From 821c26bf69154bca1799150217b8c875d00d6505 Mon Sep 17 00:00:00 2001
From: Soma Nakamura <p@centraworks.net>
Date: Thu, 10 Jul 2025 23:25:51 +0900
Subject: [PATCH] grad-repair

---
 .../training_config_gemma3_1b_8gpu_ddp.yaml   | 14 +++---
 config/training_config_gemma3_1b_minimal.yaml | 45 +++++++++++++++++++
 2 files changed, 51 insertions(+), 8 deletions(-)
 create mode 100644 config/training_config_gemma3_1b_minimal.yaml

diff --git a/config/training_config_gemma3_1b_8gpu_ddp.yaml b/config/training_config_gemma3_1b_8gpu_ddp.yaml
index ee6c2b7..5c1ed70 100644
--- a/config/training_config_gemma3_1b_8gpu_ddp.yaml
+++ b/config/training_config_gemma3_1b_8gpu_ddp.yaml
@@ -6,9 +6,7 @@ experiment:
   wandb_project: "matsuo-llm-comp-2025"
 
 model:
-  load_in_4bit: true  # Enable quantization for memory savings
-  bnb_4bit_compute_dtype: "bfloat16"
-  bnb_4bit_use_double_quant: true
+  load_in_4bit: false  # Disable quantization due to CUDA kernel issues
   device_map: "balanced"  # Distribute across all GPUs
   gradient_checkpointing: true  # Enable gradient checkpointing
   use_flash_attention_2: false
@@ -36,7 +34,7 @@ progressive_stages:
       gradient_accumulation_steps: 8  # Maintain effective batch size
       learning_rate: 5e-4
       warmup_steps: 100
-      max_length: 512  # Reduced sequence length
+      max_length: 256  # Very short sequences for memory
       bf16: true
       max_grad_norm: 1.0
       weight_decay: 0.001
@@ -53,7 +51,7 @@ progressive_stages:
       r: 8  # Minimal rank for memory
       lora_alpha: 16
       lora_dropout: 0.1
-      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
+      target_modules: ["q_proj", "v_proj"]  # Minimal modules for memory
       init_lora_weights: true
     training:
       num_epochs: 1
@@ -61,7 +59,7 @@ progressive_stages:
       gradient_accumulation_steps: 4
       learning_rate: 3e-4
       warmup_steps: 200
-      max_length: 1024  # Reduced sequence length
+      max_length: 512  # Short sequences for memory
       bf16: true
       max_grad_norm: 1.0
       weight_decay: 0.001
@@ -81,7 +79,7 @@ progressive_stages:
       r: 8  # Minimal rank for memory
       lora_alpha: 16
       lora_dropout: 0.1
-      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
+      target_modules: ["q_proj", "v_proj"]  # Minimal modules for memory
       init_lora_weights: true
     training:
       num_epochs: 1
@@ -89,7 +87,7 @@ progressive_stages:
       gradient_accumulation_steps: 2
       learning_rate: 2e-4
       warmup_steps: 300
-      max_length: 1024  # Reduced sequence length
+      max_length: 512  # Short sequences for memory
       bf16: true
       max_grad_norm: 1.0
       weight_decay: 0.001
diff --git a/config/training_config_gemma3_1b_minimal.yaml b/config/training_config_gemma3_1b_minimal.yaml
new file mode 100644
index 0000000..deed65b
--- /dev/null
+++ b/config/training_config_gemma3_1b_minimal.yaml
@@ -0,0 +1,45 @@
+experiment:
+  name: "progressive_reasoning_gemma3_1b_minimal"
+  base_model: "google/gemma-3-1b-pt"
+  output_dir: "./outputs"
+  use_wandb: true
+  wandb_project: "matsuo-llm-comp-2025"
+
+model:
+  load_in_4bit: false
+  device_map: "auto"
+  gradient_checkpointing: true
+  use_flash_attention_2: false
+  use_eager_attention: true
+
+progressive_stages:
+  - name: "basic_cot"
+    description: "Basic Chain-of-Thought reasoning"
+    dataset_path: "./data/basic_cot/"
+    adapter_config:
+      r: 4  # Extremely minimal rank
+      lora_alpha: 8
+      lora_dropout: 0.1
+      target_modules: ["q_proj"]  # Only one module
+      init_lora_weights: true
+    training:
+      num_epochs: 1  # Reduced epochs
+      per_device_batch_size: 1
+      gradient_accumulation_steps: 4
+      learning_rate: 5e-4
+      warmup_steps: 10
+      max_length: 128  # Very short sequences
+      bf16: true
+      max_grad_norm: 1.0
+      weight_decay: 0.001
+      save_steps: 100
+      logging_steps: 10
+      dataloader_num_workers: 1
+      dataloader_pin_memory: false
+
+evaluation:
+  benchmarks:
+    - "HLE"
+    - "Do-Not-Answer"
+  save_results: true
+  results_dir: "./outputs/evaluation_results"
\ No newline at end of file