#!/usr/bin/env python3 """ Gradio Interface for SmolLM3/GPT-OSS Fine-tuning Pipeline This app mirrors the core flow of launch.sh with a click-and-run UI. Tokens are read from environment variables: - HF_WRITE_TOKEN (required) - HF_READ_TOKEN (optional; used to switch the Trackio Space token after training) Key steps (configurable via UI): 1) Optional HF Dataset repo setup for Trackio 2) Optional Trackio Space deployment 3) Training (SmolLM3 or GPT-OSS) 4) Push trained model to the HF Hub 5) Optional switch Trackio HF_TOKEN to read token This uses the existing scripts in scripts/ and config/ to avoid code duplication. """ from __future__ import annotations import os import sys import time import json import shlex import traceback import importlib.util from dataclasses import dataclass from datetime import datetime from pathlib import Path from typing import Dict, Any, Generator, Optional, Tuple # Third-party try: import gradio as gr # type: ignore except Exception as _e: raise RuntimeError( "Gradio is required. Please install it first: pip install gradio" ) from _e # -------------------------------------------------------------------------------------- # Utilities # -------------------------------------------------------------------------------------- PROJECT_ROOT = Path(__file__).resolve().parent def mask_token(token: Optional[str]) -> str: if not token: return "" token = str(token) if len(token) <= 8: return "*" * len(token) return f"{token[:4]}****{token[-4:]}" def get_python() -> str: return sys.executable or "python" def get_username_from_token(token: str) -> Optional[str]: try: from huggingface_hub import HfApi # type: ignore api = HfApi(token=token) info = api.whoami() if isinstance(info, dict): return info.get("name") or info.get("username") if isinstance(info, str): return info except Exception: return None return None def detect_nvidia_driver() -> Tuple[bool, str]: """Detect NVIDIA driver/GPU presence with multiple strategies. Returns (available, human_message). """ # 1) Try torch CUDA try: import torch # type: ignore if torch.cuda.is_available(): try: num = torch.cuda.device_count() names = [torch.cuda.get_device_name(i) for i in range(num)] return True, f"NVIDIA GPU detected: {', '.join(names)}" except Exception: return True, "NVIDIA GPU detected (torch.cuda available)" except Exception: pass # 2) Try NVML via pynvml try: import pynvml # type: ignore try: pynvml.nvmlInit() cnt = pynvml.nvmlDeviceGetCount() names = [] for i in range(cnt): h = pynvml.nvmlDeviceGetHandleByIndex(i) names.append(pynvml.nvmlDeviceGetName(h).decode("utf-8", errors="ignore")) drv = pynvml.nvmlSystemGetDriverVersion().decode("utf-8", errors="ignore") pynvml.nvmlShutdown() if cnt > 0: return True, f"NVIDIA driver {drv}; GPUs: {', '.join(names)}" except Exception: pass except Exception: pass # 3) Try nvidia-smi try: import subprocess res = subprocess.run(["nvidia-smi", "-L"], capture_output=True, text=True, timeout=3) if res.returncode == 0 and res.stdout.strip(): return True, res.stdout.strip().splitlines()[0] except Exception: pass return False, "No NVIDIA driver/GPU detected" def duplicate_space_hint() -> str: space_id = os.environ.get("SPACE_ID") or os.environ.get("HF_SPACE_ID") if space_id: space_url = f"https://huggingface.co/spaces/{space_id}" dup_url = f"{space_url}?duplicate=true" return ( f"ℹ️ No NVIDIA driver detected. If you're on Hugging Face Spaces, " f"please duplicate this Space to GPU hardware: [Duplicate this Space]({dup_url})." ) return ( "ℹ️ No NVIDIA driver detected. To enable training, run on a machine with an NVIDIA GPU/driver " "or duplicate this Space on Hugging Face with GPU hardware." ) def _write_generated_config(filename: str, content: str) -> Path: """Write a generated config under config/ and return the full path.""" cfg_dir = PROJECT_ROOT / "config" cfg_dir.mkdir(parents=True, exist_ok=True) path = cfg_dir / filename with open(path, "w", encoding="utf-8") as f: f.write(content) return path def generate_medical_o1_config_file( dataset_config: str, system_message: Optional[str], developer_message: Optional[str], num_train_epochs: float, batch_size: int, gradient_accumulation_steps: int, learning_rate: float, max_seq_length: int, ) -> Path: """Create a GPT-OSS Medical o1 SFT config file from user inputs.""" # Sanitize quotes in messages def _q(s: Optional[str]) -> str: if s is None or s == "": return "None" return repr(s) py = f""" from config.train_gpt_oss_custom import GPTOSSEnhancedCustomConfig config = GPTOSSEnhancedCustomConfig( dataset_name="FreedomIntelligence/medical-o1-reasoning-SFT", dataset_config={repr(dataset_config)}, dataset_split="train", dataset_format="medical_o1_sft", # Field mapping and prefixes