added initial version

Dockerfile

@@ -0,0 +1,10 @@
+FROM python:3.9-slim
+
+WORKDIR /app
+
+COPY requirements.txt .
+RUN pip install -r requirements.txt
+
+COPY src/ .
+
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"] 

gpt-transformers.ipynb

@@ -0,0 +1 @@
+{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":10477901,"sourceType":"datasetVersion","datasetId":6487916}],"dockerImageVersionId":30840,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input/transformers-input/input.txt'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-01-15T14:34:35.524291Z","iopub.execute_input":"2025-01-15T14:34:35.524674Z","iopub.status.idle":"2025-01-15T14:34:36.015800Z","shell.execute_reply.started":"2025-01-15T14:34:35.524644Z","shell.execute_reply":"2025-01-15T14:34:36.014727Z"}},"outputs":[],"execution_count":9},{"cell_type":"code","source":"# Solving for residual std scaling issue\nimport os\nimport math\nimport time\nimport inspect\nfrom dataclasses import dataclass\nimport torch\nimport torch.nn as nn\nfrom torch.nn import functional as F","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:14.295213Z","iopub.execute_input":"2025-01-16T01:11:14.295523Z","iopub.status.idle":"2025-01-16T01:11:17.362144Z","shell.execute_reply.started":"2025-01-16T01:11:14.295499Z","shell.execute_reply":"2025-01-16T01:11:17.361475Z"}},"outputs":[],"execution_count":1},{"cell_type":"code","source":"class CausalSelfAttention(nn.Module):\n\n    def __init__(self, config):\n        super().__init__()\n        assert config.n_embd % config.n_head == 0\n        # key, query, value projections for all heads, but in a batch\n        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)\n        # output projection\n        self.c_proj = nn.Linear(config.n_embd, config.n_embd)\n        self.c_proj.NANGPT_SCALE_INIT = 1\n        # regularization\n        self.n_head = config.n_head\n        self.n_embd = config.n_embd\n        self.register_buffer(\"bias\", torch.tril(torch.ones(config.block_size, config.block_size)).view(1, 1, config.block_size, config.block_size))\n\n    def forward(self, x):\n        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)\n        # calculate query, key, values for all heads in batch and move head forward to be the batch dim\n        # nh is \"number of heads\", hs is \"head size\", and C (number of channels) = nh * hs\n        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer\n        qkv = self.c_attn(x)\n        q, k, v = qkv.split(self.n_embd, dim=2)\n        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n\n        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))\n        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))\n        att = F.softmax(att, dim=-1)\n        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n\n        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side\n        # output projection\n        y = self.c_proj(y)\n        return y\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:17.363161Z","iopub.execute_input":"2025-01-16T01:11:17.363491Z","iopub.status.idle":"2025-01-16T01:11:17.370629Z","shell.execute_reply.started":"2025-01-16T01:11:17.363471Z","shell.execute_reply":"2025-01-16T01:11:17.369710Z"}},"outputs":[],"execution_count":2},{"cell_type":"code","source":"class MLP(nn.Module):\n\n    def __init__(self, config):\n        super().__init__()\n        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd)\n        self.gelu    = nn.GELU(approximate='tanh')\n        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)\n        self.c_proj.NANOGPT_SCALE_INIT = 1\n\n    def forward(self, x):\n        x = self.c_fc(x)\n        x = self.gelu(x)\n        x = self.c_proj(x)\n        return x","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:19.837235Z","iopub.execute_input":"2025-01-16T01:11:19.837640Z","iopub.status.idle":"2025-01-16T01:11:19.843845Z","shell.execute_reply.started":"2025-01-16T01:11:19.837608Z","shell.execute_reply":"2025-01-16T01:11:19.842822Z"}},"outputs":[],"execution_count":3},{"cell_type":"code","source":"class Block(nn.Module):\n\n    def __init__(self, config):\n        super().