cleaned and scaffolded

Author: codeamt

.idea/.gitignore

@@ -0,0 +1,35 @@
+defaults:
+  - fsdp_defaults  # For FSDP jobs only
+
+training:
+  batch_size: 4            # Unsloth: per-device batch size; FSDP: may use batch_size_per_gpu
+  batch_size_per_gpu: 4    # FSDP trainer expects this
+  grad_accum_steps: 1
+  lr: 2e-5
+  max_steps: 1000
+  optimizer: adamw_torch
+
+checkpoint:
+  save_interval: 100
+  output_dir: ./outputs
+
+logging:
+  wandb_project: "unsloth-qlora"
+  log_interval: 10
+
+data:
+  name: "gbharti/finance-alpaca"
+  prompt_template: |
+    ### Instruction: {instruction}
+    ### Input: {input}
+    ### Response: {output}{eos_token}
+
+model:
+  name: "meta-llama/Llama-2-7b-hf"
+  max_length: 4096
+  load_in_4bit: true
+  hf_token: null
+
+fsdp:
+  mixed_precision: true
+  # sharding_policy and other FSDP-specific options can be provided in strategy-specific configs

.idea/FSDP-Multi-GPU-Training.iml

@@ -0,0 +1,9 @@
+# @package _global_
+fsdp:
+  sharding_strategy: "FULL_SHARD"
+  mixed_precision: true
+  activation_checkpointing: true
+
+checkpoint:
+  save_optimizer: false  # Saves VRAM
+  use_sharded_state: true

.idea/inspectionProfiles/profiles_settings.xml

@@ -0,0 +1,29 @@
+# Evaluation Protocol
+
+## Automated Metrics
+1. **Training Loss** - Tracked every `log_interval` steps  
+2. **GPU Utilization** - Via CloudWatch metrics  
+3. **Memory Usage** - Peak VRAM/CPU recorded  
+4. **Gradient Metrics** - Norm distribution, kurtosis
+
+## Manual Evaluation  
+For each checkpoint:  
+1. Run 10 prompts from `eval_samples.txt`  
+2. Score responses (1-5 scale) on:  
+   - **Accuracy**: Factual correctness  
+   - **Coherence**: Logical flow  
+   - **Conciseness**: Brevity of response  
+   - **Relevance**: Adherence to prompt  
+
+3. Use scoring template:
+```json
+{
+  "prompt": "Explain quantum computing",
+  "response": "...",
+  "scores": {
+    "accuracy": 4,
+    "coherence": 5,
+    "conciseness": 3,
+    "relevance": 5
+  }
+}