diff --git a/.gitignore b/.gitignore index f40d95d8c1..7cfe662671 100755 --- a/.gitignore +++ b/.gitignore @@ -12,12 +12,15 @@ __pycache__/ .venv/ venv/ env/ +environment.yaml # Unsloth cache unsloth_compiled_cache/ # ML artifacts (large files) outputs/ +exports/ +unsloth_training_checkpoints/ *.gguf *.safetensors @@ -28,6 +31,9 @@ outputs/ *.swp *.swo +# oh-my-codex +.omx/ + # OS .DS_Store Thumbs.db @@ -47,3 +53,5 @@ docs/canvas-lab-architecture.md studio/frontend/test/ studio/tests/ studio/backend/tests/ +log_rtx.txt +log.txt diff --git a/README.md b/README.md index 4d2092b3a9..3d2810fae5 100644 --- a/README.md +++ b/README.md @@ -45,12 +45,14 @@ This script will: ### Launch the studio ```bash -# After setup, open a new terminal (or source ~/.bashrc), then: +# After setup, open a new terminal (or source ~/.bashrc), then inside your working directory: unsloth-ui -H 0.0.0.0 -p 8000 ``` On **first launch**, a one-time setup token is printed to the console. Use it in the browser to create your admin account. +As this repo is in continuous development, please make sure to run the setup.sh file everytime you pull new changes from the repo. + ## API Reference All endpoints require a valid JWT `Authorization: Bearer ` header (except `/api/auth/*` and `/api/health`). diff --git a/Unsloth_Studio_Colab.ipynb b/Unsloth_Studio_Colab.ipynb new file mode 100644 index 0000000000..fb2ca72604 --- /dev/null +++ b/Unsloth_Studio_Colab.ipynb @@ -0,0 +1,99 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": null, + "id": "447c1156", + "metadata": {}, + "outputs": [], + "source": [ + "# ===========================================\n", + "# āš ļø GPU Check - Run This First!\n", + "# ===========================================\n", + "import torch\n", + "\n", + "print(\"šŸ” Checking for GPU...\")\n", + "if not torch.cuda.is_available():\n", + " print(\"āŒ ERROR: No GPU detected!\")\n", + " print(\"\\nšŸ“‹ To enable GPU:\")\n", + " print(\" 1. Go to: Runtime ā†’ Change runtime type\")\n", + " print(\" 2. Select: Hardware accelerator ā†’ GPU (T4 is free)\")\n", + " print(\" 3. Click: Save\")\n", + " print(\" 4. Restart and re-run all cells\")\n", + " raise RuntimeError(\"ā›” GPU required for Unsloth Studio\")\n", + "else:\n", + " gpu_name = torch.cuda.get_device_name(0)\n", + " print(f\"āœ… GPU detected: {gpu_name}\")\n", + " print(\" Ready to proceed!\")" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "f04a9b46", + "metadata": {}, + "outputs": [], + "source": [ + "# ===========================================\n", + "# GitHub Authentication (Private Repo)\n", + "# ===========================================\n", + "from getpass import getpass\n", + "import os\n", + "\n", + "print(\"šŸ” GitHub Token Required\")\n", + "print(\"Get token: https://github.com/settings/tokens\")\n", + "print(\"Scope needed: 'repo'\")\n", + "print(\"-\" * 50)\n", + "\n", + "github_token = getpass(\"Enter GitHub Token: \")\n", + "os.environ['GITHUB_TOKEN'] = github_token\n", + "print(\"āœ… Token stored\")" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "27e68f91", + "metadata": {}, + "outputs": [], + "source": [ + "# ===========================================\n", + "# Setup: Clone repo and run setup\n", + "# ===========================================\n", + "\n", + "import os\n", + "github_token = os.environ['GITHUB_TOKEN']\n", + "!git clone -b feature/colab-notebook https://{github_token}@github.com/unslothai/new-ui-prototype.git\n", + "%cd /content/new-ui-prototype\n", + "\n", + "# Run setup script\n", + "!chmod +x setup.sh\n", + "!./setup.sh" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "277e431e", + "metadata": {}, + "outputs": [], + "source": [ + "# ===========================================\n", + "# Start Unsloth Studio\n", + "# ===========================================\n", + "import sys\n", + "sys.path.insert(0, '/content/new-ui-prototype/studio/backend')\n", + "\n", + "from colab import start\n", + "start()" + ] + } + ], + "metadata": { + "language_info": { + "name": "python" + } + }, + "nbformat": 4, + "nbformat_minor": 5 +} diff --git a/cli/commands/studio.py b/cli/commands/studio.py index 4287c756a8..2489676844 100644 --- a/cli/commands/studio.py +++ b/cli/commands/studio.py @@ -1,19 +1,34 @@ +import time +from pathlib import Path +from typing import Optional + import typer def studio( port: int = typer.Option(8000, "--port", "-p", help="Port to run the UI server on."), host: str = typer.Option("0.0.0.0", "--host", "-H", help="Host address to bind to."), - share: bool = typer.Option(True, "--share", "-s", help="Create a public Gradio share link."), + frontend: Optional[Path] = typer.Option(None, "--frontend", "-f", help="Path to frontend build directory."), + silent: bool = typer.Option(False, "--silent", "-q", help="Suppress startup messages."), ): - """Launch the Unsloth web UI for training, inference, and export.""" - from app import demo, script_dir + """Launch the Unsloth web UI backend server.""" + from studio.backend.run import run_server - typer.echo(f"Starting Unsloth UI on http://{host}:{port}") + if not silent: + from studio.backend.run import _resolve_external_ip + display_host = _resolve_external_ip() if host == "0.0.0.0" else host + typer.echo(f"Starting Unsloth Studio on http://{display_host}:{port}") - demo.launch( - share=share, - server_port=port, - server_name=host, - favicon_path=f"{script_dir}/assets/favicon-32x32.png", + run_server( + host=host, + port=port, + frontend_path=frontend, + silent=silent, ) + + # Keep running until interrupted + try: + while True: + time.sleep(1) + except KeyboardInterrupt: + typer.echo("\nShutting down...") diff --git a/setup.sh b/setup.sh index e5a41d371f..2db0c7f5a5 100755 --- a/setup.sh +++ b/setup.sh @@ -24,6 +24,13 @@ echo "ā•”ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā• echo "ā•‘ Unsloth Studio Setup Script ā•‘" echo "ā•šā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•ā•" +# ā”€ā”€ Detect Colab (like unsloth does) ā”€ā”€ +IS_COLAB=false +keynames=$'\n'$(printenv | cut -d= -f1) +if [[ "$keynames" == *$'\nCOLAB_'* ]]; then + IS_COLAB=true +fi + # ā”€ā”€ 1. Check existing Node/npm versions ā”€ā”€ NEED_NODE=true if command -v node &>/dev/null && command -v npm &>/dev/null; then @@ -33,7 +40,17 @@ if command -v node &>/dev/null && command -v npm &>/dev/null; then echo "āœ… Node $(node -v) and npm $(npm -v) already meet requirements. Skipping nvm install." NEED_NODE=false else - echo "āš ļø Node $(node -v) / npm $(npm -v) too old. Installing via nvm..." + if [ "$IS_COLAB" = true ]; then + echo "āœ… Node $(node -v) and npm $(npm -v) detected in Colab." + # In Colab, just upgrade npm directly - nvm doesn't work well + if [ "$NPM_MAJOR" -lt 11 ]; then + echo " Upgrading npm to latest..." + npm install -g npm@latest > /dev/null 2>&1 + fi + NEED_NODE=false + else + echo "āš ļø Node $(node -v) / npm $(npm -v) too old. Installing via nvm..." + fi fi else echo "āš ļø Node/npm not found. Installing via nvm..." @@ -81,38 +98,179 @@ echo "āœ… Frontend built to studio/frontend/dist" # ā”€ā”€ 6. Python venv + deps ā”€ā”€ echo "" echo "Setting up Python environment..." -python3 -m venv .venv -source .venv/bin/activate -run_quiet "pip upgrade" pip install --upgrade pip -echo " Installing unsloth-zoo + unsloth..." -run_quiet "pip install unsloth" pip install unsloth-zoo unsloth -echo " Installing studio dependencies..." -run_quiet "pip install extras" pip install typer fastapi uvicorn pydantic matplotlib pandas nest_asyncio "datasets==4.3.0" pyjwt easydict addict -echo "āœ… Python dependencies installed" -# ā”€ā”€ 7. Add shell alias ā”€ā”€ +# ā”€ā”€ 6a. Discover best Python <= 3.12.x ā”€ā”€ +BEST_PY="" +BEST_MAJOR=0 +BEST_MINOR=0 + +# Collect candidate python3 binaries (python3, python3.9, python3.10, ā€¦) +for candidate in $(compgen -c python3 2>/dev/null | grep -E '^python3(\.[0-9]+)?$' | sort -u); do + if ! command -v "$candidate" &>/dev/null; then + continue + fi + # Get version string, e.g. "Python 3.11.5" + ver_str=$("$candidate" --version 2>&1 | awk '{print $2}') + py_major=$(echo "$ver_str" | cut -d. -f1) + py_minor=$(echo "$ver_str" | cut -d. -f2) + + # Skip anything that isn't Python 3 + if [ "$py_major" -ne 3 ] 2>/dev/null; then + continue + fi + + # Skip versions above 3.12 + if [ "$py_minor" -gt 12 ] 2>/dev/null; then + continue + fi + + # Keep the highest qualifying version + if [ "$py_minor" -gt "$BEST_MINOR" ]; then + BEST_PY="$candidate" + BEST_MAJOR="$py_major" + BEST_MINOR="$py_minor" + fi +done + +if [ -z "$BEST_PY" ]; then + echo "āŒ ERROR: No Python version <= 3.12.x found on this system." + echo " Detected Python 3 installations:" + for candidate in $(compgen -c python3 2>/dev/null | grep -E '^python3(\.[0-9]+)?$' | sort -u); do + if command -v "$candidate" &>/dev/null; then + echo " - $candidate ($($candidate --version 2>&1))" + fi + done + echo "" + echo " Please install Python <= 3.12.x for maximum compatibility." + echo " For example: sudo apt install python3.12 python3.12-venv" + exit 1 +fi + +BEST_VER=$("$BEST_PY" --version 2>&1 | awk '{print $2}') +echo "āœ… Using $BEST_PY ($BEST_VER) ā€” compatible (ā‰¤ 3.12.x)" + +if [ "$IS_COLAB" = true ]; then + # Colab: install packages directly without venv + run_quiet "pip upgrade" pip install --upgrade pip + echo " Installing unsloth-zoo + unsloth..." + run_quiet "pip install unsloth" pip install -r "$SCRIPT_DIR/studio/backend/requirements/base.txt" + echo " Installing additional unsloth dependencies..." + run_quiet "pip install extras" pip install --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/extras.txt" + run_quiet "pip install extras" pip install --no-deps --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/extras-no-deps.txt" + run_quiet "pip install torchao+transformers" pip install --force-reinstall --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/overrides.txt" + run_quiet "pip install triton_kernels" pip install --no-deps -r "$SCRIPT_DIR/studio/backend/requirements/triton-kernels.txt" + # Patch: override llama_cpp.py with fix from unsloth-zoo branch + LLAMA_CPP_DST="$(pip show unsloth-zoo | grep -i '^Location:' | awk '{print $2}')/unsloth_zoo/llama_cpp.py" + curl -sSL "https://raw.githubusercontent.com/unslothai/unsloth-zoo/refs/heads/main/unsloth_zoo/llama_cpp.py" \ + -o "$LLAMA_CPP_DST" + echo " Installing studio dependencies..." + run_quiet "pip install studio" pip install -r "$SCRIPT_DIR/studio/backend/requirements/studio.txt" + echo "āœ… Python dependencies installed" +else + # Local: create venv (always start fresh to preserve correct install order) + rm -rf .venv + "$BEST_PY" -m venv .venv + source .venv/bin/activate + run_quiet "pip upgrade" pip install --upgrade pip + echo " Installing unsloth-zoo + unsloth..." + run_quiet "pip install unsloth" pip install -r "$SCRIPT_DIR/studio/backend/requirements/base.txt" + echo " Installing additional unsloth dependencies..." + run_quiet "pip install extras" pip install --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/extras.txt" + run_quiet "pip install extras" pip install --no-deps --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/extras-no-deps.txt" + run_quiet "pip install torchao+transformers" pip install --force-reinstall --no-cache-dir -r "$SCRIPT_DIR/studio/backend/requirements/overrides.txt" + run_quiet "pip install triton_kernels" pip install --no-deps -r "$SCRIPT_DIR/studio/backend/requirements/triton-kernels.txt" + # Patch: override llama_cpp.py with fix from unsloth-zoo branch + LLAMA_CPP_DST="$(pip show unsloth-zoo | grep -i '^Location:' | awk '{print $2}')/unsloth_zoo/llama_cpp.py" + curl -sSL "https://raw.githubusercontent.com/unslothai/unsloth-zoo/refs/heads/main/unsloth_zoo/llama_cpp.py" \ + -o "$LLAMA_CPP_DST" + echo " Installing studio dependencies..." + run_quiet "pip install studio" pip install -r "$SCRIPT_DIR/studio/backend/requirements/studio.txt" + echo "āœ… Python dependencies installed" + + # ā”€ā”€ 7. WSL: pre-install GGUF build dependencies ā”€ā”€ + # On WSL, sudo requires a password and can't be entered during GGUF export + # (runs in a non-interactive subprocess). Install build deps here instead. + if grep -qi microsoft /proc/version 2>/dev/null; then + echo "" + echo "āš ļø WSL detected ā€” installing build dependencies for GGUF export..." + echo " You may be prompted for your password." + sudo apt-get update -y + sudo apt-get install -y build-essential cmake curl git libcurl4-openssl-dev + echo "āœ… GGUF build dependencies installed" + fi +fi + +# ā”€ā”€ 8. Add shell alias (skip in Colab) ā”€ā”€ # Note: venv activation does NOT persist across terminal sessions. # This alias hardcodes the venv python path so users don't need to activate. +if [ "$IS_COLAB" = false ]; then echo "" REPO_DIR="$SCRIPT_DIR" -if ! grep -qF "unsloth-ui" ~/.bashrc 2>/dev/null; then - cat >> ~/.bashrc </dev/null; then + mkdir -p "$(dirname "$SHELL_RC")" # needed for fish's nested config path + cat >> "$SHELL_RC" < str: + """ + Get the actual Colab proxy URL for a port. + """ + try: + from google.colab.output import eval_js + + # Use Colab's proxy mechanism + url = eval_js(f"google.colab.kernel.proxyPort({port})", timeout_sec=5) + return url if url else f"http://localhost:{port}" + except Exception as e: + print(f"Note: Could not get Colab URL ({e})") + return f"http://localhost:{port}" + + +def show_link(port: int = 8000): + """Display a styled clickable link to the UI.""" + from IPython.display import display, HTML + + # Get real Colab proxy URL + url = get_colab_url(port) + + html = f""" +
+

+ šŸ¦„ Unsloth Studio is Ready! +

+ + šŸš€ Open Unsloth Studio + +

+ {url} +

+
+ """ + display(HTML(html)) + + +def start(port: int = 8000): + """ + Start Unsloth Studio server in Colab and display the URL. + + Usage: + from colab import start + start() + """ + import sys + + print("šŸ¦„ Starting Unsloth Studio...") + + # Add backend to path + backend_path = str(Path(__file__).parent) + if backend_path not in sys.path: + sys.path.insert(0, backend_path) + + print(" Loading backend...") + from run import run_server + + # Auto-detect frontend path + repo_root = Path(__file__).parent.parent + frontend_path = repo_root / "frontend" / "dist" + + if not frontend_path.exists(): + print("āŒ Frontend not built! Please run the setup cell first.") + return + + print(" Starting server...") + # Start server silently + run_server(host="0.0.0.0", port=port, frontend_path=frontend_path, silent=True) + + print(" Server started!") + + # Show the clickable link with real URL + show_link(port) + + +if __name__ == "__main__": + start() diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py index 8977586ac3..da5b11c60d 100644 --- a/studio/backend/core/export/export.py +++ b/studio/backend/core/export/export.py @@ -18,6 +18,45 @@ from core.inference import get_inference_backend logger = logging.getLogger(__name__) + +def _is_wsl(): + """Detect if running under Windows Subsystem for Linux.""" + try: + return "microsoft" in open("/proc/version").read().lower() + except Exception: + return False + + +def _apply_wsl_sudo_patch(): + """On WSL, monkey-patch do_we_need_sudo() to return False. + + WSL doesn't have passwordless sudo, and do_we_need_sudo() runs + `sudo apt-get update` which hangs waiting for a stdin password + inside a non-interactive subprocess. setup.sh pre-installs the + build dependencies on WSL, so sudo is not needed at runtime. + """ + if not _is_wsl(): + return + + try: + import unsloth_zoo.llama_cpp as llama_cpp_module + + def _wsl_do_we_need_sudo(system_type="debian"): + logger.info( + "WSL detected ā€” skipping sudo check " + "(build deps pre-installed by setup.sh)" + ) + return False + + llama_cpp_module.do_we_need_sudo = _wsl_do_we_need_sudo + logger.info( + "Applied WSL sudo patch to " + "unsloth_zoo.llama_cpp.do_we_need_sudo" + ) + except Exception as e: + logger.warning(f"Could not apply WSL sudo patch: {e}") + + # Model card template MODEL_CARD = \ """--- @@ -80,43 +119,15 @@ class ExportBackend: logger.error(f"Error during memory cleanup: {e}") return False - def scan_checkpoints(self, outputs_dir: str = "./outputs") -> List[Tuple[str, str]]: + def scan_checkpoints(self, outputs_dir: str = "./outputs") -> List[Tuple[str, List[Tuple[str, str]]]]: """ - Scan outputs folder for model checkpoints. + Scan outputs folder for training runs and their checkpoints. Returns: - List of tuples: [(display_name, checkpoint_path), ...] + List of tuples: [(model_name, [(display_name, checkpoint_path), ...]), ...] """ - checkpoints = [] - outputs_path = Path(outputs_dir) - - if not outputs_path.exists(): - logger.warning(f"Outputs directory not found: {outputs_dir}") - return checkpoints - - try: - for item in outputs_path.iterdir(): - if item.is_dir(): - # Check if this directory contains a model - config_file = item / "config.json" - adapter_config = item / "adapter_config.json" - - if config_file.exists() or adapter_config.exists(): - # This is a valid checkpoint - display_name = item.name - checkpoint_path = str(item) - checkpoints.append((display_name, checkpoint_path)) - logger.debug(f"Found checkpoint: {display_name}") - - # Sort by modification time (newest first) - checkpoints.sort(key=lambda x: Path(x[1]).stat().st_mtime, reverse=True) - - logger.info(f"Found {len(checkpoints)} checkpoints in {outputs_dir}") - return checkpoints - - except Exception as e: - logger.error(f"Error scanning checkpoints: {e}") - return [] + from utils.models.checkpoints import scan_checkpoints + return scan_checkpoints(outputs_dir=outputs_dir) def load_checkpoint(self, checkpoint_path: str, @@ -264,7 +275,8 @@ class ExportBackend: push_to_hub: bool = False, repo_id: Optional[str] = None, hf_token: Optional[str] = None, - private: bool = False) -> Tuple[bool, str]: + private: bool = False, + base_model_id: Optional[str] = None) -> Tuple[bool, str]: """ Export base model (for non-PEFT models). @@ -294,8 +306,8 @@ class ExportBackend: logger.info(f"Pushing base model to Hub: {repo_id}") - # Get base model name - base_model = self.current_model.config._name_or_path + # Get base model name from request or model config + base_model = base_model_id or self.current_model.config._name_or_path or "unknown" # Create repo hf_api = HfApi(token=hf_token) @@ -380,6 +392,9 @@ class ExportBackend: os.chdir(save_directory) logger.info(f"Changed directory to: {save_directory}") + # On WSL, patch