__init__()\n        self.ln_1 = nn.LayerNorm(config.n_embd)\n        self.attn = CausalSelfAttention(config)\n        self.ln_2 = nn.LayerNorm(config.n_embd)\n        self.mlp = MLP(config)\n\n    def forward(self, x):\n        x = x + self.attn(self.ln_1(x))\n        x = x + self.mlp(self.ln_2(x))\n        return x","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:22.456129Z","iopub.execute_input":"2025-01-16T01:11:22.456449Z","iopub.status.idle":"2025-01-16T01:11:22.461192Z","shell.execute_reply.started":"2025-01-16T01:11:22.456421Z","shell.execute_reply":"2025-01-16T01:11:22.460364Z"}},"outputs":[],"execution_count":4},{"cell_type":"code","source":"@dataclass\nclass GPTConfig:\n    block_size: int = 1024 # max sequence length\n    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token\n    n_layer: int = 12 # number of layers\n    n_head: int = 12 # number of heads\n    n_embd: int = 768 # embedding dimension","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:24.402909Z","iopub.execute_input":"2025-01-16T01:11:24.403223Z","iopub.status.idle":"2025-01-16T01:11:24.407719Z","shell.execute_reply.started":"2025-01-16T01:11:24.403198Z","shell.execute_reply":"2025-01-16T01:11:24.406984Z"}},"outputs":[],"execution_count":5},{"cell_type":"code","source":"class GPT(nn.Module):\n\n    def __init__(self, config):\n        super().__init__()\n        self.config = config\n\n        self.transformer = nn.ModuleDict(dict(\n            wte = nn.Embedding(config.vocab_size, config.n_embd),\n            wpe = nn.Embedding(config.block_size, config.n_embd),\n            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),\n            ln_f = nn.LayerNorm(config.n_embd),\n        ))\n        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)\n\n        # weight sharing\n        self.transformer.wte.weight = self.lm_head.weight\n\n        # weight initialization\n        self.apply(self._init_weights)\n\n    def _init_weights(self, module):\n        if isinstance(module, nn.Linear):\n            std = 0.02\n            if hasattr(module, 'NANGPT_SCALE_INIT'):\n                std *= (2 * self.config.n_layer) ** -0.5\n            torch.nn.init.normal_(module.weight, mean = 0.0, std = std)\n            if module.bias is not None:\n                torch.nn.init.zeros_(module.bias)\n        elif isinstance(module, nn.Embedding):\n            torch.nn.init.normal_(module.weight, mean=0.0, std = 0.02)\n\n\n\n    def forward(self, idx, targets=None):\n        # idx is of shape (B, T)\n        B, T = idx.size()\n        assert T <= self.config.block_size, f\"Cannot forward sequence of length {T}, block size is only {self.config.block_size}\"\n        # forward the token and posisition embeddings\n        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)\n        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)\n        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)\n        x = tok_emb + pos_emb\n        # forward the blocks of the transformer\n        for block in self.transformer.h:\n            x = block(x)\n        # forward the final layernorm and the classifier\n        x = self.transformer.ln_f(x)\n        logits = self.lm_head(x) # (B, T, vocab_size)\n        loss = None\n        if targets is not None:\n            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))\n        return logits, loss\n\n    @classmethod\n    def from_pretrained(cls, model_type):\n        \"\"\"Loads pretrained GPT-2 model weights from huggingface\"\"\"\n        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}\n        from transformers import GPT2LMHeadModel\n        print(\"loading weights from pretrained gpt: %s\" % model_type)\n\n        # n_layer, n_head and n_embd are determined from model_type\n        config_args = {\n            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params\n            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params\n            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params\n            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params\n        }[model_type]\n        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints\n        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints\n        # create a from-scratch initialized minGPT model\n        config = GPTConfig(**config_args)\n        model = GPT(config)\n        sd = model.state_dict()\n        sd_keys = sd.keys()\n        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param\n\n        # init a huggingface/transformers model\n        model_hf = GPT2LMHeadModel.from_pretrained(model_type)\n        