|
||||
"grad_norm": 9.961772688539515,
|
||||
"learning_rate": 9.211518782847932e-06,
|
||||
"loss": 0.684,
|
||||
"step": 430
|
||||
},
|
||||
{
|
||||
"epoch": 0.8088235294117647,
|
||||
"grad_norm": 5.184448250813704,
|
||||
"learning_rate": 9.152884154234147e-06,
|
||||
"loss": 0.3829,
|
||||
"step": 440
|
||||
},
|
||||
{
|
||||
"epoch": 0.8272058823529411,
|
||||
"grad_norm": 16.899974205034713,
|
||||
"learning_rate": 9.092347655410818e-06,
|
||||
"loss": 0.409,
|
||||
"step": 450
|
||||
},
|
||||
{
|
||||
"epoch": 0.8455882352941176,
|
||||
"grad_norm": 9.123647569900998,
|
||||
"learning_rate": 9.029937009897176e-06,
|
||||
"loss": 0.4189,
|
||||
"step": 460
|
||||
},
|
||||
{
|
||||
"epoch": 0.8639705882352942,
|
||||
"grad_norm": 14.941661911314725,
|
||||
"learning_rate": 8.96568079950361e-06,
|
||||
"loss": 0.4846,
|
||||
"step": 470
|
||||
},
|
||||
{
|
||||
"epoch": 0.8823529411764706,
|
||||
"grad_norm": 6.576553080941865,
|
||||
"learning_rate": 8.899608451242233e-06,
|
||||
"loss": 0.4869,
|
||||
"step": 480
|
||||
},
|
||||
{
|
||||
"epoch": 0.9007352941176471,
|
||||
"grad_norm": 9.989273508498824,
|
||||
"learning_rate": 8.83175022385039e-06,
|
||||
"loss": 0.368,
|
||||
"step": 490
|
||||
},
|
||||
{
|
||||
"epoch": 0.9191176470588235,
|
||||
"grad_norm": 5.168977011450082,
|
||||
"learning_rate": 8.762137193933241e-06,
|
||||
"loss": 0.3489,
|
||||
"step": 500
|
||||
},
|
||||
{
|
||||
"epoch": 0.9375,
|
||||
"grad_norm": 9.716727823960543,
|
||||
"learning_rate": 8.690801241731818e-06,
|
||||
"loss": 0.4095,
|
||||
"step": 510
|
||||
},
|
||||
{
|
||||
"epoch": 0.9558823529411765,
|
||||
"grad_norm": 9.305144213419574,
|
||||
"learning_rate": 8.617775036523014e-06,
|
||||
"loss": 0.4641,
|
||||
"step": 520
|
||||
},
|
||||
{
|
||||
"epoch": 0.9742647058823529,
|
||||
"grad_norm": 7.973224876288721,
|
||||
"learning_rate": 8.543092021658259e-06,
|
||||
"loss": 0.3311,
|
||||
"step": 530
|
||||
},
|
||||
{
|
||||
"epoch": 0.9926470588235294,
|
||||
"grad_norm": 7.223605900810482,
|
||||
"learning_rate": 8.466786399247663e-06,
|
||||
"loss": 0.3561,
|
||||
"step": 540
|
||||
},
|
||||
{
|
||||
"epoch": 1.0110294117647058,
|
||||
"grad_norm": 7.955108957559817,
|
||||
"learning_rate": 8.388893114496706e-06,
|
||||
"loss": 0.3474,
|
||||
"step": 550
|
||||
},
|
||||
{
|
||||
"epoch": 1.0294117647058822,
|
||||
"grad_norm": 8.495071396849232,
|
||||
"learning_rate": 8.309447839702583e-06,
|
||||
"loss": 0.2188,
|
||||
"step": 560
|
||||
},
|
||||
{
|
||||
"epoch": 1.0477941176470589,
|
||||
"grad_norm": 6.043931760485918,
|
||||
"learning_rate": 8.228486957917608e-06,
|
||||
"loss": 0.2178,
|
||||
"step": 570
|
||||
},
|
||||
{
|
||||
"epoch": 1.0661764705882353,
|
||||
"grad_norm": 6.438867996730345,
|
||||
"learning_rate": 8.146047546287077e-06,
|
||||
"loss": 0.2783,
|
||||
"step": 580
|
||||
},
|
||||
{
|
||||
"epoch": 1.0845588235294117,
|
||||
"grad_norm": 3.6977346854366107,
|
||||
"learning_rate": 8.062167359069301e-06,
|
||||
"loss": 0.2463,
|
||||
"step": 590
|
||||
},
|
||||
{
|
||||
"epoch": 1.1029411764705883,
|
||||
"grad_norm": 6.660661443962452,
|
||||
"learning_rate": 7.97688481034551e-06,
|
||||
"loss": 0.2634,
|
||||
"step": 600
|
||||
},
|
||||
{
|
||||
"epoch": 1.1213235294117647,
|
||||
"grad_norm": 9.365314735040004,
|
||||
"learning_rate": 7.8902389564276e-06,
|
||||
"loss": 0.294,
|
||||
"step": 610
|
||||
},
|
||||
{
|
||||
"epoch": 1.1397058823529411,
|
||||
"grad_norm": 6.992327039164838,
|
||||
"learning_rate": 7.802269477971771e-06,
|
||||
"loss": 0.273,
|
||||
"step": 620
|
||||
},
|
||||
{
|
||||
"epoch": 1.1580882352941178,
|
||||
"grad_norm": 16.22021668710731,
|
||||
"learning_rate": 7.713016661806213e-06,
|
||||
"loss": 0.2975,
|
||||
"step": 630
|
||||
},
|
||||
{
|
||||
"epoch": 1.1764705882352942,
|
||||
"grad_norm": 5.1105720664815,
|
||||
"learning_rate": 7.622521382481208e-06,
|
||||
"loss": 0.2831,
|
||||
"step": 640
|
||||
},
|
||||
{
|
||||
"epoch": 1.1948529411764706,
|
||||
"grad_norm": 18.745915660915887,
|
||||
"learning_rate": 7.530825083550073e-06,
|
||||
"loss": 0.3548,
|
||||
"step": 650
|
||||
},
|
||||
{
|
||||
"epoch": 1.213235294117647,
|
||||
"grad_norm": 10.329160246000239,
|
||||
"learning_rate": 7.437969758589508e-06,
|
||||
"loss": 0.2454,
|
||||
"step": 660
|
||||
},
|
||||
{
|
||||
"epoch": 1.2316176470588236,
|
||||
"grad_norm": 6.408214730217885,
|
||||
"learning_rate": 7.343997931968068e-06,
|
||||
"loss": 0.2694,
|
||||
"step": 670
|
||||
},
|
||||
{
|
||||
"epoch": 1.25,
|
||||
"grad_norm": 7.486293079129486,
|
||||
"learning_rate": 7.248952639371543e-06,
|
||||
"loss": 0.3091,
|
||||
"step": 680
|
||||
},
|
||||
{
|
||||
"epoch": 1.2683823529411764,
|
||||
"grad_norm": 5.956954386852354,
|
||||
"learning_rate": 7.152877408094179e-06,
|
||||
"loss": 0.2305,
|
||||
"step": 690
|
||||
},
|
||||
{
|
||||
"epoch": 1.2867647058823528,
|
||||
"grad_norm": 8.065888848768314,
|
||||
"learning_rate": 7.055816237104753e-06,
|
||||
"loss": 0.3002,
|
||||
"step": 700
|
||||
},
|
||||
{
|
||||
"epoch": 1.3051470588235294,
|
||||
"grad_norm": 3.322320099042565,
|
||||
"learning_rate": 6.957813576896647e-06,
|
||||
"loss": 0.2868,
|
||||
"step": 710
|
||||
},
|
||||
{
|
||||
"epoch": 1.3235294117647058,
|
||||
"grad_norm": 11.20812027560389,
|
||||
"learning_rate": 6.858914309131131e-06,
|
||||
"loss": 0.2342,
|
||||
"step": 720
|
||||
},
|
||||
{
|
||||
"epoch": 1.3419117647058822,
|
||||
"grad_norm": 11.126428155503467,
|
||||
"learning_rate": 6.7591637260831925e-06,
|
||||
"loss": 0.2143,
|
||||
"step": 730
|
||||
},
|
||||
{
|
||||
"epoch": 1.3602941176470589,
|
||||
"grad_norm": 6.169632411834191,
|
||||
"learning_rate": 6.6586075098993196e-06,
|
||||
"loss": 0.3071,
|
||||
"step": 740
|
||||
},
|
||||
{
|
||||
"epoch": 1.3786764705882353,
|
||||
"grad_norm": 8.952863279335729,
|
||||
"learning_rate": 6.557291711676738e-06,
|
||||
"loss": 0.2977,
|
||||
"step": 750
|
||||
},
|
||||
{
|
||||
"epoch": 1.3970588235294117,
|
||||
"grad_norm": 4.407249523532787,
|
||||
"learning_rate": 6.455262730373673e-06,
|
||||
"loss": 0.2326,
|
||||
"step": 760
|
||||
},
|
||||
{
|
||||
"epoch": 1.4154411764705883,
|
||||
"grad_norm": 11.807741453984933,
|
||||
"learning_rate": 6.352567291560319e-06,
|
||||
"loss": 0.3158,
|
||||
"step": 770
|
||||
},
|
||||
{
|
||||
"epoch": 1.4338235294117647,
|
||||
"grad_norm": 13.900682242664102,
|
||||
"learning_rate": 6.249252426020217e-06,
|
||||
"loss": 0.3301,
|
||||
"step": 780
|
||||
},
|
||||
{
|
||||
"epoch": 1.4522058823529411,
|
||||
"grad_norm": 10.2905193894227,
|
||||
"learning_rate": 6.145365448211866e-06,
|
||||
"loss": 0.2937,
|
||||
"step": 790
|
||||
},
|
||||
{
|
||||
"epoch": 1.4705882352941178,
|
||||
"grad_norm": 5.654029084586675,
|
||||
"learning_rate": 6.040953934600425e-06,
|
||||
"loss": 0.2229,
|
||||
"step": 800
|
||||
},
|
||||
{
|
||||
"epoch": 1.4889705882352942,
|
||||
"grad_norm": 7.884909374384519,
|
||||
"learning_rate": 5.936065701869404e-06,
|
||||
"loss": 0.3169,
|
||||
"step": 810
|
||||
},
|
||||
{
|
||||
"epoch": 1.5073529411764706,
|
||||
"grad_norm": 4.885582156638979,
|
||||
"learning_rate": 5.830748785022369e-06,
|
||||
"loss": 0.257,
|
||||
"step": 820
|
||||
},
|
||||
{
|
||||
"epoch": 1.5257352941176472,
|
||||
"grad_norm": 10.069197658253753,
|
||||
"learning_rate": 5.725051415384657e-06,
|
||||
"loss": 0.2766,
|
||||
"step": 830
|
||||
},
|
||||
{
|
||||
"epoch": 1.5441176470588234,
|
||||
"grad_norm": 8.25370914927739,
|
||||
"learning_rate": 5.619021998515165e-06,
|
||||
"loss": 0.2385,
|
||||
"step": 840
|
||||
},
|
||||
{
|
||||
"epoch": 1.5625,
|
||||
"grad_norm": 4.810424736549512,
|
||||
"learning_rate": 5.51270909203838e-06,
|
||||
"loss": 0.2844,
|
||||
"step": 850
|
||||
},
|
||||
{
|
||||
"epoch": 1.5808823529411766,
|
||||
"grad_norm": 10.0208212747629,
|
||||
"learning_rate": 5.40616138340673e-06,
|
||||
"loss": 0.2924,
|
||||
"step": 860
|
||||
},
|
||||
{
|
||||
"epoch": 1.5992647058823528,
|
||||
"grad_norm": 13.748183320384355,
|
||||
"learning_rate": 5.299427667603516e-06,
|
||||
"loss": 0.2826,
|
||||
"step": 870
|
||||
},
|
||||
{
|
||||
"epoch": 1.6176470588235294,
|
||||
"grad_norm": 2.892654980772666,
|
||||
"learning_rate": 5.1925568247965686e-06,
|
||||
"loss": 0.2106,
|
||||
"step": 880
|
||||
},
|
||||
{
|
||||
"epoch": 1.6360294117647058,
|
||||
"grad_norm": 4.016652056640837,
|
||||
"learning_rate": 5.085597797952906e-06,
|
||||
"loss": 0.3306,
|
||||
"step": 890
|
||||
},
|
||||
{
|
||||
"epoch": 1.6544117647058822,
|
||||
"grad_norm": 4.348213620564617,
|
||||
"learning_rate": 4.97859957042464e-06,
|
||||
"loss": 0.2227,
|
||||
"step": 900
|
||||
},
|
||||
{
|
||||
"epoch": 1.6727941176470589,
|
||||
"grad_norm": 8.314405849156318,
|
||||
"learning_rate": 4.871611143516367e-06,
|
||||
"loss": 0.2275,
|
||||
"step": 910
|
||||
},
|
||||
{
|
||||
"epoch": 1.6911764705882353,
|
||||
"grad_norm": 3.1251931465193747,
|
||||
"learning_rate": 4.7646815140443625e-06,
|
||||
"loss": 0.2966,
|
||||
"step": 920
|
||||
},
|
||||
{
|
||||
"epoch": 1.7095588235294117,
|
||||
"grad_norm": 2.9353246185637287,
|
||||
"learning_rate": 4.657859651897807e-06,
|
||||
"loss": 0.228,
|
||||
"step": 930
|
||||
},
|
||||
{
|
||||
"epoch": 1.7279411764705883,
|
||||
"grad_norm": 4.988890457346432,
|
||||
"learning_rate": 4.551194477612351e-06,
|
||||
"loss": 0.2694,
|
||||
"step": 940
|
||||
},
|
||||
{
|
||||
"epoch": 1.7463235294117647,
|
||||
"grad_norm": 2.1319176010435252,
|
||||
"learning_rate": 4.444734839966306e-06,
|
||||
"loss": 0.1944,
|
||||
"step": 950
|
||||
},
|
||||
{
|
||||
"epoch": 1.7647058823529411,
|
||||
"grad_norm": 4.43095941592791,
|
||||
"learning_rate": 4.338529493609647e-06,
|
||||
"loss": 0.2479,
|
||||
"step": 960
|
||||
},
|
||||
{
|
||||
"epoch": 1.7830882352941178,
|
||||
"grad_norm": 4.620277796996259,
|
||||
"learning_rate": 4.232627076736181e-06,
|
||||
"loss": 0.1819,
|
||||
"step": 970
|
||||
},
|
||||
{
|
||||
"epoch": 1.8014705882352942,
|
||||
"grad_norm": 9.324548402262511,
|
||||
"learning_rate": 4.127076088809e-06,
|
||||
"loss": 0.2514,
|
||||
"step": 980
|
||||
},
|
||||
{
|
||||
"epoch": 1.8198529411764706,
|
||||
"grad_norm": 9.685553327436127,
|
||||
"learning_rate": 4.021924868349493e-06,
|
||||
"loss": 0.24,
|
||||
"step": 990
|
||||
},
|
||||
{
|
||||
"epoch": 1.8382352941176472,
|
||||
"grad_norm": 4.557597261535076,
|
||||
"learning_rate": 3.9172215708000655e-06,
|
||||
"loss": 0.2477,
|
||||
"step": 1000
|
||||
}
|
||||
],
|
||||
"logging_steps": 10,
|
||||
"max_steps": 1632,
|
||||
"num_input_tokens_seen": 0,
|
||||
"num_train_epochs": 3,
|
||||
"save_steps": 500,
|
||||
"stateful_callbacks": {
|
||||
"TrainerControl": {
|
||||
"args": {
|
||||
"should_epoch_stop": false,
|
||||
"should_evaluate": false,
|
||||
"should_log": false,
|
||||
"should_save": true,
|
||||
"should_training_stop": false
|
||||
},
|
||||
"attributes": {}
|
||||
}
|
||||
},
|
||||
"total_flos": 151077814272.0,
|
||||
"train_batch_size": 1,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoint-1000/training_args.bin
Normal file
3
checkpoint-1000/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:1f4bebf5c7141bddbed6a7feb4bf1a35ed3fbf020c7c4dc7ab21f6f30268f7e3
|
||||
size 7352
|
||||
760
checkpoint-1000/zero_to_fp32.py
Normal file
760
checkpoint-1000/zero_to_fp32.py
Normal file
@@ -0,0 +1,760 @@
|
||||
#!/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 . output_dir/
|
||||
# or
|
||||
# python zero_to_fp32.py . output_dir/ --safe_serialization
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import gc
|
||||
import json
|
||||
import numpy as np
|
||||
from tqdm import tqdm
|
||||
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, weights_only=False)
|
||||
|
||||
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 tqdm(files, desc='Loading checkpoint shards'):
|
||||
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
|
||||
# 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}")
|
||||
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
|
||||
"""
|
||||
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,
|
||||
exclude_frozen_parameters)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
||||
exclude_frozen_parameters)
|
||||
|
||||
|
||||
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 _has_callable(obj, fn):
|
||||
attr = getattr(obj, fn, None)
|
||||
return callable(attr)
|
||||
|
||||
|
||||
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() if _has_callable(shape, 'numel') else math.prod(shape)
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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")
|
||||
|
||||
|
||||
class GatheredTensor:
|
||||
"""
|
||||
A pseudo tensor that collects partitioned weights.