input_field="Question", target_field="Response", question_field="Question", reasoning_field="Complex_CoT", response_field="Response", reason_prefix="Reasoning: ", answer_prefix="Final Answer: ", # Optional context system_message={_q(system_message)}, developer_message={_q(developer_message)}, # Training hyperparameters num_train_epochs={num_train_epochs}, batch_size={batch_size}, gradient_accumulation_steps={gradient_accumulation_steps}, learning_rate={learning_rate}, min_lr=2e-5, weight_decay=0.01, warmup_ratio=0.03, # Sequence length max_seq_length={max_seq_length}, # Precision & performance fp16=False, bf16=True, dataloader_num_workers=4, dataloader_pin_memory=True, dataloader_prefetch_factor=2, group_by_length=True, remove_unused_columns=True, # LoRA & quantization use_lora=True, lora_config={ "r": 16, "lora_alpha": 32, "lora_dropout": 0.05, "target_modules": "all-linear", "target_parameters": [ "7.mlp.experts.gate_up_proj", "7.mlp.experts.down_proj", "15.mlp.experts.gate_up_proj", "15.mlp.experts.down_proj", "23.mlp.experts.gate_up_proj", "23.mlp.experts.down_proj", ], "bias": "none", "task_type": "CAUSAL_LM", }, use_quantization=True, quantization_config={ "dequantize": True, "load_in_4bit": False, }, # Logging & evaluation eval_strategy="steps", eval_steps=100, logging_steps=10, save_strategy="steps", save_steps=500, save_total_limit=3, metric_for_best_model="eval_loss", greater_is_better=False, ) """ return _write_generated_config("_generated_gpt_oss_medical_o1_sft.py", py) def generate_gpt_oss_custom_config_file( dataset_name: str, dataset_split: str, dataset_format: str, input_field: str, target_field: Optional[str], system_message: Optional[str], developer_message: Optional[str], model_identity: Optional[str], max_samples: Optional[int], min_length: int, max_length: Optional[int], num_train_epochs: float, batch_size: int, gradient_accumulation_steps: int, learning_rate: float, min_lr: float, weight_decay: float, warmup_ratio: float, max_seq_length: int, lora_r: int, lora_alpha: int, lora_dropout: float, mixed_precision: str, # "bf16"|"fp16"|"fp32" num_workers: int, quantization_type: str, # "mxfp4"|"bnb4"|"none" max_grad_norm: float, logging_steps: int, eval_steps: int, save_steps: int, ) -> Path: # Precision flags if mixed_precision.lower() == "bf16": fp16_flag = False bf16_flag = True elif mixed_precision.lower() == "fp16": fp16_flag = True bf16_flag = False else: fp16_flag = False bf16_flag = False # Quantization flags/config if quantization_type == "mxfp4": use_quant = True quant_cfg = '{"dequantize": True, "load_in_4bit": False}' elif quantization_type == "bnb4": use_quant = True quant_cfg = '{"dequantize": False, "load_in_4bit": True, "bnb_4bit_compute_dtype": "bfloat16", "bnb_4bit_use_double_quant": True, "bnb_4bit_quant_type": "nf4"}' else: use_quant = False quant_cfg = '{"dequantize": False, "load_in_4bit": False}' def _q(s: Optional[str]) -> str: if s is None or s == "": return "None" return repr(s) py = f""" from config.train_gpt_oss_custom import GPTOSSEnhancedCustomConfig config = GPTOSSEnhancedCustomConfig( # Dataset dataset_name={repr(dataset_name)}, dataset_split={repr(dataset_split)}, dataset_format={repr(dataset_format)}, input_field={repr(input_field)}, target_field={repr(target_field)} if {repr(target_field)} != 'None' else None, system_message={_q(system_message)}, developer_message={_q(developer_message)}, max_samples={repr(max_samples)} if {repr(max_samples)} != 'None' else None, min_length={min_length}, max_length={repr(max_length)} if {repr(max_length)} != 'None' else None, # Training hyperparameters num_train_epochs={num_train_epochs}, batch_size={batch_size}, gradient_accumulation_steps={gradient_accumulation_steps}, learning_rate={learning_rate}, min_lr={min_lr}, weight_decay={weight_decay}, warmup_ratio={warmup_ratio}, max_grad_norm={max_grad_norm}, # Model max_seq_length={max_seq_length}, # Precision fp16={str(fp16_flag)}, bf16={str(bf16_flag)}, # LoRA lora_config={{ "r": {lora_r}, "lora_alpha": {lora_alpha}, "lora_dropout": {lora_dropout}, "target_modules": "all-linear", "bias": "none", "task_type": "CAUSAL_LM", }}, # Quantization use_quantization={str(use_quant)}, quantization_config={quant_cfg}, # Performance dataloader_num_workers={num_workers}, dataloader_pin_memory=True, group_by_length=True, # Logging & eval logging_steps={logging_steps}, eval_steps={eval_steps}, save_steps={save_steps}, # Chat template (Harmony) chat_template_kwargs={{ "add_generation_prompt": True, "tokenize": False, "auto_insert_role": True, "reasoning_effort": "medium", "model_identity": {_q(model_identity) if _q(model_identity) != 'None' else repr('You are GPT-Tonic, a large language model trained by TonicAI.')