added training code

app.py

@@ -877,6 +877,32 @@ class LoLLMsWebUI(LoLLMsAPPI):
         return jsonify(models)
 
 
+    def train(self):
+        form_data = request.form
+
+        # Create and populate the config file
+        config = {
+            'model_name': form_data['model_name'],
+            'tokenizer_name': form_data['tokenizer_name'],
+            'dataset_path': form_data['dataset_path'],
+            'max_length': form_data['max_length'],
+            'batch_size': form_data['batch_size'],
+            'lr': form_data['lr'],
+            'num_epochs': form_data['num_epochs'],
+            'output_dir': form_data['output_dir'],
+        }
+
+        with open('train/configs/train/local_cfg.yaml', 'w') as f:
+            yaml.dump(config, f)
+
+        # Trigger the train.py script
+        # Place your code here to run the train.py script with the created config file
+        # accelerate launch --dynamo_backend=inductor --num_processes=8 --num_machines=1 --machine_rank=0 --deepspeed_multinode_launcher standard --mixed_precision=bf16  --use_deepspeed --deepspeed_config_file=configs/deepspeed/ds_config_gptj.json train.py --config configs/train/finetune_gptj.yaml
+
+        subprocess.check_call(["accelerate","launch", "--dynamo_backend=inductor", "--num_processes=8", "--num_machines=1", "--machine_rank=0", "--deepspeed_multinode_launcher standard", "--mixed_precision=bf16", "--use_deepspeed", "--deepspeed_config_file=train/configs/deepspeed/ds_config_gptj.json", "train/train.py", "--config", "train/configs/train/local_cfg.yaml"])
+
+        return jsonify({'message': 'Training started'})
+    
     def get_config(self):
         return jsonify(self.config.to_dict())
 

train/.gitignore

@@ -0,0 +1,2 @@
+output
+!output/.keep

train/configs/deepspeed/ds_config.yaml

@@ -0,0 +1,48 @@
+{
+	"train_batch_size": "auto",
+	"gradient_accumulation_steps": "auto",
+	"train_micro_batch_size_per_gpu": "auto",
+	"fp16": {
+	  "enabled": "auto",
+	  "min_loss_scale": 1,
+	  "loss_scale_window": 1000,
+	  "hysteresis": 2,
+	  "initial_scale_power": 32
+	},
+	"bf16": {
+		"enabled": "auto"
+	},
+	"gradient_clipping": 1,
+	"zero_optimization": {
+	  "stage": 2,
+	  "offload_param": {
+		"device": "none"
+	  },
+	  "offload_optimizer": {
+		"device": "none"
+	  },
+	  "allgather_partitions": true,
+	  "allgather_bucket_size": 5e8,
+	  "contiguous_gradients": true
+	},
+	"optimizer": {
+	  "type": "AdamW",
+	  "params": {
+		"lr": "auto",
+		"betas": [
+		  0.9,
+		  0.999
+		],
+		"eps": 1e-08
+	  }
+	},
+	"scheduler": {
+	  "type": "WarmupLR",
+	  "params": {
+		"warmup_min_lr": 0,
+		"warmup_max_lr": "auto",
+		"warmup_num_steps": "auto",
+		"warmup_type": "linear"
+	  }
+	}
+  }

train/configs/train/.gitignore

@@ -0,0 +1 @@
+local_cfg.yaml

train/configs/train/finetune.yaml

@@ -0,0 +1,29 @@
+# model/tokenizer
+model_name: # add model here
+tokenizer_name: # add model here
+gradient_checkpointing: true
+save_name: # CHANGE 
+
+# dataset
+streaming: false
+num_proc: 64
+dataset_path: # update
+max_length: 1024
+batch_size: 32
+
+# train dynamics
+lr: 5.0e-5
+eval_every: 800
+eval_steps: 100
+save_every: 800
+output_dir: # CHANGE
+checkpoint: null
+lora: false
+warmup_steps: 100
+num_epochs: 2
+
+# logging
+wandb: true
+wandb_entity: # update
+wandb_project_name: # update
+seed: 42