|
||||
It is more memory efficient when there are multiple groups.
|
||||
"""
|
||||
|
||||
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
|
||||
self.flat_groups = flat_groups
|
||||
self.flat_groups_offset = flat_groups_offset
|
||||
self.offset = offset
|
||||
self.partitioned_numel = partitioned_numel
|
||||
self.shape = shape
|
||||
self.dtype = self.flat_groups[0][0].dtype
|
||||
|
||||
def contiguous(self):
|
||||
"""
|
||||
Merge partitioned weights from flat_groups into a single tensor.
|
||||
"""
|
||||
end_idx = self.offset + self.partitioned_numel
|
||||
world_size = len(self.flat_groups)
|
||||
pad_flat_param_chunks = []
|
||||
|
||||
for rank_i in range(world_size):
|
||||
# for each rank, we need to collect weights from related group/groups
|
||||
flat_groups_at_rank_i = self.flat_groups[rank_i]
|
||||
start_group_id = None
|
||||
end_group_id = None
|
||||
for group_id in range(len(self.flat_groups_offset)):
|
||||
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
|
||||
start_group_id = group_id
|
||||
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
|
||||
end_group_id = group_id
|
||||
break
|
||||
# collect weights from related group/groups
|
||||
for group_id in range(start_group_id, end_group_id + 1):
|
||||
flat_tensor = flat_groups_at_rank_i[group_id]
|
||||
start_offset = self.offset - self.flat_groups_offset[group_id]
|
||||
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
|
||||
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
|
||||
|
||||
# collect weights from all ranks
|
||||
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
|
||||
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
|
||||
return param
|
||||
|
||||
|
||||
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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
|
||||
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
|
||||
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
|
||||
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}"
|
||||
)
|
||||
|
||||
# memory efficient tensor
|
||||
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
|
||||
state_dict[name] = tensor
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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 to_torch_tensor(state_dict, return_empty_tensor=False):
|
||||
"""
|
||||
Convert state_dict of GatheredTensor to torch tensor
|
||||
"""
|
||||
torch_state_dict = {}
|
||||
converted_tensors = {}
|
||||
for name, tensor in state_dict.items():
|
||||
tensor_id = id(tensor)
|
||||
if tensor_id in converted_tensors: # shared tensors
|
||||
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
|
||||
torch_state_dict[name] = shared_tensor
|
||||
else:
|
||||
converted_tensors[tensor_id] = name
|
||||
if return_empty_tensor:
|
||||
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
|
||||
else:
|
||||
torch_state_dict[name] = tensor.contiguous()
|
||||
return torch_state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False,
|
||||
lazy_mode=False):
|
||||
"""
|
||||
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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
|
||||
Convert the pesduo tensor to torch tensor by ``.contiguous()``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
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.
|
||||
|
||||
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
|
||||
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint. Or you can load state_dict in lazy mode ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
|
||||
for name, lazy_tensor in state_dict.item():
|
||||
tensor = lazy_tensor.contiguous() # to cpu
|
||||
print(name, tensor)
|
||||
# del tensor to release memory if it no longer in use
|
||||
"""
|
||||
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")
|
||||
|
||||
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
||||
if lazy_mode:
|
||||
return state_dict
|
||||
else:
|
||||
return to_torch_tensor(state_dict)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
|
||||
output_dir,
|
||||
max_shard_size="5GB",
|
||||
safe_serialization=False,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False):
|
||||
"""
|
||||
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_dir``: directory to the pytorch fp32 state_dict output files
|
||||
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
|
||||
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
||||
- ``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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
"""
|
||||
|
||||
# Dependency pre-check
|
||||
if safe_serialization:
|
||||
try:
|
||||
from safetensors.torch import save_file
|
||||
except ImportError:
|
||||
print('If you want to use `safe_serialization`, please `pip install safetensors`')
|
||||
raise
|
||||
if max_shard_size is not None:
|
||||
try:
|
||||
from huggingface_hub import split_torch_state_dict_into_shards
|
||||
except ImportError:
|
||||
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
|
||||
raise
|
||||
|
||||
# Convert zero checkpoint to state_dict
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag,
|
||||
exclude_frozen_parameters,
|
||||
lazy_mode=True)
|
||||
|
||||
# Shard the model if it is too big.
|
||||
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
|
||||
if max_shard_size is not None:
|
||||
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
|
||||
# an memory-efficient approach for sharding
|
||||
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
|
||||
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
|
||||
filename_pattern=filename_pattern,
|
||||
max_shard_size=max_shard_size)
|
||||
else:
|
||||
from collections import namedtuple
|
||||
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
|
||||
state_dict_split = StateDictSplit(is_sharded=False,
|
||||
filename_to_tensors={weights_name: list(state_dict.keys())})
|
||||
|
||||
# Save the model by shard
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
filename_to_tensors = state_dict_split.filename_to_tensors.items()
|
||||
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
|
||||
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
|
||||
shard_state_dict = to_torch_tensor(shard_state_dict)
|
||||
output_path = os.path.join(output_dir, shard_file)
|
||||
if safe_serialization:
|
||||
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
|
||||
else:
|
||||
torch.save(shard_state_dict, output_path)
|
||||
# release the memory of current shard
|
||||
for tensor_name in list(shard_state_dict.keys()):
|
||||
del state_dict[tensor_name]
|
||||
del shard_state_dict[tensor_name]
|
||||
del shard_state_dict
|
||||
gc.collect()
|
||||
|
||||
# Save index if sharded
|
||||
if state_dict_split.is_sharded:
|
||||
index = {
|
||||
"metadata": state_dict_split.metadata,
|
||||
"weight_map": state_dict_split.tensor_to_filename,
|
||||
}
|
||||
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
|
||||
save_index_file = os.path.join(output_dir, save_index_file)
|
||||
with open(save_index_file, "w", encoding="utf-8") as f:
|
||||
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
|
||||
f.write(content)
|
||||
|
||||
|
||||
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_dir",
|
||||
type=str,
|
||||
help="directory to the pytorch fp32 state_dict output files"
|
||||
"(e.g. path/checkpoint-12-output/)")
|
||||
parser.add_argument(
|
||||
"--max_shard_size",
|
||||
type=str,
|
||||
default="5GB",
|
||||
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
|
||||
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
|
||||
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
|
||||
"without CPU OOM issues.")
|
||||
parser.add_argument(
|
||||
"--safe_serialization",
|
||||
default=False,
|
||||
action='store_true',
|
||||
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
|
||||
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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
||||
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_dir,
|
||||
max_shard_size=args.max_shard_size,
|
||||
safe_serialization=args.safe_serialization,
|
||||
tag=args.tag,
|
||||
exclude_frozen_parameters=args.exclude_frozen_parameters)
|
||||
36
checkpoint-1500/config.json
Normal file
36
checkpoint-1500/config.json
Normal file
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"_name_or_path": "meta-llama/Llama-3.2-1B",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"attention_bias": false,
|
||||
"attention_dropout": 0.0,
|
||||
"bos_token_id": 128000,
|
||||
"eos_token_id": 128001,
|
||||
"head_dim": 64,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 2048,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 8192,
|
||||
"max_position_embeddings": 131072,
|
||||
"mlp_bias": false,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 16,
|
||||
"num_key_value_heads": 8,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": {
|
||||
"factor": 32.0,
|
||||
"high_freq_factor": 4.0,
|
||||
"low_freq_factor": 1.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"rope_type": "llama3"
|
||||
},
|
||||
"rope_theta": 500000.0,
|
||||
"tie_word_embeddings": true,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.46.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 128256
|
||||
}
|
||||
9
checkpoint-1500/generation_config.json
Normal file
9
checkpoint-1500/generation_config.json
Normal file
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"bos_token_id": 128000,
|
||||
"do_sample": true,
|
||||
"eos_token_id": 128001,
|
||||
"temperature": 0.6,
|
||||
"top_p": 0.9,
|
||||
"transformers_version": "4.46.1"
|
||||
}
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:f1e68b08d66e431b3cd54cd0f89659dff824fc2df3105c3b2e2180ac2da06517
|
||||
size 14829779262
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:a48e8852617ed68d2128d22e238030c2b0bba361e1cfee8f20d5f61a7370aee8
|
||||
size 76664
|
||||
1
checkpoint-1500/latest
Normal file
1
checkpoint-1500/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step1500
|
||||
3
checkpoint-1500/model.safetensors
Normal file
3
checkpoint-1500/model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:2c807005cf21d562786f30bcca039f1908d791df0c8e69013f2cad81e826a57e
|
||||
size 2471645608
|
||||
3
checkpoint-1500/rng_state.pth
Normal file
3
checkpoint-1500/rng_state.pth
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:7b5fa02ce1f2057ffdcccd86019817387a38e6b835bf0ecf5d2277b489b598e7
|
||||
size 14244
|
||||
3
checkpoint-1500/scheduler.pt
Normal file
3
checkpoint-1500/scheduler.pt
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:a7e74e96ab047aa96cdf8dce254abe6ca588ba7199dd2ebe656d636245ccf92a
|
||||
size 1064
|
||||
33
checkpoint-1500/special_tokens_map.json
Normal file
33
checkpoint-1500/special_tokens_map.json
Normal file
@@ -0,0 +1,33 @@
|
||||
{
|
||||
"additional_special_tokens": [
|
||||
{
|
||||
"content": "<|eot_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
{
|
||||
"content": "<|eom_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
],
|
||||
"bos_token": {
|
||||
"content": "<|begin_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "<|end_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<|end_of_text|>"
|
||||
}
|
||||
BIN
checkpoint-1500/tokenizer.json
(Stored with Git LFS)
Normal file
BIN
checkpoint-1500/tokenizer.json
(Stored with Git LFS)
Normal file
Binary file not shown.
2069
checkpoint-1500/tokenizer_config.json
Normal file
2069
checkpoint-1500/tokenizer_config.json
Normal file
File diff suppressed because it is too large
Load Diff
1083
checkpoint-1500/trainer_state.json
Normal file
1083
checkpoint-1500/trainer_state.json
Normal file
File diff suppressed because it is too large
Load Diff
3
checkpoint-1500/training_args.bin
Normal file
3
checkpoint-1500/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:1f4bebf5c7141bddbed6a7feb4bf1a35ed3fbf020c7c4dc7ab21f6f30268f7e3
|
||||
size 7352
|
||||
760
checkpoint-1500/zero_to_fp32.py
Normal file
760
checkpoint-1500/zero_to_fp32.py
Normal file
@@ -0,0 +1,760 @@
|
||||
#!/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 . output_dir/
|
||||
# or
|
||||
# python zero_to_fp32.py . output_dir/ --safe_serialization
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import gc
|
||||
import json
|
||||
import numpy as np
|
||||
from tqdm import tqdm
|
||||
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, weights_only=False)
|
||||
|
||||
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 tqdm(files, desc='Loading checkpoint shards'):
|
||||
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
|
||||
# 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}")
|
||||
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
|
||||
"""
|
||||
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,
|
||||
exclude_frozen_parameters)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
||||
exclude_frozen_parameters)
|
||||
|
||||
|
||||
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 _has_callable(obj, fn):
|
||||
attr = getattr(obj, fn, None)
|
||||
return callable(attr)
|
||||
|
||||
|
||||
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() if _has_callable(shape, 'numel') else math.prod(shape)
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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")
|
||||
|
||||
|
||||
class GatheredTensor:
|
||||
"""
|
||||
A pseudo tensor that collects partitioned weights.
|
||||
It is more memory efficient when there are multiple groups.
|
||||
"""
|
||||
|
||||
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
|
||||
self.flat_groups = flat_groups
|
||||
self.flat_groups_offset = flat_groups_offset
|
||||
self.offset = offset
|
||||
self.partitioned_numel = partitioned_numel
|
||||
self.shape = shape
|
||||
self.dtype = self.flat_groups[0][0].dtype
|
||||
|
||||
def contiguous(self):
|
||||
"""
|
||||
Merge partitioned weights from flat_groups into a single tensor.