out sudo check before llama.cpp build + _apply_wsl_sudo_patch() + # Now save (will save in current directory) self.current_model.save_pretrained_gguf( "model", # Base filename diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py index 23203613f2..99a7c485e7 100644 --- a/studio/backend/core/inference/inference.py +++ b/studio/backend/core/inference/inference.py @@ -38,13 +38,22 @@ class InferenceBackend: ] self.device = get_device().value - # Thread safety + # Thread safety ā€” _generation_lock serializes model.generate() calls. + # Must be a regular Lock (NOT RLock) because in async FastAPI, multiple + # requests share the same event-loop thread, so RLock reentrancy lets + # concurrent compare-mode requests race on the GPU. The lock is + # acquired by the *background generation thread*, not the event-loop. import threading - self._generation_lock = threading.RLock() + self._generation_lock = threading.Lock() self._model_state_lock = threading.Lock() logger.info(f"InferenceBackend initialized on {self.device}") + @staticmethod + def _normalize_top_k(top_k: int) -> int: + # API supports -1 as "disable top-k"; transformers expects 0 to disable. + return 0 if top_k < 0 else top_k + def load_model(self, config: ModelConfig, max_seq_length: int = 2048, @@ -443,9 +452,10 @@ class InferenceBackend: """ Apply adapter state before generation. Must be called under _generation_lock. - Uses revert_to_base_model() / activate_lora_adapter() which work correctly - for models loaded by Unsloth as complete PeftModels (via model.unload() / - model.load_adapter()), matching the proven pattern from the Gradio eval page. + Uses PEFT's disable_adapter_layers() / enable_adapter_layers() which toggle + a boolean flag on each LoRA layer. Unsloth's fast_linear_forward checks this + flag (proj.disable_adapters) and skips LoRA computation when True. + This is non-destructive ā€” no model unloading/reloading needed. Args: use_adapter: None = no change, False = disable (base model), @@ -459,53 +469,56 @@ class InferenceBackend: return model_info = self.models[base] + model = model_info.get("model") + if model is None: + return if use_adapter is False: - # Revert to pure base model by unloading adapter weights - logger.info(f"Compare mode: reverting '{base}' to base model for generation") - self.revert_to_base_model(base) + # Disable LoRA layers ā†’ base model output + if isinstance(model, (PeftModel, PeftModelForCausalLM)): + logger.info(f"Compare mode: disabling adapters on '{base}' for base model generation") + model.base_model.disable_adapter_layers() + else: + logger.info(f"Compare mode: model '{base}' is not a PeftModel, already base") elif use_adapter is True: - # Activate the LoRA adapter from the original model path - lora_path = model_info.get("model_path") - if lora_path and model_info.get("is_lora"): - logger.info(f"Compare mode: activating LoRA adapter from '{lora_path}' on '{base}'") - self.activate_lora_adapter(base, lora_path) + # Re-enable LoRA layers ā†’ adapter output + if isinstance(model, (PeftModel, PeftModelForCausalLM)): + logger.info(f"Compare mode: enabling adapters on '{base}' for LoRA generation") + model.base_model.enable_adapter_layers() else: - # Fallback for dynamically attached adapters - loaded = model_info.get("loaded_adapters", {}) - if loaded: - adapter_name = list(loaded.keys())[-1] - logger.info(f"Compare mode: enabling adapter '{adapter_name}' on '{base}'") - self.set_active_adapter(base, adapter_name) - else: - logger.warning("use_adapter=true but no adapter path/adapters on model") + logger.warning("use_adapter=true but model is not a PeftModel") elif isinstance(use_adapter, str): - # Activate a specific adapter by path - logger.info(f"Compare mode: activating specific adapter '{use_adapter}' on '{base}'") - self.activate_lora_adapter(base, use_adapter) + # Enable adapters and set the specific one active + if isinstance(model, (PeftModel, PeftModelForCausalLM)): + logger.info(f"Compare mode: enabling adapter '{use_adapter}' on '{base}'") + model.base_model.enable_adapter_layers() + self.set_active_adapter(base, use_adapter) + else: + logger.warning(f"use_adapter='{use_adapter}' but model is not a PeftModel") def generate_with_adapter_control( self, use_adapter: Optional[Union[bool, str]] = None, + cancel_event=None, **gen_kwargs, ) -> Generator[str, None, None]: """ Thread-safe generation with optional adapter toggling. - Acquires the generation lock, applies adapter state, then generates. - This ensures adapter toggle + generation are atomic ā€” critical for - compare mode where base and LoRA panes fire concurrently. + The adapter toggle + model.generate() are serialized by _generation_lock + inside the background generation thread ā€” NOT in the event-loop thread. + This prevents the RLock-reentrant race that occurs when two async SSE + handlers share the same event-loop thread. Args: use_adapter: Adapter control (None/False/True/str). See _apply_adapter_state. **gen_kwargs: Forwarded to generate_chat_response. """ - with self._generation_lock: - self._apply_adapter_state(use_adapter) - # Delegate to the lock-free generation path - yield from self._generate_chat_response_inner(**gen_kwargs) + yield from self._generate_chat_response_inner( + cancel_event=cancel_event, _adapter_state=use_adapter, **gen_kwargs + ) def generate_chat_response(self, messages: list, @@ -514,24 +527,26 @@ class InferenceBackend: temperature: float = 0.7, top_p: float = 0.9, top_k: int = 40, + min_p: float = 0.0, max_new_tokens: int = 256, - repetition_penalty: float = 1.1) -> Generator[str, None, None]: + repetition_penalty: float = 1.1, + cancel_event=None) -> Generator[str, None, None]: """ Generate response for text or vision models. - Acquires the generation lock. For adapter-controlled generation, - use generate_with_adapter_control() instead. + The generation lock is acquired by the background generation thread. """ - with self._generation_lock: - yield from self._generate_chat_response_inner( - messages=messages, - system_prompt=system_prompt, - image=image, - temperature=temperature, - top_p=top_p, - top_k=top_k, - max_new_tokens=max_new_tokens, - repetition_penalty=repetition_penalty, - ) + yield from self._generate_chat_response_inner( + messages=messages, + system_prompt=system_prompt, + image=image, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + max_new_tokens=max_new_tokens, + repetition_penalty=repetition_penalty, + cancel_event=cancel_event, + ) def _generate_chat_response_inner(self, messages: list, @@ -540,11 +555,17 @@ class InferenceBackend: temperature: float = 0.7, top_p: float = 0.9, top_k: int = 40, + min_p: float = 0.0, max_new_tokens: int = 256, - repetition_penalty: float = 1.1) -> Generator[str, None, None]: + repetition_penalty: float = 1.1, + cancel_event=None, + _adapter_state=None) -> Generator[str, None, None]: """ - Inner generation logic (no lock). Called by both generate_chat_response + Inner generation logic. Called by both generate_chat_response and generate_with_adapter_control. + + _adapter_state is passed to generate_stream/vision so the background + thread can toggle adapters under the generation lock. """ if not self.active_model_name: yield "Error: No active model" @@ -553,12 +574,14 @@ class InferenceBackend: model_info = self.models[self.active_model_name] is_vision = model_info.get("is_vision", False) tokenizer = model_info.get("tokenizer") or model_info.get("processor") + top_k = self._normalize_top_k(top_k) if is_vision: # Vision model generation yield from self._generate_vision_response( messages, system_prompt, image, - temperature, top_p, top_k, max_new_tokens, repetition_penalty + temperature, top_p, top_k, min_p, max_new_tokens, repetition_penalty, + cancel_event=cancel_event, ) else: # Text model: Use training pipeline approach @@ -600,12 +623,14 @@ class InferenceBackend: # Step 3: Generate yield from self.generate_stream( - formatted_prompt, temperature, top_p, top_k, max_new_tokens, repetition_penalty + formatted_prompt, temperature, top_p, top_k, min_p, max_new_tokens, repetition_penalty, + cancel_event=cancel_event, + _adapter_state=_adapter_state, ) def _generate_vision_response(self, messages, system_prompt, image, - temperature, top_p, top_k, max_new_tokens, - repetition_penalty) -> Generator[str, None, None]: + temperature, top_p, top_k, min_p, max_new_tokens, + repetition_penalty, cancel_event=None) -> Generator[str, None, None]: """Handle vision model generation with true token-by-token streaming.""" model_info = self.models[self.active_model_name] model = model_info["model"] @@ -651,7 +676,10 @@ class InferenceBackend: import threading streamer = TextIteratorStreamer( - processor.tokenizer, skip_prompt=True, skip_special_tokens=True + processor.tokenizer, + skip_prompt=True, + skip_special_tokens=True, + timeout=0.2, ) generation_kwargs = dict( @@ -659,28 +687,58 @@ class InferenceBackend: streamer=streamer, max_new_tokens=max_new_tokens, use_cache=True, + do_sample=temperature > 0, temperature=temperature, top_p=top_p, top_k=top_k, + min_p=min_p, ) + err: dict[str, str] = {} + def generate_fn(): - try: - model.generate(**generation_kwargs) - except Exception as e: - logger.error(f"Vision generation error in thread: {e}") + with self._generation_lock: + try: + model.generate(**generation_kwargs) + except Exception as e: + err["msg"] = str(e) + logger.error(f"Vision generation error in thread: {e}") + finally: + try: + streamer.end() + except Exception: + pass thread = threading.Thread(target=generate_fn) thread.start() output = "" - for new_token in streamer: - if new_token: - output += new_token - cleaned = self._clean_generated_text(output) - yield cleaned + from queue import Empty + try: + while True: + if cancel_event is not None and cancel_event.is_set(): + break + try: + new_token = next(streamer) + except StopIteration: + break + except Empty: + if not thread.is_alive(): + break + continue + if new_token: + output += new_token + cleaned = self._clean_generated_text(output) + yield cleaned + finally: + if cancel_event is not None: + cancel_event.set() + thread.join(timeout=10) + if thread.is_alive(): + logger.warning("Vision generation thread did not exit after cancel/join timeout") - thread.join() + if err.get("msg"): + yield f"Error: {err['msg']}" except Exception as e: logger.error(f"Vision generation error: {e}") @@ -692,9 +750,16 @@ class InferenceBackend: temperature: float = 0.7, top_p: float = 0.9, top_k: int = 40, + min_p: float = 0.0, max_new_tokens: int = 256, - repetition_penalty: float = 1.1) -> Generator[str, None, None]: - """Generate streaming text response (text models only).""" + repetition_penalty: float = 1.1, + cancel_event=None, + _adapter_state=None) -> Generator[str, None, None]: + """Generate streaming text response (text models only). + + _adapter_state: if not None, the background thread toggles adapters + before model.generate(), all under _generation_lock. + """ if not self.active_model_name: yield "Error: No active model" return @@ -709,7 +774,12 @@ class InferenceBackend: from transformers import TextIteratorStreamer import threading - streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) + streamer = TextIteratorStreamer( + tokenizer, + skip_prompt=True, + skip_special_tokens=True, + timeout=0.2, + ) generation_kwargs = dict( **inputs, @@ -718,29 +788,75 @@ class InferenceBackend: temperature=temperature, top_p=top_p, top_k=top_k, + min_p=min_p, repetition_penalty=repetition_penalty, - do_sample=True, + do_sample=temperature > 0, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id if tokenizer.pad_token_id is None else tokenizer.pad_token_id, ) + if cancel_event is not None: + from transformers.generation.stopping_criteria import ( + StoppingCriteria, + StoppingCriteriaList, + ) + + class _CancelCriteria(StoppingCriteria): + def __init__(self, ev): + self.ev = ev + + def __call__(self, input_ids, scores, **kwargs): + return self.ev.is_set() + + generation_kwargs["stopping_criteria"] = StoppingCriteriaList( + [_CancelCriteria(cancel_event)] + ) def generate_fn(): - try: - model.generate(**generation_kwargs) - except Exception as e: - logger.error(f"Generation error: {e}") + with self._generation_lock: + try: + if _adapter_state is not None: + self._apply_adapter_state(_adapter_state) + model.generate(**generation_kwargs) + except Exception as e: + err["msg"] = str(e) + logger.error(f"Generation error: {e}") + finally: + try: + streamer.end() + except Exception: + pass + err: dict[str, str] = {} thread = threading.Thread(target=generate_fn) thread.start() output = "" - for new_token in streamer: - if new_token: - output += new_token - cleaned = self._clean_generated_text(output) - yield cleaned + from queue import Empty + try: + while True: + if cancel_event is not None and cancel_event.is_set(): + break + try: + new_token = next(streamer) + except StopIteration: + break + except Empty: + if not thread.is_alive(): + break + continue + if new_token: + output += new_token + cleaned = self._clean_generated_text(output) + yield cleaned + finally: + if cancel_event is not None: + cancel_event.set() + thread.join(timeout=10) + if thread.is_alive(): + logger.warning("Generation thread did not exit after cancel/join timeout") - thread.join() + if err.get("msg"): + yield f"Error: {err['msg']}" except Exception as e: logger.error(f"Error during generation: {e}") diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py index 36067747cd..c36fa235ed 100644 --- a/studio/backend/core/training/trainer.py +++ b/studio/backend/core/training/trainer.py @@ -16,6 +16,7 @@ import json import threading import math import logging +import time from typing import Optional, Callable from dataclasses import dataclass import pandas as pd @@ -46,6 +47,11 @@ class TrainingProgress: is_completed: bool = False error: Optional[str] = None status_message: str = "Ready to train" # Current stage message + elapsed_seconds: Optional[float] = None + eta_seconds: Optional[float] = None + grad_norm: Optional[float] = None + num_tokens: Optional[int] = None + eval_loss: Optional[float] = None class UnslothTrainer: """ @@ -67,6 +73,12 @@ class UnslothTrainer: self.is_vlm = False self.model_name = None + # Training metrics tracking + self.training_start_time: Optional[float] = None + self.batch_size: Optional[int] = None + self.max_seq_length: Optional[int] = None + self.gradient_accumulation_steps: Optional[int] = None + # Thread safety self._lock = threading.Lock() @@ -99,17 +111,29 @@ class UnslothTrainer: model_name: str, max_seq_length: int = 2048, load_in_4bit: bool = True, - hf_token: Optional[str] = None) -> bool: + hf_token: Optional[str] = None, + is_dataset_multimodal: bool = False) -> bool: """Load model for training (supports both text and vision models)""" try: + if self.model is not None: + del self.model + if self.tokenizer is not None: + del self.tokenizer + + if self.trainer is not None: + del self.trainer + print("\nClearing GPU memory before training...") clear_gpu_cache() - # Detect if this is a vision model first - self.is_vlm = is_vision_model(model_name) + # Detect if this is a vision model AND dataset is multimodal + # A vision-capable model with a text-only dataset should use FastLanguageModel + self.is_vlm = is_vision_model(model_name) and is_dataset_multimodal self.model_name = model_name - logger.info(f"Model type detected: {'Vision' if self.is_vlm else 'Text'}") + logger.info(f"Model architecture is vision: {is_vision_model(model_name)}") + logger.info(f"Dataset is multimodal: {is_dataset_multimodal}") + logger.info(f"Using VLM path: {self.is_vlm}") # Reset training state for new run self._update_progress( @@ -198,7 +222,13 @@ class UnslothTrainer: return True # LoRA/QLoRA mode - apply PEFT - if target_modules is None or (isinstance(target_modules, list) and len(target_modules) == 0): + # "all-linear" is a PEFT keyword that targets every linear layer + if isinstance(target_modules, list) and "all-linear" in target_modules: + if len(target_modules) == 1: + target_modules = "all-linear" + else: + target_modules = [m for m in target_modules if m != "all-linear"] + elif target_modules is None or (isinstance(target_modules, list) and len(target_modules) == 0): target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] @@ -306,12 +336,24 @@ class UnslothTrainer: dataset_source: str, format_type: str = "auto", local_datasets: list = None, - custom_format_mapping: dict = None) -> Optional[Dataset]: + custom_format_mapping: dict = None, + subset: str = None, + train_split: str = "train", + eval_split: str = None) -> Optional[tuple]: """ - Load and prepare dataset for training + Load and prepare dataset for training. + + Strategy: format first, then split ā€” ensures both train and eval + portions are properly formatted and templated. + + Returns: + Tuple of (dataset_info, eval_dataset) or None on error. + eval_dataset may be None if no eval split is available. """ try: dataset = None + eval_dataset = None + has_separate_eval_source = False # True if eval comes from a separate HF split if local_datasets: # Load local datasets @@ -350,7 +392,10 @@ class UnslothTrainer: elif dataset_source: # Load from Hugging Face - dataset = load_dataset(dataset_source, split="train") + load_kwargs = {"path": dataset_source, "split": train_split or "train"} + if subset: + load_kwargs["name"] = subset + dataset = load_dataset(**load_kwargs) # Check if stopped during dataset loading if self.should_stop: @@ -360,6 +405,25 @@ class UnslothTrainer: self._update_progress(status_message=f"Loaded dataset from HuggingFace: {dataset_source}") print(f"Loaded dataset from Hugging Face: {dataset_source}\n") + # Resolve eval split from a separate HF split (explicit or auto-detected) + if eval_split: + # Explicit eval split provided - load it directly + print(f"Loading explicit eval split: '{eval_split}'\n") + eval_load_kwargs = {"path": dataset_source, "split": eval_split} + if subset: + eval_load_kwargs["name"] = subset + eval_dataset = load_dataset(**eval_load_kwargs) + has_separate_eval_source = True + print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n") + else: + # Auto-detect eval split from HF (returns a separate dataset, or None) + eval_dataset = self._auto_detect_eval_split_from_hf( + dataset_source=dataset_source, + subset=subset, + ) + if eval_dataset is not None: + has_separate_eval_source = True + if dataset is None: raise ValueError("No dataset provided") @@ -368,16 +432,15 @@ class UnslothTrainer: print("Stopped before applying chat template\n") return None - # NEW: Use unified format_and_template_dataset + # ========== FORMAT FIRST ========== print(f"Formatting dataset with format_type='{format_type}'...