ykarout's picture
Initial commit for initial version
87cfb14 verified
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer, BitsAndBytesConfig
from threading import Thread
import time
import logging
import gc
from pathlib import Path
import re
from huggingface_hub import HfApi, list_models
import os
import queue
import threading
from collections import deque
# Set PyTorch memory management environment variables
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
handlers=[
logging.FileHandler('gradio-chat-ui.log'),
logging.StreamHandler()
]
)
logger = logging.getLogger(__name__)
# Log memory management settings
logger.info(f"PyTorch CUDA allocation config: {os.environ.get('PYTORCH_CUDA_ALLOC_CONF')}")
logger.info(f"CUDA device count: {torch.cuda.device_count() if torch.cuda.is_available() else 'N/A'}")
# Model parameters
MODEL_NAME = "No Model Loaded"
MAX_LENGTH = 16384
DEFAULT_TEMPERATURE = 0.15
DEFAULT_TOP_P = 0.93
DEFAULT_TOP_K = 50
DEFAULT_REP_PENALTY = 1.15
# Base location for local models
LOCAL_MODELS_BASE = "/home/llm-models/"
# Global variables
model = None
tokenizer = None
hf_api = HfApi()
# Generation metadata storage with automatic cleanup
generation_metadata = deque(maxlen=100) # Fixed size deque to prevent unlimited growth
class RAMSavingIteratorStreamer:
"""
Custom streamer that saves VRAM by moving tokens to CPU and provides iteration interface for Gradio.
Combines the benefits of TextStreamer (RAM saving) with TextIteratorStreamer (iteration).
"""
def __init__(self, tokenizer, skip_special_tokens=True, skip_prompt=True, timeout=None):
self.tokenizer = tokenizer
self.skip_special_tokens = skip_special_tokens
self.skip_prompt = skip_prompt
self.timeout = timeout
# Token and text storage (CPU-based)
self.generated_tokens = []
self.generated_text = ""
self.token_cache = ""
# Queue for streaming interface
self.text_queue = queue.Queue()
self.stop_signal = threading.Event()
# Track prompt tokens to skip them
self.prompt_length = 0
self.tokens_processed = 0
# Decoding state
self.print_len = 0
def put(self, value):
"""
Receive new token(s) and process them for streaming.
This method is called by the model during generation.
"""
try:
# Handle different input types
if isinstance(value, torch.Tensor):
if value.dim() > 1:
value = value[0] # Remove batch dimension if present
token_ids = value.tolist()
# Store CPU version to save VRAM
self.generated_tokens.append(value.detach().cpu())
else:
token_ids = value if isinstance(value, list) else [value]
self.generated_tokens.append(torch.tensor(token_ids, dtype=torch.long))
# Track tokens processed
if isinstance(token_ids, list):
self.tokens_processed += len(token_ids)
else:
self.tokens_processed += 1
# Skip prompt tokens if requested
if self.skip_prompt and self.tokens_processed <= self.prompt_length:
return
# Decode incrementally for real-time streaming
try:
# Get all generated tokens so far
if self.generated_tokens:
all_tokens = []
for tokens in self.generated_tokens:
if isinstance(tokens, torch.Tensor):
if tokens.dim() == 0:
all_tokens.append(tokens.item())
else:
all_tokens.extend(tokens.tolist())
elif isinstance(tokens, list):
all_tokens.extend(tokens)
else:
all_tokens.append(tokens)
# Decode the full sequence
full_text = self.tokenizer.decode(
all_tokens,
skip_special_tokens=self.skip_special_tokens
)
# Get new text since last update
if len(full_text) > self.print_len:
new_text = full_text[self.print_len:]
self.print_len = len(full_text)
self.generated_text = full_text
# Put new text in queue for iteration
if new_text:
self.text_queue.put(new_text)
except Exception as decode_error:
logger.warning(f"Decoding error in streamer: {decode_error}")
except Exception as e:
logger.error(f"Error in RAMSavingIteratorStreamer.put: {e}")
def end(self):
"""Signal end of generation."""
self.text_queue.put(None) # Sentinel value
def __iter__(self):
"""Make this streamer iterable for Gradio compatibility."""
return self
def __next__(self):
"""Get next chunk of text for streaming."""
try:
value = self.text_queue.get(timeout=self.timeout)
if value is None: # End signal
raise StopIteration
return value
except queue.Empty:
raise StopIteration
def set_prompt_length(self, prompt_length):
"""Set the length of prompt tokens to skip."""
self.prompt_length = prompt_length
def get_generated_text(self):
"""Get the complete generated text."""
return self.generated_text
def get_generated_tokens(self):
"""Get all generated tokens as a single tensor."""
if not self.generated_tokens:
return torch.tensor([])
# Combine all tokens
all_tokens = []
for tokens in self.generated_tokens:
if isinstance(tokens, torch.Tensor):
if tokens.dim() == 0:
all_tokens.append(tokens.item())
else:
all_tokens.extend(tokens.tolist())
elif isinstance(tokens, list):
all_tokens.extend(tokens)
else:
all_tokens.append(tokens)
return torch.tensor(all_tokens, dtype=torch.long)
def cleanup(self):
"""Clean up resources."""