}, "builtin_tools": [], }}, ) """ return _write_generated_config("_generated_gpt_oss_custom.py", py) def generate_smollm3_custom_config_file( model_name: str, dataset_name: Optional[str], max_seq_length: int, batch_size: int, gradient_accumulation_steps: int, learning_rate: float, save_steps: int, eval_steps: int, logging_steps: int, filter_bad_entries: bool, input_field: str, target_field: str, sample_size: Optional[int], sample_seed: int, trainer_type: str, ) -> Path: # Create subclass to include dataset fields similar to other configs def _bool(b: bool) -> str: return "True" if b else "False" ds_section = """ # HF Dataset configuration dataset_name={} dataset_split="train" input_field={} target_field={} filter_bad_entries={} bad_entry_field="bad_entry" sample_size={} sample_seed={} """.format( repr(dataset_name) if dataset_name else "None", repr(input_field), repr(target_field), _bool(filter_bad_entries), repr(sample_size) if sample_size is not None else "None", sample_seed, ) py = f""" from dataclasses import dataclass from typing import Optional from config.train_smollm3 import SmolLM3Config @dataclass class SmolLM3GeneratedConfig(SmolLM3Config): {ds_section} config = SmolLM3GeneratedConfig( trainer_type={repr(trainer_type.lower())}, model_name={repr(model_name)}, max_seq_length={max_seq_length}, use_flash_attention=True, use_gradient_checkpointing=True, batch_size={batch_size}, gradient_accumulation_steps={gradient_accumulation_steps}, learning_rate={learning_rate}, weight_decay=0.01, warmup_steps=100, max_iters=None, eval_interval={eval_steps}, log_interval={logging_steps}, save_interval={save_steps}, optimizer="adamw", beta1=0.9, beta2=0.95, eps=1e-8, scheduler="cosine", min_lr=1e-6, fp16=True, bf16=False, save_steps={save_steps}, eval_steps={eval_steps}, logging_steps={logging_steps}, save_total_limit=3, eval_strategy="steps", metric_for_best_model="eval_loss", greater_is_better=False, load_best_model_at_end=True, ) """ return _write_generated_config("_generated_smollm3_custom.py", py) def ensure_dataset_repo(username: str, dataset_name: str, token: str) -> Tuple[str, bool, str]: """Create or ensure dataset repo exists. Returns (repo_id, created_or_exists, message).""" from huggingface_hub import create_repo # type: ignore repo_id = f"{username}/{dataset_name}" try: create_repo(repo_id=repo_id, repo_type="dataset", token=token, exist_ok=True, private=False) return repo_id, True, f"Dataset repo ready: {repo_id}" except Exception as e: return repo_id, False, f"Failed to create dataset repo {repo_id}: {e}" def import_config_object(config_path: Path) -> Optional[Any]: """Import a config file and return its 'config' object if present, else None.""" try: spec = importlib.util.spec_from_file_location("config_module", str(config_path)) if not spec or not spec.loader: return None module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) # type: ignore if hasattr(module, "config"): return getattr(module, "config") return None except Exception: return None def run_command_stream(args: list[str], env: Dict[str, str], cwd: Optional[Path] = None) -> Generator[str, None, int]: """Run a command and yield stdout/stderr lines as they arrive. Returns exit code at the end.""" import subprocess yield f"$ {' '.join(shlex.quote(a) for a in ([get_python()] + args))}" process = subprocess.Popen( [get_python()] + args, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, env=env, cwd=str(cwd or PROJECT_ROOT), bufsize=1, universal_newlines=True, ) assert process.stdout is not None for line in iter(process.stdout.readline, ""): yield line.rstrip() process.stdout.close() code = process.wait() yield f"[exit_code={code}]" return code # -------------------------------------------------------------------------------------- # Configuration Mappings (mirror launch.sh) # -------------------------------------------------------------------------------------- SMOL_CONFIGS = { "Basic Training": { "config_file": "config/train_smollm3.py", "default_model": "HuggingFaceTB/SmolLM3-3B", }, "H100 Lightweight (Rapid)": { "config_file": "config/train_smollm3_h100_lightweight.py", "default_model": "HuggingFaceTB/SmolLM3-3B", }, "A100 Large Scale": { "config_file": "config/train_smollm3_openhermes_fr_a100_large.py", "default_model": "HuggingFaceTB/SmolLM3-3B", }, "Multiple Passes": { "config_file": "config/train_smollm3_openhermes_fr_a100_multiple_passes.py", "default_model": "HuggingFaceTB/SmolLM3-3B", }, } GPT_OSS_CONFIGS = { "GPT-OSS Basic Training": { "config_file": "config/train_gpt_oss_basic.py", "default_model": "openai/gpt-oss-20b", }, "GPT-OSS H100 Optimized": { "config_file": "config/train_gpt_oss_h100_optimized.py", "default_model": "openai/gpt-oss-20b", }, "GPT-OSS Multilingual Reasoning": { "config_file": "config/train_gpt_oss_multilingual_reasoning.py", "default_model": "openai/gpt-oss-20b", }, "GPT-OSS