\n") - #breakpoint() dataset_info = format_and_template_dataset( dataset, model_name=self.model_name, - tokenizer=self.tokenizer, # Works for both text and vision models + tokenizer=self.tokenizer, is_vlm=self.is_vlm, - format_type=format_type, # "auto", "alpaca", "chatml", "sharegpt" + format_type=format_type, dataset_name=dataset_source, custom_format_mapping=custom_format_mapping, ) @@ -389,15 +452,94 @@ class UnslothTrainer: self._update_progress(status_message=f"Dataset formatted and ready for training") print(f"Dataset formatted successfully\n") - return dataset_info + + # ========== THEN SPLIT ========== + if has_separate_eval_source and eval_dataset is not None: + # Eval came from a separate HF split ā€” format it too + print(f"Formatting eval dataset ({len(eval_dataset)} rows)...\n") + eval_info = format_and_template_dataset( + eval_dataset, + model_name=self.model_name, + tokenizer=self.tokenizer, + is_vlm=self.is_vlm, + format_type=format_type, + dataset_name=dataset_source, + custom_format_mapping=custom_format_mapping, + ) + eval_dataset = eval_info["dataset"] + print(f"Eval dataset formatted successfully\n") + elif not has_separate_eval_source: + # No separate eval source ā€” split the already-formatted dataset + formatted_dataset = dataset_info["dataset"] + split_result = self._resolve_eval_split_from_dataset(formatted_dataset) + if split_result is not None: + train_portion, eval_dataset = split_result + dataset_info["dataset"] = train_portion + + return (dataset_info, eval_dataset) except Exception as e: logger.error(f"Error loading dataset: {e}") self._update_progress(error=str(e)) return None + def _auto_detect_eval_split_from_hf(self, dataset_source: str, + subset: str) -> Optional[Dataset]: + """Auto-detect an eval split from HF dataset (separate named split only).""" + try: + from datasets import get_dataset_split_names + load_kwargs = {"path": dataset_source} + if subset: + load_kwargs["name"] = subset + available_splits = get_dataset_split_names(**load_kwargs) + print(f"Available splits: {available_splits}\n") + + # Check for common eval split names + for candidate in ["eval", "validation", "valid", "val", "test"]: + if candidate in available_splits: + eval_load_kwargs = {"path": dataset_source, "split": candidate} + if subset: + eval_load_kwargs["name"] = subset + candidate_ds = load_dataset(**eval_load_kwargs) + if len(candidate_ds) >= 16: + print(f"Auto-detected eval split '{candidate}' with {len(candidate_ds)} rows\n") + return candidate_ds + else: + print(f"Found eval split '{candidate}' but only {len(candidate_ds)} rows (< 16), skipping\n") + + except Exception as e: + logger.warning(f"Could not check dataset splits: {e}") + + # No separate HF eval split found ā€” caller will handle programmatic splitting + return None + + def _resolve_eval_split_from_dataset(self, dataset) -> Optional[tuple]: + """Split a dataset into train and eval portions. + + Returns: + Tuple of (train_dataset, eval_dataset), or None if dataset too small. + """ + MIN_EVAL_ROWS = 16 + MIN_TOTAL_ROWS = 32 # Need at least 16 train + 16 eval + + n = len(dataset) + if n < MIN_TOTAL_ROWS: + print(f"Dataset too small ({n} rows) for eval split, skipping eval\n") + return None + + eval_size = max(MIN_EVAL_ROWS, min(128, int(0.05 * n))) + # Ensure we don't take more than half the dataset + eval_size = min(eval_size, n // 2) + + print(f"Auto-splitting: {eval_size} rows for eval from {n} total\n") + split_result = dataset.train_test_split(test_size=eval_size, seed=3407) + print(f"Split complete: {len(split_result['train'])} train, {len(split_result['test'])} eval\n") + return (split_result['train'], split_result['test']) + def start_training(self, dataset: Dataset, + eval_dataset: Dataset = None, + eval_steps: float = 0.01, output_dir: str = "./outputs", num_epochs: int = 3, learning_rate: float = 5e-5, @@ -451,17 +593,33 @@ class UnslothTrainer: 'wandb_token': wandb_token, 'enable_tensorboard': enable_tensorboard, 'tensorboard_dir': tensorboard_dir, + 'eval_dataset': eval_dataset, + 'eval_steps': eval_steps, **kwargs } ) self.should_stop = False - self.training_thread.start() - return True + self.is_training = True + try: + self.training_thread.start() + return True + except Exception as e: + self.is_training = False + logger.error(f"Failed to start training thread: {e}") + return False def _train_worker(self, dataset: Dataset, **training_args): """Worker function for training (runs in separate thread)""" try: + # Store training parameters for metrics calculation + self.batch_size = training_args.get('batch_size', 2) + self.max_seq_length = training_args.get('max_seq_length', 2048) + self.gradient_accumulation_steps = training_args.get('gradient_accumulation_steps', 4) + + # Set training start time + self.training_start_time = time.time() + self._update_progress(is_training=True, error=None) # Setup logging @@ -548,6 +706,8 @@ class UnslothTrainer: "seed": training_args.get('random_seed', 3407), "output_dir": output_dir, "report_to": ["wandb"] if training_args.get('enable_wandb', False) else "none", + "include_num_input_tokens_seen": True, # Enable token counting + "dataset_num_proc": max(1, os.cpu_count() // 4), } # Add warmup parameter - use warmup_ratio if provided, otherwise warmup_steps @@ -577,6 +737,17 @@ class UnslothTrainer: else: print(f"Training for {config_args['num_train_epochs']} epochs\n") + # ========== EVAL CONFIGURATION ========== + eval_dataset = training_args.get('eval_dataset', None) + eval_steps_val = training_args.get('eval_steps', 0.01) + if eval_dataset is not None: + config_args["eval_strategy"] = "steps" + config_args["eval_steps"] = eval_steps_val + print(f"Evaluation enabled: eval_steps={eval_steps_val} (fraction of total steps)\n") + print(f"Eval dataset: {len(eval_dataset)} rows\n") + else: + print("No eval dataset ā€” evaluation disabled\n") + # Add model-specific parameters # Use optim and lr_scheduler_type from training_args if provided, otherwise use defaults optim_value = training_args.get('optim', "adamw_8bit") @@ -618,21 +789,38 @@ class UnslothTrainer: print("Training configuration prepared\n") # ========== TRAINER INITIALIZATION ========== if self.is_vlm: - self.trainer = SFTTrainer( - model=self.model, - train_dataset=dataset['dataset'], - processing_class = self.tokenizer.tokenizer, - data_collator=data_collator, - args=SFTConfig(**config_args), - ) + trainer_kwargs = { + "model": self.model, + "train_dataset": dataset['dataset'], + "processing_class": self.tokenizer.tokenizer, + "data_collator": data_collator, + "args": SFTConfig(**config_args), + } + if eval_dataset is not None: + trainer_kwargs["eval_dataset"] = eval_dataset + self.trainer = SFTTrainer(**trainer_kwargs) else: - self.trainer = SFTTrainer( - model=self.model, - tokenizer=self.tokenizer, - train_dataset=dataset['dataset'], - data_collator=data_collator, - args=SFTConfig(**config_args), - ) + # For text-only training, if the tokenizer is actually a Processor + # (e.g., Gemma-3 returns a ProcessorMixin even for text), we must + # unwrap to the raw tokenizer. Otherwise Unsloth's SFTTrainer detects + # ProcessorMixin ā†’ sets _is_vlm=True ā†’ skips _prepare_dataset entirely, + # and the 'text' column never gets tokenized to 'input_ids'. + from transformers import ProcessorMixin + sft_tokenizer = self.tokenizer + if isinstance(self.tokenizer, ProcessorMixin) and hasattr(self.tokenizer, 'tokenizer'): + print(f" āš ļø Unwrapping Processor ā†’ raw tokenizer for text-only SFTTrainer") + sft_tokenizer = self.tokenizer.tokenizer + + trainer_kwargs = { + "model": self.model, + "tokenizer": sft_tokenizer, + "train_dataset": dataset['dataset'], + "data_collator": data_collator, + "args": SFTConfig(**config_args), + } + if eval_dataset is not None: + trainer_kwargs["eval_dataset"] = eval_dataset + self.trainer = SFTTrainer(**trainer_kwargs) print("Trainer initialized\n") # ========== TRAIN ON RESPONSES ONLY ========== @@ -679,6 +867,7 @@ class UnslothTrainer: self.trainer, instruction_part=instruction_part, response_part=response_part, + num_proc=config_args.get("dataset_num_proc", max(1, os.cpu_count() // 4)), ) print("Train on responses only configured successfully\n") except Exception as e: @@ -706,12 +895,41 @@ class UnslothTrainer: if logs: # Get loss from either 'loss' or 'train_loss' key loss_value = logs.get('loss', logs.get('train_loss', 0.0)) + current_step = state.global_step + + # Extract grad_norm from logs (available when gradient clipping is enabled) + grad_norm = logs.get('grad_norm', None) + + # Calculate elapsed_seconds + elapsed_seconds = None + if self.trainer_instance.training_start_time is not None: + elapsed_seconds = time.time() - self.trainer_instance.training_start_time + + # Calculate eta_seconds + eta_seconds = None + if elapsed_seconds is not None and current_step > 0: + total_steps = self.trainer_instance.training_progress.total_steps + if total_steps > 0: + steps_remaining = total_steps - current_step + if steps_remaining > 0: + time_per_step = elapsed_seconds / current_step + eta_seconds = time_per_step * steps_remaining + + # Extract num_tokens from TRL SFTTrainer state (real counter) + # Requires include_num_input_tokens_seen=True in SFTConfig + num_tokens = getattr(state, "num_input_tokens_seen", None) + self.trainer_instance._update_progress( - step=state.global_step, - epoch=round(state.epoch, 2) if state.epoch else 0, # Round epoch to 2 decimals + step=current_step, + epoch=round(state.epoch, 2) if state.epoch else 0, loss=loss_value, learning_rate=logs.get('learning_rate', 0.0), - status_message="" # Clear status message so metrics show + elapsed_seconds=elapsed_seconds, + eta_seconds=eta_seconds, + grad_norm=grad_norm, + num_tokens=num_tokens, + eval_loss=logs.get('eval_loss', None), + status_message="" ) def on_epoch_end(self, args, state, control, **kwargs): @@ -805,9 +1023,12 @@ class UnslothTrainer: print(f"\nStopping training (save={save})...") self.should_stop = True self.save_on_stop = save - self.is_training = False - # Clear the status message so timer doesn't show stale status - self._update_progress(is_training=False, status_message="") + stop_msg = ( + "Stopping training and saving checkpoint..." + if save + else "Cancelling training..." + ) + self._update_progress(status_message=stop_msg) # If trainer exists, try to stop it gracefully if self.trainer: diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py index 62aa021136..6fe08c2b9e 100644 --- a/studio/backend/core/training/training.py +++ b/studio/backend/core/training/training.py @@ -4,6 +4,7 @@ Training backend for FastAPI integration import matplotlib.pyplot as plt from typing import Any, Generator, Tuple import logging +import math from .trainer import get_trainer, TrainingProgress from utils.hardware import clear_gpu_cache @@ -28,6 +29,11 @@ class TrainingBackend: self.loss_history = [] self.lr_history = [] self.step_history = [] + self.grad_norm_history = [] + self.grad_norm_step_history = [] + self.eval_loss_history = [] + self.eval_step_history = [] + self.eval_enabled = False self.current_theme = "light" self.trainer.add_progress_callback(self._on_progress_update) @@ -40,6 +46,17 @@ class TrainingBackend: self.loss_history.append(progress.loss) self.lr_history.append(progress.learning_rate) self.step_history.append(progress.step) + if progress.step >= 0 and progress.grad_norm is not None: + try: + grad_norm = float(progress.grad_norm) + except (TypeError, ValueError): + grad_norm = None + if grad_norm is not None and math.isfinite(grad_norm): + self.grad_norm_history.append(grad_norm) + self.grad_norm_step_history.append(progress.step) + if progress.eval_loss is not None: + self.eval_loss_history.append(progress.eval_loss) + self.eval_step_history.append(progress.step) def start_training(self, # Model parameters @@ -93,8 +110,13 @@ class TrainingBackend: enable_tensorboard: bool, tensorboard_dir: str, - # Optional: user-provided column mapping - custom_format_mapping: dict = None) -> bool: + # Optional parameters + custom_format_mapping: dict = None, + subset: str = None, + train_split: str = "train", + eval_split: str = None, + eval_steps: float = 0.01, + is_dataset_multimodal: bool = False) -> bool: """ Start training. @@ -133,6 +155,11 @@ class TrainingBackend: self.loss_history = [] self.lr_history = [] self.step_history = [] + self.grad_norm_history = [] + self.grad_norm_step_history = [] + self.eval_loss_history = [] + self.eval_step_history = [] + self.eval_enabled = False import time output_dir = f"./outputs/{model_name.replace('/', '_')}_{int(time.time())}" @@ -148,7 +175,8 @@ class TrainingBackend: model_name=model_name, max_seq_length=max_seq_length, load_in_4bit=load_in_4bit if use_lora_actual else False, # Only 4bit for LoRA - hf_token=hf_token if hf_token.strip() else None + hf_token=hf_token if hf_token.strip() else None, + is_dataset_multimodal=is_dataset_multimodal, ) if not success or self.trainer.should_stop: @@ -187,13 +215,30 @@ class TrainingBackend: # ========== LOAD DATASET ========== logger.info("Loading dataset...") #breakpoint() - dataset = self.trainer.load_and_format_dataset( + dataset_result = self.trainer.load_and_format_dataset( dataset_source=hf_dataset if hf_dataset.strip() else None, format_type=format_type, local_datasets=local_datasets if local_datasets else None, custom_format_mapping=custom_format_mapping, + subset=subset, + train_split=train_split, + eval_split=eval_split, ) + # Unpack: load_and_format_dataset returns (dataset, eval_dataset) + if isinstance(dataset_result, tuple): + dataset, eval_dataset = dataset_result + else: + dataset = dataset_result + eval_dataset = None + + # If user set eval_steps to 0, disable evaluation entirely + if eval_steps is not None and float(eval_steps) <= 0: + eval_dataset = None + + # Track whether eval is enabled for status reporting + self.eval_enabled = eval_dataset is not None + if dataset is None or self.trainer.should_stop: logger.error("Failed to load dataset or stopped by user") return False @@ -213,7 +258,8 @@ class TrainingBackend: logger.info("Starting training worker thread...") success = self.trainer.start_training( dataset=dataset, - #output_dir=f"./outputs/{model_name.replace('/', '_')}_{int(__import__('time').time())}", + eval_dataset=eval_dataset, + eval_steps=eval_steps, output_dir=output_dir, num_epochs=num_epochs, learning_rate=lr_value, @@ -232,7 +278,7 @@ class TrainingBackend: wandb_token=wandb_token if wandb_token.strip() else None, enable_tensorboard=enable_tensorboard, tensorboard_dir=tensorboard_dir, - max_seq_length=max_seq_length, # Pass through for config + max_seq_length=max_seq_length, optim=optim, lr_scheduler_type=lr_scheduler_type, ) @@ -323,8 +369,13 @@ class TrainingBackend: True if training is in progress, False otherwise """ try: - # If user requested stop, training is no longer considered active - if self.trainer.should_stop: + training_thread = getattr(self.trainer, "training_thread", None) + if training_thread and training_thread.is_alive(): + return True + + # Stop requested and worker already exited => inactive. + # This allows UI to show stopped state + "Back to configuration". + if getattr(self.trainer, "should_stop", False): return False progress = self.trainer.get_training_progress() @@ -335,7 +386,23 @@ class TrainingBackend: # but haven't completed or errored yet if not is_active and not progress.is_completed and not progress.error: status = progress.status_message or "" - if any(keyword in status.lower() for keyword in ["loading", "preparing", "training"]): + status_lower = status.lower() + if any( + keyword in status_lower + for keyword in ["cancelled", "canceled", "stopped", "completed", "ready to train"] + ): + return False + if any( + keyword in status_lower + for keyword in [ + "loading", + "preparing", + "training", + "configuring", + "tokenizing", + "starting", + ] + ): is_active = True return is_active except Exception as e: diff --git a/studio/backend/main.py b/studio/backend/main.py index 6ef82234ff..1c500aea2f 100644 --- a/studio/backend/main.py +++ b/studio/backend/main.py @@ -1,6 +1,7 @@ """ Main FastAPI application for Unsloth UI Backend """ +import os import secrets import shutil from contextlib import asynccontextmanager @@ -23,7 +24,7 @@ from routes import ( training_router, ) from auth import storage -from utils.hardware import detect_hardware +from utils.hardware import detect_hardware, get_device, DeviceType import utils.hardware.hardware as _hw_module UNSLOTH_CACHE_DIR = Path(__file__).parent / "unsloth_compiled_cache" @@ -35,6 +36,18 @@ async def lifespan(app: FastAPI): # Detect hardware first ā€” sets DEVICE global used everywhere detect_hardware() + # Disable flex attention on Blackwell+ GPUs (sm_120 and above) + if get_device() == DeviceType.CUDA: + import torch + props = torch.cuda.get_device_properties(0) + sm_version = props.major * 10 + props.minor + if sm_version >= 120: + os.environ["UNSLOTH_ENABLE_FLEX_ATTENTION"] = "0" + import logging + logging.getLogger(__name__).info( + f"GPU sm_{sm_version} detected ā€” setting UNSLOTH_FLEX_ATTENTION=0" + ) + if not storage.is_initialized(): setup_token = secrets.token_urlsafe(32) storage.save_setup_token(setup_token) @@ -126,6 +139,17 @@ async def get_system_info(): } +@app.get("/api/system/hardware") +async def get_hardware_info(): + """Return GPU name, total VRAM, and key ML package versions.""" + from utils.hardware import get_gpu_summary, get_package_versions + + return { + "gpu": get_gpu_summary(), + "versions": get_package_versions(), + } + + # ============ Serve Frontend (Optional) ============ def setup_frontend(app: FastAPI, build_path: Path): diff --git a/studio/backend/models/__init__.py b/studio/backend/models/__init__.py index ce37817cc1..bafcb45a9f 100644 --- a/studio/backend/models/__init__.py +++ b/studio/backend/models/__init__.py @@ -8,7 +8,12 @@ from .training import ( TrainingProgress, ) from .models import ( + CheckpointInfo, + ModelCheckpoints, + CheckpointListResponse, ModelDetails, + LocalModelInfo, + LocalModelListResponse, LoRAInfo, LoRAScanResponse, ModelListResponse, @@ -20,8 +25,6 @@ from .auth import ( AuthStatusResponse, ) from .export import ( - CheckpointInfo, - CheckpointListResponse, LoadCheckpointRequest, ExportStatusResponse, ExportOperationResponse, @@ -65,6 +68,8 @@ __all__ = [ "TrainingProgress", # Model management schemas "ModelDetails", + "LocalModelInfo", + "LocalModelListResponse", "LoRAInfo", "LoRAScanResponse", "ModelListResponse", @@ -75,6 +80,7 @@ __all__ = [ "AuthStatusResponse", # Export schemas "CheckpointInfo", + "ModelCheckpoints", "CheckpointListResponse", "LoadCheckpointRequest", "ExportStatusResponse", diff --git a/studio/backend/models/datasets.py b/studio/backend/models/datasets.py index 530a9fa30b..81adef7577 100644 --- a/studio/backend/models/datasets.py +++ b/studio/backend/models/datasets.py @@ -1,8 +1,8 @@ """ Dataset-related Pydantic models for API requests and responses. """ -from pydantic import BaseModel -from typing import Optional, Dict, List +from pydantic import BaseModel, model_validator +from typing import Any, Optional, Dict, List class CheckFormatRequest(BaseModel): @@ -10,7 +10,16 @@ class CheckFormatRequest(BaseModel): dataset_name: str # HuggingFace dataset name or local path is_vlm: bool = False hf_token: Optional[str] = None - split: Optional[str] = "train" + subset: Optional[str] = None + train_split: Optional[str] = "train" + + @model_validator(mode="before") + @classmethod + def _compat_split(cls, values: Any) -> Any: + """Accept legacy 'split' field as alias for 'train_split'.""" + if isinstance(values, dict) and "split" in values: + values.setdefault("train_split", values.pop("split")) + return values class CheckFormatResponse(BaseModel): diff --git a/studio/backend/models/export.py b/studio/backend/models/export.py index d108c00751..a0526e80db 100644 --- a/studio/backend/models/export.py +++ b/studio/backend/models/export.py @@ -5,23 +5,6 @@ from pydantic import BaseModel, Field from typing import List, Optional, Literal, Dict, Any -class CheckpointInfo(BaseModel): - """Information about a discovered checkpoint directory.""" - - display_name: str = Field(..., description="User-friendly checkpoint name (folder name)") - path: str = Field(..., description="Full path to the checkpoint directory") - - -class CheckpointListResponse(BaseModel): - """Response for listing available checkpoints in an outputs directory.""" - - outputs_dir: str = Field(..., description="Directory that was scanned") - checkpoints: List[CheckpointInfo] = Field( - default_factory=list, - description="List of discovered checkpoints", - ) - - class LoadCheckpointRequest(BaseModel): """Request for loading a checkpoint into the export backend.""" @@ -89,6 +72,10 @@ class ExportCommonOptions(BaseModel): False, description="If True, create a private repository on the Hub (where applicable)", ) + base_model_id: Optional[str] = Field( + None, + description="HuggingFace model ID of the base model (for model card metadata)", + ) class ExportMergedModelRequest(ExportCommonOptions): diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py index b3924c5569..d2d98d7944 100644 --- a/studio/backend/models/inference.py +++ b/studio/backend/models/inference.py @@ -30,7 +30,7 @@ class GenerateRequest(BaseModel): system_prompt: str = Field("You are a helpful AI assistant.", description="System prompt") temperature: float = Field(0.7, ge=0.0, le=2.0, description="Sampling temperature") top_p: float = Field(0.9, ge=0.0, le=1.0, description="Top-p sampling") - top_k: int = Field(40, ge=1, le=100, description="Top-k sampling") + top_k: int = Field(40, ge=-1, le=100, description="Top-k sampling") max_new_tokens: int = Field(512, ge=1, le=4096, description="Maximum tokens to generate") repetition_penalty: float = Field(1.1, ge=1.0, le=2.0, description="Repetition penalty") image_base64: Optional[str] = Field(None, description="Base64 encoded image for vision models") @@ -128,7 +128,8 @@ class ChatCompletionRequest(BaseModel): max_tokens: Optional[int] = Field(512, ge=1, le=4096, description="Maximum tokens to generate") # ā”€ā”€ Unsloth extensions (ignored by standard OpenAI clients) ā”€ā”€ - top_k: int = Field(40, ge=1, le=100, description="[x-unsloth] Top-k sampling") + top_k: int = Field(40, ge=-1, le=100, description="[x-unsloth] Top-k sampling") + min_p: float = Field(0.0, ge=0.0, le=1.0, description="[x-unsloth] Min-p sampling threshold") repetition_penalty: float = Field(1.1, ge=1.0, le=2.0, description="[x-unsloth] Repetition penalty") image_base64: Optional[str] = Field(None, description="[x-unsloth] Base64-encoded image for vision models") use_adapter: Optional[Union[bool, str]] = Field( diff --git a/studio/backend/models/models.py b/studio/backend/models/models.py index 8460d94200..5542db76d5 100644 --- a/studio/backend/models/models.py +++ b/studio/backend/models/models.py @@ -2,7 +2,47 @@ Pydantic schemas for Model Management API """ from pydantic import BaseModel, Field -from typing import Optional, List, Dict, Any +from typing import Optional, List, Dict, Any, Literal + + +class CheckpointInfo(BaseModel): + """Information about a discovered checkpoint directory.""" + + display_name: str = Field(..., description="User-friendly checkpoint name (folder name)") + path: str = Field(..., description="Full path to the checkpoint directory") + loss: Optional[float] = Field(None, description="Training loss at this checkpoint") + + +class ModelCheckpoints(BaseModel): + """A training run and its associated checkpoints.""" + + name: str = Field(..., description="Training run folder name") + checkpoints: List[CheckpointInfo] = Field( + default_factory=list, + description="List of checkpoints for this training run (final + intermediate)", + ) + base_model: Optional[str] = Field( + None, + description="Base model name from adapter_config.json or config.json", + ) + peft_type: Optional[str] = Field( + None, + description="PEFT type (e.g. LORA) if adapter training, None for full fine-tune", + ) + lora_rank: Optional[int] = Field( + None, + description="LoRA rank (r) if applicable", + ) + + +class CheckpointListResponse(BaseModel): + """Response for listing available checkpoints in an outputs directory.""" + + outputs_dir: str = Field(..., description="Directory that was scanned") + models: List[ModelCheckpoints] = Field( + default_factory=list, + description="List of training runs with their checkpoints", + ) class ModelDetails(BaseModel): @@ -34,3 +74,34 @@ class ModelListResponse(BaseModel): models: List[ModelDetails] = Field(default_factory=list, description="List of models") default_models: List[str] = Field(default_factory=list, description="List of default model IDs") + +class LocalModelInfo(BaseModel): + """Discovered local model candidate.""" + id: str = Field(..., description="Identifier to use for loading/training") + display_name: str = Field(..., description="Display label") + path: str = Field(..., description="Local path where model data was discovered") + source: Literal["models_dir", "hf_cache"] = Field( + ..., + description="Discovery source", + ) + model_id: Optional[str] = Field( + None, + description="HF repo id for cached models, e.g. org/model", + ) + updated_at: Optional[float] = Field( + None, + description="Unix timestamp of latest observed update", + ) + + +class LocalModelListResponse(BaseModel): + """Response schema for listing local/cached models.""" + models_dir: str = Field(..., description="Directory scanned for custom local models") + hf_cache_dir: Optional[str] = Field( + None, + description="HF cache root that was scanned", + ) + models: List[LocalModelInfo] = Field( + default_factory=list, + description="Discovered local/cached models", + ) diff --git a/studio/backend/models/responses.py b/studio/backend/models/responses.py index 2aa798c5c9..c72dfc54dd 100644 --- a/studio/backend/models/responses.py +++ b/studio/backend/models/responses.py @@ -19,6 +19,8 @@ class TrainingMetricsResponse(BaseModel): loss_history: List[float] = Field(default_factory=list, description="Loss values per step") lr_history: List[float] = Field(default_factory=list, description="Learning rate per step") step_history: List[int] = Field(default_factory=list, description="Step numbers") + grad_norm_history: List[float] = Field(default_factory=list, description="Gradient norm values") + grad_norm_step_history: List[int] = Field(default_factory=list, description="Step numbers for gradient norm values") current_loss: Optional[float] = Field(None, description="Most recent loss value") current_lr: Optional[float] = Field(None, description="Most recent learning rate") current_step: Optional[int] = Field(None, description="Most recent step number") diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py index 2f55aeca7a..2b989e6a82 100644 --- a/studio/backend/models/training.py +++ b/studio/backend/models/training.py @@ -1,8 +1,8 @@ """ Pydantic schemas for Training API """ -from pydantic import BaseModel, Field -from typing import Optional, List, Dict, Literal +from pydantic import BaseModel, Field, model_validator +from typing import Any, Optional, List, Dict, Literal class TrainingStartRequest(BaseModel): @@ -18,6 +18,18 @@ class TrainingStartRequest(BaseModel): hf_dataset: Optional[str] = Field(None, description="HuggingFace dataset identifier") local_datasets: List[str] = Field(default_factory=list, description="List of local dataset paths") format_type: str = Field(..., description="Dataset format type") + subset: Optional[str] = None + train_split: Optional[str] = Field("train", description="Training split name") + eval_split: Optional[str] = Field(None, description="Eval split name. None = auto-detect") + eval_steps: float = Field(0.01, description="Fraction of total steps between evals (0-1)") + + @model_validator(mode="before") + @classmethod + def _compat_split(cls, values: Any) -> Any: + """Accept legacy 'split' field as alias for 'train_split'.""" + if isinstance(values, dict) and "split" in values: + values.setdefault("train_split", values.pop("split")) + return values custom_format_mapping: Optional[Dict[str, str]] = Field( None, description="User-provided column-to-role mapping, e.g. {'image': 'image', 'caption': 'text'} for VLM or {'instruction': 'user', 'output': 'assistant'} for LLM" @@ -53,6 +65,7 @@ class TrainingStartRequest(BaseModel): finetune_language_layers: bool = Field(False, description="Finetune language layers") finetune_attention_modules: bool = Field(False, description="Finetune attention modules") finetune_mlp_modules: bool = Field(False, description="Finetune MLP modules") + is_dataset_multimodal: bool = Field(False, description="Whether the dataset contains multimodal (image) data") # Logging parameters enable_wandb: bool = Field(False, description="Enable Weights & Biases logging") @@ -84,13 +97,14 @@ class TrainingStatus(BaseModel): "stopped" ] = Field(..., description="Current phase of training pipeline") is_training_running: bool = Field(..., description="True if training loop is actively running") + eval_enabled: bool = Field(False, description="True if evaluation dataset is configured for this training run") message: str = Field(..., description="Human-readable status message") error: Optional[str] = Field(None, description="Error details if phase is 'error'") details: Optional[dict] = Field(None, description="Phase-specific info, e.g. {'model_size': '8B'}") metric_history: Optional[dict] = Field( None, description="Full metric history arrays for chart recovery after SSE reconnection. " - "Keys: 'steps', 'loss', 'lr' ā€” each a list of numeric values.", + "Keys: 'steps', 'loss', 'lr', 'grad_norm', 'grad_norm_steps' ā€” each a list of numeric values.", ) @@ -107,4 +121,4 @@ class TrainingProgress(BaseModel): eta_seconds: Optional[float] = Field(None, description="Estimated time remaining") grad_norm: Optional[float] = Field(None, description="L2 norm of gradients, computed before gradient clipping") num_tokens: Optional[int] = Field(None, description="Total number of tokens processed so far") - + eval_loss: Optional[float] = Field(None, description="Eval loss from the most recent evaluation step") diff --git a/studio/backend/requirements.txt b/studio/backend/requirements.txt deleted file mode 100644 index 3f97cca66c..0000000000 --- a/studio/backend/requirements.txt +++ /dev/null @@ -1,7 +0,0 @@ -fastapi>=0.100.0 -uvicorn>=0.27.0 -pydantic>=2.0 -torch -psutil -nest-asyncio>=1.5.8 - diff --git a/studio/backend/requirements/base.txt b/studio/backend/requirements/base.txt new file mode 100644 index 0000000000..407ae01b52 --- /dev/null +++ b/studio/backend/requirements/base.txt @@ -0,0 +1,3 @@ +# Core unsloth packages +unsloth-zoo +unsloth diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt new file mode 100644 index 0000000000..c2a9c6bad8 --- /dev/null +++ b/studio/backend/requirements/extras-no-deps.txt @@ -0,0 +1,13 @@ +# Audio extras (installed with --no-deps --no-cache-dir) +descript-audio-codec +descript-audiotools +julius +torchcodec +snac + +# TRL and related packages +trl==0.23.1 +git+https://github.com/meta-pytorch/OpenEnv.git +executorch==1.0.1 +torch-c-dlpack-ext +sentence_transformers==5.2.0 diff --git a/studio/backend/requirements/extras.txt b/studio/backend/requirements/extras.txt new file mode 100644 index 0000000000..3ed20faa8b --- /dev/null +++ b/studio/backend/requirements/extras.txt @@ -0,0 +1,56 @@ +# OpenEnv dependencies +tomli +tomli-w + +# ExecuTorch dependencies +ruamel.yaml +coremltools +expecttest +flatbuffers +hydra-core +hypothesis +kgb +parameterized +pytest<9.0 +pytest-json-report +pytest-rerunfailures==15.1 +pytest-xdist +# Also needed by sentence_transformers +scikit-learn==1.7.1 + +# Additional extras +pybind11 +langid +jiwer +omegaconf +einx +pyloudnorm +openai-whisper +uroman +MeCab +loguru +flatten_dict +ffmpy +randomname +argbind +tiktoken +ftfy +importlib-resources +librosa +markdown2 +matplotlib +pystoi +soundfile +tensorboard +torch-stoi +evaluate +timm +transformers-cfg +open_spiel +addict +easydict +einops +tabulate +fastmcp>=2.0.0 +openai>=2.7.2 +websockets>=13.0,<14 diff --git a/studio/backend/requirements/overrides.txt b/studio/backend/requirements/overrides.txt new file mode 100644 index 0000000000..02770f3953 --- /dev/null +++ b/studio/backend/requirements/overrides.txt @@ -0,0 +1,7 @@ +# Torch AO overrides (installed with --force-reinstall --no-cache-dir) +torchao==0.14.0 +transformers==4.57.1 +pytorch_tokenizers + +# Kernel packages +kernels diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt new file mode 100644 index 0000000000..fd7da2626d --- /dev/null +++ b/studio/backend/requirements/studio.txt @@ -0,0 +1,13 @@ +# Studio UI backend dependencies +typer +fastapi +uvicorn +pydantic +matplotlib +pandas +nest_asyncio +datasets==4.3.0 +pyjwt +easydict +addict +gradio>=4.0.0 diff --git a/studio/backend/requirements/triton-kernels.txt b/studio/backend/requirements/triton-kernels.txt new file mode 100644 index 0000000000..17e265b35e --- /dev/null +++ b/studio/backend/requirements/triton-kernels.txt @@ -0,0 +1,2 @@ +# Triton kernels (installed with --no-deps, from source) +triton_kernels @ git+https://github.com/triton-lang/triton.git@release/3.6.x#subdirectory=python/triton_kernels diff --git a/studio/backend/routes/datasets.py b/studio/backend/routes/datasets.py index e462bbfe18..39119f1123 100644 --- a/studio/backend/routes/datasets.py +++ b/studio/backend/routes/datasets.py @@ -70,76 +70,139 @@ def _serialize_preview_rows(rows): # --- Endpoints --- +# Recognized data-file extensions for the single-file fallback approach. +DATA_EXTS = ( + '.parquet', + '.json', '.jsonl', + '.csv', '.tsv', + '.txt', + '.arrow', + '.tar', '.tar.gz', '.tgz', + '.gz', '.zst', + '.zip', +) + + @router.post("/check-format", response_model=CheckFormatResponse) -async def check_format(request: CheckFormatRequest): +def check_format(request: CheckFormatRequest): """ Check if a dataset requires manual column mapping. - - This is a lightweight check that loads only the first 10 rows, - runs format detection, and (if processable) returns processed - preview samples. The full dataset is re-processed at training time. + + Strategy for HuggingFace datasets: + 1. list_repo_files ā†’ pick the first data file ā†’ load_dataset(data_files=[ā€¦]) + Avoids resolving thousands of files; typically ~2-4 s. + 2. Full streaming load_dataset as a last-resort fallback. + + Local files are loaded directly. + + Using a plain `def` (not async) so FastAPI runs this in a thread-pool, + preventing any blocking IO from freezing the event loop. """ try: - from datasets import load_dataset + from itertools import islice + from datasets import Dataset, load_dataset from utils.datasets import format_dataset - + PREVIEW_SIZE = 10 - + logger.info(f"Checking format for dataset: {request.dataset_name}") - - # Load dataset + dataset_path = Path(request.dataset_name) - + total_rows = None + if dataset_path.exists(): - # Local dataset + # ā”€ā”€ Local file ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ if dataset_path.suffix in ['.json', '.jsonl']: - dataset = load_dataset('json', data_files=str(dataset_path), split=request.split) + dataset = load_dataset('json', data_files=str(dataset_path), split=request.train_split) elif dataset_path.suffix == '.csv': - dataset = load_dataset('csv', data_files=str(dataset_path), split=request.split) + dataset = load_dataset('csv', data_files=str(dataset_path), split=request.train_split) elif dataset_path.suffix == '.parquet': - dataset = load_dataset('parquet', data_files=str(dataset_path), split=request.split) + dataset = load_dataset('parquet', data_files=str(dataset_path), split=request.train_split) else: raise HTTPException( status_code=400, detail=f"Unsupported file format: {dataset_path.suffix}" ) + total_rows = len(dataset) + preview_slice = dataset.select(range(min(PREVIEW_SIZE, total_rows))) else: - # HuggingFace dataset - load_kwargs = {"path": request.dataset_name, "split": request.split} - if request.hf_token: - load_kwargs["token"] = request.hf_token - dataset = load_dataset(**load_kwargs) - - # Slice to top N rows ā€” all detection and preview runs on this subset - total_rows = len(dataset) - preview_slice = dataset.select(range(min(PREVIEW_SIZE, total_rows))) - + # ā”€ā”€ HuggingFace dataset ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ + # Tier 1: list_repo_files ā†’ load only the first data file + preview_slice = None + + try: + from huggingface_hub import HfApi + api = HfApi() + repo_files = api.list_repo_files( + request.dataset_name, + repo_type="dataset", + token=request.hf_token or None, + ) + data_files = [f for f in repo_files if any(f.endswith(ext) for ext in DATA_EXTS)] + + if data_files: + first_file = data_files[0] + logger.info(f"Tier 1: loading single file {first_file}") + load_kwargs = { + "path": request.dataset_name, + "data_files": [first_file], + "split": "train", + "streaming": True, + } + if request.hf_token: + load_kwargs["token"] = request.hf_token + + streamed_ds = load_dataset(**load_kwargs) + rows = list(islice(streamed_ds, PREVIEW_SIZE)) + if rows: + preview_slice = Dataset.from_list(rows) + except Exception as e: + logger.warning(f"Tier 1 (single-file) failed: {e}") + + if preview_slice is None: + # Tier 2: full streaming (resolves all files ā€” slow for large repos) + logger.info("Tier 2: falling back to full streaming load_dataset") + load_kwargs = {"path": request.dataset_name, "split": request.train_split, "streaming": True} + if request.subset: + load_kwargs["name"] = request.subset + if request.hf_token: + load_kwargs["token"] = request.hf_token + + streamed_ds = load_dataset(**load_kwargs) + + rows = list(islice(streamed_ds, PREVIEW_SIZE)) + if not rows: + raise HTTPException( + status_code=400, + detail="Dataset appears to be empty or could not be streamed" + ) + + preview_slice = Dataset.from_list(rows) + total_rows = None + # Run lightweight format check on the preview slice result = check_dataset_format(preview_slice, is_vlm=request.is_vlm) - - logger.info(f"Format check result: requires_mapping={result['requires_manual_mapping']}, format={result['detected_format']}") - + + logger.info(f"Format check result: requires_mapping={result['requires_manual_mapping']}, format={result['detected_format']}, is_multimodal={result.get('is_multimodal', False)}") + # Generate preview samples preview_samples = None if not result["requires_manual_mapping"]: - # Format detected ā€” return processed preview try: format_result = format_dataset( preview_slice, format_type="auto", custom_format_mapping=result.get("suggested_mapping"), + num_proc=1, # Only 10 preview rows ā€” no need for multiprocessing ) processed = format_result["dataset"] preview_samples = _serialize_preview_rows(processed) except Exception as e: logger.warning(f"Processed preview generation failed (non-fatal): {e}") - # Fall back to raw samples so frontend still has something preview_samples = _serialize_preview_rows(preview_slice) else: - # Format detection failed ā€” return raw samples so user can - # see actual data and map columns in the frontend preview_samples = _serialize_preview_rows(preview_slice) - + return CheckFormatResponse( requires_manual_mapping=result["requires_manual_mapping"], detected_format=result["detected_format"], @@ -152,7 +215,7 @@ async def check_format(request: CheckFormatRequest): preview_samples=preview_samples, total_rows=total_rows, ) - + except HTTPException: raise except Exception as e: diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py index b72dea48b8..6616c9fbd8 100644 --- a/studio/backend/routes/export.py +++ b/studio/backend/routes/export.py @@ -26,8 +26,6 @@ except ImportError: # Import Pydantic models from models import ( - CheckpointInfo, - CheckpointListResponse, LoadCheckpointRequest, ExportStatusResponse, ExportOperationResponse, @@ -50,38 +48,6 @@ if not logger.handlers: logger.setLevel(logging.INFO) -@router.get("/checkpoints", response_model=CheckpointListResponse) -async def list_checkpoints( - outputs_dir: str = Query( - default="./outputs", - description="Directory to scan for checkpoints", - ), - current_subject: str = Depends(get_current_subject), -): - """ - List available checkpoints in the outputs directory. - - Wraps ExportBackend.scan_checkpoints. - """ - try: - backend = get_export_backend() - raw_checkpoints = backend.scan_checkpoints(outputs_dir=outputs_dir) - - checkpoints = [ - CheckpointInfo(display_name=display_name, path=path) - for display_name, path in raw_checkpoints - ] - - return CheckpointListResponse( - outputs_dir=outputs_dir, - checkpoints=checkpoints, - ) - except Exception as e: - logger.error(f"Error listing checkpoints: {e}", exc_info=True) - raise HTTPException( - status_code=500, - detail=f"Failed to list checkpoints: {str(e)}", - ) @router.post("/load-checkpoint", response_model=ExportOperationResponse) @@ -224,6 +190,7 @@ async def export_base_model( repo_id=request.repo_id, hf_token=request.hf_token, private=request.private, + base_model_id=request.base_model_id, ) if not success: diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index ae8fc31a46..c30a1638d4 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -5,11 +5,13 @@ import sys import time import uuid from pathlib import Path -from fastapi import APIRouter, HTTPException +from fastapi import APIRouter, HTTPException, Request from fastapi.responses import StreamingResponse, JSONResponse from typing import Optional import json import logging +import asyncio +import threading @@ -304,7 +306,7 @@ def _extract_content_parts( @router.post("/chat/completions") -async def openai_chat_completions(request: ChatCompletionRequest): +async def openai_chat_completions(payload: ChatCompletionRequest, request: Request): """ OpenAI-compatible chat completions endpoint. @@ -324,7 +326,7 @@ async def openai_chat_completions(request: ChatCompletionRequest): # ā”€ā”€ Parse messages (handles multimodal content parts) ā”€ā”€ā”€ā”€ā”€ system_prompt, chat_messages, extracted_image_b64 = _extract_content_parts( - request.messages + payload.messages ) # If no non-system messages were provided, error out @@ -336,7 +338,7 @@ async def openai_chat_completions(request: ChatCompletionRequest): # ā”€ā”€ Decode image (from content parts OR legacy field) ā”€ā”€ā”€ā”€ā”€ # Content-part images take priority; fall back to legacy field - image_b64 = extracted_image_b64 or request.image_base64 + image_b64 = extracted_image_b64 or payload.image_base64 image = None if image_b64: @@ -366,31 +368,36 @@ async def openai_chat_completions(request: ChatCompletionRequest): messages=chat_messages, system_prompt=system_prompt, image=image, - temperature=request.temperature, - top_p=request.top_p, - top_k=request.top_k, - max_new_tokens=request.max_tokens or 512, - repetition_penalty=request.repetition_penalty, + temperature=payload.temperature, + top_p=payload.top_p, + top_k=payload.top_k, + min_p=payload.min_p, + max_new_tokens=payload.max_tokens or 512, + repetition_penalty=payload.repetition_penalty, ) # ā”€ā”€ Choose generation path (adapter-controlled or standard) ā”€ā”€ - if request.use_adapter is not None: + cancel_event = threading.Event() + + if payload.use_adapter is not None: # Compare mode: toggle adapter state atomically with generation def generate(): return backend.generate_with_adapter_control( - use_adapter=request.use_adapter, **gen_kwargs + use_adapter=payload.use_adapter, + cancel_event=cancel_event, + **gen_kwargs, ) else: # Standard path: no adapter toggling def generate(): - return backend.generate_chat_response(**gen_kwargs) + return backend.generate_chat_response(cancel_event=cancel_event, **gen_kwargs) - model_name = backend.active_model_name or request.model + model_name = backend.active_model_name or payload.model completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}" created = int(time.time()) # ā”€ā”€ Streaming response ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ - if request.stream: + if payload.stream: async def stream_chunks(): try: # First chunk: send the role @@ -409,6 +416,10 @@ async def openai_chat_completions(request: ChatCompletionRequest): # text, so we diff to get incremental deltas. prev_text = "" for cumulative in generate(): + if await request.is_disconnected(): + cancel_event.set() + backend.reset_generation_state() + return new_text = cumulative[len(prev_text):] prev_text = cumulative if not new_text: @@ -437,6 +448,10 @@ async def openai_chat_completions(request: ChatCompletionRequest): yield f"data: {final_chunk.model_dump_json(exclude_none=True)}\n\n" yield "data: [DONE]\n\n" + except asyncio.CancelledError: + cancel_event.set() + backend.reset_generation_state() + raise except Exception as e: backend.reset_generation_state() logger.error(f"Error during OpenAI streaming: {e}", exc_info=True) @@ -477,4 +492,3 @@ async def openai_chat_completions(request: ChatCompletionRequest): backend.reset_generation_state() logger.error(f"Error during OpenAI completion: {e}", exc_info=True) raise HTTPException(status_code=500, detail=str(e)) - diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py index 1bfd08f38c..761c04d3e7 100644 --- a/studio/backend/routes/models.py +++ b/studio/backend/routes/models.py @@ -21,6 +21,7 @@ try: load_model_defaults, get_base_model_from_lora, is_vision_model, + scan_checkpoints, ModelConfig, ) from core.inference import get_inference_backend @@ -34,11 +35,17 @@ except ImportError: load_model_defaults, get_base_model_from_lora, is_vision_model, + scan_checkpoints, ModelConfig, ) from core.inference import get_inference_backend from models import ( + CheckpointInfo, + CheckpointListResponse, + LocalModelInfo, + LocalModelListResponse, + ModelCheckpoints, ModelDetails, LoRAScanResponse, LoRAInfo, @@ -59,6 +66,116 @@ if not logger.handlers: logger.setLevel(logging.INFO) +def _resolve_hf_cache_dir() -> Path: + """Resolve local HF cache root used by hub downloads.""" + try: + from huggingface_hub.constants import HF_HUB_CACHE + return Path(HF_HUB_CACHE) + except Exception: + return Path.home() / ".cache" / "huggingface" / "hub" + + +def _scan_models_dir(models_dir: Path) -> List[LocalModelInfo]: + if not models_dir.exists() or not models_dir.is_dir(): + return [] + + found: List[LocalModelInfo] = [] + for child in models_dir.iterdir(): + if not child.is_dir(): + continue + has_model_files = ( + (child / "config.json").exists() + or (child / "adapter_config.json").exists() + or any(child.glob("*.safetensors")) + or any(child.glob("*.bin")) + ) + if not has_model_files: + continue + try: + updated_at = child.stat().st_mtime + except OSError: + updated_at = None + found.append( + LocalModelInfo( + id=str(child), + display_name=child.name, + path=str(child), + source="models_dir", + updated_at=updated_at, + ), + ) + return found + + +def _scan_hf_cache(cache_dir: Path) -> List[LocalModelInfo]: + if not cache_dir.exists() or not cache_dir.is_dir(): + return [] + + found: List[LocalModelInfo] = [] + for repo_dir in cache_dir.glob("models--*"): + if not repo_dir.is_dir(): + continue + + repo_name = repo_dir.name[len("models--"):] + if not repo_name: + continue + model_id = repo_name.replace("--", "/") + + try: + updated_at = repo_dir.stat().st_mtime + except OSError: + updated_at = None + + found.append( + LocalModelInfo( + id=model_id, + model_id=model_id, + display_name=model_id.split("/")[-1], + path=str(repo_dir), + source="hf_cache", + updated_at=updated_at, + ), + ) + return found + + +@router.get("/local", response_model=LocalModelListResponse) +async def list_local_models( + models_dir: str = Query(default="./models", description="Directory to scan for local model folders"), + current_subject: str = Depends(get_current_subject), +): + """ + List local model candidates from custom models dir and HF cache. + """ + try: + models_root = Path(models_dir).expanduser().resolve() + hf_cache_dir = _resolve_hf_cache_dir() + local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir) + + deduped: dict[str, LocalModelInfo] = {} + for model in local_models: + if model.id not in deduped: + deduped[model.id] = model + + models = sorted( + deduped.values(), + key=lambda item: (item.updated_at or 0), + reverse=True, + ) + + return LocalModelListResponse( + models_dir=str(models_root), + hf_cache_dir=str(hf_cache_dir), + models=models, + ) + except Exception as e: + logger.error(f"Error listing local models: {e}", exc_info=True) + raise HTTPException( + status_code=500, + detail=f"Failed to list local models: {str(e)}", + ) + + @router.get("/list") @@ -131,6 +248,7 @@ async def get_model_config( This endpoint wraps the backend load_model_defaults function. """ try: + logger.info(f"Getting model config for: {model_name}") # Load model defaults from backend config_dict = load_model_defaults(model_name) @@ -150,6 +268,7 @@ async def get_model_config( # If ModelConfig creation fails, use defaults pass + logger.info(f"Model config result for {model_name}: is_vision={is_vision}, is_lora={is_lora}, base_model={base_model}") return ModelDetails( id=model_name, model_name=model_name, @@ -252,8 +371,10 @@ async def check_vision_model( This endpoint wraps the backend is_vision_model function. """ try: + logger.info(f"Checking if vision model: {model_name}") is_vision = is_vision_model(model_name) + logger.info(f"Vision check result for {model_name}: is_vision={is_vision}") return VisionCheckResponse( model_name=model_name, is_vision=is_vision, @@ -266,3 +387,43 @@ async def check_vision_model( detail=f"Failed to check vision model: {str(e)}" ) +@router.get("/checkpoints", response_model=CheckpointListResponse) +async def list_checkpoints( + outputs_dir: str = Query( + default="./outputs", + description="Directory to scan for checkpoints", + ), + current_subject: str = Depends(get_current_subject), +): + """ + List available checkpoints in the outputs directory. + + Scans the outputs folder for training runs and their checkpoints. + """ + try: + raw_models = scan_checkpoints(outputs_dir=outputs_dir) + + models = [ + ModelCheckpoints( + name=model_name, + checkpoints=[ + CheckpointInfo(display_name=display_name, path=path, loss=loss) + for display_name, path, loss in checkpoints + ], + base_model=metadata.get("base_model"), + peft_type=metadata.get("peft_type"), + lora_rank=metadata.get("lora_rank"), + ) + for model_name, checkpoints, metadata in raw_models + ] + + return CheckpointListResponse( + outputs_dir=outputs_dir, + models=models, + ) + except Exception as e: + logger.error(f"Error listing checkpoints: {e}", exc_info=True) + raise HTTPException( + status_code=500, + detail=f"Failed to list checkpoints: {str(e)}", + ) diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py index 1a904cd371..f8de2f639f 100644 --- a/studio/backend/routes/training.py +++ b/studio/backend/routes/training.py @@ -5,7 +5,7 @@ import sys from pathlib import Path from fastapi import APIRouter, Depends, HTTPException, Request from fastapi.responses import StreamingResponse -from typing import Dict, Optional +from typing import Dict, Optional, Any import logging import asyncio from datetime import datetime @@ -56,6 +56,22 @@ if not logger.handlers: logger.setLevel(logging.INFO) +@router.get("/hardware") +async def get_hardware_utilization( + current_subject: str = Depends(get_current_subject), +): + """ + Get a live snapshot of GPU hardware utilization. + + Designed to be polled by the frontend during training. + Returns GPU utilization %, temperature, VRAM usage, and power draw + via nvidia-smi for maximum accuracy. + """ + from utils.hardware import get_gpu_utilization + + return get_gpu_utilization() + + @router.post("/start") async def start_training( request: TrainingStartRequest, @@ -129,6 +145,10 @@ async def start_training( "hf_dataset": request.hf_dataset or "", "local_datasets": request.local_datasets, "format_type": request.format_type, + "subset": request.subset, + "train_split": request.train_split, + "eval_split": request.eval_split, + "eval_steps": request.eval_steps, "custom_format_mapping": request.custom_format_mapping, "num_epochs": request.num_epochs, "learning_rate": request.learning_rate, @@ -158,6 +178,7 @@ async def start_training( "finetune_language_layers": request.finetune_language_layers, "finetune_attention_modules": request.finetune_attention_modules, "finetune_mlp_modules": request.finetune_mlp_modules, + "is_dataset_multimodal": request.is_dataset_multimodal, "enable_wandb": request.enable_wandb, "wandb_token": request.wandb_token or "", "wandb_project": request.wandb_project or "", @@ -267,21 +288,31 @@ async def stop_training( """ try: backend = get_training_backend() - - if not backend.is_training_active(): + trainer_thread = getattr(getattr(backend, "trainer", None), "training_thread", None) + thread_alive = bool(trainer_thread and trainer_thread.is_alive()) + is_active = backend.is_training_active() + logger.info( + "Stop requested: save=%s is_active=%s thread_alive=%s should_stop=%s", + body.save, + is_active, + thread_alive, + getattr(getattr(backend, "trainer", None), "should_stop", None), + ) + + if not is_active and not thread_alive: return TrainingStopResponse( status="idle", message="No training job is currently running" ) - + # Call backend stop method backend.stop_training(save=body.save) - + return TrainingStopResponse( status="stopped", - message="Training job stopped successfully" + message="Stop requested. Training will stop at the next safe step." ) - + except Exception as e: logger.error(f"Error stopping training: {e}", exc_info=True) raise HTTPException( @@ -299,12 +330,33 @@ async def reset_training( """ try: backend = get_training_backend() + trainer_thread = getattr(getattr(backend, "trainer", None), "training_thread", None) + thread_alive = bool(trainer_thread and trainer_thread.is_alive()) + is_active = backend.is_training_active() + + if is_active or thread_alive: + logger.warning( + "Rejected reset while training active: is_active=%s thread_alive=%s should_stop=%s", + is_active, + thread_alive, + getattr(getattr(backend, "trainer", None), "should_stop", None), + ) + raise HTTPException( + status_code=409, + detail="Training is still running. Stop training and wait for it to finish before resetting.", + ) + + logger.info("Reset training state: clearing runtime + metric history") backend.trainer.should_stop = False backend.trainer.training_progress = backend.trainer.training_progress.