self.generated_tokens.clear()
self.generated_text = ""
self.token_cache = ""
# Clear queue
while not self.text_queue.empty():
try:
self.text_queue.get_nowait()
except queue.Empty:
break
self.stop_signal.set()
def scan_local_models(base_path=LOCAL_MODELS_BASE):
"""Scan for valid models in the local models directory"""
try:
base_path = Path(base_path)
if not base_path.exists():
logger.warning(f"Base path does not exist: {base_path}")
return []
valid_models = []
# Scan subdirectories (depth 1 only)
for item in base_path.iterdir():
if item.is_dir():
# Check if directory contains required model files
config_file = item / "config.json"
# Look for model weight files (safetensors or bin)
safetensors_files = list(item.glob("*.safetensors"))
bin_files = list(item.glob("*.bin"))
# Check if it's a valid model directory
if config_file.exists() and (safetensors_files or bin_files):
valid_models.append(str(item))
logger.info(f"Found valid model: {item}")
# Sort models for consistent ordering
valid_models.sort()
logger.info(f"Found {len(valid_models)} valid models in {base_path}")
return valid_models
except Exception as e:
logger.error(f"Error scanning local models: {e}")
return []
def update_local_models_dropdown(base_path):
"""Update the local models dropdown based on base path"""
if not base_path or not base_path.strip():
return gr.Dropdown(choices=[], value=None, interactive=True)
models = scan_local_models(base_path)
model_choices = [Path(model).name for model in models] # Show just the model name
model_paths = models # Keep full paths for internal use
# Create a mapping for display name to full path
if model_choices:
return gr.Dropdown(
choices=list(zip(model_choices, model_paths)),
value=model_paths[0] if model_paths else None,
label="๐Ÿ“‹ Available Local Models",
interactive=True,
allow_custom_value=False, # Don't allow custom for local models
filterable=True
)
else:
return gr.Dropdown(
choices=[],
value=None,
label="๐Ÿ“‹ Available Local Models (None found)",
interactive=True,
allow_custom_value=False,
filterable=True
)
def search_hf_models(query, limit=20):
"""Enhanced search for models on Hugging Face Hub with better coverage"""
if not query or len(query.strip()) < 2:
return []
try:
query = query.strip()
model_choices = []
# Strategy 1: Direct model ID search (if query looks like a model ID)
if '/' in query:
try:
# Try to get the specific model
model_info = hf_api.model_info(query)
if model_info and hasattr(model_info, 'id'):
model_choices.append(model_info.id)
logger.info(f"Found direct model: {model_info.id}")
except Exception as direct_error:
logger.debug(f"Direct model search failed: {direct_error}")
# Strategy 2: Search with different parameters
search_strategies = [
# Exact search
{"search": query, "sort": "downloads", "direction": -1, "limit": limit//2},
# Author search (if query contains /)
{"author": query.split('/')[0] if '/' in query else query, "sort": "downloads", "direction": -1, "limit": limit//4} if '/' in query else None,
# Broader search
{"search": query, "sort": "trending", "direction": -1, "limit": limit//4},
]
for strategy in search_strategies:
if strategy is None:
continue
try:
models = list_models(
task="text-generation",
**strategy
)
for model in models:
if model.id not in model_choices:
model_choices.append(model.id)
except Exception as strategy_error:
logger.debug(f"Search strategy failed: {strategy_error}")
# Remove duplicates while preserving order
seen = set()
unique_choices = []
for choice in model_choices:
if choice not in seen:
seen.add(choice)
unique_choices.append(choice)
# Limit results
final_choices = unique_choices[:limit]
logger.info(f"HF search for '{query}' returned {len(final_choices)} models")
return final_choices
except Exception as e:
logger.error(f"Error searching models: {str(e)}")
return []
def update_model_dropdown(query):
"""Update dropdown with enhanced search results"""
if not query or len(query.strip()) < 2:
return gr.Dropdown(choices=[], value=None, interactive=True)
choices = search_hf_models(query, limit=20)
return gr.Dropdown(
choices=choices,
value=choices[0] if choices else None,
interactive=True,
allow_custom_value=True, # Allow manual typing
filterable=True
)
def load_model_with_progress(model_source, hf_model, local_path, local_model_selection, quantization, memory_optimization):
"""Load model with progress tracking and memory optimization"""
global model, tokenizer, MODEL_NAME
# Determine model path based on source
if model_source == "Hugging Face Model":
if not hf_model:
return "โŒ Error: Please select a model from the dropdown"
model_path = hf_model
else:
# Use selected local model if available, otherwise use manual path
if local_model_selection:
model_path = local_model_selection
else:
model_path = local_path
if not Path(model_path).exists():
logger.error(f"Local path does not exist: {model_path}")
return f"โŒ Error: Local path does not exist: {model_path}"
MODEL_NAME = model_path.split("/")[-1] if "/" in model_path else model_path
logger.info(f"Loading model from {model_path} with memory optimization: {memory_optimization}")
try:
# Yield progress updates
yield "๐Ÿ”„ Initializing model loading..."
# Setup memory configuration (GPU-only, generous allocation)
if torch.cuda.is_available():
device_properties = torch.cuda.get_device_properties(0)
total_memory_gb = device_properties.total_memory / (1024**3)
# Set max memory to 11GB as requested (GPU-bound)
max_memory_val = 11.5 # Fixed 11GB allocation
max_memory = f"{max_memory_val}GB"
logger.info(f"Setting max GPU memory to {max_memory} (Total available: {total_memory_gb:.2f}GB)")
else:
max_memory = "11GB"
logger.info("CUDA not available. Using CPU fallback.")
yield "๐Ÿ”„ Configuring quantization settings..."
# Configure quantization (removed CPU offloading)
bnb_config = BitsAndBytesConfig(
load_in_4bit=quantization == "4bit",
load_in_8bit=quantization == "8bit",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_quant_type="nf4",
)
yield "๐Ÿ”„ Loading tokenizer..."