|
||||
"""
|
||||
end_idx = self.offset + self.partitioned_numel
|
||||
world_size = len(self.flat_groups)
|
||||
pad_flat_param_chunks = []
|
||||
|
||||
for rank_i in range(world_size):
|
||||
# for each rank, we need to collect weights from related group/groups
|
||||
flat_groups_at_rank_i = self.flat_groups[rank_i]
|
||||
start_group_id = None
|
||||
end_group_id = None
|
||||
for group_id in range(len(self.flat_groups_offset)):
|
||||
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
|
||||
start_group_id = group_id
|
||||
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
|
||||
end_group_id = group_id
|
||||
break
|
||||
# collect weights from related group/groups
|
||||
for group_id in range(start_group_id, end_group_id + 1):
|
||||
flat_tensor = flat_groups_at_rank_i[group_id]
|
||||
start_offset = self.offset - self.flat_groups_offset[group_id]
|
||||
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
|
||||
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
|
||||
|
||||
# collect weights from all ranks
|
||||
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
|
||||
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
|
||||
return param
|
||||
|
||||
|
||||
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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
|
||||
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
|
||||
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
|
||||
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}"
|
||||
)
|
||||
|
||||
# memory efficient tensor
|
||||
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
|
||||
state_dict[name] = tensor
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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 to_torch_tensor(state_dict, return_empty_tensor=False):
|
||||
"""
|
||||
Convert state_dict of GatheredTensor to torch tensor
|
||||
"""
|
||||
torch_state_dict = {}
|
||||
converted_tensors = {}
|
||||
for name, tensor in state_dict.items():
|
||||
tensor_id = id(tensor)
|
||||
if tensor_id in converted_tensors: # shared tensors
|
||||
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
|
||||
torch_state_dict[name] = shared_tensor
|
||||
else:
|
||||
converted_tensors[tensor_id] = name
|
||||
if return_empty_tensor:
|
||||
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
|
||||
else:
|
||||
torch_state_dict[name] = tensor.contiguous()
|
||||
return torch_state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False,
|
||||
lazy_mode=False):
|
||||
"""
|
||||
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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
|
||||
Convert the pesduo tensor to torch tensor by ``.contiguous()``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
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.
|
||||
|
||||
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
|
||||
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint. Or you can load state_dict in lazy mode ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
|
||||
for name, lazy_tensor in state_dict.item():
|
||||
tensor = lazy_tensor.contiguous() # to cpu
|
||||
print(name, tensor)
|
||||
# del tensor to release memory if it no longer in use
|
||||
"""
|
||||
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")
|
||||
|
||||
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
||||
if lazy_mode:
|
||||
return state_dict
|
||||
else:
|
||||
return to_torch_tensor(state_dict)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
|
||||
output_dir,
|
||||
max_shard_size="5GB",
|
||||
safe_serialization=False,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False):
|
||||
"""
|
||||
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_dir``: directory to the pytorch fp32 state_dict output files
|
||||
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
|
||||
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
||||
- ``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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
"""
|
||||
|
||||
# Dependency pre-check
|
||||
if safe_serialization:
|
||||
try:
|
||||
from safetensors.torch import save_file
|
||||
except ImportError:
|
||||
print('If you want to use `safe_serialization`, please `pip install safetensors`')
|
||||
raise
|
||||
if max_shard_size is not None:
|
||||
try:
|
||||
from huggingface_hub import split_torch_state_dict_into_shards
|
||||
except ImportError:
|
||||
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
|
||||
raise
|
||||
|
||||
# Convert zero checkpoint to state_dict
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag,
|
||||
exclude_frozen_parameters,
|
||||
lazy_mode=True)
|
||||
|
||||
# Shard the model if it is too big.
|
||||
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
|
||||
if max_shard_size is not None:
|
||||
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
|
||||
# an memory-efficient approach for sharding
|
||||
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
|
||||
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
|
||||
filename_pattern=filename_pattern,
|
||||
max_shard_size=max_shard_size)
|
||||
else:
|
||||
from collections import namedtuple
|
||||
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
|
||||
state_dict_split = StateDictSplit(is_sharded=False,
|
||||
filename_to_tensors={weights_name: list(state_dict.keys())})
|
||||
|
||||
# Save the model by shard
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
filename_to_tensors = state_dict_split.filename_to_tensors.items()
|
||||
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
|
||||
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
|
||||
shard_state_dict = to_torch_tensor(shard_state_dict)
|
||||
output_path = os.path.join(output_dir, shard_file)
|
||||
if safe_serialization:
|
||||
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
|
||||
else:
|
||||
torch.save(shard_state_dict, output_path)
|
||||
# release the memory of current shard
|
||||
for tensor_name in list(shard_state_dict.keys()):
|
||||
del state_dict[tensor_name]
|
||||
del shard_state_dict[tensor_name]
|
||||
del shard_state_dict
|
||||
gc.collect()
|
||||
|
||||
# Save index if sharded
|
||||
if state_dict_split.is_sharded:
|
||||
index = {
|
||||
"metadata": state_dict_split.metadata,
|
||||
"weight_map": state_dict_split.tensor_to_filename,
|
||||
}
|
||||
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
|
||||
save_index_file = os.path.join(output_dir, save_index_file)
|
||||
with open(save_index_file, "w", encoding="utf-8") as f:
|
||||
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
|
||||
f.write(content)
|
||||
|
||||
|
||||
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_dir",
|
||||
type=str,
|
||||
help="directory to the pytorch fp32 state_dict output files"
|
||||
"(e.g. path/checkpoint-12-output/)")
|
||||
parser.add_argument(
|
||||
"--max_shard_size",
|
||||
type=str,
|
||||
default="5GB",
|
||||
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
|
||||
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
|
||||
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
|
||||
"without CPU OOM issues.")
|
||||
parser.add_argument(
|
||||
"--safe_serialization",
|
||||
default=False,
|
||||
action='store_true',
|
||||
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
|
||||
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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
||||
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_dir,
|
||||
max_shard_size=args.max_shard_size,
|
||||
safe_serialization=args.safe_serialization,
|
||||
tag=args.tag,
|
||||
exclude_frozen_parameters=args.exclude_frozen_parameters)
|
||||
36
checkpoint-1632/config.json
Normal file
36
checkpoint-1632/config.json
Normal file
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"_name_or_path": "meta-llama/Llama-3.2-1B",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"attention_bias": false,
|
||||
"attention_dropout": 0.0,
|
||||
"bos_token_id": 128000,
|
||||
"eos_token_id": 128001,
|
||||
"head_dim": 64,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 2048,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 8192,
|
||||
"max_position_embeddings": 131072,
|
||||
"mlp_bias": false,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 16,
|
||||
"num_key_value_heads": 8,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": {
|
||||
"factor": 32.0,
|
||||
"high_freq_factor": 4.0,
|
||||
"low_freq_factor": 1.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"rope_type": "llama3"
|
||||
},
|
||||
"rope_theta": 500000.0,
|
||||
"tie_word_embeddings": true,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.46.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 128256
|
||||
}
|
||||
9
checkpoint-1632/generation_config.json
Normal file
9
checkpoint-1632/generation_config.json
Normal file
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"bos_token_id": 128000,
|
||||
"do_sample": true,
|
||||
"eos_token_id": 128001,
|
||||
"temperature": 0.6,
|
||||
"top_p": 0.9,
|
||||
"transformers_version": "4.46.1"
|
||||
}
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:46b7ba153e6a773ec5bb4f19f823a41bf3262a6e785936ec134ed20e02588fdf
|
||||
size 14829779262
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:5acab3879626811ad85f5df10c5139869df0cb796888beee5202ff8554e87206
|
||||
size 76664
|
||||
1
checkpoint-1632/latest
Normal file
1
checkpoint-1632/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step1632
|
||||
3
checkpoint-1632/model.safetensors
Normal file
3
checkpoint-1632/model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:297c740fd49c32381d3300f0f09c71cffc651269dd89120d8e28cf8032173667
|
||||
size 2471645608
|
||||
3
checkpoint-1632/rng_state.pth
Normal file
3
checkpoint-1632/rng_state.pth
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:7b5fa02ce1f2057ffdcccd86019817387a38e6b835bf0ecf5d2277b489b598e7
|
||||
size 14244
|
||||
3
checkpoint-1632/scheduler.pt
Normal file
3
checkpoint-1632/scheduler.pt
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:b641a7a33fc204d057f4018cff79d83370a7ec087a5e1a7e9641a81a0c087fd4
|
||||
size 1064
|
||||
33
checkpoint-1632/special_tokens_map.json
Normal file
33
checkpoint-1632/special_tokens_map.json
Normal file
@@ -0,0 +1,33 @@
|
||||
{
|
||||
"additional_special_tokens": [
|
||||
{
|
||||
"content": "<|eot_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
{
|
||||
"content": "<|eom_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
],
|
||||
"bos_token": {
|
||||
"content": "<|begin_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "<|end_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<|end_of_text|>"
|
||||
}
|
||||
BIN
checkpoint-1632/tokenizer.json
(Stored with Git LFS)
Normal file
BIN
checkpoint-1632/tokenizer.json
(Stored with Git LFS)
Normal file
Binary file not shown.
2069
checkpoint-1632/tokenizer_config.json
Normal file
2069
checkpoint-1632/tokenizer_config.json
Normal file
File diff suppressed because it is too large
Load Diff
1174
checkpoint-1632/trainer_state.json
Normal file
1174
checkpoint-1632/trainer_state.json
Normal file
File diff suppressed because it is too large
Load Diff
3
checkpoint-1632/training_args.bin
Normal file
3
checkpoint-1632/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:1f4bebf5c7141bddbed6a7feb4bf1a35ed3fbf020c7c4dc7ab21f6f30268f7e3
|
||||
size 7352
|
||||
760
checkpoint-1632/zero_to_fp32.py
Normal file
760
checkpoint-1632/zero_to_fp32.py
Normal file
@@ -0,0 +1,760 @@
|
||||
#!/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 . output_dir/
|
||||
# or
|
||||
# python zero_to_fp32.py . output_dir/ --safe_serialization
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import gc
|
||||
import json
|
||||
import numpy as np
|
||||
from tqdm import tqdm
|
||||
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, weights_only=False)
|
||||
|
||||
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 tqdm(files, desc='Loading checkpoint shards'):
|
||||
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
|
||||
# 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}")
|
||||
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
|
||||
"""
|
||||
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,
|
||||
exclude_frozen_parameters)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
||||
exclude_frozen_parameters)
|
||||
|
||||
|
||||
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 _has_callable(obj, fn):
|
||||
attr = getattr(obj, fn, None)
|
||||
return callable(attr)
|
||||
|
||||
|
||||
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() if _has_callable(shape, 'numel') else math.prod(shape)
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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")
|
||||
|
||||
|
||||
class GatheredTensor:
|
||||
"""
|
||||
A pseudo tensor that collects partitioned weights.
|
||||
It is more memory efficient when there are multiple groups.
|
||||
"""
|
||||
|
||||
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
|
||||
self.flat_groups = flat_groups
|
||||
self.flat_groups_offset = flat_groups_offset
|
||||
self.offset = offset
|
||||
self.partitioned_numel = partitioned_numel
|
||||
self.shape = shape
|
||||
self.dtype = self.flat_groups[0][0].dtype
|
||||
|
||||
def contiguous(self):
|
||||
"""
|
||||
Merge partitioned weights from flat_groups into a single tensor.
|
||||
"""
|
||||
end_idx = self.offset + self.partitioned_numel
|
||||
world_size = len(self.flat_groups)
|
||||
pad_flat_param_chunks = []
|
||||
|
||||
for rank_i in range(world_size):
|
||||
# for each rank, we need to collect weights from related group/groups
|
||||
flat_groups_at_rank_i = self.flat_groups[rank_i]
|
||||
start_group_id = None
|
||||
end_group_id = None
|
||||
for group_id in range(len(self.flat_groups_offset)):
|
||||
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
|
||||
start_group_id = group_id
|
||||
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
|
||||
end_group_id = group_id
|
||||
break
|
||||
# collect weights from related group/groups
|
||||
for group_id in range(start_group_id, end_group_id + 1):
|
||||
flat_tensor = flat_groups_at_rank_i[group_id]
|
||||
start_offset = self.offset - self.flat_groups_offset[group_id]
|
||||
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
|
||||
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
|
||||
|
||||
# collect weights from all ranks
|
||||
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
|
||||
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
|
||||
return param
|
||||
|
||||
|
||||
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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
|
||||
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
|
||||
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
|
||||
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}"
|
||||
)
|
||||
|
||||
# memory efficient tensor
|
||||
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
|
||||
state_dict[name] = tensor
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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 to_torch_tensor(state_dict, return_empty_tensor=False):
|
||||
"""
|
||||
Convert state_dict of GatheredTensor to torch tensor
|
||||
"""
|
||||
torch_state_dict = {}
|
||||
converted_tensors = {}
|
||||
for name, tensor in state_dict.items():
|
||||
tensor_id = id(tensor)
|
||||
if tensor_id in converted_tensors: # shared tensors
|
||||
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
|
||||
torch_state_dict[name] = shared_tensor
|
||||
else:
|
||||
converted_tensors[tensor_id] = name
|
||||
if return_empty_tensor:
|
||||
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
|
||||
else:
|
||||
torch_state_dict[name] = tensor.contiguous()
|
||||
return torch_state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False,
|
||||
lazy_mode=False):
|
||||
"""
|
||||
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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
|
||||
Convert the pesduo tensor to torch tensor by ``.contiguous()``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
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.
|
||||
|
||||
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
|
||||
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint. Or you can load state_dict in lazy mode ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
|
||||
for name, lazy_tensor in state_dict.item():
|
||||
tensor = lazy_tensor.contiguous() # to cpu
|
||||
print(name, tensor)
|
||||
# del tensor to release memory if it no longer in use
|
||||
"""
|
||||
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")
|
||||
|
||||
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
||||
if lazy_mode:
|
||||
return state_dict
|
||||
else:
|
||||
return to_torch_tensor(state_dict)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
|
||||
output_dir,
|
||||
max_shard_size="5GB",
|
||||
safe_serialization=False,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False):
|
||||
"""
|
||||
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_dir``: directory to the pytorch fp32 state_dict output files
|
||||
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
|
||||
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
||||
- ``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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
"""
|
||||
|
||||
# Dependency pre-check
|
||||
if safe_serialization:
|
||||
try:
|
||||
from safetensors.torch import save_file
|
||||
except ImportError:
|
||||
print('If you want to use `safe_serialization`, please `pip install safetensors`')
|
||||
raise
|
||||
if max_shard_size is not None:
|
||||
try:
|
||||
from huggingface_hub import split_torch_state_dict_into_shards
|
||||
except ImportError:
|
||||
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
|
||||
raise
|
||||
|
||||
# Convert zero checkpoint to state_dict
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag,
|
||||
exclude_frozen_parameters,
|
||||
lazy_mode=True)
|
||||
|
||||
# Shard the model if it is too big.
|
||||
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
|
||||
if max_shard_size is not None:
|
||||
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
|
||||
# an memory-efficient approach for sharding
|
||||
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
|
||||
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
|
||||
filename_pattern=filename_pattern,
|
||||
max_shard_size=max_shard_size)
|
||||
else:
|
||||
from collections import namedtuple
|
||||
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
|
||||
state_dict_split = StateDictSplit(is_sharded=False,
|
||||
filename_to_tensors={weights_name: list(state_dict.keys())})
|
||||
|
||||
# Save the model by shard
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
filename_to_tensors = state_dict_split.filename_to_tensors.items()
|
||||
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
|
||||
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
|
||||
shard_state_dict = to_torch_tensor(shard_state_dict)
|
||||
output_path = os.path.join(output_dir, shard_file)
|
||||
if safe_serialization:
|
||||
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
|
||||
else:
|
||||
torch.save(shard_state_dict, output_path)
|
||||
# release the memory of current shard
|
||||
for tensor_name in list(shard_state_dict.keys()):
|
||||
del state_dict[tensor_name]
|
||||
del shard_state_dict[tensor_name]
|
||||
del shard_state_dict
|
||||
gc.collect()
|
||||
|
||||
# Save index if sharded
|
||||
if state_dict_split.is_sharded:
|
||||
index = {
|
||||
"metadata": state_dict_split.metadata,
|
||||
"weight_map": state_dict_split.tensor_to_filename,
|
||||
}
|
||||
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
|
||||
save_index_file = os.path.join(output_dir, save_index_file)
|
||||
with open(save_index_file, "w", encoding="utf-8") as f:
|
||||
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
|
||||
f.write(content)
|
||||
|
||||
|
||||
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_dir",
|
||||
type=str,
|
||||
help="directory to the pytorch fp32 state_dict output files"
|
||||
"(e.g. path/checkpoint-12-output/)")
|
||||
parser.add_argument(
|
||||
"--max_shard_size",
|
||||
type=str,
|
||||
default="5GB",
|
||||
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
|
||||
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
|
||||
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
|
||||
"without CPU OOM issues.")
