From 6f631f00aed486a94fbb86f94a41b68e3afc711b Mon Sep 17 00:00:00 2001
From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com>
Date: Wed, 1 Jul 2026 07:29:32 -0400
Subject: [PATCH] fix: use FP8 model with 32GB CPU offload

---
 training/configs/ornith-35b-lora.yaml | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/training/configs/ornith-35b-lora.yaml b/training/configs/ornith-35b-lora.yaml
index 2ee0360..2dec4ec 100644
--- a/training/configs/ornith-35b-lora.yaml
+++ b/training/configs/ornith-35b-lora.yaml
@@ -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
+base_model: /data/models/Ornith-1.0-35B-FP8
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 
@@ -59,9 +59,12 @@ deepspeed_config:
     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