# Load tokenizer
if model_source == "Local Path":
tokenizer = AutoTokenizer.from_pretrained(
model_path,
trust_remote_code=True,
local_files_only=True
)
else:
tokenizer = AutoTokenizer.from_pretrained(
model_path,
trust_remote_code=True
)
yield "๐Ÿ”„ Cleaning memory cache..."
# Clean memory
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Determine torch dtype
if quantization in ["4bit", "8bit"]:
torch_dtype = torch.bfloat16
elif quantization == "f16":
torch_dtype = torch.float16
else: # bf16
torch_dtype = torch.bfloat16
yield "๐Ÿ”„ Loading model weights (this may take a while)..."
# Simple GPU-only model loading parameters
model_kwargs = {
"device_map": "auto",
"max_memory": {0: max_memory} if torch.cuda.is_available() else None,
"torch_dtype": torch_dtype,
"quantization_config": bnb_config if quantization in ["4bit", "8bit"] else None,
"trust_remote_code": True,
}
# Memory optimization specific settings (GPU-only)
if memory_optimization:
model_kwargs.update({
"attn_implementation": "flash_attention_2" if torch.cuda.is_available() else "sdpa",
"use_cache": False, # Disable cache by default for memory optimization
})
else:
model_kwargs.update({
"attn_implementation": "flash_attention_2" if torch.cuda.is_available() else "sdpa",
#"use_cache": True, # Enable cache for performance
})
# Add local files only for local models
if model_source == "Local Path":
model_kwargs["local_files_only"] = True
# Load model
model = AutoModelForCausalLM.from_pretrained(model_path, **model_kwargs)
# Post-loading memory optimization
if memory_optimization:
yield "๐Ÿ”„ Applying memory optimizations..."
# Additional memory cleanup after loading
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.synchronize()
logger.info("Model loaded successfully with memory optimization")
yield "โœ… Model loaded successfully with memory optimization!" if memory_optimization else "โœ… Model loaded successfully!"
except Exception as e:
logger.error(f"Error loading model: {str(e)}", exc_info=True)
yield f"โŒ Error loading model: {str(e)}"
def unload_model():
"""Unload the model and free memory with aggressive cleanup"""
global model, tokenizer, MODEL_NAME
if model is None:
return "No model loaded"
try:
logger.info("Unloading model with aggressive memory cleanup...")
# Step 1: Move model to CPU first (if it was on GPU)
if torch.cuda.is_available() and hasattr(model, 'device'):
try:
model.cpu()
logger.info("Model moved to CPU")
except Exception as cpu_error:
logger.warning(f"Could not move model to CPU: {cpu_error}")
# Step 2: Clear model cache if available
if hasattr(model, 'clear_cache'):
model.clear_cache()
# Step 3: Delete model and tokenizer references
del model
del tokenizer
model = None
tokenizer = None
# Step 4: Reset model name
MODEL_NAME = "No Model Loaded"
# Step 5: Clear metadata deque
generation_metadata.clear()
# Step 6: Aggressive garbage collection (multiple rounds)
for i in range(5): # More aggressive - 5 rounds
gc.collect()
time.sleep(0.1) # Small delay between rounds
# Step 7: Aggressive CUDA cleanup
if torch.cuda.is_available():
logger.info("Performing aggressive CUDA cleanup...")
# Multiple rounds of cache clearing
for i in range(5):
torch.cuda.empty_cache()
torch.cuda.synchronize()
# Additional PyTorch CUDA cleanup
if hasattr(torch.cuda, 'ipc_collect'):
torch.cuda.ipc_collect()
# Reset memory stats
if hasattr(torch.cuda, 'reset_peak_memory_stats'):
torch.cuda.reset_peak_memory_stats()
if hasattr(torch.cuda, 'reset_accumulated_memory_stats'):
torch.cuda.reset_accumulated_memory_stats()
time.sleep(0.1)
# Step 8: Force PyTorch to release all unused memory
if torch.cuda.is_available():
try:
# Try to trigger the memory pool cleanup
torch.cuda.empty_cache()
# Force a small allocation and deallocation to trigger cleanup
dummy_tensor = torch.zeros(1, device='cuda')
del dummy_tensor
torch.cuda.empty_cache()
logger.info("Forced memory pool cleanup")
except Exception as cleanup_error:
logger.warning(f"Advanced cleanup failed: {cleanup_error}")
# Step 9: Final garbage collection
gc.collect()
logger.info("Model unloaded successfully with aggressive cleanup")
return "โœ… Model unloaded with aggressive memory cleanup"
except Exception as e:
logger.error(f"Error unloading model: {str(e)}", exc_info=True)
# Emergency cleanup even if unload fails
model = None
tokenizer = None
MODEL_NAME = "No Model Loaded"
generation_metadata.clear()
# Emergency memory cleanup
for _ in range(3):
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return f"โŒ Error unloading model: {str(e)} (Emergency cleanup performed)"
def cleanup_memory():
"""Enhanced memory cleanup function with PyTorch optimizations"""
try:
# Clear Python garbage
gc.collect()
# Clear CUDA cache if available
if torch.cuda.is_available():
# Multiple aggressive cleanup rounds
for i in range(3):
torch.cuda.empty_cache()
torch.cuda.synchronize()
if hasattr(torch.cuda, 'ipc_collect'):
torch.cuda.ipc_collect()
# PyTorch specific memory management
if hasattr(torch.cuda, 'reset_peak_memory_stats'):
torch.cuda.reset_peak_memory_stats()
if hasattr(torch.cuda, 'reset_accumulated_memory_stats'):
torch.cuda.reset_accumulated_memory_stats()
# Brief pause between cleanup rounds
time.sleep(0.1)
# Clear metadata deque
generation_metadata.clear()
# Force garbage collection again
gc.collect()
logger.info("Enhanced memory cleanup completed")
return "๐Ÿงน Enhanced memory cleanup completed"
except Exception as e:
logger.error(f"Memory cleanup error: {e}")
return f"Memory cleanup error: {e}"
def nuclear_memory_cleanup():
"""Nuclear option: Complete VRAM reset (use if normal unload doesn't work)"""
global model, tokenizer, MODEL_NAME
try:
logger.info("Performing nuclear memory cleanup...")