train/configs/train/finetune_lora.yaml

@@ -0,0 +1,31 @@
+# model/tokenizer
+model_name: # update
+tokenizer_name: # update
+gradient_checkpointing: false
+save_name: # CHANGE
+
+# dataset
+streaming: false
+num_proc: 64
+dataset_path: # CHANGE
+max_length: 1024
+batch_size: 4
+
+# train dynamics
+lr: 5.0e-5
+min_lr: 0
+weight_decay: 0.0
+eval_every: 2000
+eval_steps: 100
+save_every: 2000
+output_dir: # CHANGE
+checkpoint: null
+lora: true
+warmup_steps: 100
+num_epochs: 2
+
+# logging
+wandb: true
+wandb_entity: # update
+wandb_project_name: # update
+seed: 42

train/configs/train/finetune_lora_ airoboros-7b-gpt4.yaml

@@ -0,0 +1,31 @@
+# model/tokenizer
+model_name: jondurbin/airoboros-7b-gpt4 # update
+tokenizer_name: jondurbin/airoboros-7b-gpt4 # update
+gradient_checkpointing: false
+save_name: parisneo-7b_gpt42_lora # CHANGE
+
+# dataset
+streaming: false
+num_proc: 64
+dataset_path: # CHANGE
+max_length: 1024
+batch_size: 4
+
+# train dynamics
+lr: 5.0e-5
+min_lr: 0
+weight_decay: 0.0
+eval_every: 2000
+eval_steps: 100
+save_every: 2000
+output_dir: output # CHANGE
+checkpoint: null
+lora: true
+warmup_steps: 100
+num_epochs: 2
+
+# logging
+wandb: false # update if you want to use weights and biases
+wandb_entity: # update
+wandb_project_name: # update
+seed: 42

train/requirements.txt

@@ -0,0 +1,15 @@
+accelerate
+datasets
+torchmetrics
+evaluate
+transformers>=4.28.0
+wandb
+pip
+peft
+nodelist-inflator
+deepspeed
+sentencepiece
+jsonlines
+nomic
+scikit-learn
+matplotlib

