Upload folder using huggingface_hub

README.md

@@ -16,21 +16,19 @@ This space is dedicated to training Microsoft's Phi-4 model using Unsloth optimi
 
 ## Installation
 
-Before running the training, ensure you have all required dependencies installed:
+This Hugging Face Space automatically installs dependencies from requirements.txt. The following packages are included:
 
-```bash
-pip install -r requirements.txt
-```
-
-### Critical Dependencies
+### Essential Dependencies
 
 - **unsloth** (>=2024.3): Required for optimized 4-bit training
 - **peft** (>=0.9.0): Required for parameter-efficient fine-tuning
-- **transformers** (>=4.36.0): Required for model architecture and tokenization
+- **transformers** (>=4.36.0): For model architecture and tokenization
+- **einops**: Required by Unsloth for tensor manipulation
+- **sentencepiece**: Required for tokenization
 
-### Optional but Recommended
+### Optional Dependencies
 
-- **flash-attn** (>=2.5.0): Significantly speeds up attention computations
+- **flash-attn**: Optional for faster attention computation (not included by default as it can cause build issues)
 
 ## Features
 

requirements.txt

@@ -3,14 +3,12 @@ bitsandbytes>=0.41.0
 datasets>=2.15.0
 einops>=0.7.0
 filelock>=3.13.1
-flash-attn>=2.5.0
 gradio>=5.17.0
 huggingface-hub>=0.19.0
 matplotlib>=3.7.0
 numpy>=1.24.0
 packaging>=23.0
 peft>=0.9.0
-protobuf>=4.23.4
 psutil>=5.9.0
 python-dotenv>=1.0.0
 pyyaml>=6.0.1

run_transformers_training.py

@@ -129,16 +129,25 @@ def load_model_and_tokenizer(config):
     try:
         if not unsloth_available:
             logger.error("Unsloth is required for training with pre-quantized model")
-            logger.error("Please install required packages with: pip install -r requirements.txt")
-            logger.error("Or directly install with: pip install unsloth>=2024.3 peft>=0.9.0")
-            raise ImportError("Required packages missing. See log for installation instructions.")
+            logger.error("Please ensure unsloth is in requirements.txt")
+            raise ImportError("Unsloth is required for this training setup")
             
         logger.info("Using Unsloth optimizations with pre-quantized model")
+        # Check for flash attention without importing it directly
+        use_flash_attention = config.get("use_flash_attention", True)
+        try:
+            import flash_attn
+            logger.info("Flash attention detected and will be used")
+        except ImportError:
+            use_flash_attention = False
+            logger.warning("Flash attention not available, falling back to standard attention")
+            
         model, tokenizer = FastLanguageModel.from_pretrained(
             model_name=config.get("model_name"),
             max_seq_length=config.get("max_seq_length", 2048),
             dtype=None,  # Let Unsloth choose optimal dtype
             device_map="auto",
+            # Don't explicitly use flash attention config here, let Unsloth handle it
         )
         
         # Apply Unsloth's training optimizations with config parameters
@@ -170,7 +179,7 @@ def load_model_and_tokenizer(config):
     
     except Exception as e:
         logger.error(f"Error in model/tokenizer loading: {str(e)}")
-        logger.error("If missing dependencies, install with: pip install -r requirements.txt")
+        logger.error("If missing dependencies, check the requirements.txt file")
         raise
 
 def load_dataset_with_mapping(dataset_config):
@@ -373,30 +382,22 @@ def check_dependencies():
     if not peft_available:
         missing_packages.append("peft>=0.9.0")
     
-    # Optional but recommended packages
+    # Optional packages - don't add to missing list, just log
     try:
         import flash_attn
+        logger.info("flash-attn found. Flash attention will be used for faster training.")
     except ImportError:
-        logger.warning("flash-attn not found. Flash attention will not be used.")
-        missing_packages.append("flash-attn>=2.5.0 (optional)")
+        logger.warning("flash-attn not found. Training will work but may be slower.")
+        # Don't add to missing packages since it's optional and can cause build issues
     
     # If critical packages are missing, exit with instructions
-    critical_missing = [pkg for pkg in missing_packages if "optional" not in pkg]
-    if critical_missing:
+    if missing_packages:
         logger.error("Critical dependencies missing:")
-        for pkg in critical_missing:
+        for pkg in missing_packages:
             logger.error(f"  - {pkg}")
-        logger.error("Please install required packages with: pip install -r requirements.txt")
-        logger.error(f"Or directly with: pip install {' '.join(critical_missing)}")
+        logger.error("Please ensure the space has these packages in requirements.txt")
         return False
     
-    # If optional packages are missing, just warn
-    optional_missing = [pkg for pkg in missing_packages if "optional" in pkg]
-    if optional_missing:
-        logger.warning("Optional dependencies missing (recommended for best performance):")
-        for pkg in optional_missing:
-            logger.warning(f"  - {pkg.split(' ')[0]}")
-    
     return True
 
 def main():