sd_hf = model_hf.state_dict()\n\n        # copy while ensuring all of the parameters are aligned and match in names and shapes\n        sd_keys_hf = sd_hf.keys()\n        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer\n        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)\n        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']\n        # basically the openai checkpoints use a \"Conv1D\" module, but we only want to use a vanilla Linear\n        # this means that we have to transpose these weights when we import them\n        assert len(sd_keys_hf) == len(sd_keys), f\"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}\"\n        for k in sd_keys_hf:\n            if any(k.endswith(w) for w in transposed):\n                # special treatment for the Conv1D weights we need to transpose\n                assert sd_hf[k].shape[::-1] == sd[k].shape\n                with torch.no_grad():\n                    sd[k].copy_(sd_hf[k].t())\n            else:\n                # vanilla copy over the other parameters\n                assert sd_hf[k].shape == sd[k].shape\n                with torch.no_grad():\n                    sd[k].copy_(sd_hf[k])\n\n        return model","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:26.530758Z","iopub.execute_input":"2025-01-16T01:11:26.531068Z","iopub.status.idle":"2025-01-16T01:11:26.543967Z","shell.execute_reply.started":"2025-01-16T01:11:26.531045Z","shell.execute_reply":"2025-01-16T01:11:26.542879Z"}},"outputs":[],"execution_count":6},{"cell_type":"code","source":"# model = GPT.from_pretrained('gpt2')\n\ndevice = 'cpu'\nif torch.cuda.is_available():\n    device = 'cuda'\nelif hasattr(torch.backends, \"mps\") and torch.backends.mps.is_available():\n    device = \"mps\"\nprint(f\"using device: {device}\")\n\n# SEED\ntorch.manual_seed(1337)\nif torch.cuda.is_available():\n    torch.cuda.manual_seed(1337)\n\n# STOP\nnum_return_sequences = 5\nmax_length = 30","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:31.096650Z","iopub.execute_input":"2025-01-16T01:11:31.096934Z","iopub.status.idle":"2025-01-16T01:11:31.154262Z","shell.execute_reply.started":"2025-01-16T01:11:31.096913Z","shell.execute_reply":"2025-01-16T01:11:31.153422Z"}},"outputs":[{"name":"stdout","text":"using device: cuda\n","output_type":"stream"}],"execution_count":7},{"cell_type":"code","source":"!pip install tiktoken","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:33.747825Z","iopub.execute_input":"2025-01-16T01:11:33.748149Z","iopub.status.idle":"2025-01-16T01:11:38.202233Z","shell.execute_reply.started":"2025-01-16T01:11:33.748124Z","shell.execute_reply":"2025-01-16T01:11:38.201182Z"}},"outputs":[{"name":"stdout","text":"Requirement already satisfied: tiktoken in /usr/local/lib/python3.10/dist-packages (0.8.0)\nRequirement already satisfied: regex>=2022.1.18 in /usr/local/lib/python3.10/dist-packages (from tiktoken) (2024.11.6)\nRequirement already satisfied: requests>=2.26.0 in /usr/local/lib/python3.10/dist-packages (from tiktoken) (2.32.3)\nRequirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests>=2.26.0->tiktoken) (3.4.0)\nRequirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests>=2.26.0->tiktoken) (3.10)\nRequirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests>=2.26.0->tiktoken) (2.2.3)\nRequirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests>=2.26.0->tiktoken) (2024.12.14)\n","output_type":"stream"}],"execution_count":8},{"cell_type":"code","source":"import tiktoken\n\nclass DataLoaderLite:\n    def __init__(self, B, T):\n        self.B = B\n        self.T = T\n\n        # at init load tokens from disk and store them in memory\n        with open('/kaggle/input/transformers-input/input.txt', 'r') as f:\n            text = f.read()\n        enc = tiktoken.get_encoding('gpt2') \n        tokens = enc.encode(text)\n        self.tokens = torch.tensor(tokens)\n        print(f'loaded {len(self.tokens)} tokens')\n        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')\n\n        # state\n        self.current_position = 0\n    \n    def next_batch(self):\n        B, T = self.B, self.T\n        buf = self.tokens[self.current_position: self.current_position + B * T + 1]\n        x = (buf[:-1]).view(B, T) # inputs\n        y = (buf[1:]).view(B, T) # targets\n        # advance the position in the tensor\n        