__class__() backend.loss_history = [] backend.lr_history = [] backend.step_history = [] + backend.grad_norm_history = [] + backend.grad_norm_step_history = [] return {"status": "ok"} + except HTTPException: + raise except Exception as e: logger.error(f"Error resetting training: {e}", exc_info=True) raise HTTPException( @@ -387,12 +439,17 @@ async def get_training_status( "steps": list(backend.step_history), "loss": list(backend.loss_history), "lr": list(backend.lr_history), + "grad_norm": list(getattr(backend, "grad_norm_history", [])), + "grad_norm_steps": list(getattr(backend, "grad_norm_step_history", [])), + "eval_loss": list(backend.eval_loss_history), + "eval_steps": list(backend.eval_step_history), } return TrainingStatus( job_id=job_id, phase=phase, is_training_running=is_active, + eval_enabled=backend.eval_enabled, message=status_message, error=error_message, details=details, @@ -421,6 +478,8 @@ async def get_training_metrics( loss_history = backend.loss_history lr_history = backend.lr_history step_history = backend.step_history + grad_norm_history = getattr(backend, "grad_norm_history", []) + grad_norm_step_history = getattr(backend, "grad_norm_step_history", []) # Get current values current_loss = loss_history[-1] if loss_history else None @@ -431,6 +490,8 @@ async def get_training_metrics( loss_history=loss_history, lr_history=lr_history, step_history=step_history, + grad_norm_history=grad_norm_history, + grad_norm_step_history=grad_norm_step_history, current_loss=current_loss, current_lr=current_lr, current_step=current_step, @@ -480,6 +541,9 @@ async def stream_training_progress( learning_rate: float, total_steps: int, epoch: Optional[float] = None, + progress: Optional[Any] = None, + grad_norm_override: Optional[float] = None, + eval_loss_override: Optional[float] = None, ) -> TrainingProgress: total = max(total_steps, 0) if step < 0 or total == 0: @@ -489,6 +553,17 @@ async def stream_training_progress( float(step) / float(total) * 100.0 if total > 0 else 0.0 ) + # Get actual values from progress object if available + elapsed_seconds = getattr(progress, 'elapsed_seconds', None) if progress else None + eta_seconds = getattr(progress, 'eta_seconds', None) if progress else None + grad_norm = grad_norm_override + if grad_norm is None and progress: + grad_norm = getattr(progress, 'grad_norm', None) + num_tokens = getattr(progress, 'num_tokens', None) if progress else None + eval_loss = eval_loss_override + if eval_loss is None and progress: + eval_loss = getattr(progress, 'eval_loss', None) + return TrainingProgress( job_id=job_id, step=step, @@ -497,10 +572,11 @@ async def stream_training_progress( learning_rate=learning_rate, progress_percent=progress_percent, epoch=epoch, - elapsed_seconds=None, - eta_seconds=None, - grad_norm=None, - num_tokens=None, + elapsed_seconds=elapsed_seconds, + eta_seconds=eta_seconds, + grad_norm=grad_norm, + num_tokens=num_tokens, + eval_loss=eval_loss, ) def format_sse( @@ -525,6 +601,13 @@ async def stream_training_progress( # ā”€ā”€ Replay missed steps on reconnect ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ if resume_from_step is not None and backend.step_history: replayed = 0 + grad_norm_by_step = { + step_val: grad_val + for step_val, grad_val in zip( + getattr(backend, "grad_norm_step_history", []), + getattr(backend, "grad_norm_history", []), + ) + } for i, step_val in enumerate(backend.step_history): if step_val > resume_from_step: loss_val = backend.loss_history[i] if i < len(backend.loss_history) else 0.0 @@ -534,7 +617,15 @@ async def stream_training_progress( ) total_replay = getattr(tp_replay, "total_steps", step_val) if tp_replay else step_val epoch_replay = getattr(tp_replay, "epoch", None) if tp_replay else None - payload = build_progress(step_val, loss_val, lr_val, total_replay, epoch_replay) + payload = build_progress( + step_val, + loss_val, + lr_val, + total_replay, + epoch_replay, + progress=tp_replay, + grad_norm_override=grad_norm_by_step.get(step_val), + ) yield format_sse(payload.model_dump_json(), event="progress", event_id=step_val) replayed += 1 if replayed: @@ -553,6 +644,7 @@ async def stream_training_progress( learning_rate=0.0, total_steps=initial_total_steps, epoch=initial_epoch, + progress=tp, ) yield format_sse(initial_progress.model_dump_json(), event="progress", event_id=0) @@ -566,11 +658,11 @@ async def stream_training_progress( getattr(tp, "total_steps", final_step) if tp else final_step ) final_epoch = getattr(tp, "epoch", None) if tp else None - payload = build_progress(final_step, final_loss, final_lr, final_total_steps, final_epoch) + payload = build_progress(final_step, final_loss, final_lr, final_total_steps, final_epoch, progress=tp) yield format_sse(payload.model_dump_json(), event="complete", event_id=final_step) else: yield format_sse( - build_progress(-1, 0.0, 0.0, 0).model_dump_json(), + build_progress(-1, 0.0, 0.0, 0, progress=tp).model_dump_json(), event="complete", event_id=0, ) @@ -605,6 +697,7 @@ async def stream_training_progress( current_lr, current_total_steps, current_epoch, + progress=tp_inner, ) yield format_sse( progress_payload.model_dump_json(), @@ -623,6 +716,7 @@ async def stream_training_progress( current_lr, current_total_steps, current_epoch, + progress=tp_inner, ) yield format_sse( heartbeat_payload.model_dump_json(), @@ -644,7 +738,7 @@ async def stream_training_progress( if tp_prep else 0 ) preparing_payload = build_progress( - 0, 0.0, 0.0, prep_total, + 0, 0.0, 0.0, prep_total, progress=tp_prep, ) yield format_sse( preparing_payload.model_dump_json(), @@ -655,7 +749,8 @@ async def stream_training_progress( # Timeout check if no_update_count > max_no_updates: logger.warning("Progress stream timeout - no updates received") - timeout_payload = build_progress(last_step, 0.0, 0.0, 0) + tp_timeout = getattr(getattr(backend, "trainer", None), "training_progress", None) + timeout_payload = build_progress(last_step, 0.0, 0.0, 0, progress=tp_timeout) yield format_sse( timeout_payload.model_dump_json(), event="error", @@ -667,7 +762,8 @@ async def stream_training_progress( except Exception as e: logger.error(f"Error in progress stream: {e}", exc_info=True) - error_payload = build_progress(0, 0.0, 0.0, 0) + tp_error = getattr(getattr(backend, "trainer", None), "training_progress", None) + error_payload = build_progress(0, 0.0, 0.0, 0, progress=tp_error) yield format_sse( error_payload.model_dump_json(), event="error", @@ -692,6 +788,7 @@ async def stream_training_progress( final_lr, final_total_steps, final_epoch, + progress=final_tp, ) yield format_sse( final_payload.model_dump_json(), diff --git a/studio/backend/utils/datasets/chat_templates.py b/studio/backend/utils/datasets/chat_templates.py index a868109f5d..6420aa899d 100644 --- a/studio/backend/utils/datasets/chat_templates.py +++ b/studio/backend/utils/datasets/chat_templates.py @@ -60,7 +60,24 @@ def get_tokenizer_chat_template(tokenizer, model_name): print(f"āš ļø Failed to apply Unsloth template '{matched_template}': {e}") print(f" Falling back to tokenizer's default chat template") else: - print(f"šŸ“ Using tokenizer's default chat template (no Unsloth template match)") + # Check if tokenizer actually has a chat_template set + has_chat_template = ( + hasattr(tokenizer, 'chat_template') + and tokenizer.chat_template is not None + ) + if has_chat_template: + print(f"šŸ“ Using tokenizer's own chat template (no Unsloth template match)") + else: + # Base model with no chat template ā€” apply default ChatML + print(f"šŸ“ No chat template found ā€” applying default ChatML template (base model)") + try: + tokenizer = get_chat_template( + tokenizer, + chat_template="chatml", + ) + except Exception as e: + print(f"āš ļø Failed to apply default ChatML template: {e}") + print(f" Falling back to tokenizer as-is") return tokenizer @@ -227,6 +244,16 @@ def apply_chat_template_to_dataset( # ALPACA FORMAT if final_format == "alpaca": + # Set alpaca chat template on tokenizer for saving (if not already set) + # This ensures the template is saved with the model for inference + if not (hasattr(tokenizer, 'chat_template') and tokenizer.chat_template): + try: + from unsloth.chat_templates import get_chat_template + tokenizer = get_chat_template(tokenizer, chat_template="alpaca") + print(f"šŸ“ Set alpaca chat template on tokenizer for model saving") + except Exception as e: + print(f"āš ļø Could not set alpaca template on tokenizer: {e}") + # Use custom template if provided def _format_alpaca_custom(examples): texts = [] @@ -258,7 +285,7 @@ def apply_chat_template_to_dataset( if not isinstance(dataset, IterableDataset): from multiprocessing import cpu_count if num_proc is None or type(num_proc) is not int: - num_proc = cpu_count() + num_proc = max(1, cpu_count() // 3) dataset_map_kwargs['num_proc'] = num_proc dataset_map_kwargs['desc'] = "Applying template to Alpaca format" @@ -322,7 +349,7 @@ def apply_chat_template_to_dataset( if not isinstance(dataset, IterableDataset): from multiprocessing import cpu_count if num_proc is None or type(num_proc) is not int: - num_proc = cpu_count() + num_proc = max(1, cpu_count() // 3) dataset_map_kwargs['num_proc'] = num_proc dataset_map_kwargs['desc'] = f"Applying chat template to {final_format}" diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py index 9367741e8e..a97db20ba4 100644 --- a/studio/backend/utils/datasets/format_conversion.py +++ b/studio/backend/utils/datasets/format_conversion.py @@ -110,7 +110,7 @@ def standardize_chat_format( from multiprocessing import cpu_count if num_proc is None or type(num_proc) is not int: - num_proc = cpu_count() + num_proc = max(1, cpu_count() // 3) dataset_map_kwargs['num_proc'] = num_proc dataset_map_kwargs['desc'] = "Standardizing chat format" @@ -176,7 +176,7 @@ def convert_chatml_to_alpaca(dataset, batch_size=1000, num_proc=None): from multiprocessing import cpu_count if num_proc is None or type(num_proc) is not int: - num_proc = cpu_count() + num_proc = max(1, cpu_count() // 3) dataset_map_kwargs['num_proc'] = num_proc dataset_map_kwargs['desc'] = "Converting ChatML to Alpaca format" @@ -224,7 +224,7 @@ def convert_alpaca_to_chatml(dataset, batch_size=1000, num_proc=None): from multiprocessing import cpu_count if num_proc is None or type(num_proc) is not int: - num_proc = cpu_count() + num_proc = max(1, cpu_count() // 3) dataset_map_kwargs['num_proc'] = num_proc dataset_map_kwargs['desc'] = "Converting Alpaca to ChatML format" diff --git a/studio/backend/utils/datasets/format_detection.py b/studio/backend/utils/datasets/format_detection.py index ae1b6a388d..9283ea5d55 100644 --- a/studio/backend/utils/datasets/format_detection.py +++ b/studio/backend/utils/datasets/format_detection.py @@ -328,6 +328,11 @@ def detect_multimodal_dataset(dataset): """ Detects if dataset contains multimodal data (images/vision). + Two-pass approach: + 1. Column-name heuristic (fast): checks for keywords like 'image', 'img', 'pixel'. + 2. Value-type inspection (reliable): checks if actual values are PIL Images, + bytes with image headers, or HF Image-feature dicts. + Returns: dict: { "is_multimodal": bool, @@ -339,11 +344,16 @@ def detect_multimodal_dataset(dataset): column_names = list(sample.keys()) # Keywords that indicate multimodal/image data - multimodal_keywords = ['image', 'img', 'pixel'] + multimodal_keywords = [ + 'image', 'img', 'pixel', + 'jpg', 'jpeg', 'png', 'webp', 'bmp', 'gif', 'tiff', 'svg', + 'photo', 'pic', 'picture', 'visual', + ] multimodal_columns = [] modality_types = set() + # ā”€ā”€ Pass 1: column-name heuristic ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ for col_name in column_names: col_lower = col_name.lower() @@ -353,6 +363,17 @@ def detect_multimodal_dataset(dataset): modality_types.add(keyword) break # Don't check other keywords for this column + # ā”€ā”€ Pass 2: inspect actual values ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ + # Catches columns with non-obvious names (e.g. "jpg", "photo", "pic") + already_detected = set(multimodal_columns) + for col_name in column_names: + if col_name in already_detected: + continue + value = sample[col_name] + if _is_image_value(value): + multimodal_columns.append(col_name) + modality_types.add("image") + return { "is_multimodal": len(multimodal_columns) > 0, "multimodal_columns": multimodal_columns, @@ -360,6 +381,54 @@ def detect_multimodal_dataset(dataset): } +def _is_image_value(value) -> bool: + """Check if a single sample value looks like image data.""" + if value is None: + return False + + # PIL Image instance + try: + from PIL.Image import Image as PILImage + if isinstance(value, PILImage): + return True + except ImportError: + pass + + # HF datasets Image feature stores decoded images as PIL or dicts with + # {"bytes": b"...", "path": "..."} when not yet decoded + if isinstance(value, dict): + if "bytes" in value and "path" in value: + return True + + # Raw bytes with a known image magic header + if isinstance(value, (bytes, bytearray)): + return _has_image_header(value) + + return False + + +def _has_image_header(data: bytes) -> bool: + """Quick magic-byte check for common image formats.""" + if len(data) < 4: + return False + # JPEG + if data[:2] == b'\xff\xd8': + return True + # PNG + if data[:4] == b'\x89PNG': + return True + # GIF + if data[:3] == b'GIF': + return True + # WebP + if data[:4] == b'RIFF' and len(data) >= 12 and data[8:12] == b'WEBP': + return True + # BMP + if data[:2] == b'BM': + return True + return False + + def detect_vlm_dataset_structure(dataset): """ Detects if VLM dataset is: diff --git a/studio/backend/utils/hardware/__init__.py b/studio/backend/utils/hardware/__init__.py index b992fc191d..667466efea 100644 --- a/studio/backend/utils/hardware/__init__.py +++ b/studio/backend/utils/hardware/__init__.py @@ -10,6 +10,9 @@ from .hardware import ( clear_gpu_cache, get_gpu_memory_info, log_gpu_memory, + get_gpu_summary, + get_package_versions, + get_gpu_utilization, ) __all__ = [ @@ -21,4 +24,7 @@ __all__ = [ 'clear_gpu_cache', 'get_gpu_memory_info', 'log_gpu_memory', + 'get_gpu_summary', + 'get_package_versions', + 'get_gpu_utilization', ] diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py index 25320a25e1..0d7cc97cfb 100644 --- a/studio/backend/utils/hardware/hardware.py +++ b/studio/backend/utils/hardware/hardware.py @@ -120,6 +120,7 @@ def clear_gpu_cache(): if device == DeviceType.CUDA: import torch + torch.cuda.synchronize() torch.cuda.empty_cache() torch.cuda.ipc_collect() elif device == DeviceType.MLX: @@ -206,3 +207,181 @@ def log_gpu_memory(context: str): ) else: logger.info(f"GPU Memory [{context}]: No GPU available (CPU-only)") + + +# ========== GPU Summary & Package Versions ========== + +def get_gpu_summary() -> Dict[str, Any]: + """ + Return a compact summary of the primary GPU. + + Returns dict with keys: + gpu_name ā€“ e.g. "NVIDIA L4" (or None) + vram_total_gb ā€“ e.g. 22.17 (or None) + """ + mem = get_gpu_memory_info() + if mem.get("available"): + return { + "gpu_name": mem.get("device_name"), + "vram_total_gb": round(mem.get("total_gb", 0), 2), + } + return {"gpu_name": None, "vram_total_gb": None} + + +def get_package_versions() -> Dict[str, Optional[str]]: + """ + Return the installed versions of key ML packages. + + Uses importlib.metadata (stdlib) so no subprocess is needed. + CUDA version comes from torch.version.cuda. + + Returns dict with keys: unsloth, torch, transformers, cuda. + Missing packages yield None. + """ + from importlib.metadata import version as pkg_version, PackageNotFoundError + + packages = ("unsloth", "torch", "transformers") + versions: Dict[str, Optional[str]] = {} + + for name in packages: + try: + versions[name] = pkg_version(name) + except PackageNotFoundError: + versions[name] = None + + # CUDA toolkit version bundled with torch + try: + import torch + versions["cuda"] = getattr(torch.version, "cuda", None) + except Exception: + versions["cuda"] = None + + return versions + + +# ========== Live GPU Utilization (nvidia-smi) ========== + +def get_gpu_utilization() -> Dict[str, Any]: + """ + Return a live snapshot of GPU utilization via ``nvidia-smi``. + + Designed to be polled by the frontend during training (not streaming). + Uses ``nvidia-smi --query-gpu`` which is the most accurate source for + utilization %, temperature, and power draw ā€“ stats that PyTorch does + not expose. + + Returns dict with keys: + available ā€“ bool, whether stats could be retrieved + gpu_utilization_pct ā€“ GPU core utilization % + temperature_c ā€“ GPU temperature in Ā°C + vram_used_gb ā€“ VRAM currently used (GiB) + vram_total_gb ā€“ VRAM total (GiB) + vram_utilization_pct ā€“ VRAM used / total * 100 + power_draw_w ā€“ current power draw (W) + power_limit_w ā€“ power limit (W) + power_utilization_pct ā€“ power draw / limit * 100 + """ + device = get_device() + + if device != DeviceType.CUDA: + return {"available": False, "backend": device.value} + + def _parse_smi_value(raw: str): + """Parse a single nvidia-smi CSV value. Returns float or None for [N/A].""" + raw = raw.strip() + if not raw or raw == "[N/A]": + return None + try: + return float(raw) + except (ValueError, TypeError): + return None + + # ā”€ā”€ nvidia-smi (most complete source) ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ + smi_data = {} + try: + import subprocess + + result = subprocess.run( + [ + "nvidia-smi", + "--query-gpu=utilization.gpu,temperature.gpu," + "memory.used,memory.total,power.draw,power.limit", + "--format=csv,noheader,nounits", + ], + capture_output=True, + text=True, + timeout=5, + ) + + if result.returncode == 0 and result.stdout.strip(): + # nvidia-smi outputs one line per GPU; take GPU 0 + first_line = result.stdout.strip().splitlines()[0] + parts = [p.strip() for p in first_line.split(",")] + if len(parts) >= 6: + smi_data = { + "gpu_util": _parse_smi_value(parts[0]), + "temp": _parse_smi_value(parts[1]), + "vram_used_mb": _parse_smi_value(parts[2]), + "vram_total_mb": _parse_smi_value(parts[3]), + "power_draw": _parse_smi_value(parts[4]), + "power_limit": _parse_smi_value(parts[5]), + } + + except FileNotFoundError: + logger.debug("nvidia-smi not found, falling back to torch.cuda") + except Exception as e: + logger.warning(f"nvidia-smi query failed: {e}") + + # ā”€ā”€ Backfill VRAM from torch.cuda if nvidia-smi returned [N/A] ā”€ā”€ + vram_used_mb = smi_data.get("vram_used_mb") + vram_total_mb = smi_data.get("vram_total_mb") + + if vram_used_mb is None or vram_total_mb is None: + try: + import torch + + idx = torch.cuda.current_device() + props = torch.cuda.get_device_properties(idx) + if vram_total_mb is None: + vram_total_mb = props.total_memory / (1024**2) # bytes