Memory Optimized": { "config_file": "config/train_gpt_oss_memory_optimized.py", "default_model": "openai/gpt-oss-20b", }, "GPT-OSS OpenHermes-FR (Recommended)": { "config_file": "config/train_gpt_oss_openhermes_fr.py", "default_model": "openai/gpt-oss-20b", }, "GPT-OSS OpenHermes-FR Memory Optimized": { "config_file": "config/train_gpt_oss_openhermes_fr_memory_optimized.py", "default_model": "openai/gpt-oss-20b", }, # Custom dataset and medical SFT can be added later as advanced UI panels } def get_config_map(family: str) -> Dict[str, Dict[str, str]]: return SMOL_CONFIGS if family == "SmolLM3" else GPT_OSS_CONFIGS # -------------------------------------------------------------------------------------- # Pipeline Orchestration # -------------------------------------------------------------------------------------- @dataclass class PipelineInputs: model_family: str config_choice: str trainer_type: str # "SFT" | "DPO" monitoring_mode: str # "both" | "trackio" | "dataset" | "none" experiment_name: str repo_short: str author_name: str model_description: str trackio_space_name: Optional[str] deploy_trackio_space: bool create_dataset_repo: bool push_to_hub: bool switch_to_read_after: bool scheduler_override: Optional[str] min_lr: Optional[float] min_lr_rate: Optional[float] def make_defaults(model_family: str) -> Tuple[str, str]: ts = datetime.now().strftime("%Y%m%d_%H%M%S") family_slug = "gpt-oss" if model_family == "GPT-OSS" else "smollm3" exp = f"smolfactory-{family_slug}_{ts}" repo_short = f"smolfactory-{datetime.now().strftime('%Y%m%d')}" return exp, repo_short def run_pipeline(params: PipelineInputs) -> Generator[str, None, None]: # Tokens from environment write_token = os.environ.get("HF_WRITE_TOKEN") or os.environ.get("HF_TOKEN") read_token = os.environ.get("HF_READ_TOKEN") if not write_token: yield "❌ HF_WRITE_TOKEN (or HF_TOKEN) is not set in the environment." return # Resolve username username = get_username_from_token(write_token) or os.environ.get("HF_USERNAME") if not username: yield "❌ Could not resolve Hugging Face username from token." return yield f"✅ Authenticated as: {username}" # Compute Trackio URL if applicable trackio_url: Optional[str] = None if params.monitoring_mode != "none" and params.trackio_space_name: trackio_url = f"https://huggingface.co/spaces/{username}/{params.trackio_space_name}" yield f"Trackio Space URL: {trackio_url}" # Decide space deploy token per monitoring mode space_deploy_token = write_token if params.monitoring_mode in ("both", "trackio") else (read_token or write_token) # Dataset repo setup dataset_repo = f"{username}/trackio-experiments" if params.create_dataset_repo and params.monitoring_mode != "none": yield f"Creating/ensuring dataset repo exists: {dataset_repo}" rid, ok, msg = ensure_dataset_repo(username, "trackio-experiments", write_token) yield ("✅ " if ok else "⚠️ ") + msg dataset_repo = rid # Resolve config file and model name conf_map = get_config_map(params.model_family) if params.config_choice not in conf_map: yield f"❌ Unknown config choice: {params.config_choice}" return config_file = PROJECT_ROOT / conf_map[params.config_choice]["config_file"] base_model_fallback = conf_map[params.config_choice]["default_model"] if not config_file.exists(): yield f"❌ Config file not found: {config_file}" return cfg_obj = import_config_object(config_file) base_model = getattr(cfg_obj, "model_name", base_model_fallback) if cfg_obj else base_model_fallback dataset_name = getattr(cfg_obj, "dataset_name", None) if cfg_obj else None batch_size = getattr(cfg_obj, "batch_size", None) if cfg_obj else None learning_rate = getattr(cfg_obj, "learning_rate", None) if cfg_obj else None max_seq_length = getattr(cfg_obj, "max_seq_length", None) if cfg_obj else None # Prepare env for subprocesses env = os.environ.copy() env["HF_TOKEN"] = write_token env["HUGGING_FACE_HUB_TOKEN"] = write_token env["HF_USERNAME"] = username env["TRACKIO_DATASET_REPO"] = dataset_repo env["MONITORING_MODE"] = params.monitoring_mode # Optional Trackio Space deployment if params.deploy_trackio_space and params.monitoring_mode != "none" and params.trackio_space_name: yield f"\n=== Deploying Trackio Space: {params.trackio_space_name} ===" # deploy_trackio_space.py expects: space_name, token, git_email, git_name, dataset_repo args = [ str(PROJECT_ROOT / "scripts/trackio_tonic/deploy_trackio_space.py"), params.trackio_space_name, space_deploy_token, f"{username}@users.noreply.hf.co", username, dataset_repo, ] for line in run_command_stream(args, env, cwd=PROJECT_ROOT / "scripts/trackio_tonic"): yield line # Training output directory out_dir = PROJECT_ROOT / "outputs" / f"{params.experiment_name}_{datetime.now().strftime('%Y%m%d_%H%M%S')}" out_dir.mkdir(parents=True, exist_ok=True) yield