Bug Description
When metrics are being evaluated using online aggregation, it is desirable to see real-time updates to the estimates and confidence intervals. Right now it only updates at shard boundaries. That means, if I have num_shards=4, I have to wait for 25% (1/4) of the data to be processed before I can see the first estimate.
To Reproduce
- Set online aggregation in the config.
"online_strategy_kwargs": {
"strategy_name": "normal",
"confidence_level": 0.95,
"use_fpc": True,
}
- Run any of the context engineering notebooks -- I'm running GSM8K notebook.
Expected Behavior
Ideally it should update after every mini-batch of inference. But if not, at least update with a reasonable wall clock frequency such as every 30/60 sec
Screenshots
Environment
(infer) palebluedot@palebluedotcloud:~/infer$ rapidfireai doctor
🔍 RapidFire AI System Diagnostics
==================================================
🐍 Python Environment:
------------------------------
Version: 3.12.12
Implementation: CPython
Executable: /home/palebluedot/miniconda3/envs/infer/bin/python3.12
Conda Environment: infer
Virtual Environment: no
📦 Installed Packages:
------------------------------
bitsandbytes 0.48.2
evaluate 0.4.6
flashinfer-python 0.2.5+cu124torch2.5
Flask 3.1.2
flask-cors 6.0.1
gunicorn 23.0.0
mlflow 3.6.0
mlflow-skinny 3.6.0
mlflow-tracing 3.6.0
nltk 3.9.2
peft 0.18.0
rapidfireai 0.12.6rc1
sentence-transformers 5.1.0
sentencepiece 0.2.1
torch 2.5.1+cu124
torchaudio 2.5.1+cu124
torchvision 0.20.1+cu124
transformers 4.56.1
trl 0.21.0
... (showing only relevant packages)
🚀 GPU & CUDA Information:
------------------------------
nvidia-smi: found
Driver Version: 575.64.03
CUDA Runtime: 12.9
GPU Count: 1
GPU Model: NVIDIA A100-SXM4-80GB
Total VRAM: 80.0 GB
nvcc: not found
CUDA Installation: present
CUDA on PATH: no
💻 System Information:
------------------------------
Platform: Linux-6.8.0-64-generic-x86_64-with-glibc2.35
Architecture: x86_64
Processor: x86_64
🔧 Environment Variables:
------------------------------
CUDA_HOME: not set
CUDA_PATH: not set
LD_LIBRARY_PATH: not set
PATH: /home/palebluedot/miniconda3/envs/infer/bin:/home/palebluedot/miniconda3/condabin:/home/palebluedot/...
✅ Diagnostics complete!
Additional Context
Add any other context about the problem here.
Error Logs
If applicable, paste any error logs or stack traces here.
Bug Description
When metrics are being evaluated using online aggregation, it is desirable to see real-time updates to the estimates and confidence intervals. Right now it only updates at shard boundaries. That means, if I have
num_shards=4, I have to wait for 25% (1/4) of the data to be processed before I can see the first estimate.To Reproduce
"online_strategy_kwargs": {
"strategy_name": "normal",
"confidence_level": 0.95,
"use_fpc": True,
}
Expected Behavior
Ideally it should update after every mini-batch of inference. But if not, at least update with a reasonable wall clock frequency such as every 30/60 sec
Screenshots
Environment
Additional Context
Add any other context about the problem here.
Error Logs
If applicable, paste any error logs or stack traces here.