train/train.py

@@ -0,0 +1,233 @@
+import os
+from transformers import AutoModelForCausalLM, AutoTokenizer, get_scheduler, LlamaForCausalLM
+import torch
+from torch.optim import AdamW
+from argparse import ArgumentParser
+from read import read_config
+from accelerate import Accelerator
+from accelerate.utils import DummyScheduler, DummyOptim, set_seed
+from peft import get_peft_model, LoraConfig, TaskType
+from data import load_data
+from torchmetrics import MeanMetric
+from tqdm import tqdm
+import wandb
+
+torch.backends.cuda.matmul.allow_tf32 = True
+
+def format_metrics(metrics, split, prefix=""):
+    log = f"[{split}]" + prefix
+    log += " ".join([f"{key}: {value:.4f}" for key, value in metrics.items()])
+
+    return log
+
+
+def evaluate(model, val_dataloader):
+    model.eval()
+    val_loss = MeanMetric(nan_strategy="error").to(model.device)
+
+    with torch.no_grad():
+        for batch in tqdm(val_dataloader):
+            loss = model(**batch).loss
+
+            loss_values = accelerator.gather_for_metrics({"loss": loss.detach()})
+
+            val_loss.update(loss_values["loss"])
+
+    return val_loss
+
+
+def train(accelerator, config):
+    set_seed(config['seed'])
+
+    accelerator.print(config)
+    accelerator.print(f"Using {accelerator.num_processes} GPUs")
+
+    tokenizer = AutoTokenizer.from_pretrained(config['tokenizer_name'], model_max_length=config['max_length'])
+    # if no pad token, set it to eos
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+        
+    with accelerator.main_process_first():
+        train_dataloader, val_dataloader = load_data(config, tokenizer) 
+
+
+    checkpoint = config["gradient_checkpointing"]
+    model = AutoModelForCausalLM.from_pretrained(config["model_name"], 
+                                                    use_cache=False if checkpoint else True,
+                                                    trust_remote_code=True) 
+    if checkpoint:
+        model.gradient_checkpointing_enable()
+
+    if config["lora"]:
+        peft_config = LoraConfig(
+            # should R be configurable?
+            task_type=TaskType.CAUSAL_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1
+        )
+        model = get_peft_model(model, peft_config)
+        model.print_trainable_parameters()
+
+    optimizer_cls = (
+        AdamW
+        if accelerator.state.deepspeed_plugin is None
+        or "optimizer" not in accelerator.state.deepspeed_plugin.deepspeed_config
+        else DummyOptim
+    )
+
+    # karpathy doesn't decay embeddding, maybe we should exclude
+    # https://github.com/karpathy/minGPT/commit/bbbdac74fa9b2e55574d70056163ffbae42310c1#diff-2075fa9c224b395be5bda85544dd36572b59c76c54562819eadadbf268602834R157s
+    optimizer = optimizer_cls(model.parameters(), lr=config["lr"], weight_decay=config["weight_decay"])
+
+    if accelerator.state.deepspeed_plugin is not None:
+        gradient_accumulation_steps = accelerator.state.deepspeed_plugin.deepspeed_config[
+            "gradient_accumulation_steps"
+        ]
+
+    # decay to min_lr instead of 0
+    lr_ratio = config["min_lr"] / config["lr"]
+    accelerator.print(f"Len of train_dataloader: {len(train_dataloader)}")
+    total_num_steps = (len(train_dataloader) / gradient_accumulation_steps) * config["num_epochs"]
+    # instead of decaying to zero, decay to ratio of min_lr / lr
+    total_num_steps += int(total_num_steps * lr_ratio) + config["warmup_steps"]
+    accelerator.print(f"Total training steps: {total_num_steps}")
+
+    # Creates Dummy Scheduler if `scheduler` was specified in the config file else creates `args.lr_scheduler_type` Scheduler
+    if (
+        accelerator.state.deepspeed_plugin is None
+        or "scheduler" not in accelerator.state.deepspeed_plugin.deepspeed_config
+    ):
+        scheduler = get_scheduler(
+            name="cosine",
+            optimizer=optimizer,
+            num_warmup_steps=config["warmup_steps"] * accelerator.num_processes,
+            num_training_steps=total_num_steps,
+        )
+    else:
+        scheduler = DummyScheduler(
+            optimizer, total_num_steps=config["warmup_steps"], warmup_num_steps=config["warmup_steps"]
+        )
+
+    model, optimizer, train_dataloader, val_dataloader, scheduler = accelerator.prepare(
+            model, optimizer, train_dataloader, val_dataloader, scheduler
+    )
+
+    # setup for saving training states in case preemption
+    accelerator.register_for_checkpointing(scheduler)
+
+    if config["checkpoint"]:
+        accelerator.load_state(config["checkpoint"])