ā†’ MiB + if vram_used_mb is None: + vram_used_mb = torch.cuda.memory_allocated(idx) / (1024**2) + except Exception as e: + logger.debug(f"torch.cuda VRAM backfill failed: {e}") + + # ā”€ā”€ Build response ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ + gpu_util = smi_data.get("gpu_util") + temp = smi_data.get("temp") + power_draw = smi_data.get("power_draw") + power_limit = smi_data.get("power_limit") + + vram_used_gb = round(vram_used_mb / 1024, 2) if vram_used_mb is not None else None + vram_total_gb = round(vram_total_mb / 1024, 2) if vram_total_mb is not None else None + vram_pct = ( + round((vram_used_mb / vram_total_mb) * 100, 1) + if vram_used_mb is not None and vram_total_mb and vram_total_mb > 0 + else None + ) + power_pct = ( + round((power_draw / power_limit) * 100, 1) + if power_draw is not None and power_limit and power_limit > 0 + else None + ) + + # If we got at least something useful, report available + has_any = any(v is not None for v in [gpu_util, temp, vram_used_gb, power_draw]) + if not has_any: + return {"available": False, "backend": device.value} + + return { + "available": True, + "backend": device.value, + "gpu_utilization_pct": gpu_util, + "temperature_c": temp, + "vram_used_gb": vram_used_gb, + "vram_total_gb": vram_total_gb, + "vram_utilization_pct": vram_pct, + "power_draw_w": power_draw, + "power_limit_w": power_limit, + "power_utilization_pct": power_pct, + } diff --git a/studio/backend/utils/models/__init__.py b/studio/backend/utils/models/__init__.py index 8344d21a89..505fd35edd 100644 --- a/studio/backend/utils/models/__init__.py +++ b/studio/backend/utils/models/__init__.py @@ -11,6 +11,7 @@ from .model_config import ( MODEL_NAME_MAPPING, UI_STATUS_INDICATORS, ) +from .checkpoints import scan_checkpoints __all__ = [ 'ModelConfig', @@ -21,4 +22,5 @@ __all__ = [ 'load_model_config', 'MODEL_NAME_MAPPING', 'UI_STATUS_INDICATORS', + 'scan_checkpoints', ] diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py new file mode 100644 index 0000000000..7a50d0900d --- /dev/null +++ b/studio/backend/utils/models/checkpoints.py @@ -0,0 +1,120 @@ +""" +Checkpoint scanning utilities for discovering training runs and their checkpoints. +""" +import json +import logging +from pathlib import Path +from typing import List, Optional, Tuple + +logger = logging.getLogger(__name__) + + +def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]: + """ + Read the training loss from a checkpoint's trainer_state.json. + + Returns the loss from the last log_history entry, or None if unavailable. + """ + trainer_state = checkpoint_path / "trainer_state.json" + if not trainer_state.exists(): + return None + try: + with open(trainer_state) as f: + state = json.load(f) + log_history = state.get("log_history", []) + if log_history: + return log_history[-1].get("loss") + except Exception as e: + logger.debug(f"Could not read loss from {trainer_state}: {e}") + return None + + +def scan_checkpoints( + outputs_dir: str = "./outputs", +) -> List[Tuple[str, List[Tuple[str, str, Optional[float]]], dict]]: + """ + Scan outputs folder for training runs and their checkpoints. + + Returns: + List of tuples: [(model_name, [(display_name, checkpoint_path, loss), ...], metadata), ...] + metadata keys: base_model, peft_type, lora_rank (all optional) + The first entry in each checkpoint list is the main adapter; its loss is + set to the loss of the last (highest-step) intermediate checkpoint. + """ + models = [] + outputs_path = Path(outputs_dir) + + if not outputs_path.exists(): + logger.warning(f"Outputs directory not found: {outputs_dir}") + return models + + try: + for item in outputs_path.iterdir(): + if not item.is_dir(): + continue + + config_file = item / "config.json" + adapter_config = item / "adapter_config.json" + + if not (config_file.exists() or adapter_config.exists()): + continue + + # Extract training metadata from adapter_config.json / config.json + metadata: dict = {} + try: + if adapter_config.exists(): + cfg = json.loads(adapter_config.read_text()) + metadata["base_model"] = cfg.get("base_model_name_or_path") + metadata["peft_type"] = cfg.get("peft_type") + metadata["lora_rank"] = cfg.get("r") + elif config_file.exists(): + cfg = json.loads(config_file.read_text()) + metadata["base_model"] = cfg.get("_name_or_path") + except Exception: + pass + + # Fallback: extract base model name from folder name + # e.g. "unsloth_Llama-3.2-3B-Instruct_1771227800" ā†’ "unsloth/Llama-3.2-3B-Instruct" + if not metadata.get("base_model"): + parts = item.name.rsplit("_", 1) + if len(parts) == 2 and parts[1].isdigit(): + name_part = parts[0] + idx = name_part.find("_") + if idx > 0: + metadata["base_model"] = name_part[:idx] + "/" + name_part[idx + 1:] + else: + metadata["base_model"] = name_part + + # This is a valid training run + checkpoints = [] + + # Placeholder for the main adapter ā€” loss filled from last checkpoint below + checkpoints.append((item.name, str(item), None)) + + # Scan for intermediate checkpoints (checkpoint-N subdirs) + for sub in sorted(item.iterdir()): + if not sub.is_dir() or not sub.name.startswith("checkpoint-"): + continue + sub_config = sub / "config.json" + sub_adapter = sub / "adapter_config.json" + if sub_config.exists() or sub_adapter.exists(): + loss = _read_checkpoint_loss(sub) + checkpoints.append((sub.name, str(sub), loss)) + + # Assign the last checkpoint's loss to the main adapter entry + if len(checkpoints) > 1: + last_checkpoint_loss = checkpoints[-1][2] + checkpoints[0] = (checkpoints[0][0], checkpoints[0][1], last_checkpoint_loss) + + models.append((item.name, checkpoints, metadata)) + logger.debug(f"Found model: {item.name} with {len(checkpoints)} checkpoint(s)") + + # Sort by modification time (newest first) + models.sort(key=lambda x: Path(x[1][0][1]).stat().st_mtime, reverse=True) + + logger.info(f"Found {len(models)} training runs in {outputs_dir}") + return models + + except Exception as e: + logger.error(f"Error scanning checkpoints: {e}") + return [] diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py index 5882f41e2a..fdf89fce39 100644 --- a/studio/backend/utils/models/model_config.py +++ b/studio/backend/utils/models/model_config.py @@ -369,7 +369,7 @@ pass def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool: """ - Detect vision models by checking architecture in config. + Detect vision-language models (VLMs) by checking architecture in config. Works for fine-tuned models since they inherit the base architecture. Args: @@ -377,23 +377,43 @@ def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool: hf_token: Optional HF token for accessing gated/private models """ try: - config = load_model_config(model_name, token=hf_token) + config = load_model_config(model_name, use_auth=True, token=hf_token) - # Check vision arch + # Check 1: Architecture class name patterns if hasattr(config, 'architectures'): is_vlm = any( x.endswith(("ForConditionalGeneration", "ForVisionText2Text")) for x in config.architectures ) if is_vlm: - logger.info(f"Model {model_name} detected as vision model: architecture {config.architectures}") + logger.info(f"Model {model_name} detected as VLM: architecture {config.architectures}") return True - # Quick check for vision config as backup + # Check 2: Has vision_config (most VLMs: LLaVA, Gemma-3, Qwen2-VL, etc.) if hasattr(config, 'vision_config'): - logger.info(f"Model {model_name} detected as vision model: has vision_config") + logger.info(f"Model {model_name} detected as VLM: has vision_config") return True + # Check 3: Has img_processor (Phi-3.5 Vision uses this instead of vision_config) + if hasattr(config, 'img_processor'): + logger.info(f"Model {model_name} detected as VLM: has img_processor") + return True + + # Check 4: Has image_token_index (common in VLMs for image placeholder tokens) + if hasattr(config, 'image_token_index'): + logger.info(f"Model {model_name} detected as VLM: has image_token_index") + return True + + # Check 5: Known VLM model_type values that may not match above checks + if hasattr(config, 'model_type'): + vlm_model_types = { + 'phi3_v', 'llava', 'llava_next', 'llava_onevision', + 'internvl_chat', 'cogvlm2', 'minicpmv', + } + if config.model_type in vlm_model_types: + logger.info(f"Model {model_name} detected as VLM: model_type={config.model_type}") + return True + return False except Exception as e: diff --git a/studio/backend/utils/paths/path_utils.py b/studio/backend/utils/paths/path_utils.py index 7743952b6b..856e478bc2 100644 --- a/studio/backend/utils/paths/path_utils.py +++ b/studio/backend/utils/paths/path_utils.py @@ -41,6 +41,13 @@ def is_local_path(path: str) -> bool: if not path: return False + # If it exists on disk, treat as local (covers relative paths like "outputs/foo"). + try: + if Path(normalize_path(path)).expanduser().exists(): + return True + except Exception: + pass + # Obvious HF patterns if path.count('/') == 1 and not path.startswith(('/', '.', '~')): return False # Looks like org/model format diff --git a/studio/frontend/bun.lock b/studio/frontend/bun.lock index 9859b3aae5..d20305d662 100644 --- a/studio/frontend/bun.lock +++ b/studio/frontend/bun.lock @@ -5,10 +5,10 @@ "": { "name": "unsloth-theme", "dependencies": { - "@assistant-ui/react": "^0.12.10", - "@assistant-ui/react-markdown": "^0.12.3", - "@assistant-ui/react-streamdown": "^0.1.2", - "@base-ui/react": "^1.2.0", + "@assistant-ui/react": "^0.12.3", + "@assistant-ui/react-markdown": "^0.12.1", + "@assistant-ui/react-streamdown": "^0.1.0", + "@base-ui/react": "^1.1.0", "@dagrejs/dagre": "^2.0.4", "@dagrejs/graphlib": "^3.0.4", "@fontsource-variable/figtree": "^5.2.10", diff --git a/studio/frontend/package.json b/studio/frontend/package.json index 9a10e8e639..bd087b6748 100644 --- a/studio/frontend/package.json +++ b/studio/frontend/package.json @@ -72,13 +72,13 @@ "zustand": "^5.0.11" }, "devDependencies": { - "@eslint/js": "^9.39.2", - "@types/node": "^24.10.13", - "@types/react": "^19.2.14", - "@types/react-dom": "^19.2.3", - "@vitejs/plugin-react": "^5.1.4", "@biomejs/biome": "^1.9.4", - "eslint": "^9.39.2", + "@eslint/js": "^9.39.1", + "@types/node": "^24.10.1", + "@types/react": "^19.2.5", + "@types/react-dom": "^19.2.3", + "@vitejs/plugin-react": "^5.1.1", + "eslint": "^9.39.1", "eslint-plugin-react-hooks": "^7.0.1", "eslint-plugin-react-refresh": "^0.4.26", "globals": "^16.5.0", diff --git a/studio/frontend/src/app/router.tsx b/studio/frontend/src/app/router.tsx index 16152ccb78..dbbd9b1148 100644 --- a/studio/frontend/src/app/router.tsx +++ b/studio/frontend/src/app/router.tsx @@ -5,14 +5,14 @@ import { Route as dataRecipeRoute } from "./routes/data-recipes.$recipeId"; import { Route as chatRoute } from "./routes/chat"; import { Route as exportRoute } from "./routes/export"; import { Route as gridTestRoute } from "./routes/grid-test"; -import { Route as homeRoute } from "./routes/home"; +import { Route as indexRoute } from "./routes/index"; import { Route as loginRoute } from "./routes/login"; import { Route as onboardingRoute } from "./routes/onboarding"; import { Route as signupRoute } from "./routes/signup"; import { Route as studioRoute } from "./routes/studio"; const routeTree = rootRoute.addChildren([ - homeRoute, + indexRoute, onboardingRoute, loginRoute, signupRoute, diff --git a/studio/frontend/src/app/routes/home.tsx b/studio/frontend/src/app/routes/home.tsx deleted file mode 100644 index bf2f3a6b58..0000000000 --- a/studio/frontend/src/app/routes/home.tsx +++ /dev/null @@ -1,15 +0,0 @@ -import { ComponentExample } from "@/components/component-example"; -import { createRoute } from "@tanstack/react-router"; -import { requireAuth } from "../auth-guards"; -import { Route as rootRoute } from "./__root"; - -export const Route = createRoute({ - getParentRoute: () => rootRoute, - path: "/", - beforeLoad: () => requireAuth(), - component: HomePage, -}); - -function HomePage() { - return ; -} diff --git a/studio/frontend/src/app/routes/index.tsx b/studio/frontend/src/app/routes/index.tsx new file mode 100644 index 0000000000..4ce567fbec --- /dev/null +++ b/studio/frontend/src/app/routes/index.tsx @@ -0,0 +1,14 @@ +import { createRoute, redirect } from "@tanstack/react-router"; +import { getPostAuthRoute } from "@/features/auth"; +import { requireAuth } from "../auth-guards"; +import { Route as rootRoute } from "./__root"; + +export const Route = createRoute({ + getParentRoute: () => rootRoute, + path: "/", + beforeLoad: async () => { + await requireAuth(); + throw redirect({ to: getPostAuthRoute() }); + }, + component: () => null, +}); diff --git a/studio/frontend/src/components/assistant-ui/model-selector.tsx b/studio/frontend/src/components/assistant-ui/model-selector.tsx index 85bc1cacdb..bf5f219558 100644 --- a/studio/frontend/src/components/assistant-ui/model-selector.tsx +++ b/studio/frontend/src/components/assistant-ui/model-selector.tsx @@ -1,39 +1,26 @@ "use client"; -import { Input } from "@/components/ui/input"; import { Popover, PopoverContent, PopoverTrigger, } from "@/components/ui/popover"; -import { Spinner } from "@/components/ui/spinner"; import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs"; -import { useDebouncedValue, useHfModelSearch, useInfiniteScroll } from "@/hooks"; -import { cn, formatCompact } from "@/lib/utils"; +import { cn } from "@/lib/utils"; import { ArrowDown01Icon, Logout01Icon, - Search01Icon, } from "@hugeicons/core-free-icons"; import { HugeiconsIcon } from "@hugeicons/react"; -import { type ReactNode, useMemo, useState } from "react"; +import { useMemo, useState } from "react"; +import type { + LoraModelOption, + ModelOption, + ModelSelectorChangeMeta, +} from "./model-selector/types"; +import { HubModelPicker, LoraModelPicker } from "./model-selector/pickers"; -export interface ModelOption { - id: string; - name: string; - description?: string; - icon?: ReactNode; -} - -export interface LoraModelOption extends ModelOption { - baseModel?: string; - updatedAt?: number; -} - -export interface ModelSelectorChangeMeta { - source: "hub" | "lora"; - isLora: boolean; -} +export type { LoraModelOption, ModelOption, ModelSelectorChangeMeta } from "./model-selector/types"; interface ModelSelectorProps { models: ModelOption[]; @@ -46,10 +33,10 @@ interface ModelSelectorProps { size?: "sm" | "default" | "lg"; className?: string; contentClassName?: string; -} - -function dedupe(values: string[]): string[] { - return [...new Set(values.filter(Boolean))]; + open?: boolean; + onOpenChange?: (open: boolean) => void; + triggerDataTour?: string; + contentDataTour?: string; } function ModelSelectorTrigger({ @@ -58,17 +45,20 @@ function ModelSelectorTrigger({ variant = "outline", size = "default", className, + dataTour, }: { currentModel?: ModelOption; isLoaded: boolean; variant?: "outline" | "ghost" | "muted"; size?: "sm" | "default" | "lg"; className?: string; + dataTour?: string; }) { return ( - ); -} - -function HubModelPicker({ - models, - value, - onSelect, -}: { - models: ModelOption[]; - value?: string; - onSelect: (id: string, meta: ModelSelectorChangeMeta) => void; -}) { - const [query, setQuery] = useState(""); - const debouncedQuery = useDebouncedValue(query); - const { results, isLoading, isLoadingMore, fetchMore } = useHfModelSearch( - debouncedQuery, - ); - - const recommendedIds = useMemo( - () => dedupe([...models.map((model) => model.id), value ?? ""]), - [models, value], - ); - - const showHfSection = debouncedQuery.trim().length > 0; - const recommendedSet = useMemo( - () => new Set(recommendedIds), - [recommendedIds], - ); - - const hfIds = useMemo(() => { - if (!showHfSection) { - return []; - } - return results - .map((result) => result.id) - .filter((id) => !recommendedSet.has(id)); - }, [recommendedSet, results, showHfSection]); - - const metricsById = useMemo( - () => - new Map( - results.map((result) => [ - result.id, - result.totalParams - ? formatCompact(result.totalParams) - : `ā†“${formatCompact(result.downloads)}`, - ]), - ), - [results], - ); - - const { scrollRef, sentinelRef } = useInfiniteScroll(fetchMore, results.length); - - return ( -
-
- - setQuery(event.target.value)} - placeholder="Search Hugging Face models" - className="h-9 pl-8 pr-8" - /> - {isLoading && ( - - )} -
- -
-
- {!showHfSection ? ( - <> - Recommended - {recommendedIds.length === 0 ? ( -
- No default models. -
- ) : ( - recommendedIds.map((id) => ( - onSelect(id, { source: "hub", isLora: false })} - /> - )) - )} - - ) : null} - - {showHfSection ? ( - <> - Hugging Face - {hfIds.length === 0 && !isLoading ? ( -
- No matching models. -
- ) : ( - hfIds.map((id) => ( - onSelect(id, { source: "hub", isLora: false })} - /> - )) - )} -
- {isLoadingMore ? ( -
- -
- ) : null} - - ) : null} -
-
-
- ); -} - -function LoraModelPicker({ - loraModels, - value, - onSelect, -}: { - loraModels: LoraModelOption[]; - value?: string; - onSelect: (id: string, meta: ModelSelectorChangeMeta) => void; -}) { - const [query, setQuery] = useState(""); - - const normalized = useMemo( - () => - loraModels - .map((model) => ({ - ...model, - baseModel: model.baseModel || model.description || "Unknown base model", - })) - .sort((a, b) => { - const aTime = a.updatedAt ?? -1; - const bTime = b.updatedAt ?? -1; - if (aTime !== bTime) { - return bTime - aTime; - } - const baseCmp = a.baseModel.localeCompare(b.baseModel); - if (baseCmp !== 0) { - return baseCmp; - } - return a.name.localeCompare(b.name); - }), - [loraModels], - ); - - const grouped = useMemo(() => { - const needle = query.trim().toLowerCase(); - const out = new Map(); - - for (const model of normalized) { - const searchText = `${model.name} ${model.baseModel} ${model.id}`.toLowerCase(); - if (needle && !searchText.includes(needle)) { - continue; - } - - const key = model.baseModel || "Unknown base model"; - const prev = out.get(key) ?? []; - prev.push(model); - out.set(key, prev); - } - - return [...out.entries()].sort((a, b) => { - const aLatest = Math.max(...a[1].map((model) => model.updatedAt ?? -1)); - const bLatest = Math.max(...b[1].map((model) => model.updatedAt ?? -1)); - if (aLatest !== bLatest) { - return bLatest - aLatest; - } - return a[0].localeCompare(b[0]); - }); - }, [normalized, query]); - - return ( -
-
- - setQuery(event.target.value)} - placeholder="Search local adapters" - className="h-9 pl-8" - /> -
- -
-
- {grouped.length === 0 ? ( -
No adapters found.
