cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: implement QLoRA with 4-bit BitsAndBytes quantization

Browse Source
This commit is contained in:
Christian Medina
2026-07-01 07:45:33 -04:00
parent f0ee6bc9a2
commit da5eb3abed
2 changed files with 14 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
training/configs/ornith-35b-lora.yaml
View File

@@ -1,7 +1,7 @@
# LoRA Training Configuration for Llama-2-7b
# Dataset: cyron_summary_lora_dataset (20k examples)
base_model: /data/models/Ornith-1.0-35B-FP8
base_model: /data/models/Ornith-1.0-35B
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
@@ -53,18 +53,9 @@ huggingface_hub:
deepspeed_config:
zero_optimization:
stage: 3
offload_optimizer:
device: cpu
pin_memory: true
offload_param:
device: cpu
pin_memory: true
offload_params_device: cpu
gradient_clipping: 1.0
train_batch_size: auto
train_micro_batch_size_per_gpu: auto
# Offload 32GB to RAM
zero_hierarchical_offload: true
# Evaluation
eval_strategy: steps