self.current_position += B*T\n        # if loading the next batch would be out of bounds, reset\n        if self.current_position + (B * T + 1) > len(self.tokens):\n            self.current_position = 0\n        return x, y\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:40.943438Z","iopub.execute_input":"2025-01-16T01:11:40.943746Z","iopub.status.idle":"2025-01-16T01:11:40.988785Z","shell.execute_reply.started":"2025-01-16T01:11:40.943722Z","shell.execute_reply":"2025-01-16T01:11:40.988104Z"}},"outputs":[],"execution_count":9},{"cell_type":"code","source":"import logging\n\n# Configure logging\nlogging.basicConfig(filename='/kaggle/working/training_log.txt', level=logging.INFO, \n                    format='%(asctime)s - %(levelname)s - %(message)s', force=True)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:43.834394Z","iopub.execute_input":"2025-01-16T01:11:43.834693Z","iopub.status.idle":"2025-01-16T01:11:43.839266Z","shell.execute_reply.started":"2025-01-16T01:11:43.834671Z","shell.execute_reply":"2025-01-16T01:11:43.838356Z"}},"outputs":[],"execution_count":10},{"cell_type":"code","source":"model = GPT(GPTConfig())\nmodel.to(device)\n\n# ... existing model initialization ...\n\n# Get model summary and parameter count\ntotal_params = sum(p.numel() for p in model.parameters())\ntrainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)\nlogging.info(f'Total Parameters: {total_params:,}')\nlogging.info(f'Trainable Parameters: {trainable_params:,}')\n\n\ntrain_loader = DataLoaderLite(B = 4, T = 256)\noptimizer = torch.optim.AdamW(model.parameters(), lr = 3e-4)\n\n# Track best loss for model saving\nbest_loss = float('inf')\nbest_model_path = '/kaggle/working/best_model.pth'\ntarget_loss = 0.099999\nmax_iterations = 1000000  # Safety limit to prevent infinite loops\n\ni = 0\nwhile True:\n    x, y = train_loader.next_batch()\n    x, y = x.to(device), y.to(device)\n    optimizer.zero_grad()\n    logits, loss = model(x, y)\n    loss.backward()\n    optimizer.step()\n    \n    # Save best model\n    if loss.item() < best_loss:\n        best_loss = loss.item()\n        torch.save({\n            'epoch': i,\n            'model_state_dict': model.state_dict(),\n            'optimizer_state_dict': optimizer.state_dict(),\n            'loss': best_loss,\n        }, best_model_path)\n    \n    # Replace print statements with logging\n    logging.info(f'step {i}, loss: {loss.item():.6f}, best loss: {best_loss:.6f}')\n    \n    # Check if target loss is reached\n    if loss.item() <= target_loss:\n        logging.info(f'\\nTarget loss reached! Training completed at step {i}')\n        break\n    \n    # Safety check to prevent infinite training\n    if i >= max_iterations:\n        logging.info('\\nMaximum iterations reached without hitting target loss')\n        break\n        \n    i += 1\n\nlogging.info(f'\\nTraining completed!')\nlogging.info(f'Final loss: {loss.item():.6f}')\nlogging.info(f'Best loss achieved: {best_loss:.6f}')\nlogging.info(f'Best model saved to: {best_model_path}')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T01:11:51.706858Z","iopub.execute_input":"2025-01-16T01:11:51.707283Z","iopub.status.idle":"2025-01-16T02:14:36.050844Z","shell.execute_reply.started":"2025-01-16T01:11:51.707251Z","shell.execute_reply":"2025-01-16T02:14:36.049681Z"}},"outputs":[{"name":"stdout","text":"loaded 338025 tokens\n1 epoch = 330 batches\n","output_type":"stream"}],"execution_count":11},{"cell_type":"code","source":"import sys;\n\ntorch.manual_seed(42)\ntorch.cuda.manual_seed(42)\nwhile x.size(1) < max_length:\n    # forward the model to get the logits\n    with torch.no_grad():\n        logits = model(x)[0] # (B, T, vocab_size)\n        # take the logits at the last position\n        logits = logits[:, -1, :] # (B, vocab_size)\n        # get the probabilities\n        probs = F.softmax(logits, dim=-1)\n        # do top-k sampling of 50 (huggingface pipeline default)\n        # topk_probs here becomes (5, 50), topk_indices is (5, 50)\n        topk_probs, topk_indices = torch.topk(probs, 50, dim=-1)\n        # select a token from the top-k probabilities\n        # note: multinomial does not demand the input to sum to 1\n        ix = torch.multinomial(topk_probs, 1) # (B, 1)\n        # gather the corresponding indices\n        xcol = torch.gather(topk_indices, -1, ix) # (B, 1)\n        # append to the