# Force unload everything
model = None
tokenizer = None
MODEL_NAME = "No Model Loaded"
generation_metadata.clear()
# Import PyTorch again to reset some internal states
import torch
# Multiple aggressive cleanup rounds
for round_num in range(10): # Very aggressive - 10 rounds
gc.collect()
if torch.cuda.is_available():
# Multiple types of CUDA cleanup
torch.cuda.empty_cache()
torch.cuda.synchronize()
# Try to reset CUDA context
try:
if hasattr(torch.cuda, 'ipc_collect'):
torch.cuda.ipc_collect()
if hasattr(torch.cuda, 'memory_summary'):
logger.info(f"Round {round_num + 1}: {torch.cuda.memory_summary()}")
except Exception:
pass
# Reset memory stats
try:
if hasattr(torch.cuda, 'reset_peak_memory_stats'):
torch.cuda.reset_peak_memory_stats()
if hasattr(torch.cuda, 'reset_accumulated_memory_stats'):
torch.cuda.reset_accumulated_memory_stats()
except Exception:
pass
time.sleep(0.1)
# Final attempt: allocate and free a small tensor to trigger cleanup
if torch.cuda.is_available():
try:
for _ in range(5):
dummy = torch.zeros(1024, 1024, device='cuda') # 4MB tensor
del dummy
torch.cuda.empty_cache()
torch.cuda.synchronize()
except Exception as nuclear_error:
logger.warning(f"Nuclear tensor cleanup failed: {nuclear_error}")
logger.info("Nuclear memory cleanup completed")
return "โ˜ข๏ธ Nuclear memory cleanup completed! VRAM should be minimal now."
except Exception as e:
logger.error(f"Nuclear cleanup error: {e}")
return f"โ˜ข๏ธ Nuclear cleanup error: {e}"
def get_memory_stats():
"""Get comprehensive VRAM usage information"""
if not torch.cuda.is_available():
return """
<div style="text-align: center; padding: 15px; background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); border-radius: 10px; color: white;">
<h3 style="margin: 0; font-size: 16px;">๐Ÿ’ป CPU Mode</h3>
<p style="margin: 5px 0; opacity: 0.9;">GPU not available</p>
</div>
"""
try:
torch.cuda.synchronize()
total = torch.cuda.get_device_properties(0).total_memory / (1024**3)
allocated = torch.cuda.memory_allocated(0) / (1024**3)
reserved = torch.cuda.memory_reserved(0) / (1024**3)
free = total - reserved
usage_percent = (reserved/total)*100
# Get peak memory if available
peak_allocated = 0
if hasattr(torch.cuda, 'max_memory_allocated'):
peak_allocated = torch.cuda.max_memory_allocated(0) / (1024**3)
# Dynamic color based on usage
if usage_percent < 50:
color = "#10b981" # Green
elif usage_percent < 80:
color = "#f59e0b" # Orange
else:
color = "#ef4444" # Red
return f"""
<div style="text-align: center; padding: 15px; background: linear-gradient(135deg, {color}22 0%, {color}44 100%); border: 2px solid {color}; border-radius: 10px;">
<h3 style="margin: 0; font-size: 16px; color: {color};">๐ŸŽฎ VRAM Usage</h3>
<div style="margin: 10px 0;">
<div style="background: #f3f4f6; border-radius: 8px; height: 8px; overflow: hidden;">
<div style="width: {usage_percent}%; height: 100%; background: {color}; transition: width 0.3s ease;"></div>
</div>
</div>
<p style="margin: 5px 0; font-weight: 600;">Total: {total:.2f} GB</p>
<p style="margin: 5px 0;">Allocated: {allocated:.2f} GB ({usage_percent:.1f}%)</p>
<p style="margin: 5px 0;">Reserved: {reserved:.2f} GB</p>
<p style="margin: 5px 0;">Free: {free:.2f} GB</p>
<p style="margin: 5px 0; font-size: 12px; opacity: 0.8;">Peak: {peak_allocated:.2f} GB</p>
<p style="margin: 5px 0; font-size: 10px; opacity: 0.6;">RAM-Saving Streamer Active</p>
</div>
"""
except Exception as e:
logger.error(f"Error getting memory stats: {str(e)}")
return f"""
<div style="text-align: center; padding: 15px; background: #fee2e2; border: 2px solid #ef4444; border-radius: 10px;">
<h3 style="margin: 0; color: #ef4444;">โŒ Error</h3>
<p style="margin: 5px 0;">{str(e)}</p>
</div>
"""
def process_latex_content(text):
"""Enhanced LaTeX processing for streaming without UI glitches"""
# Don't process LaTeX here - let Gradio handle it natively
# Just return the text as-is for now
return text
def process_think_tags(text):
"""Process thinking tags with progressive streaming support"""
# Check if we're in the middle of generating a think section
if '<think>' in text and '</think>' not in text:
# We're currently generating inside a think section
parts = text.split('<think>')
if len(parts) == 2:
before_think = parts[0]
thinking_content = parts[1]
# Create a progressive thinking display
formatted_thinking = f"""
<div style="background: linear-gradient(135deg, #e0e7ff 0%, #c7d2fe 100%); border-left: 4px solid #6366f1; padding: 12px; margin: 8px 0; border-radius: 8px;">
<div style="display: flex; align-items: center; margin-bottom: 8px;">
<span style="font-size: 16px; margin-right: 8px;">๐Ÿค”</span>
<strong style="color: #4338ca;">Thinking...</strong>
</div>
<div style="color: #475569; font-style: italic;">{thinking_content}</div>
</div>
"""
return before_think + formatted_thinking
# Handle completed think sections
think_pattern = re.compile(r'<think>(.*?)</think>', re.DOTALL)
def replace_think(match):