f"\nOutput directory: {out_dir}" # Scheduler overrides (GPT-OSS only) if params.model_family == "GPT-OSS" and params.scheduler_override: env["GPT_OSS_SCHEDULER"] = params.scheduler_override if params.min_lr is not None: env["GPT_OSS_MIN_LR"] = str(params.min_lr) if params.min_lr_rate is not None: env["GPT_OSS_MIN_LR_RATE"] = str(params.min_lr_rate) # Start training yield f"\n=== Starting Training ({params.model_family}) ===" if params.model_family == "GPT-OSS": args = [ str(PROJECT_ROOT / "scripts/training/train_gpt_oss.py"), "--config", str(config_file), "--experiment-name", params.experiment_name, "--output-dir", str(out_dir), "--trackio-url", trackio_url or "", "--trainer-type", params.trainer_type.lower(), ] else: args = [ str(PROJECT_ROOT / "scripts/training/train.py"), "--config", str(config_file), "--experiment-name", params.experiment_name, "--output-dir", str(out_dir), "--trackio-url", trackio_url or "", "--trainer-type", params.trainer_type.lower(), ] # Stream training logs train_failed = False for line in run_command_stream(args, env): yield line if line.strip().startswith("[exit_code=") and not line.strip().endswith("[exit_code=0]"): train_failed = True if train_failed: yield "❌ Training failed. Aborting remaining steps." return # Push to Hub if params.push_to_hub: yield "\n=== Pushing Model to Hugging Face Hub ===" repo_name = f"{username}/{params.repo_short}" if params.model_family == "GPT-OSS": push_args = [ str(PROJECT_ROOT / "scripts/model_tonic/push_gpt_oss_to_huggingface.py"), str(out_dir), repo_name, "--token", write_token, "--trackio-url", trackio_url or "", "--experiment-name", params.experiment_name, "--dataset-repo", dataset_repo, "--author-name", params.author_name or username, "--model-description", params.model_description, "--training-config-type", params.config_choice, "--model-name", base_model, ] if dataset_name: push_args += ["--dataset-name", str(dataset_name)] if batch_size is not None: push_args += ["--batch-size", str(batch_size)] if learning_rate is not None: push_args += ["--learning-rate", str(learning_rate)] if max_seq_length is not None: push_args += ["--max-seq-length", str(max_seq_length)] push_args += ["--trainer-type", params.trainer_type] else: push_args = [ str(PROJECT_ROOT / "scripts/model_tonic/push_to_huggingface.py"), str(out_dir), repo_name, "--token", write_token, "--trackio-url", trackio_url or "", "--experiment-name", params.experiment_name, "--dataset-repo", dataset_repo, "--author-name", params.author_name or username, "--model-description", params.model_description, "--training-config-type", params.config_choice, "--model-name", base_model, ] if dataset_name: push_args += ["--dataset-name", str(dataset_name)] if batch_size is not None: push_args += ["--batch-size", str(batch_size)] if learning_rate is not None: push_args += ["--learning-rate", str(learning_rate)] if max_seq_length is not None: push_args += ["--max-seq-length", str(max_seq_length)] push_args += ["--trainer-type", params.trainer_type] for line in run_command_stream(push_args, env): yield line # Switch Space token to read-only (security) if params.switch_to_read_after and params.monitoring_mode in ("both", "trackio") and params.trackio_space_name and read_token: yield "\n=== Switching Trackio Space HF_TOKEN to READ token ===" space_id = f"{username}/{params.trackio_space_name}" sw_args = [ str(PROJECT_ROOT / "scripts/trackio_tonic/switch_to_read_token.py"), space_id, read_token, write_token, ] for line in run_command_stream(sw_args, env, cwd=PROJECT_ROOT / "scripts/trackio_tonic"): yield line elif params.switch_to_read_after and not read_token: yield "⚠️ HF_READ_TOKEN not set; skipping token switch." # Final summary yield "\n🎉 Pipeline completed." if params.monitoring_mode != "none" and trackio_url: yield f"Trackio: {trackio_url}" yield f"Model repo (if pushed): https://huggingface.co/{username}/{params.repo_short}" yield f"Outputs: {out_dir}" # -------------------------------------------------------------------------------------- # Gradio UI # -------------------------------------------------------------------------------------- MODEL_FAMILIES = ["SmolLM3", "GPT-OSS"] TRAINER_CHOICES = ["SFT", "DPO"] MONITORING_CHOICES = ["both", "trackio", "dataset", "none"] SCHEDULER_CHOICES = [None, "linear", "cosine", "cosine_with_min_lr", "constant"] def ui_defaults(family: str) -> Tuple[str, str, str, str]: exp, repo_short = make_defaults(family) default_desc = ( "A fine-tuned GPT-OSS-20B model optimized for multilingual reasoning and instruction following." if family == "GPT-OSS" else "A fine-tuned SmolLM3-3B model optimized for instruction following and French language tasks." ) trackio_space_name = f"trackio-monitoring-{datetime.now().strftime('%Y%m%d')}" return exp, repo_short, default_desc, trackio_space_name def on_family_change(family: str) -> Tuple[list[str], str, str, str, str]: confs = list(get_config_map(family).keys()) exp, repo_short, desc, space = ui_defaults(family) return confs, confs[0] if confs else "", exp, repo_short, desc def start_pipeline( model_family: str, config_choice: str, trainer_type: str, monitoring_mode: str, experiment_name: str, repo_short: str, author_name: str, model_description: str, trackio_space_name: str, deploy_trackio_space: bool, create_dataset_repo: bool, push_to_hub: bool, switch_to_read_after: bool, scheduler_override: Optional[str], min_lr: Optional[float], min_lr_rate: Optional[float], ) -> Generator[str, None, None]: try: params = PipelineInputs( model_family=model_family, config_choice=config_choice, trainer_type=trainer_type, monitoring_mode=monitoring_mode, experiment_name=experiment_name, repo_short=repo_short, author_name=author_name, model_description=model_description, trackio_space_name=trackio_space_name or None, deploy_trackio_space=deploy_trackio_space, create_dataset_repo=create_dataset_repo, push_to_hub=push_to_hub, switch_to_read_after=switch_to_read_after, scheduler_override=(scheduler_override or None), min_lr=min_lr, min_lr_rate=min_lr_rate, ) # Show token presence write_token = os.environ.get("HF_WRITE_TOKEN") or os.environ.get("HF_TOKEN") read_token = os.environ.get("HF_READ_TOKEN") yield f"HF_WRITE_TOKEN: {mask_token(write_token)}" yield f"HF_READ_TOKEN: {mask_token(read_token)}" # Run the orchestrated pipeline for line in run_pipeline(params): yield line # Small delay for smoother streaming time.sleep(0.01) except Exception as e: yield f"❌ Error: {e}" tb = traceback.format_exc(limit=2) yield tb with gr.Blocks(title="SmolLM3 / GPT-OSS Fine-tuning Pipeline") as demo: # GPU/driver detection banner has_gpu, gpu_msg = detect_nvidia_driver() if has_gpu: gr.Markdown(f""" **SmolLM3 / GPT-OSS Fine-tuning Pipeline** - {gpu_msg} — training is available on this runtime. - Reads tokens from environment: `HF_WRITE_TOKEN` (required), `HF_READ_TOKEN` (optional) - Select a config and run training; optionally deploy Trackio and push to Hub """) else: gr.Markdown(f""" **SmolLM3 / GPT-OSS Fine-tuning Pipeline** - {duplicate_space_hint()} - Reads tokens from environment: `HF_WRITE_TOKEN` (required), `HF_READ_TOKEN` (optional) - You can still configure and push, but training requires a GPU runtime. """) with gr.Row(): model_family = gr.Dropdown(choices=MODEL_FAMILIES, value="SmolLM3", label="Model family") trainer_type = gr.Radio(choices=TRAINER_CHOICES, value="SFT", label="Trainer type") monitoring_mode = gr.Dropdown(choices=MONITORING_CHOICES, value="both", label="Monitoring mode") config_choice = gr.Dropdown(choices=list(get_config_map("SmolLM3").keys()), value="Basic Training", label="Training configuration") exp_default, repo_default, desc_default, trackio_space_default = ui_defaults("SmolLM3") with gr.Row(): experiment_name = gr.Textbox(value=exp_default, label="Experiment name") repo_short = gr.Textbox(value=repo_default, label="Model repo (short name)") with gr.Row(): author_name = gr.Textbox(value=os.environ.get("HF_USERNAME", ""), label="Author name") model_description = gr.Textbox(value=desc_default, label="Model description") with gr.Row(): trackio_space_name = gr.Textbox(value=trackio_space_default, label="Trackio Space name (used when monitoring != none)") deploy_trackio_space = gr.Checkbox(value=True, label="Deploy Trackio Space") create_dataset_repo = gr.Checkbox(value=True, label="Create/ensure HF Dataset repo") with gr.Row(): push_to_hub = gr.Checkbox(value=True, label="Push model to Hugging Face Hub") switch_to_read_after = gr.Checkbox(value=True, label="Switch Space token to READ after training") with gr.Tabs(): with gr.Tab("Run"): with gr.Row(): model_family = gr.Dropdown(choices=MODEL_FAMILIES, value="SmolLM3", label="Model family") trainer_type = gr.Radio(choices=TRAINER_CHOICES, value="SFT", label="Trainer type") monitoring_mode = gr.Dropdown(choices=MONITORING_CHOICES, value="both", label="Monitoring mode") config_choice = gr.Dropdown(choices=list(get_config_map("SmolLM3").keys()), value="Basic Training", label="Training configuration") exp_default, repo_default, desc_default, trackio_space_default = ui_defaults("SmolLM3") with gr.Row(): experiment_name = gr.Textbox(value=exp_default, label="Experiment name") repo_short = gr.Textbox(value=repo_default, label="Model repo (short name)") with gr.Row(): author_name = gr.Textbox(value=os.environ.get("HF_USERNAME", ""), label="Author name") model_description = gr.Textbox(value=desc_default, label="Model description") with gr.Row(): trackio_space_name = gr.Textbox(value=trackio_space_default, label="Trackio Space name (used when monitoring != none)") deploy_trackio_space = gr.Checkbox(value=True, label="Deploy Trackio Space") create_dataset_repo = gr.Checkbox(value=True, label="Create/ensure HF Dataset repo") with gr.Row(): push_to_hub = gr.Checkbox(value=True, label="Push model to Hugging Face Hub") switch_to_read_after = gr.Checkbox(value=True, label="Switch Space token to READ after training") gr.Markdown("### Medical SFT (GPT-OSS o1)") gr.Markdown("Configure GPT-OSS Medical o1 SFT (FreedomIntelligence/medical-o1-reasoning-SFT)") med_dataset_config = gr.Dropdown(choices=["en", "en_mix", "zh", "zh_mix"], value="en", label="Dataset config") med_system = gr.Textbox(value="You are GPT-Tonic, a large language model trained by TonicAI.", label="System message", lines=2) med_developer = gr.Textbox(value="You are are GPT-Tonic, an intelligent assistant that always answers health-related queries scientifically.", label="Developer message", lines=3) with gr.Row(): med_epochs = gr.Number(value=2.0, precision=2, label="Epochs") med_bs = gr.Number(value=4, precision=0, label="Batch size") med_gas = gr.Number(value=4, precision=0, label="Grad accumulation") med_lr = gr.Number(value=2e-4, precision=6, label="Learning rate") med_msl = gr.Number(value=2048, precision=0, label="Max seq length") med_generate = gr.Button("Generate Medical Config") med_status = gr.Textbox(label="Generated config path", interactive=False) logs = gr.Textbox(value="", label="Logs", lines=20) start_btn = gr.Button("Run Pipeline") with gr.Tab("Advanced Config"): with gr.Accordion("GPT-OSS Scheduler Overrides", open=False): scheduler_override = gr.Dropdown(choices=[c for c in SCHEDULER_CHOICES if c is not None], value=None, allow_custom_value=True, label="Scheduler override") min_lr = gr.Number(value=None, precision=6, label="min_lr (when cosine_with_min_lr)") min_lr_rate = gr.Number(value=None, precision=6, label="min_lr_rate (when cosine_with_min_lr)") gr.Markdown("### GPT-OSS Custom Dataset") with gr.Row(): cds_dataset = gr.Textbox(value="legmlai/openhermes-fr", label="Dataset name") cds_split = gr.Textbox(value="train", label="Split") cds_format = gr.Dropdown(choices=["openhermes_fr", "messages", "text", "medical_o1_sft", "custom", "preference"], value="openhermes_fr", label="Format") with gr.Row(): cds_input = gr.Textbox(value="prompt", label="Input field") cds_target = gr.Textbox(value="accepted_completion", label="Target field (optional, blank for None)") with gr.Row(): cds_sys = gr.Textbox(value="", label="System message (optional)") cds_dev = gr.Textbox(value="", label="Developer message (optional)") with gr.Row(): cds_identity = gr.Textbox(value="You are GPT-Tonic, a large language model trained by TonicAI.", label="Model identity (chat_template_kwargs.model_identity)") with gr.Row(): cds_max_samples = gr.Number(value=None, precision=0, label="Max samples (optional)") cds_min_len = gr.Number(value=10, precision=0, label="Min length") cds_max_len = gr.Number(value=None, precision=0, label="Max length (optional)") gr.Markdown("#### Training Hyperparameters") with gr.Row(): cds_epochs = gr.Number(value=1.0, precision=2, label="Epochs") cds_bs = gr.Number(value=4, precision=0, label="Batch size") cds_gas = gr.Number(value=4, precision=0, label="Grad accumulation") cds_lr = gr.Number(value=2e-4, precision=6, label="Learning rate") cds_minlr = gr.Number(value=2e-5, precision=6, label="Min LR") with gr.Row(): cds_wd = gr.Number(value=0.01, precision=6, label="Weight decay") cds_warm = gr.Number(value=0.03, precision=6, label="Warmup ratio") cds_msl = gr.Number(value=2048, precision=0, label="Max seq length") gr.Markdown("#### LoRA / Precision / Quantization / Perf") with gr.Row(): cds_lora_r = gr.Number(value=16, precision=0, label="LoRA r") cds_lora_alpha = gr.Number(value=32, precision=0, label="LoRA alpha") cds_lora_dropout = gr.Number(value=0.05, precision=4, label="LoRA dropout") with gr.Row(): cds_precision = gr.Dropdown(choices=["bf16", "fp16", "fp32"], value="bf16", label="Mixed precision") cds_workers = gr.Number(value=4, precision=0, label="Data workers") cds_quant = gr.Dropdown(choices=["mxfp4", "bnb4", "none"], value="mxfp4", label="Quantization") with gr.Row(): cds_mgn = gr.Number(value=1.0, precision=4, label="Max grad norm") cds_log_steps = gr.Number(value=10, precision=0, label="Logging steps") cds_eval_steps = gr.Number(value=100, precision=0, label="Eval steps") cds_save_steps = gr.Number(value=500, precision=0, label="Save steps") cds_generate = gr.Button("Generate GPT-OSS Custom Config") cds_status = gr.Textbox(label="Generated config path", interactive=False) gr.Markdown("### SmolLM3 Custom Configuration") with gr.Row(): sm_model = gr.Textbox(value="HuggingFaceTB/SmolLM3-3B", label="Model name") sm_dataset = gr.Textbox(value="legmlai/openhermes-fr", label="Dataset (optional; leave blank for local)") with gr.Row(): sm_msl = gr.Number(value=4096, precision=0, label="Max seq length") sm_bs = gr.Number(value=2, precision=0, label="Batch size") sm_gas = gr.Number(value=8, precision=0, label="Grad accumulation") sm_lr = gr.Number(value=5e-6, precision=8, label="Learning rate") with gr.Row(): sm_save = gr.Number(value=500, precision=0, label="Save steps") sm_eval = gr.Number(value=100, precision=0, label="Eval steps") sm_log = gr.Number(value=10, precision=0, label="Logging steps") with gr.Row(): sm_filter = gr.Checkbox(value=False, label="Filter bad entries") sm_in = gr.Textbox(value="prompt", label="Input field") sm_out = gr.Textbox(value="accepted_completion", label="Target field") with gr.Row(): sm_sample = gr.Number(value=None, precision=0, label="Sample size (optional)") sm_seed = gr.Number(value=42, precision=0, label="Sample seed") sm_trainer = gr.Dropdown(choices=["SFT", "DPO"], value="SFT", label="Trainer type") sm_generate = gr.Button("Generate SmolLM3 Custom Config") sm_status = gr.Textbox(label="Generated config path", interactive=False) logs = gr.Textbox(value="", label="Logs", lines=20) start_btn = gr.Button("Run Pipeline") # Events model_family.change(on_family_change, inputs=model_family, outputs=[config_choice, config_choice, experiment_name, repo_short, model_description]) # Generate config handlers med_generate.click( lambda dc, sysm, devm, ep, bs, gas, lr, msl: str( generate_medical_o1_config_file( dataset_config=dc, system_message=sysm, developer_message=devm, num_train_epochs=float(ep or 2.0), batch_size=int(bs or 4), gradient_accumulation_steps=int(gas or 4), learning_rate=float(lr or 2e-4), max_seq_length=int(msl or 2048), ) ), inputs=[med_dataset_config, med_system, med_developer, med_epochs, med_bs, med_gas, med_lr, med_msl], outputs=[med_status], ) cds_generate.click( lambda dname, dsplit, dformat, ifld, tfld, sm, dm, ident, ms, minl, maxl, ep, bs, gas, lr, minlr, wd, warm, msl, lr_, la, ld, prec, nw, q, mgn, logst, evst, savst: str( generate_gpt_oss_custom_config_file( dataset_name=dname, dataset_split=dsplit, dataset_format=dformat, input_field=ifld, target_field=(tfld or None), system_message=sm, developer_message=dm, model_identity=ident, max_samples=(int(ms) if ms is not None else None), min_length=int(minl or 10), max_length=(int(maxl) if maxl is not None else None), num_train_epochs=float(ep or 1.0), batch_size=int(bs or 4), gradient_accumulation_steps=int(gas or 4), learning_rate=float(lr or 2e-4), min_lr=float(minlr or 2e-5), weight_decay=float(wd or 0.01), warmup_ratio=float(warm or 0.03), max_seq_length=int(msl or 2048), lora_r=int(lr_), lora_alpha=int(la), lora_dropout=float(ld), mixed_precision=prec, num_workers=int(nw or 4), quantization_type=q, max_grad_norm=float(mgn or 1.0), logging_steps=int(logst or 10), eval_steps=int(evst or 100), save_steps=int(savst or 500), ) ), inputs=[ cds_dataset, cds_split, cds_format, cds_input, cds_target, cds_sys, cds_dev, cds_identity, cds_max_samples, cds_min_len, cds_max_len, cds_epochs, cds_bs, cds_gas, cds_lr, cds_minlr, cds_wd, cds_warm, cds_msl, cds_lora_r, cds_lora_alpha, cds_lora_dropout, cds_precision, cds_workers, cds_quant, cds_mgn, cds_log_steps, cds_eval_steps, cds_save_steps ], outputs=[cds_status], ) sm_generate.click( lambda mn, dn, msl, bs, gas, lr, sst, est, lst, fbe, ifld, tfld, ss, seed, tt: str( generate_smollm3_custom_config_file( model_name=mn, dataset_name=(dn or None), max_seq_length=int(msl or 4096), batch_size=int(bs or 2), gradient_accumulation_steps=int(gas or 8), learning_rate=float(lr or 5e-6), save_steps=int(sst or 500), eval_steps=int(est or 100), logging_steps=int(lst or 10), filter_bad_entries=bool(fbe), input_field=ifld, target_field=tfld, sample_size=(int(ss) if ss is not None else None), sample_seed=int(seed or 42), trainer_type=tt, ) ), inputs=[ sm_model, sm_dataset, sm_msl, sm_bs, sm_gas, sm_lr, sm_save, sm_eval, sm_log, sm_filter, sm_in, sm_out, sm_sample, sm_seed, sm_trainer, ], outputs=[sm_status], ) start_btn.click( start_pipeline, inputs=[ model_family, config_choice, trainer_type, monitoring_mode, experiment_name, repo_short, author_name, model_description, trackio_space_name, deploy_trackio_space, create_dataset_repo, push_to_hub, switch_to_read_after, scheduler_override, min_lr, min_lr_rate, ], outputs=[logs], ) if __name__ == "__main__": # Optional: allow setting server parameters via env server_port = int(os.environ.get("INTERFACE_PORT", "7860")) server_name = os.environ.get("INTERFACE_HOST", "0.0.0.0") demo.queue().launch(server_name=server_name, server_port=server_port)