sequence\n        x = torch.cat((x, xcol), dim=1)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T02:21:27.100740Z","iopub.execute_input":"2025-01-16T02:21:27.101343Z","iopub.status.idle":"2025-01-16T02:21:27.109634Z","shell.execute_reply.started":"2025-01-16T02:21:27.101308Z","shell.execute_reply":"2025-01-16T02:21:27.108902Z"}},"outputs":[],"execution_count":12},{"cell_type":"code","source":"enc = tiktoken.get_encoding('gpt2') \n# print the generated text\nfor i in range(num_return_sequences-1):\n    tokens = x[i, :max_length].tolist()\n    decoded = enc.decode(tokens)\n    print(\">\", decoded)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-16T02:29:15.075790Z","iopub.execute_input":"2025-01-16T02:29:15.076153Z","iopub.status.idle":"2025-01-16T02:29:15.082856Z","shell.execute_reply.started":"2025-01-16T02:29:15.076121Z","shell.execute_reply":"2025-01-16T02:29:15.082202Z"}},"outputs":[{"name":"stdout","text":">  braved me in this matter so?\n\nBAPTISTA:\nWhy, tell me, is not this my Cambio?\n\n> ; thy father will not frown.\n\nGREMIO:\nMy cake is dough; but I'll in among the rest,\nOut of\n>  scapes and perils overblown.\nMy fair Bianca, bid my father welcome,\nWhile I with self-same kindness welcome thine\n>  afeard of you.\n\nWidow:\nHe that is giddy thinks the world turns round.\n\nPETRUCH\n","output_type":"stream"}],"execution_count":18},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}

requirements.txt

@@ -0,0 +1,8 @@
+fastapi==0.68.0
+uvicorn==0.15.0
+jinja2==3.0.1
+tiktoken
+torch
+transformers
+aiofiles
+fastapi

src/app.py

@@ -0,0 +1,111 @@
+import torch
+import os
+from fastapi import FastAPI, Request
+from pydantic import BaseModel
+from transformers import PreTrainedModel, AutoConfig
+from huggingface_hub import hf_hub_download
+import tiktoken
+from model import GPT, GPTConfig
+from fastapi.templating import Jinja2Templates
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import HTMLResponse
+from fastapi.staticfiles import StaticFiles
+
+# Get the absolute path to the templates directory
+TEMPLATES_DIR = os.path.join(os.path.dirname(__file__), "templates")
+
+MODEL_ID = "sagargurujula/text-generator"
+
+# Initialize FastAPI
+app = FastAPI(title="GPT Text Generator")
+
+# Templates with absolute path
+templates = Jinja2Templates(directory=TEMPLATES_DIR)
+
+# Add CORS middleware
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# Set device
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+# Load model from Hugging Face Hub
+def load_model():
+    try:
+        # Download the model file from HF Hub
+        model_path = hf_hub_download(
+            repo_id=MODEL_ID,
+            filename="best_model.pth"
+        )
+        
+        # Initialize our custom GPT model
+        model = GPT(GPTConfig())
+        
+        # Load the state dict
+        checkpoint = torch.load(model_path, map_location=device, weights_only=True)
+        model.load_state_dict(checkpoint['model_state_dict'])
+        
+        model.to(device)
+        model.eval()
+        return model
+    
+    except Exception as e:
+        print(f"Error loading model: {e}")
+        raise
+
+# Load the model
+model = load_model()
+
+# Define the request body
+class TextInput(BaseModel):
+    text: str
+
+@app.post("/generate/")
+async def generate_text(input: TextInput):
+    # Prepare input tensor
+    enc = tiktoken.get_encoding('gpt2')
+    input_ids = torch.tensor([enc.encode(input.text)]).to(device)
+    
+    # Generate multiple tokens
+    generated_tokens = []
+    num_tokens_to_generate = 50  # Generate 20 new tokens
+    
+    with torch.no_grad():
+        current_ids = input_ids
+        
+        for _ in range(num_tokens_to_generate):
+            # Get model predictions
+            logits, _ = model(current_ids)
+            next_token = logits[0, -1, :].argmax().item()
+            generated_tokens.append(next_token)
+            
+            # Add the new token to our current sequence
+            current_ids = torch.cat([current_ids, torch.tensor([[next_token]]).to(device)], dim=1)
+    
+    # Decode all generated tokens
+    generated_text = enc.decode(generated_tokens)
+    
+    # Return both input and generated text
+    return {
+        "input_text": input.text,
+        "generated_text": generated_text
+    }
+
+# Modify the root route to serve the template