|
||||
parser.add_argument(
|
||||
"--safe_serialization",
|
||||
default=False,
|
||||
action='store_true',
|
||||
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
|
||||
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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
||||
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_dir,
|
||||
max_shard_size=args.max_shard_size,
|
||||
safe_serialization=args.safe_serialization,
|
||||
tag=args.tag,
|
||||
exclude_frozen_parameters=args.exclude_frozen_parameters)
|
||||
36
checkpoint-500/config.json
Normal file
36
checkpoint-500/config.json
Normal file
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"_name_or_path": "meta-llama/Llama-3.2-1B",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"attention_bias": false,
|
||||
"attention_dropout": 0.0,
|
||||
"bos_token_id": 128000,
|
||||
"eos_token_id": 128001,
|
||||
"head_dim": 64,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 2048,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 8192,
|
||||
"max_position_embeddings": 131072,
|
||||
"mlp_bias": false,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 16,
|
||||
"num_key_value_heads": 8,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": {
|
||||
"factor": 32.0,
|
||||
"high_freq_factor": 4.0,
|
||||
"low_freq_factor": 1.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"rope_type": "llama3"
|
||||
},
|
||||
"rope_theta": 500000.0,
|
||||
"tie_word_embeddings": true,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.46.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 128256
|
||||
}
|
||||
9
checkpoint-500/generation_config.json
Normal file
9
checkpoint-500/generation_config.json
Normal file
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"bos_token_id": 128000,
|
||||
"do_sample": true,
|
||||
"eos_token_id": 128001,
|
||||
"temperature": 0.6,
|
||||
"top_p": 0.9,
|
||||
"transformers_version": "4.46.1"
|
||||
}
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:f961f3448064dc752f94d8dfd289d74cce46be39b3ec34d5dd06a04b9648ecf6
|
||||
size 14829779262
|
||||
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:39f51449e1fd4b5a48b41f33e928e36d9d078efd55264156360b3ee751e865d0
|
||||
size 76664
|
||||
1
checkpoint-500/latest
Normal file
1
checkpoint-500/latest
Normal file
@@ -0,0 +1 @@
|
||||
global_step500
|
||||
3
checkpoint-500/model.safetensors
Normal file
3
checkpoint-500/model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:fefe2d5a3d03415cdfd9e766365e52e825289ffd058ac4b8e7bd97ead3599ad0
|
||||
size 2471645608
|
||||
3
checkpoint-500/rng_state.pth
Normal file
3
checkpoint-500/rng_state.pth
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:835f869ea325fd6edf27b48b589309fb66641cb92b45f2fc13d1bb6e8814106c
|
||||
size 14244
|
||||
3
checkpoint-500/scheduler.pt
Normal file
3
checkpoint-500/scheduler.pt
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:f7f7a6ec5b6b336abd2fdc77b6cf79ecd28a2794fe661fa2e10edb2faaed0f3a
|
||||
size 1064
|
||||
33
checkpoint-500/special_tokens_map.json
Normal file
33
checkpoint-500/special_tokens_map.json
Normal file
@@ -0,0 +1,33 @@
|
||||
{
|
||||
"additional_special_tokens": [
|
||||
{
|
||||
"content": "<|eot_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
{
|
||||
"content": "<|eom_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
],
|
||||
"bos_token": {
|
||||
"content": "<|begin_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "<|end_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<|end_of_text|>"
|
||||
}
|
||||
BIN
checkpoint-500/tokenizer.json
(Stored with Git LFS)
Normal file
BIN
checkpoint-500/tokenizer.json
(Stored with Git LFS)
Normal file
Binary file not shown.
2069
checkpoint-500/tokenizer_config.json
Normal file
2069
checkpoint-500/tokenizer_config.json
Normal file
File diff suppressed because it is too large
Load Diff
383
checkpoint-500/trainer_state.json
Normal file
383
checkpoint-500/trainer_state.json
Normal file
@@ -0,0 +1,383 @@
|
||||
{
|
||||
"best_metric": null,
|
||||
"best_model_checkpoint": null,
|
||||
"epoch": 0.9191176470588235,
|
||||
"eval_steps": 500,
|
||||
"global_step": 500,
|
||||
"is_hyper_param_search": false,
|
||||
"is_local_process_zero": true,
|
||||
"is_world_process_zero": true,
|
||||
"log_history": [
|
||||
{
|
||||
"epoch": 0.01838235294117647,
|
||||
"grad_norm": 24.461347652326168,
|
||||
"learning_rate": 6.097560975609757e-07,
|
||||
"loss": 1.2036,
|
||||
"step": 10
|
||||
},
|
||||
{
|
||||
"epoch": 0.03676470588235294,
|
||||
"grad_norm": 17.308880580961944,
|
||||
"learning_rate": 1.2195121951219514e-06,
|
||||
"loss": 1.2064,
|
||||
"step": 20
|
||||
},
|
||||
{
|
||||
"epoch": 0.05514705882352941,
|
||||
"grad_norm": 29.601720546719697,
|
||||
"learning_rate": 1.8292682926829268e-06,
|
||||
"loss": 1.2741,
|
||||
"step": 30
|
||||
},
|
||||
{
|
||||
"epoch": 0.07352941176470588,
|
||||
"grad_norm": 14.115818236245202,
|
||||
"learning_rate": 2.4390243902439027e-06,
|
||||
"loss": 0.7596,
|
||||
"step": 40
|
||||
},
|
||||
{
|
||||
"epoch": 0.09191176470588236,
|
||||
"grad_norm": 21.73939509164672,
|
||||
"learning_rate": 3.0487804878048782e-06,
|
||||
"loss": 0.6171,
|
||||
"step": 50
|
||||
},
|
||||
{
|
||||
"epoch": 0.11029411764705882,
|
||||
"grad_norm": 16.96116387460838,
|
||||
"learning_rate": 3.6585365853658537e-06,
|
||||
"loss": 0.794,
|
||||
"step": 60
|
||||
},
|
||||
{
|
||||
"epoch": 0.12867647058823528,
|
||||
"grad_norm": 15.372652400517632,
|
||||
"learning_rate": 4.268292682926829e-06,
|
||||
"loss": 0.6233,
|
||||
"step": 70
|
||||
},
|
||||
{
|
||||
"epoch": 0.14705882352941177,
|
||||
"grad_norm": 15.062285237489078,
|
||||
"learning_rate": 4.8780487804878055e-06,
|
||||
"loss": 0.5441,
|
||||
"step": 80
|
||||
},
|
||||
{
|
||||
"epoch": 0.16544117647058823,
|
||||
"grad_norm": 18.632304280702545,
|
||||
"learning_rate": 5.487804878048781e-06,
|
||||
"loss": 0.5784,
|
||||
"step": 90
|
||||
},
|
||||
{
|
||||
"epoch": 0.18382352941176472,
|
||||
"grad_norm": 13.945519044585177,
|
||||
"learning_rate": 6.0975609756097564e-06,
|
||||
"loss": 0.4844,
|
||||
"step": 100
|
||||
},
|
||||
{
|
||||
"epoch": 0.20220588235294118,
|
||||
"grad_norm": 6.9751627817275,
|
||||
"learning_rate": 6.707317073170733e-06,
|
||||
"loss": 0.6535,
|
||||
"step": 110
|
||||
},
|
||||
{
|
||||
"epoch": 0.22058823529411764,
|
||||
"grad_norm": 25.188175477401735,
|
||||
"learning_rate": 7.317073170731707e-06,
|
||||
"loss": 0.6897,
|
||||
"step": 120
|
||||
},
|
||||
{
|
||||
"epoch": 0.23897058823529413,
|
||||
"grad_norm": 14.572936912172068,
|
||||
"learning_rate": 7.926829268292685e-06,
|
||||
"loss": 0.4217,
|
||||
"step": 130
|
||||
},
|
||||
{
|
||||
"epoch": 0.25735294117647056,
|
||||
"grad_norm": 14.934836192557746,
|
||||
"learning_rate": 8.536585365853658e-06,
|
||||
"loss": 0.5868,
|
||||
"step": 140
|
||||
},
|
||||
{
|
||||
"epoch": 0.2757352941176471,
|
||||
"grad_norm": 11.450309910668999,
|
||||
"learning_rate": 9.146341463414635e-06,
|
||||
"loss": 0.53,
|
||||
"step": 150
|
||||
},
|
||||
{
|
||||
"epoch": 0.29411764705882354,
|
||||
"grad_norm": 21.677329690439954,
|
||||
"learning_rate": 9.756097560975611e-06,
|
||||
"loss": 0.5538,
|
||||
"step": 160
|
||||
},
|
||||
{
|
||||
"epoch": 0.3125,
|
||||
"grad_norm": 13.221017712308962,
|
||||
"learning_rate": 9.999587822598768e-06,
|
||||
"loss": 0.4616,
|
||||
"step": 170
|
||||
},
|
||||
{
|
||||
"epoch": 0.33088235294117646,
|
||||
"grad_norm": 17.10580370679984,
|
||||
"learning_rate": 9.997069206794246e-06,
|
||||
"loss": 0.5633,
|
||||
"step": 180
|
||||
},
|
||||
{
|
||||
"epoch": 0.3492647058823529,
|
||||
"grad_norm": 12.4225156252439,
|
||||
"learning_rate": 9.992262114666653e-06,
|
||||
"loss": 0.6035,
|
||||
"step": 190
|
||||
},
|
||||
{
|
||||
"epoch": 0.36764705882352944,
|
||||
"grad_norm": 9.012606555821149,
|
||||
"learning_rate": 9.985168747689706e-06,
|
||||
"loss": 0.5822,
|
||||
"step": 200
|
||||
},
|
||||
{
|
||||
"epoch": 0.3860294117647059,
|
||||
"grad_norm": 7.876601539287651,
|
||||
"learning_rate": 9.975792354368019e-06,
|
||||
"loss": 0.3755,
|
||||
"step": 210
|
||||
},
|
||||
{
|
||||
"epoch": 0.40441176470588236,
|
||||
"grad_norm": 6.10882571982772,
|
||||
"learning_rate": 9.964137228749409e-06,
|
||||
"loss": 0.4914,
|
||||
"step": 220
|
||||
},
|
||||
{
|
||||
"epoch": 0.4227941176470588,
|
||||
"grad_norm": 6.537974128609375,
|
||||
"learning_rate": 9.950208708458371e-06,
|
||||
"loss": 0.4604,
|
||||
"step": 230
|
||||
},
|
||||
{
|
||||
"epoch": 0.4411764705882353,
|
||||
"grad_norm": 7.108775542718406,
|
||||
"learning_rate": 9.934013172251654e-06,
|
||||
"loss": 0.5029,
|
||||
"step": 240
|
||||
},
|
||||
{
|
||||
"epoch": 0.45955882352941174,
|
||||
"grad_norm": 13.652621643454376,
|
||||
"learning_rate": 9.915558037097003e-06,
|
||||
"loss": 0.5404,
|
||||
"step": 250
|
||||
},
|
||||
{
|
||||
"epoch": 0.47794117647058826,
|
||||
"grad_norm": 9.685271320615122,
|
||||
"learning_rate": 9.894851754776473e-06,
|
||||
"loss": 0.3555,
|
||||
"step": 260
|
||||
},
|
||||
{
|
||||
"epoch": 0.4963235294117647,
|
||||
"grad_norm": 10.504810722270348,
|
||||
"learning_rate": 9.871903808015812e-06,
|
||||
"loss": 0.4092,
|
||||
"step": 270
|
||||
},
|
||||
{
|
||||
"epoch": 0.5147058823529411,
|
||||
"grad_norm": 12.557835687927717,
|
||||
"learning_rate": 9.846724706141718e-06,
|
||||
"loss": 0.4489,
|
||||
"step": 280
|
||||
},
|
||||
{
|
||||
"epoch": 0.5330882352941176,
|
||||
"grad_norm": 19.32420837447236,
|
||||
"learning_rate": 9.819325980268946e-06,
|
||||
"loss": 0.5297,
|
||||
"step": 290
|
||||
},
|
||||
{
|
||||
"epoch": 0.5514705882352942,
|
||||
"grad_norm": 12.56688753149935,
|
||||
"learning_rate": 9.789720178019483e-06,
|
||||
"loss": 0.4301,
|
||||
"step": 300
|
||||
},
|
||||
{
|
||||
"epoch": 0.5698529411764706,
|
||||
"grad_norm": 8.769304725602016,
|
||||
"learning_rate": 9.757920857776188e-06,
|
||||
"loss": 0.5499,
|
||||
"step": 310
|
||||
},
|
||||
{
|
||||
"epoch": 0.5882352941176471,
|
||||
"grad_norm": 6.742725850222398,
|
||||
"learning_rate": 9.723942582473545e-06,
|
||||
"loss": 0.3723,
|
||||
"step": 320
|
||||
},
|
||||
{
|
||||
"epoch": 0.6066176470588235,
|
||||
"grad_norm": 32.81061857477797,
|
||||
"learning_rate": 9.687800912928362e-06,
|
||||
"loss": 0.6103,
|
||||
"step": 330
|
||||
},
|
||||
{
|
||||
"epoch": 0.625,
|
||||
"grad_norm": 7.250155807815409,
|
||||
"learning_rate": 9.649512400713497e-06,
|
||||
"loss": 0.5561,
|
||||
"step": 340
|
||||
},
|
||||
{
|
||||
"epoch": 0.6433823529411765,
|
||||
"grad_norm": 8.449295772111478,
|
||||
"learning_rate": 9.609094580577825e-06,
|
||||
"loss": 0.4384,
|
||||
"step": 350
|
||||
},
|
||||
{
|
||||
"epoch": 0.6617647058823529,
|
||||
"grad_norm": 8.965901929446623,
|
||||
"learning_rate": 9.566565962415958e-06,
|
||||
"loss": 0.492,
|
||||
"step": 360
|
||||
},
|
||||
{
|
||||
"epoch": 0.6801470588235294,
|
||||
"grad_norm": 10.783592668291883,
|
||||
"learning_rate": 9.521946022791402e-06,
|
||||
"loss": 0.5596,
|
||||
"step": 370
|
||||
},
|
||||
{
|
||||
"epoch": 0.6985294117647058,
|
||||
"grad_norm": 6.524203722153494,
|
||||
"learning_rate": 9.475255196016972e-06,
|
||||
"loss": 0.4657,
|
||||
"step": 380
|
||||
},
|
||||
{
|
||||
"epoch": 0.7169117647058824,
|
||||
"grad_norm": 6.844495651307719,
|
||||
"learning_rate": 9.426514864796648e-06,
|
||||
"loss": 0.547,
|
||||
"step": 390
|
||||
},
|
||||
{
|
||||
"epoch": 0.7352941176470589,
|
||||
"grad_norm": 7.093995062440304,
|
||||
"learning_rate": 9.375747350433044e-06,
|
||||
"loss": 0.4821,
|
||||
"step": 400
|
||||
},
|
||||
{
|
||||
"epoch": 0.7536764705882353,
|
||||
"grad_norm": 13.008619062718342,
|
||||
"learning_rate": 9.322975902605082e-06,
|
||||
"loss": 0.3627,
|
||||
"step": 410
|
||||
},
|
||||
{
|
||||
"epoch": 0.7720588235294118,
|
||||
"grad_norm": 7.658314017103231,
|
||||
"learning_rate": 9.268224688720475e-06,
|
||||
"loss": 0.2779,
|
||||
"step": 420
|
||||
},
|
||||
{
|
||||
"epoch": 0.7904411764705882,
|
||||
"grad_norm": 9.961772688539515,
|
||||
"learning_rate": 9.211518782847932e-06,
|
||||
"loss": 0.684,
|
||||
"step": 430
|
||||
},
|
||||
{
|
||||
"epoch": 0.8088235294117647,
|
||||
"grad_norm": 5.184448250813704,
|
||||
"learning_rate": 9.152884154234147e-06,
|
||||
"loss": 0.3829,
|
||||
"step": 440
|
||||
},
|
||||
{
|
||||
"epoch": 0.8272058823529411,
|
||||
"grad_norm": 16.899974205034713,
|
||||
"learning_rate": 9.092347655410818e-06,
|
||||
"loss": 0.409,
|
||||
"step": 450
|
||||
},
|
||||
{
|
||||
"epoch": 0.8455882352941176,
|
||||
"grad_norm": 9.123647569900998,
|
||||
"learning_rate": 9.029937009897176e-06,
|
||||
"loss": 0.4189,
|
||||
"step": 460
|
||||
},
|
||||
{
|
||||
"epoch": 0.8639705882352942,
|
||||
"grad_norm": 14.941661911314725,
|
||||
"learning_rate": 8.96568079950361e-06,
|
||||
"loss": 0.4846,
|
||||
"step": 470
|
||||
},
|
||||
{
|
||||
"epoch": 0.8823529411764706,
|
||||
"grad_norm": 6.576553080941865,
|
||||
"learning_rate": 8.899608451242233e-06,
|
||||
"loss": 0.4869,
|
||||
"step": 480
|
||||
},
|
||||
{
|
||||
"epoch": 0.9007352941176471,
|
||||
"grad_norm": 9.989273508498824,
|
||||
"learning_rate": 8.83175022385039e-06,
|
||||
"loss": 0.368,
|
||||
"step": 490
|
||||
},
|
||||
{
|
||||
"epoch": 0.9191176470588235,
|
||||
"grad_norm": 5.168977011450082,
|
||||
"learning_rate": 8.762137193933241e-06,
|
||||
"loss": 0.3489,
|
||||
"step": 500
|
||||
}
|
||||
],
|
||||
