fix: use CompressedTensors for Ornith, add torch import, expand LoRA targets

training/configs/llama2-7b-lora.yaml

@@ -5,11 +5,8 @@ base_model: deepreinforce-ai/Ornith-1.0-35B-FP8
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 
-# Quantization (QLoRA)
-load_in_4bit: true
-bnb_4bit_compute_dtype: bfloat16
-bnb_4bit_quant_type: nf4
-use_nested_quant: false
+# Model is already quantized (Ornith uses CompressedTensors)
+# No need for BitsAndBytes configuration
 
 # LoRA Configuration
 lora_r: 16
@@ -20,6 +17,9 @@ target_modules:
   - v_proj
   - k_proj
   - o_proj
+  - gate_proj
+  - up_proj
+  - down_proj
 lora_task_type: CAUSAL_LM
 
 # Dataset