+@app.get("/", response_class=HTMLResponse)
+async def home(request: Request):
+    return templates.TemplateResponse(
+        "index.html", 
+        {"request": request, "title": "GPT Text Generator"}
+    )
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="127.0.0.1", port=8080)
+
+# To run the app, use the command: uvicorn app:app --reload

src/model.py

@@ -0,0 +1,298 @@
+# Solving for residual std scaling issue
+import os
+import math
+import time
+import inspect
+from dataclasses import dataclass
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+import tiktoken
+import logging
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.c_proj.NANGPT_SCALE_INIT = 1
+        # regularization
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size)).view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
+        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+        # output projection
+        y = self.c_proj(y)
+        return y
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd)
+        self.gelu    = nn.GELU(approximate='tanh')
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)
+        self.c_proj.NANOGPT_SCALE_INIT = 1
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        return x
+    
+class Block(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+    
+@dataclass
+class GPTConfig:
+    block_size: int = 1024 # max sequence length
+    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
+    n_layer: int = 12 # number of layers
+    n_head: int = 12 # number of heads
+    n_embd: int = 768 # embedding dimension
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
+            ln_f = nn.LayerNorm(config.n_embd),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # weight sharing
+        self.transformer.wte.weight = self.lm_head.weight
+
+        # weight initialization
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            std = 0.02
+            if hasattr(module, 'NANGPT_SCALE_INIT'):
+                std *= (2 * self.config.n_layer) ** -0.5
+            torch.nn.init.normal_(module.weight, mean = 0.0, std = std)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std = 0.02)
+
+
+
+    def forward(self, idx, targets=None):
+        # idx is of shape (B, T)
+        B, T = idx.size()
+        assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
+        # forward the token and posisition embeddings
+        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
+        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
+        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
+        x = tok_emb + pos_emb
+        # forward the blocks of the transformer
+        for block in self.transformer.h:
+            x = block(x)
+        # forward the final layernorm and the classifier
+        x = self.transformer.ln_f(x)
+        logits = self.lm_head(x) # (B, T, vocab_size)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+
+    @classmethod
+    def from_pretrained(cls, model_type):
+        """Loads pretrained GPT-2 model weights from huggingface"""
+        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
+        from transformers import GPT2LMHeadModel
+        print("loading weights from pretrained gpt: %s" % model_type)
+
+        # n_layer, n_head and n_embd are determined from model_type
+        config_args = {
+            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
+            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
+            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
+            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
+        }[model_type]
+        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
+        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
+        # create a from-scratch initialized minGPT model
+        config = GPTConfig(**config_args)
+        model = GPT(config)
+        sd = model.state_dict()
+        sd_keys = sd.keys()
+        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param
+
+        # init a huggingface/transformers model
+        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
+        sd_hf = model_hf.state_dict()
+
+        # copy while ensuring all of the parameters are aligned and match in names and shapes
+        sd_keys_hf = sd_hf.keys()
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
+        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