"logging_steps": 10,
|
||||
"max_steps": 1632,
|
||||
"num_input_tokens_seen": 0,
|
||||
"num_train_epochs": 3,
|
||||
"save_steps": 500,
|
||||
"stateful_callbacks": {
|
||||
"TrainerControl": {
|
||||
"args": {
|
||||
"should_epoch_stop": false,
|
||||
"should_evaluate": false,
|
||||
"should_log": false,
|
||||
"should_save": true,
|
||||
"should_training_stop": false
|
||||
},
|
||||
"attributes": {}
|
||||
}
|
||||
},
|
||||
"total_flos": 75890884608.0,
|
||||
"train_batch_size": 1,
|
||||
"trial_name": null,
|
||||
"trial_params": null
|
||||
}
|
||||
3
checkpoint-500/training_args.bin
Normal file
3
checkpoint-500/training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:1f4bebf5c7141bddbed6a7feb4bf1a35ed3fbf020c7c4dc7ab21f6f30268f7e3
|
||||
size 7352
|
||||
760
checkpoint-500/zero_to_fp32.py
Normal file
760
checkpoint-500/zero_to_fp32.py
Normal file
@@ -0,0 +1,760 @@
|
||||
#!/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 . output_dir/
|
||||
# or
|
||||
# python zero_to_fp32.py . output_dir/ --safe_serialization
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
import glob
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import gc
|
||||
import json
|
||||
import numpy as np
|
||||
from tqdm import tqdm
|
||||
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, weights_only=False)
|
||||
|
||||
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 tqdm(files, desc='Loading checkpoint shards'):
|
||||
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
|
||||
# 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}")
|
||||
|
||||
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
|
||||
"""
|
||||
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,
|
||||
exclude_frozen_parameters)
|
||||
elif zero_stage == 3:
|
||||
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
||||
exclude_frozen_parameters)
|
||||
|
||||
|
||||
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 _has_callable(obj, fn):
|
||||
attr = getattr(obj, fn, None)
|
||||
return callable(attr)
|
||||
|
||||
|
||||
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() if _has_callable(shape, 'numel') else math.prod(shape)
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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")
|
||||
|
||||
|
||||
class GatheredTensor:
|
||||
"""
|
||||
A pseudo tensor that collects partitioned weights.
|
||||
It is more memory efficient when there are multiple groups.
|
||||
"""
|
||||
|
||||
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
|
||||
self.flat_groups = flat_groups
|
||||
self.flat_groups_offset = flat_groups_offset
|
||||
self.offset = offset
|
||||
self.partitioned_numel = partitioned_numel
|
||||
self.shape = shape
|
||||
self.dtype = self.flat_groups[0][0].dtype
|
||||
|
||||
def contiguous(self):
|
||||
"""
|
||||
Merge partitioned weights from flat_groups into a single tensor.
|
||||
"""
|
||||
end_idx = self.offset + self.partitioned_numel
|
||||
world_size = len(self.flat_groups)
|
||||
pad_flat_param_chunks = []
|
||||
|
||||
for rank_i in range(world_size):
|
||||
# for each rank, we need to collect weights from related group/groups
|
||||
flat_groups_at_rank_i = self.flat_groups[rank_i]
|
||||
start_group_id = None
|
||||
end_group_id = None
|
||||
for group_id in range(len(self.flat_groups_offset)):
|
||||
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
|
||||
start_group_id = group_id
|
||||
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
|
||||
end_group_id = group_id
|
||||
break
|
||||
# collect weights from related group/groups
|
||||
for group_id in range(start_group_id, end_group_id + 1):
|
||||
flat_tensor = flat_groups_at_rank_i[group_id]
|
||||
start_offset = self.offset - self.flat_groups_offset[group_id]
|
||||
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
|
||||
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
|
||||
|
||||
# collect weights from all ranks
|
||||
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
|
||||
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
|
||||
return param
|
||||
|
||||
|
||||
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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
|
||||
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
|
||||
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
|
||||
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}"
|
||||
)
|
||||
|
||||
# memory efficient tensor
|
||||
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
|
||||
state_dict[name] = tensor
|
||||
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,
|
||||
exclude_frozen_parameters):
|
||||
state_dict = OrderedDict()
|
||||
|
||||
# buffers
|
||||
buffers = zero_model_states[0].buffers
|
||||
state_dict.update(buffers)
|
||||
if debug:
|
||||
print(f"added {len(buffers)} buffers")
|
||||
|
||||
if not exclude_frozen_parameters:
|
||||
_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 to_torch_tensor(state_dict, return_empty_tensor=False):
|
||||
"""
|
||||
Convert state_dict of GatheredTensor to torch tensor
|
||||
"""
|
||||
torch_state_dict = {}
|
||||
converted_tensors = {}
|
||||
for name, tensor in state_dict.items():
|
||||
tensor_id = id(tensor)
|
||||
if tensor_id in converted_tensors: # shared tensors
|
||||
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
|
||||
torch_state_dict[name] = shared_tensor
|
||||
else:
|
||||
converted_tensors[tensor_id] = name
|
||||
if return_empty_tensor:
|
||||
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
|
||||
else:
|
||||
torch_state_dict[name] = tensor.contiguous()
|
||||
return torch_state_dict
|
||||
|
||||
|
||||
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False,
|
||||
lazy_mode=False):
|
||||
"""
|
||||
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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
|
||||
Convert the pesduo tensor to torch tensor by ``.contiguous()``
|
||||
|
||||
Returns:
|
||||
- pytorch ``state_dict``
|
||||
|
||||
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.
|
||||
|
||||
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
|
||||
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
||||
the checkpoint. Or you can load state_dict in lazy mode ::
|
||||
|
||||
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
|
||||
for name, lazy_tensor in state_dict.item():
|
||||
tensor = lazy_tensor.contiguous() # to cpu
|
||||
print(name, tensor)
|
||||
# del tensor to release memory if it no longer in use
|
||||
"""
|
||||
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")
|
||||
|
||||
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
||||
if lazy_mode:
|
||||
return state_dict
|
||||
else:
|
||||
return to_torch_tensor(state_dict)
|
||||
|
||||
|
||||
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
|
||||
output_dir,
|
||||
max_shard_size="5GB",
|
||||
safe_serialization=False,
|
||||
tag=None,
|
||||
exclude_frozen_parameters=False):
|
||||
"""
|
||||
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_dir``: directory to the pytorch fp32 state_dict output files
|
||||
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
|
||||
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
||||
- ``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``
|
||||
- ``exclude_frozen_parameters``: exclude frozen parameters
|
||||
"""
|
||||
|
||||
# Dependency pre-check
|
||||
if safe_serialization:
|
||||
try:
|
||||
from safetensors.torch import save_file
|
||||
except ImportError:
|
||||
print('If you want to use `safe_serialization`, please `pip install safetensors`')
|
||||
raise
|
||||
if max_shard_size is not None:
|
||||
try:
|
||||
from huggingface_hub import split_torch_state_dict_into_shards
|
||||
except ImportError:
|
||||
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
|
||||
raise
|
||||
|
||||
# Convert zero checkpoint to state_dict
|
||||
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
||||
tag,
|
||||
exclude_frozen_parameters,
|
||||
lazy_mode=True)
|
||||
|
||||
# Shard the model if it is too big.
|
||||
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
|
||||
if max_shard_size is not None:
|
||||
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
|
||||
# an memory-efficient approach for sharding
|
||||
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
|
||||
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
|
||||
filename_pattern=filename_pattern,
|
||||
max_shard_size=max_shard_size)
|
||||
else:
|
||||
from collections import namedtuple
|
||||
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
|
||||
state_dict_split = StateDictSplit(is_sharded=False,
|
||||
filename_to_tensors={weights_name: list(state_dict.keys())})
|
||||
|
||||
# Save the model by shard
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
filename_to_tensors = state_dict_split.filename_to_tensors.items()
|
||||
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
|
||||
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
|
||||
shard_state_dict = to_torch_tensor(shard_state_dict)
|
||||
output_path = os.path.join(output_dir, shard_file)
|
||||
if safe_serialization:
|
||||
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
|
||||
else:
|
||||
torch.save(shard_state_dict, output_path)
|
||||
# release the memory of current shard
|
||||
for tensor_name in list(shard_state_dict.keys()):
|
||||
del state_dict[tensor_name]
|
||||
del shard_state_dict[tensor_name]
|
||||
del shard_state_dict
|
||||
gc.collect()
|
||||
|
||||
# Save index if sharded
|
||||
if state_dict_split.is_sharded:
|
||||
index = {
|
||||
"metadata": state_dict_split.metadata,
|
||||
"weight_map": state_dict_split.tensor_to_filename,
|
||||
}
|
||||
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
|
||||
save_index_file = os.path.join(output_dir, save_index_file)
|
||||
with open(save_index_file, "w", encoding="utf-8") as f:
|
||||
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
|
||||
f.write(content)
|
||||
|
||||
|
||||
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_dir",
|
||||
type=str,
|
||||
help="directory to the pytorch fp32 state_dict output files"
|
||||
"(e.g. path/checkpoint-12-output/)")
|
||||
parser.add_argument(
|
||||
"--max_shard_size",
|
||||
type=str,
|
||||
default="5GB",
|
||||
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
|
||||
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
|
||||
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
|
||||
"without CPU OOM issues.")
|
||||
parser.add_argument(
|
||||
"--safe_serialization",
|
||||
default=False,
|
||||
action='store_true',
|
||||
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
|
||||
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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
||||
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_dir,
|
||||
max_shard_size=args.max_shard_size,
|
||||
safe_serialization=args.safe_serialization,
|
||||
tag=args.tag,
|
||||
exclude_frozen_parameters=args.exclude_frozen_parameters)
|
||||
36
config.json
Normal file
36
config.json
Normal file
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"_name_or_path": "meta-llama/Llama-3.2-1B",
|
||||
"architectures": [
|
||||
"LlamaForCausalLM"
|
||||
],
|
||||
"attention_bias": false,
|
||||
"attention_dropout": 0.0,
|
||||
"bos_token_id": 128000,
|
||||
"eos_token_id": 128001,
|
||||
"head_dim": 64,
|
||||
"hidden_act": "silu",
|
||||
"hidden_size": 2048,
|
||||
"initializer_range": 0.02,
|
||||
"intermediate_size": 8192,
|
||||
"max_position_embeddings": 131072,
|
||||
"mlp_bias": false,
|
||||
"model_type": "llama",
|
||||
"num_attention_heads": 32,
|
||||
"num_hidden_layers": 16,
|
||||
"num_key_value_heads": 8,
|
||||
"pretraining_tp": 1,
|
||||
"rms_norm_eps": 1e-05,
|
||||
"rope_scaling": {
|
||||
"factor": 32.0,
|
||||
"high_freq_factor": 4.0,
|
||||
"low_freq_factor": 1.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"rope_type": "llama3"
|
||||
},
|
||||
"rope_theta": 500000.0,
|
||||
"tie_word_embeddings": true,
|
||||
"torch_dtype": "bfloat16",
|
||||
"transformers_version": "4.46.1",
|
||||
"use_cache": false,
|
||||
"vocab_size": 128256
|
||||
}
|
||||
9
generation_config.json
Normal file
9
generation_config.json
Normal file
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"bos_token_id": 128000,
|
||||
"do_sample": true,
|
||||
"eos_token_id": 128001,
|
||||
"temperature": 0.6,
|
||||
"top_p": 0.9,
|
||||
"transformers_version": "4.46.1"
|
||||
}
|
||||
3
model.safetensors
Normal file
3
model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:297c740fd49c32381d3300f0f09c71cffc651269dd89120d8e28cf8032173667
|
||||
size 2471645608
|
||||
33
special_tokens_map.json
Normal file
33
special_tokens_map.json
Normal file
@@ -0,0 +1,33 @@
|
||||
{
|
||||
"additional_special_tokens": [
|
||||
{
|
||||
"content": "<|eot_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
{
|
||||
"content": "<|eom_id|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
}
|
||||
],
|
||||
"bos_token": {
|
||||
"content": "<|begin_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"eos_token": {
|
||||
"content": "<|end_of_text|>",
|
||||
"lstrip": false,
|
||||
"normalized": false,
|
||||
"rstrip": false,
|
||||
"single_word": false
|
||||
},
|
||||
"pad_token": "<|end_of_text|>"
|
||||
}
|
||||
BIN
tokenizer.json
(Stored with Git LFS)
Normal file
BIN
tokenizer.json
(Stored with Git LFS)
Normal file
Binary file not shown.