think_content = match.group(1).strip()
return f"""
<div style="background: linear-gradient(135deg, #e0e7ff 0%, #c7d2fe 100%); border-left: 4px solid #6366f1; padding: 12px; margin: 8px 0; border-radius: 8px;">
<div style="display: flex; align-items: center; margin-bottom: 8px;">
<span style="font-size: 16px; margin-right: 8px;">๐Ÿค”</span>
<strong style="color: #4338ca;">Thinking...</strong>
</div>
<div style="color: #475569; font-style: italic;">{think_content}</div>
</div>
"""
# Replace completed <think> tags with formatted version
processed_text = think_pattern.sub(replace_think, text)
return processed_text
def calculate_generation_metrics(start_time, total_tokens):
"""Calculate generation metrics"""
end_time = time.time()
generation_time = end_time - start_time
tokens_per_second = total_tokens / generation_time if generation_time > 0 else 0
return {
"generation_time": generation_time,
"total_tokens": total_tokens,
"tokens_per_second": tokens_per_second,
"model_name": MODEL_NAME
}
def format_metadata_tooltip(metadata):
"""Format metadata for tooltip display"""
return f"""Model: {metadata['model_name']}
Tokens: {metadata['total_tokens']}
Speed: {metadata['tokens_per_second']:.2f} tok/s
Time: {metadata['generation_time']:.2f}s"""
def add_metadata_to_response(response_text, metadata):
"""Add metadata icon with tooltip to the response"""
tooltip_content = format_metadata_tooltip(metadata)
# Create a metadata icon with tooltip using HTML
metadata_html = f"""
<div style="position: relative; display: inline-block; margin-left: 8px;">
<span class="metadata-icon" style="cursor: help; opacity: 0.6; font-size: 14px;" title="{tooltip_content}">โ„น๏ธ</span>
</div>
"""
# Add metadata icon at the end of the response
return response_text + "\n\n" + metadata_html
def chat_with_model(message, history, system_prompt, temp, top_p_val, top_k_val, rep_penalty_val, memory_opt):
"""
Enhanced chat function with RAM-saving streamer and improved memory management.
Uses direct generation approach for better memory control and VRAM efficiency.
"""
global model, tokenizer, generation_metadata
# Check if model is loaded
if model is None or tokenizer is None:
return "โŒ Model not loaded. Please load the model first."
# Initialize variables for cleanup
input_ids = None
streamer = None
try:
# Record start time for metrics
start_time = time.time()
token_count = 0
# Format conversation for model
messages = [{"role": "system", "content": system_prompt}]
# Add chat history - HANDLE BOTH FORMATS (tuples from original and dicts from new)
for h in history:
if isinstance(h, dict):
# New dict format
if h.get("role") == "user":
messages.append({"role": "user", "content": h["content"]})
elif h.get("role") == "assistant":
messages.append({"role": "assistant", "content": h["content"]})
else:
# Original tuple format (user_msg, bot_msg)
if len(h) >= 2:
messages.append({"role": "user", "content": h[0]})
if h[1] is not None:
messages.append({"role": "assistant", "content": h[1]})
# Add the current message
messages.append({"role": "user", "content": message})
# Wrap generation in torch.no_grad() to prevent gradient accumulation
with torch.no_grad():
# Create model input with memory-efficient approach
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
)
# Handle edge case
if input_ids.ndim == 1:
input_ids = input_ids.unsqueeze(0)
# Move to device
input_ids = input_ids.to(model.device)
# Setup RAM-saving streamer
streamer = RAMSavingIteratorStreamer(
tokenizer,
skip_special_tokens=True,
skip_prompt=True,
timeout=1.0
)
# Set prompt length for the streamer
streamer.set_prompt_length(input_ids.shape[1])
# Pre-generation memory cleanup (only if memory optimization is on)
if memory_opt:
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Conditional generation parameters based on memory optimization
gen_kwargs = {
"input_ids": input_ids,
"max_new_tokens": MAX_LENGTH,
"temperature": temp,
"top_p": top_p_val,
"top_k": top_k_val,
"repetition_penalty": rep_penalty_val,
"do_sample": temp > 0,
"streamer": streamer,
"use_cache": not memory_opt, # Disable cache only if memory optimization is on
}
# Generate in a thread for real-time streaming
thread = Thread(
target=model.generate,
kwargs=gen_kwargs,
daemon=True
)
thread.start()
# Stream the response with conditional memory management
partial_text = ""
try:
for new_text in streamer:
partial_text += new_text
token_count += 1
# Process the text to handle think tags while preserving LaTeX
processed_text = process_think_tags(partial_text)
yield processed_text
# Conditional cleanup based on memory optimization setting (less frequent)
if memory_opt and token_count % 150 == 0: # Reduced frequency for performance
gc.collect() # Only light cleanup if memory optimization is on
except StopIteration:
# Normal end of generation
pass
except Exception as stream_error:
logger.error(f"Streaming error: {stream_error}")
yield f"โŒ Streaming error: {stream_error}"
return
finally:
# Add metadata to final response
try:
metrics = calculate_generation_metrics(start_time, token_count)
partial_text = add_metadata_to_response(partial_text, metrics)
except Exception as e:
logger.warning(f"Couldn't add metadata: {str(e)}")
yield partial_text
# Ensure thread completion
if thread.is_alive():
thread.join(timeout=5.0)
if thread.is_alive():
logger.warning("Generation thread did not complete in time")
# Calculate generation metrics
try:
metrics = calculate_generation_metrics(start_time, token_count)
# Store metadata (using deque with max size to prevent memory leaks)
generation_metadata.append(metrics)
# Log the metrics
logger.info(f"Generation metrics - Tokens: {metrics['total_tokens']}, Speed: {metrics['tokens_per_second']:.2f} tok/s, Time: {metrics['generation_time']:.2f}s")
except Exception as metrics_error:
logger.warning(f"Error calculating metrics: {metrics_error}")
# Final cleanup
try:
# Clean up streamer
if streamer:
streamer.cleanup()
del streamer
streamer = None
# Clean up input tensors
if input_ids is not None:
del input_ids
input_ids = None
# Conditional cleanup based on memory optimization setting
if memory_opt:
# Aggressive cleanup only if memory optimization is enabled
if torch.cuda.is_available():
for _ in range(2): # Reduced rounds for performance
torch.cuda.empty_cache()
torch.cuda.synchronize()
# Force garbage collection
for _ in range(2):
gc.collect()
else:
# Light cleanup for performance mode
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
logger.info(f"Generation completed, {token_count} tokens, memory_opt: {memory_opt}, VRAM saved with RAM-saving streamer")
except Exception as cleanup_error:
logger.warning(f"Final cleanup warning: {cleanup_error}")
except Exception as e:
logger.error(f"Error in chat_with_model: {str(e)}", exc_info=True)
# Emergency cleanup
try:
if streamer:
streamer.cleanup()
del streamer
if input_ids is not None:
del input_ids
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
except Exception as emergency_cleanup_error:
logger.error(f"Emergency cleanup failed: {emergency_cleanup_error}")
yield f"โŒ Error: {str(e)}"
def update_model_name():
"""Update the displayed model name"""
return f"๐Ÿ”ฎ AI Chat Assistant ({MODEL_NAME})"
def add_page_refresh_warning():
"""Add JavaScript to warn about page refresh when model is loaded"""
return """
<script>
window.addEventListener('beforeunload', function (e) {
// Check if model is loaded by looking for specific text in the page
const statusElements = document.querySelectorAll('input[type="text"], textarea');
let modelLoaded = false;
statusElements.forEach(element => {
if (element.value && element.value.includes('Model loaded successfully')) {
modelLoaded = true;
}
});
if (modelLoaded) {
e.preventDefault();
e.returnValue = 'A model is currently loaded. Are you sure you want to leave?';
return 'A model is currently loaded. Are you sure you want to leave?';
}
});
</script>
"""
# Custom CSS for elegant styling with fixed dropdown behavior
custom_css = """
/* Main container styling */
.gradio-container {
font-family: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', sans-serif !important;
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
min-height: 100vh;
}
/* Header styling */
.header-text {
background: rgba(255, 255, 255, 0.95);
backdrop-filter: blur(10px);
border-radius: 15px;
padding: 20px;
margin: 20px 0;
text-align: center;
box-shadow: 0 8px 32px rgba(0, 0, 0, 0.1);
border: 1px solid rgba(255, 255, 255, 0.2);
}
/* Chat interface styling */
.chat-container {
background: rgba(255, 255, 255, 0.95) !important;
border-radius: 20px !important;
box-shadow: 0 20px 40px rgba(0, 0, 0, 0.1) !important;
border: 1px solid rgba(255, 255, 255, 0.2) !important;
backdrop-filter: blur(10px) !important;
}
/* Control panel styling */
.control-panel {
background: rgba(255, 255, 255, 0.9) !important;
border-radius: 15px !important;
padding: 20px !important;
box-shadow: 0 10px 30px rgba(0, 0, 0, 0.1) !important;
border: 1px solid rgba(255, 255, 255, 0.3) !important;
backdrop-filter: blur(10px) !important;
overflow: visible !important; /* Allow dropdowns to overflow */
}
/* Button styling */
.btn-primary {
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
border: none !important;
border-radius: 10px !important;
color: white !important;
font-weight: 600 !important;
transition: all 0.3s ease !important;
box-shadow: 0 4px 15px rgba(102, 126, 234, 0.4) !important;
}
.btn-primary:hover {
transform: translateY(-2px) !important;
box-shadow: 0 8px 25px rgba(102, 126, 234, 0.6) !important;
}
.btn-secondary {
background: linear-gradient(135deg, #f093fb 0%, #f5576c 100%) !important;
border: none !important;
border-radius: 10px !important;
color: white !important;
font-weight: 600 !important;
transition: all 0.3s ease !important;
}
/* Input field styling */
.input-field {
border-radius: 10px !important;
border: 2px solid rgba(102, 126, 234, 0.2) !important;
transition: all 0.3s ease !important;
