Delete README.md

README.md

@@ -1,118 +0,0 @@
-# Checkpoint Compatibility Information
-
-## Checkpoint Location
-`/scratch/zsh/shuhongz_adobe_ckpts/ckpt_for_single_lap/`
-
-## Source
-Converted from: `/datasets/objaverse/shuhongz_adobe_ckpts/1023_generated_830k_lap_0_28_only_t2i/checkpoint-18000`
-
-## Changes Applied
-- **Removed**: All `image_layerwise_attention_pooling.*` keys (18 keys)
-- **Kept**: All other modules including `text_layerwise_attention_pooling`
-
-## Compatibility
-This checkpoint is compatible with models using the **HYBRID STRATEGY**:
-- Text tokens: Processed through `text_layerwise_attention_pooling`
-- Image tokens: Use ViT features directly (no LAP)
-
-Target model: `uno_debug/1029_mixed_text_lap_internvl_s2i_train_mllm_only_masked_loss_clip_lora.py`
-
-## Files Included
-1. ✅ `dit_lora.safetensors` - Cleaned model weights (334 keys, ~2.2GB)
-2. ✅ `scheduler.bin` - Learning rate scheduler state
-3. ❌ `optimizer.bin` - NOT INCLUDED (see below)
-
-## ⚠️  Important: optimizer.bin NOT Included
-
-**Why optimizer.bin is NOT compatible:**
-
-The optimizer.bin from the original checkpoint stores optimizer states (momentum, variance, etc.) 
-for all 352 parameters, including the 18 `image_layerwise_attention_pooling` parameters that 
-have been removed.
-
-**Problem:**
-- Optimizer states are indexed by parameter position/ID, not by name
-- The cleaned model has 334 parameters (18 fewer than the original 352)
-- Using the old optimizer.bin would cause parameter ID mismatches
-- This leads to training errors or incorrect optimizer state application
-
-**Solutions:**
-
-### Option 1: Start Fresh (RECOMMENDED)
-```python
-# In your training config, set:
-resume_from_checkpoint = "/scratch/zsh/shuhongz_adobe_ckpts/ckpt_for_single_lap"
-
-# The training script will:
-# ✅ Load model weights from dit_lora.safetensors
-# ✅ Load scheduler state from scheduler.bin  
-# ✅ Initialize a fresh optimizer (no momentum/variance carried over)
-```
-
-**Pros:**
-- Clean start with no parameter mismatches
-- Model weights are preserved
-- Safe and reliable
-
-**Cons:**
-- Loses optimizer momentum/variance accumulated during previous training
-- May need a brief warm-up period (but usually minimal impact)
-
-### Option 2: Keep Original Checkpoint
-If you absolutely need the optimizer state, use the original checkpoint:
-```python
-resume_from_checkpoint = "/datasets/objaverse/shuhongz_adobe_ckpts/1023_generated_830k_lap_0_28_only_t2i/checkpoint-18000"
-```
-
-But you'll need to modify the loading code to skip the incompatible keys:
-```python
-# In resume_from_checkpoint function:
-lora_state = load_file(path, device=device)
-# Filter out image_layerwise_attention_pooling keys
-lora_state = {k: v for k, v in lora_state.items() 
-              if not k.startswith('image_layerwise_attention_pooling.')}
-unwarp_dit.load_state_dict(lora_state, strict=False)
-```
-
-## Verification
-
-To verify the checkpoint structure:
-```bash
-python3 -c "
-from safetensors.torch import load_file
-state_dict = load_file('/scratch/zsh/shuhongz_adobe_ckpts/ckpt_for_single_lap/dit_lora.safetensors')
-modules = {}
-for key in state_dict.keys():
-    module = key.split('.')[0]
-    modules[module] = modules.get(module, 0) + 1
-print('Modules in checkpoint:')
-for m, count in sorted(modules.items()):
-    print(f'  {m}: {count} keys')
-"
-```
-
-Expected output:
-- double_blocks: 152 keys
-- internvl_projector: 8 keys
-- single_blocks: 152 keys
-- text_layerwise_attention_pooling: 18 keys
-- vector_in: 4 keys
-
-**Total: 334 keys** (vs 352 in original)
-
-## Training Command Example
-
-```bash
-# Using the cleaned checkpoint without optimizer state
-accelerate launch --config_file config/accelerate/default_config.yaml \
-    uno_debug/1029_mixed_text_lap_internvl_s2i_train_mllm_only_masked_loss_clip_lora.py \
-    --config config/train_config.yaml \
-    --resume_from_checkpoint "/scratch/zsh/shuhongz_adobe_ckpts/ckpt_for_single_lap"
-```
-
-The training will automatically:
-1. Load `dit_lora.safetensors` with 334 parameters
-2. Load `scheduler.bin` for learning rate schedule
-3. Initialize fresh optimizer for all trainable parameters
-4. Continue training from step 18000
-