2069
tokenizer_config.json
Normal file
2069
tokenizer_config.json
Normal file
File diff suppressed because it is too large
Load Diff
8
train_results.json
Normal file
8
train_results.json
Normal file
@@ -0,0 +1,8 @@
|
||||
{
|
||||
"epoch": 3.0,
|
||||
"total_flos": 246796222464.0,
|
||||
"train_loss": 0.3023774076490572,
|
||||
"train_runtime": 943.8917,
|
||||
"train_samples_per_second": 3.458,
|
||||
"train_steps_per_second": 1.729
|
||||
}
|
||||
164
trainer_log.jsonl
Normal file
164
trainer_log.jsonl
Normal file
@@ -0,0 +1,164 @@
|
||||
{"current_steps": 10, "total_steps": 1632, "loss": 1.2036, "lr": 6.097560975609757e-07, "epoch": 0.01838235294117647, "percentage": 0.61, "elapsed_time": "0:00:06", "remaining_time": "0:17:09"}
|
||||
{"current_steps": 20, "total_steps": 1632, "loss": 1.2064, "lr": 1.2195121951219514e-06, "epoch": 0.03676470588235294, "percentage": 1.23, "elapsed_time": "0:00:11", "remaining_time": "0:15:49"}
|
||||
{"current_steps": 30, "total_steps": 1632, "loss": 1.2741, "lr": 1.8292682926829268e-06, "epoch": 0.05514705882352941, "percentage": 1.84, "elapsed_time": "0:00:17", "remaining_time": "0:15:16"}
|
||||
{"current_steps": 40, "total_steps": 1632, "loss": 0.7596, "lr": 2.4390243902439027e-06, "epoch": 0.07352941176470588, "percentage": 2.45, "elapsed_time": "0:00:22", "remaining_time": "0:14:58"}
|
||||
{"current_steps": 50, "total_steps": 1632, "loss": 0.6171, "lr": 3.0487804878048782e-06, "epoch": 0.09191176470588236, "percentage": 3.06, "elapsed_time": "0:00:27", "remaining_time": "0:14:44"}
|
||||
{"current_steps": 60, "total_steps": 1632, "loss": 0.794, "lr": 3.6585365853658537e-06, "epoch": 0.11029411764705882, "percentage": 3.68, "elapsed_time": "0:00:33", "remaining_time": "0:14:33"}
|
||||
{"current_steps": 70, "total_steps": 1632, "loss": 0.6233, "lr": 4.268292682926829e-06, "epoch": 0.12867647058823528, "percentage": 4.29, "elapsed_time": "0:00:38", "remaining_time": "0:14:25"}
|
||||
{"current_steps": 80, "total_steps": 1632, "loss": 0.5441, "lr": 4.8780487804878055e-06, "epoch": 0.14705882352941177, "percentage": 4.9, "elapsed_time": "0:00:44", "remaining_time": "0:14:16"}
|
||||
{"current_steps": 90, "total_steps": 1632, "loss": 0.5784, "lr": 5.487804878048781e-06, "epoch": 0.16544117647058823, "percentage": 5.51, "elapsed_time": "0:00:49", "remaining_time": "0:14:08"}
|
||||
{"current_steps": 100, "total_steps": 1632, "loss": 0.4844, "lr": 6.0975609756097564e-06, "epoch": 0.18382352941176472, "percentage": 6.13, "elapsed_time": "0:00:54", "remaining_time": "0:14:00"}
|
||||
{"current_steps": 110, "total_steps": 1632, "loss": 0.6535, "lr": 6.707317073170733e-06, "epoch": 0.20220588235294118, "percentage": 6.74, "elapsed_time": "0:01:00", "remaining_time": "0:13:53"}
|
||||
{"current_steps": 120, "total_steps": 1632, "loss": 0.6897, "lr": 7.317073170731707e-06, "epoch": 0.22058823529411764, "percentage": 7.35, "elapsed_time": "0:01:05", "remaining_time": "0:13:47"}
|
||||
{"current_steps": 130, "total_steps": 1632, "loss": 0.4217, "lr": 7.926829268292685e-06, "epoch": 0.23897058823529413, "percentage": 7.97, "elapsed_time": "0:01:11", "remaining_time": "0:13:41"}
|
||||
{"current_steps": 140, "total_steps": 1632, "loss": 0.5868, "lr": 8.536585365853658e-06, "epoch": 0.25735294117647056, "percentage": 8.58, "elapsed_time": "0:01:16", "remaining_time": "0:13:35"}
|
||||
{"current_steps": 150, "total_steps": 1632, "loss": 0.53, "lr": 9.146341463414635e-06, "epoch": 0.2757352941176471, "percentage": 9.19, "elapsed_time": "0:01:21", "remaining_time": "0:13:29"}
|
||||
{"current_steps": 160, "total_steps": 1632, "loss": 0.5538, "lr": 9.756097560975611e-06, "epoch": 0.29411764705882354, "percentage": 9.8, "elapsed_time": "0:01:27", "remaining_time": "0:13:23"}
|
||||
{"current_steps": 170, "total_steps": 1632, "loss": 0.4616, "lr": 9.999587822598768e-06, "epoch": 0.3125, "percentage": 10.42, "elapsed_time": "0:01:32", "remaining_time": "0:13:17"}
|
||||
{"current_steps": 180, "total_steps": 1632, "loss": 0.5633, "lr": 9.997069206794246e-06, "epoch": 0.33088235294117646, "percentage": 11.03, "elapsed_time": "0:01:38", "remaining_time": "0:13:11"}
|
||||
{"current_steps": 190, "total_steps": 1632, "loss": 0.6035, "lr": 9.992262114666653e-06, "epoch": 0.3492647058823529, "percentage": 11.64, "elapsed_time": "0:01:43", "remaining_time": "0:13:05"}
|
||||
{"current_steps": 200, "total_steps": 1632, "loss": 0.5822, "lr": 9.985168747689706e-06, "epoch": 0.36764705882352944, "percentage": 12.25, "elapsed_time": "0:01:48", "remaining_time": "0:13:00"}
|
||||
{"current_steps": 210, "total_steps": 1632, "loss": 0.3755, "lr": 9.975792354368019e-06, "epoch": 0.3860294117647059, "percentage": 12.87, "elapsed_time": "0:01:54", "remaining_time": "0:12:54"}
|
||||
{"current_steps": 220, "total_steps": 1632, "loss": 0.4914, "lr": 9.964137228749409e-06, "epoch": 0.40441176470588236, "percentage": 13.48, "elapsed_time": "0:01:59", "remaining_time": "0:12:49"}
|
||||
{"current_steps": 230, "total_steps": 1632, "loss": 0.4604, "lr": 9.950208708458371e-06, "epoch": 0.4227941176470588, "percentage": 14.09, "elapsed_time": "0:02:05", "remaining_time": "0:12:43"}
|
||||
{"current_steps": 240, "total_steps": 1632, "loss": 0.5029, "lr": 9.934013172251654e-06, "epoch": 0.4411764705882353, "percentage": 14.71, "elapsed_time": "0:02:10", "remaining_time": "0:12:37"}
|
||||
{"current_steps": 250, "total_steps": 1632, "loss": 0.5404, "lr": 9.915558037097003e-06, "epoch": 0.45955882352941174, "percentage": 15.32, "elapsed_time": "0:02:16", "remaining_time": "0:12:32"}
|
||||
{"current_steps": 260, "total_steps": 1632, "loss": 0.3555, "lr": 9.894851754776473e-06, "epoch": 0.47794117647058826, "percentage": 15.93, "elapsed_time": "0:02:21", "remaining_time": "0:12:26"}
|
||||
{"current_steps": 270, "total_steps": 1632, "loss": 0.4092, "lr": 9.871903808015812e-06, "epoch": 0.4963235294117647, "percentage": 16.54, "elapsed_time": "0:02:26", "remaining_time": "0:12:21"}
|
||||
{"current_steps": 280, "total_steps": 1632, "loss": 0.4489, "lr": 9.846724706141718e-06, "epoch": 0.5147058823529411, "percentage": 17.16, "elapsed_time": "0:02:32", "remaining_time": "0:12:15"}
|
||||
{"current_steps": 290, "total_steps": 1632, "loss": 0.5297, "lr": 9.819325980268946e-06, "epoch": 0.5330882352941176, "percentage": 17.77, "elapsed_time": "0:02:37", "remaining_time": "0:12:10"}
|
||||
{"current_steps": 300, "total_steps": 1632, "loss": 0.4301, "lr": 9.789720178019483e-06, "epoch": 0.5514705882352942, "percentage": 18.38, "elapsed_time": "0:02:43", "remaining_time": "0:12:04"}
|
||||
{"current_steps": 310, "total_steps": 1632, "loss": 0.5499, "lr": 9.757920857776188e-06, "epoch": 0.5698529411764706, "percentage": 19.0, "elapsed_time": "0:02:48", "remaining_time": "0:11:59"}
|
||||
{"current_steps": 320, "total_steps": 1632, "loss": 0.3723, "lr": 9.723942582473545e-06, "epoch": 0.5882352941176471, "percentage": 19.61, "elapsed_time": "0:02:54", "remaining_time": "0:11:54"}
|
||||
{"current_steps": 330, "total_steps": 1632, "loss": 0.6103, "lr": 9.687800912928362e-06, "epoch": 0.6066176470588235, "percentage": 20.22, "elapsed_time": "0:02:59", "remaining_time": "0:11:48"}
|
||||
{"current_steps": 340, "total_steps": 1632, "loss": 0.5561, "lr": 9.649512400713497e-06, "epoch": 0.625, "percentage": 20.83, "elapsed_time": "0:03:04", "remaining_time": "0:11:42"}
|
||||
{"current_steps": 350, "total_steps": 1632, "loss": 0.4384, "lr": 9.609094580577825e-06, "epoch": 0.6433823529411765, "percentage": 21.45, "elapsed_time": "0:03:10", "remaining_time": "0:11:37"}
|
||||
{"current_steps": 360, "total_steps": 1632, "loss": 0.492, "lr": 9.566565962415958e-06, "epoch": 0.6617647058823529, "percentage": 22.06, "elapsed_time": "0:03:15", "remaining_time": "0:11:31"}
|
||||
{"current_steps": 370, "total_steps": 1632, "loss": 0.5596, "lr": 9.521946022791402e-06, "epoch": 0.6801470588235294, "percentage": 22.67, "elapsed_time": "0:03:21", "remaining_time": "0:11:26"}
|
||||
{"current_steps": 380, "total_steps": 1632, "loss": 0.4657, "lr": 9.475255196016972e-06, "epoch": 0.6985294117647058, "percentage": 23.28, "elapsed_time": "0:03:26", "remaining_time": "0:11:20"}
|
||||
{"current_steps": 390, "total_steps": 1632, "loss": 0.547, "lr": 9.426514864796648e-06, "epoch": 0.7169117647058824, "percentage": 23.9, "elapsed_time": "0:03:32", "remaining_time": "0:11:15"}
|
||||
{"current_steps": 400, "total_steps": 1632, "loss": 0.4821, "lr": 9.375747350433044e-06, "epoch": 0.7352941176470589, "percentage": 24.51, "elapsed_time": "0:03:37", "remaining_time": "0:11:09"}
|
||||
{"current_steps": 410, "total_steps": 1632, "loss": 0.3627, "lr": 9.322975902605082e-06, "epoch": 0.7536764705882353, "percentage": 25.12, "elapsed_time": "0:03:42", "remaining_time": "0:11:04"}
|
||||
{"current_steps": 420, "total_steps": 1632, "loss": 0.2779, "lr": 9.268224688720475e-06, "epoch": 0.7720588235294118, "percentage": 25.74, "elapsed_time": "0:03:48", "remaining_time": "0:10:58"}
|
||||
{"current_steps": 430, "total_steps": 1632, "loss": 0.684, "lr": 9.211518782847932e-06, "epoch": 0.7904411764705882, "percentage": 26.35, "elapsed_time": "0:03:53", "remaining_time": "0:10:53"}
|
||||
{"current_steps": 440, "total_steps": 1632, "loss": 0.3829, "lr": 9.152884154234147e-06, "epoch": 0.8088235294117647, "percentage": 26.96, "elapsed_time": "0:03:59", "remaining_time": "0:10:47"}
|
||||
{"current_steps": 450, "total_steps": 1632, "loss": 0.409, "lr": 9.092347655410818e-06, "epoch": 0.8272058823529411, "percentage": 27.57, "elapsed_time": "0:04:04", "remaining_time": "0:10:42"}
|
||||
{"current_steps": 460, "total_steps": 1632, "loss": 0.4189, "lr": 9.029937009897176e-06, "epoch": 0.8455882352941176, "percentage": 28.19, "elapsed_time": "0:04:09", "remaining_time": "0:10:36"}
|
||||
{"current_steps": 470, "total_steps": 1632, "loss": 0.4846, "lr": 8.96568079950361e-06, "epoch": 0.8639705882352942, "percentage": 28.8, "elapsed_time": "0:04:15", "remaining_time": "0:10:31"}
|
||||
{"current_steps": 480, "total_steps": 1632, "loss": 0.4869, "lr": 8.899608451242233e-06, "epoch": 0.8823529411764706, "percentage": 29.41, "elapsed_time": "0:04:20", "remaining_time": "0:10:25"}
|
||||
{"current_steps": 490, "total_steps": 1632, "loss": 0.368, "lr": 8.83175022385039e-06, "epoch": 0.9007352941176471, "percentage": 30.02, "elapsed_time": "0:04:26", "remaining_time": "0:10:20"}
|
||||
{"current_steps": 500, "total_steps": 1632, "loss": 0.3489, "lr": 8.762137193933241e-06, "epoch": 0.9191176470588235, "percentage": 30.64, "elapsed_time": "0:04:31", "remaining_time": "0:10:14"}
|
||||
{"current_steps": 510, "total_steps": 1632, "loss": 0.4095, "lr": 8.690801241731818e-06, "epoch": 0.9375, "percentage": 31.25, "elapsed_time": "0:04:52", "remaining_time": "0:10:42"}
|
||||
{"current_steps": 520, "total_steps": 1632, "loss": 0.4641, "lr": 8.617775036523014e-06, "epoch": 0.9558823529411765, "percentage": 31.86, "elapsed_time": "0:04:57", "remaining_time": "0:10:36"}
|
||||
{"current_steps": 530, "total_steps": 1632, "loss": 0.3311, "lr": 8.543092021658259e-06, "epoch": 0.9742647058823529, "percentage": 32.48, "elapsed_time": "0:05:03", "remaining_time": "0:10:30"}
|
||||
{"current_steps": 540, "total_steps": 1632, "loss": 0.3561, "lr": 8.466786399247663e-06, "epoch": 0.9926470588235294, "percentage": 33.09, "elapsed_time": "0:05:08", "remaining_time": "0:10:23"}
