deepspeedai / deepspeedai/DeepSpeedExamples
“Current loss scale already at minimum - cannot decrease scale anymore” when I enabled ZERO++
Open
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 6.8k
- Forks
- 1.1k
- Avg merge
- 2d 16h
- Merged PRs (30d)
- 1
Description
deepspeed 0.10.0
here is my config:
json = {
"train_batch_size": 2,
"train_micro_batch_size_per_gpu": 2,
"steps_per_print": 20,
"zero_allow_untested_optimizer": true,
"zero_optimization": {
"stage": 3,
"reduce_bucket_size": 1.000000e+07,
"reduce_scatter": true,
"zero_quantized_weights": true,
"zero_hpz_partition_size": 16,
"zero_quantized_gradients": true,
"contiguous_gradients": true,
"overlap_comm": true,
"offload_param": {
"device": "none"
},
"offload_optimizer": {
"device": "none"
},
"stage3_param_persistence_threshold": 1.000000e+04,
"stage3_max_live_parameters": 3.000000e+07,
"stage3_prefetch_bucket_size": 3.000000e+07,
"memory_efficient_linear": false
},
"fp16": {
"enabled": true,
"loss_scale_window": 100
},
"gradient_clipping": 1.0,
"prescale_gradients": false,
"wall_clock_breakdown": false,
"hybrid_engine": {
"enabled": false,
"max_out_tokens": 2.048000e+03,
"inference_tp_size": 1,
"release_inference_cache": false,
"pin_parameters": true,
"tp_gather_partition_size": 8
}
}
model and params:
model=LlamaForCausalLM.from_pretrained("decapoda-research/llama-13b-hf")
args.zero_stage=3
args.num_train_epochs=1
args.per_device_train_batch_size = 2
args.gradient_accumulation_steps =1
args.seed=1234
args.weight_decay=0.01
args.lr_scheduler_type="cosine"
args.num_warmup_steps=500
args.learning_rate=1e-5
args.output_dir = "."
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Reproduce the report using DeepSpeed 0.10.0, the supplied ZERO++/fp16 configuration, and the LlamaForCausalLM model from decapoda-research/llama-13b-hf. Trace when the loss scale reaches its minimum and determine whether the reported configuration causes it; done should identify the cause and provide a confirmed resolution or workaround.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100