+        accelerator.print(f"Resumed from checkpoint: {config['checkpoint']}")
+        path = os.path.basename(config["train_args"]["resume_from_checkpoint"])
+        training_difference = os.path.splitext(path)[0]
+        resume_step = int(training_difference.replace("step_", ""))
+        accelerator.skip_first_batches(train_dataloader, resume_step)
+        accelerator.print(f"Resuming from step {resume_step}")
+
+
+    # log gradients
+    if accelerator.is_main_process and config["wandb"]:
+        wandb.watch(model, log_freq=config["log_grads_every"], log="all")
+
+    for epoch in range(config["num_epochs"]):
+        train_loss = MeanMetric(nan_strategy="error").to(model.device)
+        for step, batch in enumerate(tqdm(train_dataloader)):
+            model.train()
+            outputs = model(**batch)
+            loss = outputs.loss
+
+            # gather loss before backprop in case of gradient accumulation
+            loss_values = accelerator.gather_for_metrics({"loss": loss.detach().float()})
+            train_loss.update(loss_values["loss"])
+
+            loss = loss / gradient_accumulation_steps
+            accelerator.backward(loss)
+            # get gradient norm of all params
+
+            # log LR in case something weird happens 
+            if step > 0 and step % (config["eval_every"] // 10) == 0:
+                if config["wandb"]:
+                    curr_step = step + epoch * len(train_dataloader)
+                    accelerator.log({"lr": scheduler.get_last_lr()[0]}, step=curr_step)
+
+            if (step + 1) % gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
+                optimizer.step()
+                scheduler.step()
+                optimizer.zero_grad()
+
+
+            if step > 0 and step % config["save_every"] == 0:
+                curr_step = step + epoch * len(train_dataloader)
+                accelerator.save_state(f"{config['output_dir']}/step_{curr_step}")
+
+            if step > 0 and (step % config["eval_every"] == 0 or step == len(train_dataloader) - 1):
+                val_loss = evaluate(model, val_dataloader)
+
+                log_train = {
+                        "train_loss": train_loss.compute()
+                    }
+                log_val = {
+                    "val_loss": val_loss.compute()
+                }
+
+                if config["wandb"]:
+                    curr_step = step + epoch * len(train_dataloader)
+                    accelerator.log({**log_train, **log_val}, step=curr_step)
+
+                accelerator.print(f"Current LR: {scheduler.get_last_lr()[0]}")
+                accelerator.print(format_metrics(log_train, "train", f" step {step} "))
+                accelerator.print(format_metrics(log_val, "val", f" step {step} "))
+
+                train_loss.reset()
+
+        accelerator.print(f"Epoch {epoch} finished")
+        accelerator.print(f"Pushing to HF hub")
+        accelerator.wait_for_everyone()
+        unwrapped_model = accelerator.unwrap_model(model)
+        try:
+            if accelerator.is_main_process:
+                unwrapped_model.push_to_hub(config["save_name"] + f"-epoch_{epoch}", private=True)
+
+        except Exception as e:
+            accelerator.print(e)
+            accelerator.print(f"Failed to push to hub")
+
+        unwrapped_model.save_pretrained(
+            f"{config['output_dir']}/epoch_{epoch}",
+            is_main_process=accelerator.is_main_process,
+            save_function=accelerator.save,
+            state_dict=accelerator.get_state_dict(model),
+        )
+            
+    accelerator.wait_for_everyone()
+    unwrapped_model = accelerator.unwrap_model(model)
+    unwrapped_model.save_pretrained(
+        f"{config['output_dir']}/final",
+        is_main_process=accelerator.is_main_process,
+        save_function=accelerator.save,
+        state_dict=accelerator.get_state_dict(model),
+    )
+
+    accelerator.end_training()
+
+    
+
+if __name__ == "__main__":
+    # parse arguments by reading in a config
+    parser = ArgumentParser()
+    parser.add_argument("--config", type=str, default="config.yaml")
+
+    args = parser.parse_args()
+
+    config = read_config(args.config)
+
+    if config["wandb"]:
+        accelerator = Accelerator(log_with="wandb")
+        accelerator.init_trackers(
+            project_name=config["wandb_project_name"],
+            config=config,
+            init_kwargs={"wandb": {"entity": config["wandb_entity"]}},
+        )
+    else:
+        accelerator = Accelerator()
+
+    train(accelerator, config=config)