|
||||
{"current_steps": 550, "total_steps": 1632, "loss": 0.3474, "lr": 8.388893114496706e-06, "epoch": 1.0110294117647058, "percentage": 33.7, "elapsed_time": "0:05:13", "remaining_time": "0:10:17"}
|
||||
{"current_steps": 560, "total_steps": 1632, "loss": 0.2188, "lr": 8.309447839702583e-06, "epoch": 1.0294117647058822, "percentage": 34.31, "elapsed_time": "0:05:19", "remaining_time": "0:10:11"}
|
||||
{"current_steps": 570, "total_steps": 1632, "loss": 0.2178, "lr": 8.228486957917608e-06, "epoch": 1.0477941176470589, "percentage": 34.93, "elapsed_time": "0:05:24", "remaining_time": "0:10:05"}
|
||||
{"current_steps": 580, "total_steps": 1632, "loss": 0.2783, "lr": 8.146047546287077e-06, "epoch": 1.0661764705882353, "percentage": 35.54, "elapsed_time": "0:05:30", "remaining_time": "0:09:58"}
|
||||
{"current_steps": 590, "total_steps": 1632, "loss": 0.2463, "lr": 8.062167359069301e-06, "epoch": 1.0845588235294117, "percentage": 36.15, "elapsed_time": "0:05:35", "remaining_time": "0:09:52"}
|
||||
{"current_steps": 600, "total_steps": 1632, "loss": 0.2634, "lr": 7.97688481034551e-06, "epoch": 1.1029411764705883, "percentage": 36.76, "elapsed_time": "0:05:41", "remaining_time": "0:09:46"}
|
||||
{"current_steps": 610, "total_steps": 1632, "loss": 0.294, "lr": 7.8902389564276e-06, "epoch": 1.1213235294117647, "percentage": 37.38, "elapsed_time": "0:05:46", "remaining_time": "0:09:40"}
|
||||
{"current_steps": 620, "total_steps": 1632, "loss": 0.273, "lr": 7.802269477971771e-06, "epoch": 1.1397058823529411, "percentage": 37.99, "elapsed_time": "0:05:51", "remaining_time": "0:09:34"}
|
||||
{"current_steps": 630, "total_steps": 1632, "loss": 0.2975, "lr": 7.713016661806213e-06, "epoch": 1.1580882352941178, "percentage": 38.6, "elapsed_time": "0:05:57", "remaining_time": "0:09:28"}
|
||||
{"current_steps": 640, "total_steps": 1632, "loss": 0.2831, "lr": 7.622521382481208e-06, "epoch": 1.1764705882352942, "percentage": 39.22, "elapsed_time": "0:06:02", "remaining_time": "0:09:22"}
|
||||
{"current_steps": 650, "total_steps": 1632, "loss": 0.3548, "lr": 7.530825083550073e-06, "epoch": 1.1948529411764706, "percentage": 39.83, "elapsed_time": "0:06:08", "remaining_time": "0:09:16"}
|
||||
{"current_steps": 660, "total_steps": 1632, "loss": 0.2454, "lr": 7.437969758589508e-06, "epoch": 1.213235294117647, "percentage": 40.44, "elapsed_time": "0:06:13", "remaining_time": "0:09:10"}
|
||||
{"current_steps": 670, "total_steps": 1632, "loss": 0.2694, "lr": 7.343997931968068e-06, "epoch": 1.2316176470588236, "percentage": 41.05, "elapsed_time": "0:06:19", "remaining_time": "0:09:04"}
|
||||
{"current_steps": 680, "total_steps": 1632, "loss": 0.3091, "lr": 7.248952639371543e-06, "epoch": 1.25, "percentage": 41.67, "elapsed_time": "0:06:24", "remaining_time": "0:08:58"}
|
||||
{"current_steps": 690, "total_steps": 1632, "loss": 0.2305, "lr": 7.152877408094179e-06, "epoch": 1.2683823529411764, "percentage": 42.28, "elapsed_time": "0:06:29", "remaining_time": "0:08:52"}
|
||||
{"current_steps": 700, "total_steps": 1632, "loss": 0.3002, "lr": 7.055816237104753e-06, "epoch": 1.2867647058823528, "percentage": 42.89, "elapsed_time": "0:06:35", "remaining_time": "0:08:46"}
|
||||
{"current_steps": 710, "total_steps": 1632, "loss": 0.2868, "lr": 6.957813576896647e-06, "epoch": 1.3051470588235294, "percentage": 43.5, "elapsed_time": "0:06:40", "remaining_time": "0:08:40"}
|
||||
{"current_steps": 720, "total_steps": 1632, "loss": 0.2342, "lr": 6.858914309131131e-06, "epoch": 1.3235294117647058, "percentage": 44.12, "elapsed_time": "0:06:45", "remaining_time": "0:08:34"}
|
||||
{"current_steps": 730, "total_steps": 1632, "loss": 0.2143, "lr": 6.7591637260831925e-06, "epoch": 1.3419117647058822, "percentage": 44.73, "elapsed_time": "0:06:51", "remaining_time": "0:08:28"}
|
||||
{"current_steps": 740, "total_steps": 1632, "loss": 0.3071, "lr": 6.6586075098993196e-06, "epoch": 1.3602941176470589, "percentage": 45.34, "elapsed_time": "0:06:56", "remaining_time": "0:08:22"}
|
||||
{"current_steps": 750, "total_steps": 1632, "loss": 0.2977, "lr": 6.557291711676738e-06, "epoch": 1.3786764705882353, "percentage": 45.96, "elapsed_time": "0:07:02", "remaining_time": "0:08:16"}
|
||||
{"current_steps": 760, "total_steps": 1632, "loss": 0.2326, "lr": 6.455262730373673e-06, "epoch": 1.3970588235294117, "percentage": 46.57, "elapsed_time": "0:07:07", "remaining_time": "0:08:10"}
|
||||
{"current_steps": 770, "total_steps": 1632, "loss": 0.3158, "lr": 6.352567291560319e-06, "epoch": 1.4154411764705883, "percentage": 47.18, "elapsed_time": "0:07:12", "remaining_time": "0:08:04"}
|
||||
{"current_steps": 780, "total_steps": 1632, "loss": 0.3301, "lr": 6.249252426020217e-06, "epoch": 1.4338235294117647, "percentage": 47.79, "elapsed_time": "0:07:18", "remaining_time": "0:07:58"}
|
||||
{"current_steps": 790, "total_steps": 1632, "loss": 0.2937, "lr": 6.145365448211866e-06, "epoch": 1.4522058823529411, "percentage": 48.41, "elapsed_time": "0:07:23", "remaining_time": "0:07:53"}
|
||||
{"current_steps": 800, "total_steps": 1632, "loss": 0.2229, "lr": 6.040953934600425e-06, "epoch": 1.4705882352941178, "percentage": 49.02, "elapsed_time": "0:07:29", "remaining_time": "0:07:47"}
|
||||
{"current_steps": 810, "total_steps": 1632, "loss": 0.3169, "lr": 5.936065701869404e-06, "epoch": 1.4889705882352942, "percentage": 49.63, "elapsed_time": "0:07:34", "remaining_time": "0:07:41"}
|
||||
{"current_steps": 820, "total_steps": 1632, "loss": 0.257, "lr": 5.830748785022369e-06, "epoch": 1.5073529411764706, "percentage": 50.25, "elapsed_time": "0:07:40", "remaining_time": "0:07:35"}
|
||||
{"current_steps": 830, "total_steps": 1632, "loss": 0.2766, "lr": 5.725051415384657e-06, "epoch": 1.5257352941176472, "percentage": 50.86, "elapsed_time": "0:07:45", "remaining_time": "0:07:29"}
|
||||
{"current_steps": 840, "total_steps": 1632, "loss": 0.2385, "lr": 5.619021998515165e-06, "epoch": 1.5441176470588234, "percentage": 51.47, "elapsed_time": "0:07:50", "remaining_time": "0:07:23"}
|
||||
{"current_steps": 850, "total_steps": 1632, "loss": 0.2844, "lr": 5.51270909203838e-06, "epoch": 1.5625, "percentage": 52.08, "elapsed_time": "0:07:56", "remaining_time": "0:07:18"}
|
||||
{"current_steps": 860, "total_steps": 1632, "loss": 0.2924, "lr": 5.40616138340673e-06, "epoch": 1.5808823529411766, "percentage": 52.7, "elapsed_time": "0:08:01", "remaining_time": "0:07:12"}
|
||||
{"current_steps": 870, "total_steps": 1632, "loss": 0.2826, "lr": 5.299427667603516e-06, "epoch": 1.5992647058823528, "percentage": 53.31, "elapsed_time": "0:08:07", "remaining_time": "0:07:06"}
|
||||
{"current_steps": 880, "total_steps": 1632, "loss": 0.2106, "lr": 5.1925568247965686e-06, "epoch": 1.6176470588235294, "percentage": 53.92, "elapsed_time": "0:08:12", "remaining_time": "0:07:00"}
|
||||
{"current_steps": 890, "total_steps": 1632, "loss": 0.3306, "lr": 5.085597797952906e-06, "epoch": 1.6360294117647058, "percentage": 54.53, "elapsed_time": "0:08:18", "remaining_time": "0:06:55"}
|
||||
{"current_steps": 900, "total_steps": 1632, "loss": 0.2227, "lr": 4.97859957042464e-06, "epoch": 1.6544117647058822, "percentage": 55.15, "elapsed_time": "0:08:23", "remaining_time": "0:06:49"}
|
||||
{"current_steps": 910, "total_steps": 1632, "loss": 0.2275, "lr": 4.871611143516367e-06, "epoch": 1.6727941176470589, "percentage": 55.76, "elapsed_time": "0:08:28", "remaining_time": "0:06:43"}
|
||||
{"current_steps": 920, "total_steps": 1632, "loss": 0.2966, "lr": 4.7646815140443625e-06, "epoch": 1.6911764705882353, "percentage": 56.37, "elapsed_time": "0:08:34", "remaining_time": "0:06:38"}
|
||||
{"current_steps": 930, "total_steps": 1632, "loss": 0.228, "lr": 4.657859651897807e-06, "epoch": 1.7095588235294117, "percentage": 56.99, "elapsed_time": "0:08:39", "remaining_time": "0:06:32"}
|
||||
{"current_steps": 940, "total_steps": 1632, "loss": 0.2694, "lr": 4.551194477612351e-06, "epoch": 1.7279411764705883, "percentage": 57.6, "elapsed_time": "0:08:45", "remaining_time": "0:06:26"}
|
||||
{"current_steps": 950, "total_steps": 1632, "loss": 0.1944, "lr": 4.444734839966306e-06, "epoch": 1.7463235294117647, "percentage": 58.21, "elapsed_time": "0:08:50", "remaining_time": "0:06:20"}
|
||||
{"current_steps": 960, "total_steps": 1632, "loss": 0.2479, "lr": 4.338529493609647e-06, "epoch": 1.7647058823529411, "percentage": 58.82, "elapsed_time": "0:08:55", "remaining_time": "0:06:15"}
|
||||
{"current_steps": 970, "total_steps": 1632, "loss": 0.1819, "lr": 4.232627076736181e-06, "epoch": 1.7830882352941178, "percentage": 59.44, "elapsed_time": "0:09:01", "remaining_time": "0:06:09"}
|
||||
{"current_steps": 980, "total_steps": 1632, "loss": 0.2514, "lr": 4.127076088809e-06, "epoch": 1.8014705882352942, "percentage": 60.05, "elapsed_time": "0:09:06", "remaining_time": "0:06:03"}
|
||||
{"current_steps": 990, "total_steps": 1632, "loss": 0.24, "lr": 4.021924868349493e-06, "epoch": 1.8198529411764706, "percentage": 60.66, "elapsed_time": "0:09:12", "remaining_time": "0:05:57"}
|
||||
{"current_steps": 1000, "total_steps": 1632, "loss": 0.2477, "lr": 3.9172215708000655e-06, "epoch": 1.8382352941176472, "percentage": 61.27, "elapsed_time": "0:09:17", "remaining_time": "0:05:52"}
|
||||
{"current_steps": 1010, "total_steps": 1632, "loss": 0.2207, "lr": 3.8130141464706806e-06, "epoch": 1.8566176470588234, "percentage": 61.89, "elapsed_time": "0:09:36", "remaining_time": "0:05:55"}
|
||||
{"current_steps": 1020, "total_steps": 1632, "loss": 0.2634, "lr": 3.709350318579371e-06, "epoch": 1.875, "percentage": 62.5, "elapsed_time": "0:09:41", "remaining_time": "0:05:49"}
|
||||
{"current_steps": 1030, "total_steps": 1632, "loss": 0.236, "lr": 3.606277561396726e-06, "epoch": 1.8933823529411766, "percentage": 63.11, "elapsed_time": "0:09:47", "remaining_time": "0:05:43"}
|
||||
{"current_steps": 1040, "total_steps": 1632, "loss": 0.2132, "lr": 3.503843078504405e-06, "epoch": 1.9117647058823528, "percentage": 63.73, "elapsed_time": "0:09:52", "remaining_time": "0:05:37"}
|
||||
{"current_steps": 1050, "total_steps": 1632, "loss": 0.2088, "lr": 3.4020937811776154e-06, "epoch": 1.9301470588235294, "percentage": 64.34, "elapsed_time": "0:09:58", "remaining_time": "0:05:31"}
|
||||
{"current_steps": 1060, "total_steps": 1632, "loss": 0.2413, "lr": 3.301076266901435e-06, "epoch": 1.9485294117647058, "percentage": 64.95, "elapsed_time": "0:10:03", "remaining_time": "0:05:25"}
|
||||
{"current_steps": 1070, "total_steps": 1632, "loss": 0.2271, "lr": 3.2008367980308737e-06, "epoch": 1.9669117647058822, "percentage": 65.56, "elapsed_time": "0:10:08", "remaining_time": "0:05:19"}
|
||||
{"current_steps": 1080, "total_steps": 1632, "loss": 0.2285, "lr": 3.1014212806043794e-06, "epoch": 1.9852941176470589, "percentage": 66.18, "elapsed_time": "0:10:14", "remaining_time": "0:05:13"}
|
||||
{"current_steps": 1090, "total_steps": 1632, "loss": 0.1852, "lr": 3.0028752433205475e-06, "epoch": 2.0036764705882355, "percentage": 66.79, "elapsed_time": "0:10:19", "remaining_time": "0:05:08"}
|
||||
{"current_steps": 1100, "total_steps": 1632, "loss": 0.1519, "lr": 2.9052438166876305e-06, "epoch": 2.0220588235294117, "percentage": 67.4, "elapsed_time": "0:10:25", "remaining_time": "0:05:02"}
|