+        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
+        # this means that we have to transpose these weights when we import them
+        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
+        for k in sd_keys_hf:
+            if any(k.endswith(w) for w in transposed):
+                # special treatment for the Conv1D weights we need to transpose
+                assert sd_hf[k].shape[::-1] == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k].t())
+            else:
+                # vanilla copy over the other parameters
+                assert sd_hf[k].shape == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k])
+
+        return model        
+    
+class DataLoaderLite:
+    def __init__(self, B, T):
+        self.B = B
+        self.T = T
+
+        # at init load tokens from disk and store them in memory
+        with open('/kaggle/input/transformers-input/input.txt', 'r') as f:
+            text = f.read()
+        enc = tiktoken.get_encoding('gpt2') 
+        tokens = enc.encode(text)
+        self.tokens = torch.tensor(tokens)
+        print(f'loaded {len(self.tokens)} tokens')
+        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')
+
+        # state
+        self.current_position = 0
+    
+    def next_batch(self):
+        B, T = self.B, self.T
+        buf = self.tokens[self.current_position: self.current_position + B * T + 1]
+        x = (buf[:-1]).view(B, T) # inputs
+        y = (buf[1:]).view(B, T) # targets
+        # advance the position in the tensor
+        self.current_position += B*T
+        # if loading the next batch would be out of bounds, reset
+        if self.current_position + (B * T + 1) > len(self.tokens):
+            self.current_position = 0
+        return x, y
+
+if __name__ == "__main__":
+    # model = GPT.from_pretrained('gpt2')
+    # Configure logging
+    logging.basicConfig(filename='/kaggle/working/training_log.txt', level=logging.INFO, 
+                        format='%(asctime)s - %(levelname)s - %(message)s', force=True)
+
+    device = 'cpu'
+    if torch.cuda.is_available():
+        device = 'cuda'
+    elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        device = "mps"
+    print(f"using device: {device}")
+
+    # SEED
+    torch.manual_seed(1337)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(1337)
+
+    # STOP
+    num_return_sequences = 5
+    max_length = 30
+
+    model = GPT(GPTConfig())
+    model.to(device)
+
+    # ... existing model initialization ...
+
+    # Get model summary and parameter count
+    total_params = sum(p.numel() for p in model.parameters())
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    logging.info(f'Total Parameters: {total_params:,}')
+    logging.info(f'Trainable Parameters: {trainable_params:,}')
+
+
+    train_loader = DataLoaderLite(B = 4, T = 256)
+    optimizer = torch.optim.AdamW(model.parameters(), lr = 3e-4)
+
+    # Track best loss for model saving
+    best_loss = float('inf')
+    best_model_path = '/kaggle/working/best_model.pth'
+    target_loss = 0.099999
+    max_iterations = 1000000  # Safety limit to prevent infinite loops
+
+    i = 0
+    while True:
+        x, y = train_loader.next_batch()
+        x, y = x.to(device), y.to(device)
+        optimizer.zero_grad()
+        logits, loss = model(x, y)
+        loss.backward()
+        optimizer.step()
+        
+        # Save best model
+        if loss.item() < best_loss:
+            best_loss = loss.item()
+            torch.save({
+                'epoch': i,
+                'model_state_dict': model.state_dict(),
+                'optimizer_state_dict': optimizer.state_dict(),
+                'loss': best_loss,
+            }, best_model_path)
+        
+        # Replace print statements with logging
+        logging.info(f'step {i}, loss: {loss.item():.6f}, best loss: {best_loss:.6f}')
+        
+        # Check if target loss is reached
+        if loss.item() <= target_loss:
+            logging.info(f'\nTarget loss reached! Training completed at step {i}')
+            break
+        
+        # Safety check to prevent infinite training
+        if i >= max_iterations:
+            logging.info('\nMaximum iterations reached without hitting target loss')
+            break
+            
+        i += 1
+
+    logging.info(f'\nTraining completed!')