- ) : ( - grouped.map(([baseModel, adapters], index) => ( -
- {index > 0 ?
: null} - {baseModel} - {adapters.map((adapter) => ( - onSelect(adapter.id, { source: "lora", isLora: true })} - /> - ))} -
- )) - )} -
-
-
- ); -} - function ModelSelectorContent({ models, loraModels, @@ -368,6 +96,7 @@ function ModelSelectorContent({ onSelect, onEject, className, + dataTour, }: { models: ModelOption[]; loraModels: LoraModelOption[]; @@ -375,13 +104,18 @@ function ModelSelectorContent({ onSelect: (id: string, meta: ModelSelectorChangeMeta) => void; onEject?: () => void; className?: string; + dataTour?: string; }) { const hasSelection = Boolean(value); return ( @@ -430,8 +164,14 @@ export function ModelSelector({ size = "default", className, contentClassName, + open: controlledOpen, + onOpenChange, + triggerDataTour, + contentDataTour, }: ModelSelectorProps) { - const [open, setOpen] = useState(false); + const [uncontrolledOpen, setUncontrolledOpen] = useState(false); + const open = controlledOpen ?? uncontrolledOpen; + const setOpen = onOpenChange ?? setUncontrolledOpen; const [uncontrolled, setUncontrolled] = useState(defaultValue ?? ""); const selected = value ?? uncontrolled; @@ -477,6 +217,7 @@ export function ModelSelector({ variant={variant} size={size} className={className} + dataTour={triggerDataTour} /> ); diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx new file mode 100644 index 0000000000..b90cdd0e3b --- /dev/null +++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx @@ -0,0 +1,265 @@ +import { Input } from "@/components/ui/input"; +import { Spinner } from "@/components/ui/spinner"; +import { useDebouncedValue, useHfModelSearch, useInfiniteScroll } from "@/hooks"; +import { cn, formatCompact } from "@/lib/utils"; +import { Search01Icon } from "@hugeicons/core-free-icons"; +import { HugeiconsIcon } from "@hugeicons/react"; +import { useMemo, useState, type ReactNode } from "react"; +import type { + LoraModelOption, + ModelOption, + ModelSelectorChangeMeta, +} from "./types"; + +function dedupe(values: string[]): string[] { + return [...new Set(values.filter(Boolean))]; +} + +function ListLabel({ children }: { children: ReactNode }) { + return ( +
+ {children} +
+ ); +} + +function ModelRow({ + label, + meta, + selected, + onClick, +}: { + label: string; + meta?: string; + selected?: boolean; + onClick: () => void; +}) { + return ( + + ); +} + +export function HubModelPicker({ + models, + value, + onSelect, +}: { + models: ModelOption[]; + value?: string; + onSelect: (id: string, meta: ModelSelectorChangeMeta) => void; +}) { + const [query, setQuery] = useState(""); + const debouncedQuery = useDebouncedValue(query); + const { results, isLoading, isLoadingMore, fetchMore } = useHfModelSearch( + debouncedQuery, + ); + + const recommendedIds = useMemo( + () => dedupe([...models.map((model) => model.id), value ?? ""]), + [models, value], + ); + + const showHfSection = debouncedQuery.trim().length > 0; + const recommendedSet = useMemo(() => new Set(recommendedIds), [recommendedIds]); + + const hfIds = useMemo(() => { + if (!showHfSection) return []; + return results + .map((result) => result.id) + .filter((id) => !recommendedSet.has(id)); + }, [recommendedSet, results, showHfSection]); + + const metricsById = useMemo( + () => + new Map( + results.map((result) => [ + result.id, + result.totalParams + ? formatCompact(result.totalParams) + : `ā†“${formatCompact(result.downloads)}`, + ]), + ), + [results], + ); + + const { scrollRef, sentinelRef } = useInfiniteScroll(fetchMore, results.length); + + return ( +
+
+ + setQuery(event.target.value)} + placeholder="Search Hugging Face models" + className="h-9 pl-8 pr-8" + /> + {isLoading && ( + + )} +
+ +
+
+ {!showHfSection ? ( + <> + Recommended + {recommendedIds.length === 0 ? ( +
+ No default models. +
+ ) : ( + recommendedIds.map((id) => ( + onSelect(id, { source: "hub", isLora: false })} + /> + )) + )} + + ) : null} + + {showHfSection ? ( + <> + Hugging Face + {hfIds.length === 0 && !isLoading ? ( +
+ No matching models. +
+ ) : ( + hfIds.map((id) => ( + onSelect(id, { source: "hub", isLora: false })} + /> + )) + )} +
+ {isLoadingMore ? ( +
+ +
+ ) : null} + + ) : null} +
+
+
+ ); +} + +export function LoraModelPicker({ + loraModels, + value, + onSelect, +}: { + loraModels: LoraModelOption[]; + value?: string; + onSelect: (id: string, meta: ModelSelectorChangeMeta) => void; +}) { + const [query, setQuery] = useState(""); + + const normalized = useMemo( + () => + loraModels + .map((model) => ({ + ...model, + baseModel: model.baseModel || model.description || "Unknown base model", + })) + .sort((a, b) => { + const aTime = a.updatedAt ?? -1; + const bTime = b.updatedAt ?? -1; + if (aTime !== bTime) return bTime - aTime; + const baseCmp = a.baseModel.localeCompare(b.baseModel); + if (baseCmp !== 0) return baseCmp; + return a.name.localeCompare(b.name); + }), + [loraModels], + ); + + const grouped = useMemo(() => { + const needle = query.trim().toLowerCase(); + const out = new Map(); + + for (const model of normalized) { + const searchText = `${model.name} ${model.baseModel} ${model.id}`.toLowerCase(); + if (needle && !searchText.includes(needle)) continue; + + const key = model.baseModel || "Unknown base model"; + const prev = out.get(key) ?? []; + prev.push(model); + out.set(key, prev); + } + + return [...out.entries()].sort((a, b) => { + const aLatest = Math.max(...a[1].map((model) => model.updatedAt ?? -1)); + const bLatest = Math.max(...b[1].map((model) => model.updatedAt ?? -1)); + if (aLatest !== bLatest) return bLatest - aLatest; + return a[0].localeCompare(b[0]); + }); + }, [normalized, query]); + + return ( +
+
+ + setQuery(event.target.value)} + placeholder="Search local adapters" + className="h-9 pl-8" + /> +
+ +
+
+ {grouped.length === 0 ? ( +
+ No adapters found. +
+ ) : ( + grouped.map(([baseModel, adapters], index) => ( +
+ {index > 0 ?
: null} + {baseModel} + {adapters.map((adapter) => ( + onSelect(adapter.id, { source: "lora", isLora: true })} + /> + ))} +
+ )) + )} +
+
+
+ ); +} + diff --git a/studio/frontend/src/components/assistant-ui/model-selector/types.ts b/studio/frontend/src/components/assistant-ui/model-selector/types.ts new file mode 100644 index 0000000000..dcf110bfb7 --- /dev/null +++ b/studio/frontend/src/components/assistant-ui/model-selector/types.ts @@ -0,0 +1,19 @@ +import type { ReactNode } from "react"; + +export interface ModelOption { + id: string; + name: string; + description?: string; + icon?: ReactNode; +} + +export interface LoraModelOption extends ModelOption { + baseModel?: string; + updatedAt?: number; +} + +export interface ModelSelectorChangeMeta { + source: "hub" | "lora"; + isLora: boolean; +} + diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx index 9c94e4eb22..c109c92ead 100644 --- a/studio/frontend/src/components/assistant-ui/thread.tsx +++ b/studio/frontend/src/components/assistant-ui/thread.tsx @@ -7,9 +7,7 @@ import { MarkdownText } from "@/components/assistant-ui/markdown-text"; import { Reasoning, ReasoningGroup } from "@/components/assistant-ui/reasoning"; import { ToolFallback } from "@/components/assistant-ui/tool-fallback"; import { TooltipIconButton } from "@/components/assistant-ui/tooltip-icon-button"; -import { AnimatedShinyText } from "@/components/ui/animated-shiny-text"; import { Button } from "@/components/ui/button"; -import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store"; import { cn } from "@/lib/utils"; import { ActionBarMorePrimitive, @@ -73,7 +71,6 @@ export const Thread: FC<{ hideComposer?: boolean; hideWelcome?: boolean }> = ({ - !thread.isEmpty}> {!hideComposer && } @@ -83,28 +80,6 @@ export const Thread: FC<{ hideComposer?: boolean; hideWelcome?: boolean }> = ({ ); }; -const WarmupIndicator: FC = () => { - const threadId = useAuiState(({ threads }) => threads.mainThreadId); - const isRunning = useAuiState(({ thread }) => thread.isRunning); - const isWarmingUp = useChatRuntimeStore((state) => - Boolean(state.warmingByThreadId[threadId ?? "__default"]), - ); - - if (!isRunning || !isWarmingUp) { - return null; - } - - return ( -
-
- - Warming up model... - -
-
- ); -}; - const ThreadScrollToBottom: FC = () => { return ( diff --git a/studio/frontend/src/components/component-example.tsx b/studio/frontend/src/components/component-example.tsx deleted file mode 100644 index 3171e1abc0..0000000000 --- a/studio/frontend/src/components/component-example.tsx +++ /dev/null @@ -1,1318 +0,0 @@ -import * as React from "react"; - -import { Example, ExampleWrapper } from "@/components/example"; -import { - Accordion, - AccordionContent, - AccordionItem, - AccordionTrigger, -} from "@/components/ui/accordion"; -import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert"; -import { - AlertDialog, - AlertDialogAction, - AlertDialogCancel, - AlertDialogContent, - AlertDialogDescription, - AlertDialogFooter, - AlertDialogHeader, - AlertDialogMedia, - AlertDialogTitle, - AlertDialogTrigger, -} from "@/components/ui/alert-dialog"; -import { - Avatar, - AvatarBadge, - AvatarFallback, - AvatarGroup, - AvatarGroupCount, - AvatarImage, -} from "@/components/ui/avatar"; -import { Badge } from "@/components/ui/badge"; -import { - Breadcrumb, - BreadcrumbItem, - BreadcrumbLink, - BreadcrumbList, - BreadcrumbPage, - BreadcrumbSeparator, -} from "@/components/ui/breadcrumb"; -import { Button } from "@/components/ui/button"; -import { - Card, - CardAction, - CardContent, - CardDescription, - CardFooter, - CardHeader, - CardTitle, -} from "@/components/ui/card"; -import { Checkbox } from "@/components/ui/checkbox"; -import { - Combobox, - ComboboxContent, - ComboboxEmpty, - ComboboxInput, - ComboboxItem, - ComboboxList, -} from "@/components/ui/combobox"; -import { - Dialog, - DialogContent, - DialogDescription, - DialogFooter, - DialogHeader, - DialogTitle, - DialogTrigger, -} from "@/components/ui/dialog"; -import { - DropdownMenu, - DropdownMenuCheckboxItem, - DropdownMenuContent, - DropdownMenuGroup, - DropdownMenuItem, - DropdownMenuLabel, - DropdownMenuPortal, - DropdownMenuRadioGroup, - DropdownMenuRadioItem, - DropdownMenuSeparator, - DropdownMenuShortcut, - DropdownMenuSub, - DropdownMenuSubContent, - DropdownMenuSubTrigger, - DropdownMenuTrigger, -} from "@/components/ui/dropdown-menu"; -import { Field, FieldGroup, FieldLabel } from "@/components/ui/field"; -import { Input } from "@/components/ui/input"; -import { Label } from "@/components/ui/label"; -import { - Pagination, - PaginationContent, - PaginationItem, - PaginationLink, - PaginationNext, - PaginationPrevious, -} from "@/components/ui/pagination"; -import { Progress } from "@/components/ui/progress"; -import { RadioGroup, RadioGroupItem } from "@/components/ui/radio-group"; -import { - Select, - SelectContent, - SelectGroup, - SelectItem, - SelectTrigger, - SelectValue, -} from "@/components/ui/select"; -import { - Sheet, - SheetContent, - SheetDescription, - SheetFooter, - SheetHeader, - SheetTitle, - SheetTrigger, -} from "@/components/ui/sheet"; -import { Skeleton } from "@/components/ui/skeleton"; -import { Slider } from "@/components/ui/slider"; -import { Switch } from "@/components/ui/switch"; -import { - Table, - TableBody, - TableCell, - TableHead, - TableHeader, - TableRow, -} from "@/components/ui/table"; -import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs"; -import { Textarea } from "@/components/ui/textarea"; -import { Toggle } from "@/components/ui/toggle"; -import { ToggleGroup, ToggleGroupItem } from "@/components/ui/toggle-group"; -import { - Tooltip, - TooltipContent, - TooltipTrigger, -} from "@/components/ui/tooltip"; -import { - AlertCircleIcon, - BluetoothIcon, - CodeIcon, - ComputerIcon, - CreditCardIcon, - DownloadIcon, - EyeIcon, - File01Icon, - FileIcon, - FloppyDiskIcon, - FolderIcon, - FolderOpenIcon, - HelpCircleIcon, - InformationCircleIcon, - KeyboardIcon, - LanguageCircleIcon, - LayoutIcon, - LogoutIcon, - MailIcon, - MoonIcon, - MoreHorizontalCircle01Icon, - MoreVerticalCircle01Icon, - NotificationIcon, - PaintBoardIcon, - PanelRightIcon, - PlusSignIcon, - SearchIcon, - SettingsIcon, - ShieldIcon, - SunIcon, - TextBoldIcon, - TextItalicIcon, - TextUnderlineIcon, - UserIcon, -} from "@hugeicons/core-free-icons"; -import { HugeiconsIcon } from "@hugeicons/react"; - -export function ComponentExample() { - return ( - - - - - - - - - - - - - - ); -} - -function CardExample() { - return ( - - -
- mymind on Unsplash - - Observability Plus is replacing Monitoring - - Switch to the improved way to explore your data, with natural - language. Monitoring will no longer be available on the Pro plan in - November, 2025 - - - - - - - - - - - - - Allow accessory to connect? - - Do you want to allow the USB accessory to connect to this - device? - - - - Don't allow - Allow - - - - - Warning - - - - - ); -} - -const frameworks = [ - "Next.js", - "SvelteKit", - "Nuxt.js", - "Remix", - "Astro", -] as const; - -function FormExample() { - const [notifications, setNotifications] = React.useState({ - email: true, - sms: false, - push: true, - }); - const [theme, setTheme] = React.useState("light"); - - return ( - - - - User Information - Please fill in your details below - - - - - - - - File - - - New File - āŒ˜N - - - - New Folder - ā‡§āŒ˜N - - - - - Open Recent - - - - - Recent Projects - - - Project Alpha - - - - Project Beta - - - - - More Projects - - - - - - Project Gamma - - - - Project Delta - - - - - - - - - - Browse... - - - - - - - - - Save - āŒ˜S - - - - Export - ā‡§āŒ˜E - - - - - View - - setNotifications({ - ...notifications, - email: checked === true, - }) - } - > - - Show Sidebar - - - setNotifications({ - ...notifications, - sms: checked === true, - }) - } - > - - Show Status Bar - - - - - Theme - - - - - Appearance - - - - Light - - - - Dark - - - - System - - - - - - - - - - Account - - - Profile - ā‡§āŒ˜P - - - - Billing - - - - - Settings - - - - - Preferences - - - Keyboard Shortcuts - - - - Language - - - - - Notifications - - - - - - Notification Types - - - setNotifications({ - ...notifications, - push: checked === true, - }) - } - > - - Push Notifications - - - setNotifications({ - ...notifications, - email: checked === true, - }) - } - > - - Email Notifications - - - - - - - - - - - Privacy & Security - - - - - - - - - - - Help & Support - - - - Documentation - - - - - - - Sign Out - ā‡§āŒ˜Q - - - - - - - -
- -
- - Name - - - - Role - - -
- - - Framework - - - - - No frameworks found. - - {(item) => ( - - {item} - - )} - - - - - - Comments -