||||
{"current_steps": 1110, "total_steps": 1632, "loss": 0.1601, "lr": 2.808571712355389e-06, "epoch": 2.0404411764705883, "percentage": 68.01, "elapsed_time": "0:10:30", "remaining_time": "0:04:56"}
|
||||
{"current_steps": 1120, "total_steps": 1632, "loss": 0.1264, "lr": 2.7129032026388046e-06, "epoch": 2.0588235294117645, "percentage": 68.63, "elapsed_time": "0:10:35", "remaining_time": "0:04:50"}
|
||||
{"current_steps": 1130, "total_steps": 1632, "loss": 0.137, "lr": 2.618282100242935e-06, "epoch": 2.077205882352941, "percentage": 69.24, "elapsed_time": "0:10:41", "remaining_time": "0:04:44"}
|
||||
{"current_steps": 1140, "total_steps": 1632, "loss": 0.1085, "lr": 2.5247517381983137e-06, "epoch": 2.0955882352941178, "percentage": 69.85, "elapsed_time": "0:10:46", "remaining_time": "0:04:39"}
|
||||
{"current_steps": 1150, "total_steps": 1632, "loss": 0.1258, "lr": 2.4323549500159806e-06, "epoch": 2.113970588235294, "percentage": 70.47, "elapsed_time": "0:10:52", "remaining_time": "0:04:33"}
|
||||
{"current_steps": 1160, "total_steps": 1632, "loss": 0.1046, "lr": 2.3411340500712833e-06, "epoch": 2.1323529411764706, "percentage": 71.08, "elapsed_time": "0:10:57", "remaining_time": "0:04:27"}
|
||||
{"current_steps": 1170, "total_steps": 1632, "loss": 0.092, "lr": 2.251130814225449e-06, "epoch": 2.150735294117647, "percentage": 71.69, "elapsed_time": "0:11:03", "remaining_time": "0:04:21"}
|
||||
{"current_steps": 1180, "total_steps": 1632, "loss": 0.0868, "lr": 2.16238646069373e-06, "epoch": 2.1691176470588234, "percentage": 72.3, "elapsed_time": "0:11:08", "remaining_time": "0:04:16"}
|
||||
{"current_steps": 1190, "total_steps": 1632, "loss": 0.1059, "lr": 2.0749416311689845e-06, "epoch": 2.1875, "percentage": 72.92, "elapsed_time": "0:11:13", "remaining_time": "0:04:10"}
|
||||
{"current_steps": 1200, "total_steps": 1632, "loss": 0.098, "lr": 1.9888363722092376e-06, "epoch": 2.2058823529411766, "percentage": 73.53, "elapsed_time": "0:11:19", "remaining_time": "0:04:04"}
|
||||
{"current_steps": 1210, "total_steps": 1632, "loss": 0.108, "lr": 1.9041101168978094e-06, "epoch": 2.224264705882353, "percentage": 74.14, "elapsed_time": "0:11:24", "remaining_time": "0:03:58"}
|
||||
{"current_steps": 1220, "total_steps": 1632, "loss": 0.1039, "lr": 1.8208016667844153e-06, "epoch": 2.2426470588235294, "percentage": 74.75, "elapsed_time": "0:11:30", "remaining_time": "0:03:53"}
|
||||
{"current_steps": 1230, "total_steps": 1632, "loss": 0.1132, "lr": 1.7389491741154373e-06, "epoch": 2.261029411764706, "percentage": 75.37, "elapsed_time": "0:11:35", "remaining_time": "0:03:47"}
|
||||
{"current_steps": 1240, "total_steps": 1632, "loss": 0.0999, "lr": 1.6585901243616044e-06, "epoch": 2.2794117647058822, "percentage": 75.98, "elapsed_time": "0:11:41", "remaining_time": "0:03:41"}
|
||||
{"current_steps": 1250, "total_steps": 1632, "loss": 0.1059, "lr": 1.5797613190509908e-06, "epoch": 2.297794117647059, "percentage": 76.59, "elapsed_time": "0:11:46", "remaining_time": "0:03:35"}
|
||||
{"current_steps": 1260, "total_steps": 1632, "loss": 0.1186, "lr": 1.502498858915254e-06, "epoch": 2.3161764705882355, "percentage": 77.21, "elapsed_time": "0:11:51", "remaining_time": "0:03:30"}
|
||||
{"current_steps": 1270, "total_steps": 1632, "loss": 0.1018, "lr": 1.4268381273568232e-06, "epoch": 2.3345588235294117, "percentage": 77.82, "elapsed_time": "0:11:57", "remaining_time": "0:03:24"}
|
||||
{"current_steps": 1280, "total_steps": 1632, "loss": 0.0953, "lr": 1.352813774244565e-06, "epoch": 2.3529411764705883, "percentage": 78.43, "elapsed_time": "0:12:02", "remaining_time": "0:03:18"}
|
||||
{"current_steps": 1290, "total_steps": 1632, "loss": 0.1211, "lr": 1.2804597000454217e-06, "epoch": 2.3713235294117645, "percentage": 79.04, "elapsed_time": "0:12:08", "remaining_time": "0:03:13"}
|
||||
{"current_steps": 1300, "total_steps": 1632, "loss": 0.1821, "lr": 1.2098090402992085e-06, "epoch": 2.389705882352941, "percentage": 79.66, "elapsed_time": "0:12:13", "remaining_time": "0:03:07"}
|
||||
{"current_steps": 1310, "total_steps": 1632, "loss": 0.1263, "lr": 1.1408941504437533e-06, "epoch": 2.4080882352941178, "percentage": 80.27, "elapsed_time": "0:12:18", "remaining_time": "0:03:01"}
|
||||
{"current_steps": 1320, "total_steps": 1632, "loss": 0.0792, "lr": 1.0737465909972778e-06, "epoch": 2.426470588235294, "percentage": 80.88, "elapsed_time": "0:12:24", "remaining_time": "0:02:55"}
|
||||
{"current_steps": 1330, "total_steps": 1632, "loss": 0.1395, "lr": 1.0083971131048158e-06, "epoch": 2.4448529411764706, "percentage": 81.5, "elapsed_time": "0:12:29", "remaining_time": "0:02:50"}
|
||||
{"current_steps": 1340, "total_steps": 1632, "loss": 0.1168, "lr": 9.448756444553226e-07, "epoch": 2.463235294117647, "percentage": 82.11, "elapsed_time": "0:12:35", "remaining_time": "0:02:44"}
|
||||
{"current_steps": 1350, "total_steps": 1632, "loss": 0.0876, "lr": 8.832112755758598e-07, "epoch": 2.4816176470588234, "percentage": 82.72, "elapsed_time": "0:12:40", "remaining_time": "0:02:38"}
|
||||
{"current_steps": 1360, "total_steps": 1632, "loss": 0.1019, "lr": 8.234322465092049e-07, "epoch": 2.5, "percentage": 83.33, "elapsed_time": "0:12:46", "remaining_time": "0:02:33"}
|
||||
{"current_steps": 1370, "total_steps": 1632, "loss": 0.1377, "lr": 7.655659338809329e-07, "epoch": 2.5183823529411766, "percentage": 83.95, "elapsed_time": "0:12:51", "remaining_time": "0:02:27"}
|
||||
{"current_steps": 1380, "total_steps": 1632, "loss": 0.0879, "lr": 7.09638838361908e-07, "epoch": 2.536764705882353, "percentage": 84.56, "elapsed_time": "0:12:56", "remaining_time": "0:02:21"}
|
||||
{"current_steps": 1390, "total_steps": 1632, "loss": 0.0958, "lr": 6.556765725319525e-07, "epoch": 2.5551470588235294, "percentage": 85.17, "elapsed_time": "0:13:02", "remaining_time": "0:02:16"}
|
||||
{"current_steps": 1400, "total_steps": 1632, "loss": 0.0994, "lr": 6.037038491501978e-07, "epoch": 2.5735294117647056, "percentage": 85.78, "elapsed_time": "0:13:07", "remaining_time": "0:02:10"}
|
||||
{"current_steps": 1410, "total_steps": 1632, "loss": 0.0818, "lr": 5.53744469837551e-07, "epoch": 2.5919117647058822, "percentage": 86.4, "elapsed_time": "0:13:13", "remaining_time": "0:02:04"}
|
||||
{"current_steps": 1420, "total_steps": 1632, "loss": 0.1018, "lr": 5.058213141764151e-07, "epoch": 2.610294117647059, "percentage": 87.01, "elapsed_time": "0:13:18", "remaining_time": "0:01:59"}
|
||||
{"current_steps": 1430, "total_steps": 1632, "loss": 0.0874, "lr": 4.599563292326592e-07, "epoch": 2.6286764705882355, "percentage": 87.62, "elapsed_time": "0:13:23", "remaining_time": "0:01:53"}
|
||||
{"current_steps": 1440, "total_steps": 1632, "loss": 0.1395, "lr": 4.1617051950467613e-07, "epoch": 2.6470588235294117, "percentage": 88.24, "elapsed_time": "0:13:29", "remaining_time": "0:01:47"}
|
||||
{"current_steps": 1450, "total_steps": 1632, "loss": 0.1316, "lr": 3.744839373040682e-07, "epoch": 2.6654411764705883, "percentage": 88.85, "elapsed_time": "0:13:34", "remaining_time": "0:01:42"}
|
||||
{"current_steps": 1460, "total_steps": 1632, "loss": 0.1149, "lr": 3.3491567357242736e-07, "epoch": 2.6838235294117645, "percentage": 89.46, "elapsed_time": "0:13:40", "remaining_time": "0:01:36"}
|
||||
{"current_steps": 1470, "total_steps": 1632, "loss": 0.0907, "lr": 2.9748384913837525e-07, "epoch": 2.702205882352941, "percentage": 90.07, "elapsed_time": "0:13:45", "remaining_time": "0:01:30"}
|
||||
{"current_steps": 1480, "total_steps": 1632, "loss": 0.1253, "lr": 2.6220560641887385e-07, "epoch": 2.7205882352941178, "percentage": 90.69, "elapsed_time": "0:13:50", "remaining_time": "0:01:25"}
|
||||
{"current_steps": 1490, "total_steps": 1632, "loss": 0.098, "lr": 2.2909710156863275e-07, "epoch": 2.7389705882352944, "percentage": 91.3, "elapsed_time": "0:13:56", "remaining_time": "0:01:19"}
|
||||
{"current_steps": 1500, "total_steps": 1632, "loss": 0.1071, "lr": 1.981734970811644e-07, "epoch": 2.7573529411764706, "percentage": 91.91, "elapsed_time": "0:14:01", "remaining_time": "0:01:14"}
|
||||
{"current_steps": 1510, "total_steps": 1632, "loss": 0.097, "lr": 1.6944895484492075e-07, "epoch": 2.775735294117647, "percentage": 92.52, "elapsed_time": "0:14:21", "remaining_time": "0:01:09"}
|
||||
{"current_steps": 1520, "total_steps": 1632, "loss": 0.0984, "lr": 1.429366296576623e-07, "epoch": 2.7941176470588234, "percentage": 93.14, "elapsed_time": "0:14:26", "remaining_time": "0:01:03"}
|
||||
{"current_steps": 1530, "total_steps": 1632, "loss": 0.0858, "lr": 1.1864866320203116e-07, "epoch": 2.8125, "percentage": 93.75, "elapsed_time": "0:14:32", "remaining_time": "0:00:58"}
|
||||
{"current_steps": 1540, "total_steps": 1632, "loss": 0.1211, "lr": 9.659617848510883e-08, "epoch": 2.8308823529411766, "percentage": 94.36, "elapsed_time": "0:14:37", "remaining_time": "0:00:52"}
|
||||
{"current_steps": 1550, "total_steps": 1632, "loss": 0.1214, "lr": 7.678927474447817e-08, "epoch": 2.849264705882353, "percentage": 94.98, "elapsed_time": "0:14:43", "remaining_time": "0:00:46"}
|
||||
{"current_steps": 1560, "total_steps": 1632, "loss": 0.0839, "lr": 5.9237022823140924e-08, "epoch": 2.8676470588235294, "percentage": 95.59, "elapsed_time": "0:14:48", "remaining_time": "0:00:41"}
|
||||
{"current_steps": 1570, "total_steps": 1632, "loss": 0.1084, "lr": 4.394746101540115e-08, "epoch": 2.8860294117647056, "percentage": 96.2, "elapsed_time": "0:14:54", "remaining_time": "0:00:35"}
|
||||
{"current_steps": 1580, "total_steps": 1632, "loss": 0.109, "lr": 3.092759138561607e-08, "epoch": 2.9044117647058822, "percentage": 96.81, "elapsed_time": "0:14:59", "remaining_time": "0:00:29"}
|
||||
{"current_steps": 1590, "total_steps": 1632, "loss": 0.0956, "lr": 2.0183376561507263e-08, "epoch": 2.922794117647059, "percentage": 97.43, "elapsed_time": "0:15:04", "remaining_time": "0:00:23"}
|
||||
{"current_steps": 1600, "total_steps": 1632, "loss": 0.119, "lr": 1.171973700349216e-08, "epoch": 2.9411764705882355, "percentage": 98.04, "elapsed_time": "0:15:10", "remaining_time": "0:00:18"}
|
||||
{"current_steps": 1610, "total_steps": 1632, "loss": 0.1018, "lr": 5.540548751292174e-09, "epoch": 2.9595588235294117, "percentage": 98.65, "elapsed_time": "0:15:15", "remaining_time": "0:00:12"}
|
||||
{"current_steps": 1620, "total_steps": 1632, "loss": 0.0922, "lr": 1.6486416488459279e-09, "epoch": 2.9779411764705883, "percentage": 99.26, "elapsed_time": "0:15:21", "remaining_time": "0:00:06"}
|
||||
{"current_steps": 1630, "total_steps": 1632, "loss": 0.0803, "lr": 4.579804834703438e-11, "epoch": 2.9963235294117645, "percentage": 99.88, "elapsed_time": "0:15:26", "remaining_time": "0:00:01"}
|
||||
{"current_steps": 1632, "total_steps": 1632, "epoch": 3.0, "percentage": 100.0, "elapsed_time": "0:15:43", "remaining_time": "0:00:00"}
|
||||
1183
trainer_state.json
Normal file
1183
trainer_state.json
Normal file
File diff suppressed because it is too large
Load Diff
3
training_args.bin
Normal file
3
training_args.bin
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:1f4bebf5c7141bddbed6a7feb4bf1a35ed3fbf020c7c4dc7ab21f6f30268f7e3
|
||||
size 7352
|
||||
BIN
training_loss.png
Normal file
BIN
training_loss.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 44 KiB |
Reference in New Issue
Block a user