}
.input-field:focus {
border-color: #667eea !important;
box-shadow: 0 0 0 3px rgba(102, 126, 234, 0.1) !important;
}
/* Dropdown fixes */
.dropdown-container {
position: relative !important;
z-index: 1000 !important;
overflow: visible !important;
}
/* Fix dropdown menu positioning and styling */
.dropdown select,
.dropdown-menu,
.svelte-select,
.svelte-select-list {
position: relative !important;
z-index: 1001 !important;
background: white !important;
border: 2px solid rgba(102, 126, 234, 0.2) !important;
border-radius: 10px !important;
box-shadow: 0 4px 20px rgba(0, 0, 0, 0.15) !important;
max-height: 200px !important;
overflow-y: auto !important;
}
/* Fix dropdown option styling */
.dropdown option,
.svelte-select-option {
padding: 8px 12px !important;
background: white !important;
color: #333 !important;
border: none !important;
}
.dropdown option:hover,
.svelte-select-option:hover {
background: #f0f0f0 !important;
color: #667eea !important;
}
/* Ensure dropdown arrow is clickable */
.dropdown::after,
.dropdown-arrow {
pointer-events: none !important;
z-index: 1002 !important;
}
/* Fix any overflow issues in parent containers */
.gradio-group,
.gradio-column {
overflow: visible !important;
}
/* Accordion styling */
.accordion {
border-radius: 10px !important;
border: 1px solid rgba(102, 126, 234, 0.2) !important;
overflow: visible !important; /* Allow dropdowns to overflow accordion */
}
/* Status indicators */
.status-success {
color: #10b981 !important;
font-weight: 600 !important;
}
.status-error {
color: #ef4444 !important;
font-weight: 600 !important;
}
/* Reduced transition frequency to avoid conflicts */
.gradio-container * {
transition: background-color 0.3s ease, border-color 0.3s ease !important;
}
/* Chat bubble styling */
.message {
border-radius: 18px !important;
padding: 12px 16px !important;
margin: 8px 0 !important;
max-width: 80% !important;
}
.user-message {
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
color: white !important;
margin-left: auto !important;
}
.bot-message {
background: #f8fafc !important;
border: 1px solid #e2e8f0 !important;
}
/* Metadata tooltip styling - Enhanced */
.metadata-icon {
display: inline-block;
margin-left: 8px;
cursor: help;
opacity: 0.6;
transition: opacity 0.3s ease, transform 0.2s ease;
font-size: 14px;
user-select: none;
vertical-align: middle;
}
.metadata-icon:hover {
opacity: 1;
transform: scale(1.1);
}
/* Enhanced tooltip styling */
.metadata-icon[title]:hover::after {
content: attr(title);
position: absolute;
bottom: 100%;
left: 50%;
transform: translateX(-50%);
background: rgba(0, 0, 0, 0.9);
color: white;
padding: 8px 12px;
border-radius: 6px;
font-size: 12px;
white-space: pre-line;
z-index: 1000;
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.3);
margin-bottom: 5px;
min-width: 200px;
text-align: left;
}
.metadata-icon[title]:hover::before {
content: '';
position: absolute;
bottom: 100%;
left: 50%;
transform: translateX(-50%);
border: 5px solid transparent;
border-top-color: rgba(0, 0, 0, 0.9);
z-index: 1001;
}
/* Compact system prompt */
.compact-prompt {
min-height: 40px !important;
transition: min-height 0.3s ease !important;
}
.compact-prompt:focus {
min-height: 80px !important;
}
"""
# Main application
with gr.Blocks(css=custom_css, title="๐Ÿ”ฎ AI Chat Assistant") as demo:
# Add page refresh warning script
gr.HTML(add_page_refresh_warning())
# Header
with gr.Row():
title = gr.Markdown("# ๐Ÿ”ฎ AI Chat Assistant (No Model Loaded)", elem_classes="header-text")
with gr.Row(equal_height=True):
# Main chat area (left side - 70% width)
with gr.Column(scale=7, elem_classes="chat-container"):
# Compact system prompt (changed from 4 lines to 1)
system_prompt = gr.Textbox(
label="๐ŸŽฏ System Prompt",
value="You are a helpful AI assistant.",
lines=1, # Changed from 4 to 1
elem_classes="input-field compact-prompt"
)
# Generation settings in accordion
with gr.Accordion("โš™๏ธ Generation Settings", open=False, elem_classes="accordion"):
with gr.Row():
temperature = gr.Slider(0.0, 2.0, DEFAULT_TEMPERATURE, step=0.05, label="๐ŸŒก๏ธ Temperature")
top_p = gr.Slider(0.0, 1.0, DEFAULT_TOP_P, step=0.01, label="๐ŸŽฏ Top-p")
with gr.Row():
top_k = gr.Slider(1, 200, DEFAULT_TOP_K, step=1, label="๐Ÿ” Top-k")
rep_penalty = gr.Slider(1.0, 2.0, DEFAULT_REP_PENALTY, step=0.01, label="๐Ÿ”„ Repetition Penalty")
# Memory optimization for chat (moved here to be defined before use)
memory_opt_chat = gr.Checkbox(
label="๐Ÿง  Memory Optimization for Chat",
value=True,
info="Use memory optimization during chat generation (disables KV cache)"
)
# Chat interface using original gr.ChatInterface for fast streaming and stop button
chatbot = gr.Chatbot(
height=500,
latex_delimiters=[
{"left": "$", "right": "$", "display": True},
{"left": "$", "right": "$", "display": False},
{"left": "\\(", "right": "\\)", "display": False},
{"left": "\\[", "right": "\\]", "display": True}
],
show_copy_button=True,
