deepspeedai / deepspeedai/DeepSpeedExamples

Overflow in deepspeed-chat LoRA and BF16 mode

Open
#783 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
6.8k
Forks
1.1k
Avg merge
2d 16h
Merged PRs (30d)
1

Description

  • Example: Deepspeed-chat
  • Model: Llama2-7b-hf
  • Mode: LoRA, lora_dim=128
  • precision: FP16
  • Output log as below:
  • Question: Does the log mean it's training correctly? I found the log is different from the log of SFT and LoRA only mode, which can output loss in each step. If not correct, how to make LoRA mode run correcly?
Model Parameters: 6.927 B, Latency: 6.08s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.06s, TFLOPs: 1.73, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.73, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
[2023-10-20 09:49:17,004] [INFO] [logging.py:96:log_dist] [Rank 0] step=40, skipped=6, lr=[9.618683345445294e-06, 0.0004983773754116733], mom=[(0.9, 0.95), (0.9, 0.95)]                                                                                                                   
[2023-10-20 09:49:17,004] [INFO] [timer.py:260:stop] epoch=0/micro_step=40/global_step=40, RunningAvgSamplesPerSec=5.187625237365746, CurrSamplesPerSec=5.008550092748559, MemAllocated=3.9GB, MaxMemAllocated=6.71GB                                                                      
Model Parameters: 6.927 B, Latency: 6.39s, TFLOPs: 1.64, Samples/sec: 0.63, Time/seq 1.60s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.06s, TFLOPs: 1.73, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.09s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.08s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512                                                                                                                                                            
Model Parameters: 6.927 B, Latency: 6.07s, TFLOPs: 1.72, Samples/sec: 0.66, Time/seq 1.52s, Batch Size: 4, Sequence Length: 512 

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the Deepspeed-chat example and compare its LoRA and BF16/FP16 execution with the SFT and LoRA-only modes mentioned in the report. Determine whether the repeated performance logs indicate correct training and whether loss output is expected; done means the behavior is explained or the underlying logging or training issue is resolved.

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
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.