web/dist/index.html

@@ -6,8 +6,8 @@
     
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <title>GPT4All - WEBUI</title>
-    <script type="module" crossorigin src="/assets/index-0344eb9b.js"></script>
-    <link rel="stylesheet" href="/assets/index-488cca87.css">
+    <script type="module" crossorigin src="/assets/index-f5f472ed.js"></script>
+    <link rel="stylesheet" href="/assets/index-54621153.css">
   </head>
   <body>
     <div id="app"></div>

web/src/views/TrainingView.vue

@@ -1,16 +1,159 @@
 <template>
-    <div>
-        Training
+    <div class="container overflow-y-scroll flex flex-col no-scrollbar shadow-lg p-10 pt-0">
+      <form @submit.prevent="submitForm" class="max-w-md mx-auto">
+        <!-- Model/Tokenizer -->
+        <div class="mb-4">
+          <label for="model_name" class="text-sm">Model Name:</label>
+          <input
+            type="text"
+            id="model_name"
+            v-model="model_name"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+        <div class="mb-4">
+          <label for="tokenizer_name" class="text-sm">Tokenizer Name:</label>
+          <input
+            type="text"
+            id="tokenizer_name"
+            v-model="tokenizer_name"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+  
+        <!-- Dataset -->
+        <div class="mb-4">
+          <label for="dataset_path" class="text-sm">Dataset:</label>
+          <input
+            type="file"
+            id="dataset_path"
+            ref="dataset_path"
+            accept=".parquet"
+            v-on:change="selectDatasetPath"
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+          <p class="mt-2 text-xs">Selected File: {{ selectedDatasetPath }}</p>
+        </div>
+        <div class="mb-4">
+          <label for="max_length" class="text-sm">Max Length:</label>
+          <input
+            type="number"
+            id="max_length"
+            v-model.number="max_length"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+        <div class="mb-4">
+          <label for="batch_size" class="text-sm">Batch Size:</label>
+          <input
+            type="number"
+            id="batch_size"
+            v-model.number="batch_size"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+  
+        <!-- Train Dynamics -->
+        <div class="mb-4">
+          <label for="lr" class="text-sm">Learning Rate:</label>
+          <input
+            type="number"
+            id="lr"
+            v-model.number="lr"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+        <div class="mb-4">
+          <label for="num_epochs" class="text-sm">Number of Epochs:</label>
+          <input
+            type="number"
+            id="num_epochs"
+            v-model.number="num_epochs"
+            required
+            class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          >
+        </div>
+  
+        <!-- Logging -->
+        <div class="mb-4">
+        <label for="output_dir" class="text-sm">Output Directory:</label>
+        <input
+          type="text"
+          id="output_dir"
+          v-model="selectedFolder"
+          class="w-full mt-1 px-2 py-1 border border-gray-300 rounded"
+          placeholder="Enter or select the output folder"
+        >
+        <input
+          type="file"
+          id="folder_selector"
+          ref="folder_selector"
+          style="display: none"
+          webkitdirectory
+          v-on:change="selectOutputDirectory"
+        >
+        <button type="button" @click="openFolderSelector" class="bg-blue-500 text-white px-4 py-2 rounded">Select Folder</button>
+      </div>
+  
+        <button type="submit" class="bg-blue-500 text-white px-4 py-2 rounded">Train LLM</button>
+      </form>
     </div>
-</template>
+  </template>
   
-<script>
-export default {
-    setup () {
+  <script>
+  export default {
+    data() {
+        return {
+            model_name: 'jondurbin/airoboros-7b-gpt4',
+            tokenizer_name: 'jondurbin/airoboros-7b-gpt4',
+            dataset_path: '',
+            max_length: 1024,
+            batch_size: 4,
+            lr: 5.0e-5,
+            num_epochs: 2,
+            selectedFolder: '',
+            selectedDatasetPath: '',
+        };
+    },
+    methods: {
+      submitForm() {
+        const formData = {
+          model_name: this.model_name,
+          tokenizer_name: this.tokenizer_name,
+          dataset_path: this.selectedDatasetPath,
+          max_length: this.max_length,
+          batch_size: this.batch_size,
+          lr: this.lr,
+          num_epochs: this.num_epochs,
+          output_dir: this.selectedFolder,
+        };
   
-
-        return {}
-    }
-}
-</script>
+        // Send the form data to the backend
+        // ...
+      },
+      openFolderSelector() {
+      this.$refs.folder_selector.click();
+      },
+      selectOutputDirectory(event) {
+        console.log("here")
+        const folderPath = event.target.files[0]?.path;
+        console.log(folderPath)
+        if (folderPath) {
+            this.selectedFolder = folderPath;
+        }
+      },
+      selectDatasetPath(event) {
+        const files = event.target.files;
+        if (files.length > 0) {
+            this.selectedDatasetPath = files[0].webkitRelativePath;
+        }
+      },
+    },
+  };
+  </script>