avatar_images=("๐Ÿ‘ค", "๐Ÿค–"),
type="messages",
render_markdown=True
)
chat_interface = gr.ChatInterface(
fn=chat_with_model,
chatbot=chatbot,
additional_inputs=[system_prompt, temperature, top_p, top_k, rep_penalty, memory_opt_chat],
type="messages",
submit_btn="Send ๐Ÿ“ค",
stop_btn="โน๏ธ Stop"
)
# Control panel (right side - 30% width)
with gr.Column(scale=3, elem_classes="control-panel"):
# Model status and controls
with gr.Group():
gr.Markdown("### ๐Ÿš€ Model Controls")
with gr.Row():
load_btn = gr.Button("๐Ÿš€ Load Model", variant="primary", elem_classes="btn-primary")
unload_btn = gr.Button("๐Ÿ—‘๏ธ Unload", variant="secondary", elem_classes="btn-secondary")
model_status = gr.Textbox(
label="๐Ÿ“Š Status",
value="Model not loaded",
interactive=False,
elem_classes="input-field"
)
progress_display = gr.Textbox(
label="๐Ÿ“ˆ Progress",
value="Ready to load model",
interactive=False,
elem_classes="input-field"
)
# Model selection
with gr.Group():
gr.Markdown("### ๐ŸŽ›๏ธ Model Selection")
model_source = gr.Radio(
choices=["Hugging Face Model", "Local Path"],
value="Local Path", # Changed default to Local Path
label="๐Ÿ“ Model Source"
)
# HF Model search and selection (initially hidden)
with gr.Group(visible=False) as hf_group:
model_search = gr.Textbox(
label="๐Ÿ” Search Models",
placeholder="e.g., microsoft/Phi-3, meta-llama/Llama-3, ykarout/your-model",
elem_classes="input-field"
)
hf_model = gr.Dropdown(
label="๐Ÿ“‹ Select Model",
choices=[],
interactive=True,
elem_classes="input-field dropdown-container",
allow_custom_value=True, # Allow typing custom model names
filterable=True # Enable filtering
)
# Local path group (visible by default)
with gr.Group(visible=True) as local_group:
local_path = gr.Textbox(
value=LOCAL_MODELS_BASE, # Changed default to new base location
label="๐Ÿ“ Local Models Base Path",
elem_classes="input-field"
)
# Button to refresh local models
refresh_local_btn = gr.Button("๐Ÿ”„ Scan Local Models", elem_classes="btn-secondary")
# Dropdown for local models with better configuration
local_models_dropdown = gr.Dropdown(
label="๐Ÿ“‹ Available Local Models",
choices=[],
interactive=True,
elem_classes="input-field dropdown-container",
allow_custom_value=False, # Don't allow custom for local models
filterable=True # Enable filtering
)
quantization = gr.Radio(
choices=["4bit", "8bit", "bf16", "f16"],
value="4bit",
label="โšก Quantization"
)
# Advanced memory optimization toggle
memory_optimization = gr.Checkbox(
label="๐Ÿง  Advanced Memory Optimization",
value=True,
info="Reduces VRAM usage but may slightly impact speed"
)
# Note: Memory optimization for chat is now in Generation Settings
# Memory stats with cleanup buttons
with gr.Group():
gr.Markdown("### ๐Ÿ’พ System Status")
memory_info = gr.HTML()
with gr.Row():
refresh_btn = gr.Button("โ†ป Refresh Stats", elem_classes="btn-secondary")
cleanup_btn = gr.Button("๐Ÿงน Clean Memory", elem_classes="btn-secondary")
with gr.Row():
nuclear_btn = gr.Button("โ˜ข๏ธ Nuclear Cleanup", elem_classes="btn-secondary", variant="stop")
# Event handlers
# Model search functionality for HF
model_search.change(
update_model_dropdown,
inputs=[model_search],
outputs=[hf_model]
)
# Show/hide model selection based on source
def toggle_model_source(choice):
return (
gr.Group(visible=choice == "Hugging Face Model"),
gr.Group(visible=choice == "Local Path")
)
model_source.change(
toggle_model_source,
inputs=[model_source],
outputs=[hf_group, local_group]
)
# Local model scanning
refresh_local_btn.click(
update_local_models_dropdown,
inputs=[local_path],
outputs=[local_models_dropdown]
)
# Auto-scan on path change
local_path.change(
update_local_models_dropdown,
inputs=[local_path],
outputs=[local_models_dropdown]
)
# Model loading with progress
load_btn.click(
load_model_with_progress,
inputs=[model_source, hf_model, local_path, local_models_dropdown, quantization, memory_optimization],
outputs=[progress_display]
).then(
lambda: "โœ… Model loaded successfully!" if model is not None else "โŒ Model loading failed",
outputs=[model_status]
).then(
get_memory_stats,
outputs=[memory_info]
).then(
update_model_name,
outputs=[title]
)
# Model unloading
unload_btn.click(
unload_model,
outputs=[model_status]
).then(
lambda: "Ready to load model",
outputs=[progress_display]
).then(
get_memory_stats,
outputs=[memory_info]
).then(
lambda: "# ๐Ÿ”ฎ AI Chat Assistant (No Model Loaded)",
outputs=[title]
)
# Refresh memory stats
refresh_btn.click(get_memory_stats, outputs=[memory_info])
# Manual memory cleanup
cleanup_btn.click(cleanup_memory, outputs=[]).then(
get_memory_stats, outputs=[memory_info]
)
# Nuclear memory cleanup
nuclear_btn.click(nuclear_memory_cleanup, outputs=[]).then(
get_memory_stats, outputs=[memory_info]
)
# Initialize on load
demo.load(get_memory_stats, outputs=[memory_info])
demo.load(
lambda: update_local_models_dropdown(LOCAL_MODELS_BASE),
outputs=[local_models_dropdown]
)
# Enable queue for streaming
demo.queue()