+    logging.info(f'Final loss: {loss.item():.6f}')
+    logging.info(f'Best loss achieved: {best_loss:.6f}')
+    logging.info(f'Best model saved to: {best_model_path}')

src/templates/index.html

@@ -0,0 +1,98 @@
+<!DOCTYPE html>
+<html>
+<head>
+    <title>GPT Text Generator</title>
+    <style>
+        body {
+            font-family: Arial, sans-serif;
+            max-width: 800px;
+            margin: 0 auto;
+            padding: 20px;
+        }
+        textarea {
+            width: 100%;
+            height: 100px;
+            margin: 10px 0;
+            padding: 10px;
+            font-size: 16px;
+            text-align: left;
+        }
+        #result {
+            margin-top: 20px;
+            padding: 15px;
+            border: 1px solid #ddd;
+            border-radius: 5px;
+            min-height: 50px;
+            white-space: pre-wrap;
+            background-color: #f9f9f9;
+            text-align: left;
+            font-size: 16px;
+            line-height: 1.5;
+        }
+        .loading {
+            opacity: 0.5;
+        }
+        button {
+            padding: 10px 20px;
+            font-size: 16px;
+            cursor: pointer;
+            display: block;
+            margin-bottom: 20px;
+        }
+        .label {
+            font-weight: bold;
+            display: block;
+            margin-bottom: 10px;
+        }
+    </style>
+</head>
+<body>
+    <h1>GPT Text Generator</h1>
+    <form id="generateForm">
+        <textarea id="inputText" placeholder="Enter your text here..."></textarea>
+        <button type="submit">Generate</button>
+    </form>
+    <div id="result"></div>
+
+    <script>
+        document.getElementById('generateForm').addEventListener('submit', async (e) => {
+            e.preventDefault();
+            
+            const inputText = document.getElementById('inputText').value;
+            const resultDiv = document.getElementById('result');
+            const submitButton = document.querySelector('button[type="submit"]');
+            
+            // Show loading state
+            submitButton.disabled = true;
+            resultDiv.classList.add('loading');
+            resultDiv.textContent = 'Generating...';
+            
+            try {
+                const response = await fetch('/generate/', {
+                    method: 'POST',
+                    headers: {
+                        'Content-Type': 'application/json',
+                    },
+                    body: JSON.stringify({ text: inputText })
+                });
+                
+                const data = await response.json();
+                resultDiv.innerHTML = `
+                    <div class="label">Input:</div>
+                    ${data.input_text}
+                    
+                    <div class="label" style="margin-top: 20px;">Generated continuation:</div>
+                    ${data.generated_text}
+                `;
+            } catch (error) {
+                console.error('Error:', error);
+                resultDiv.textContent = 'Error generating text. Please try again.';
+            } finally {
+                // Reset loading state
+                submitButton.disabled = false;
+                resultDiv.classList.remove('loading');
+            }
+        });
+    </script>
+</body>
+</html>

train_get2-8-init.py

@@ -0,0 +1,287 @@
+# Solving for residual std scaling issue
+import os
+import math
+import time
+import inspect
+from dataclasses import dataclass
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.c_proj.NANGPT_SCALE_INIT = 1
+        # regularization
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size)).view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
+        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+        # output projection
+        y = self.c_proj(y)
+        return y
+
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd)
+        self.gelu    = nn.GELU(approximate='tanh')
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)
+        self.c_proj.NANOGPT_SCALE_INIT = 1
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024 # max sequence length
+    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
+    n_layer: int = 12 # number of layers
+    n_head: int = 12 # number of heads
+    n_embd: int = 768 # embedding dimension
+
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
+            ln_f = nn.LayerNorm(config.n_embd),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # weight sharing
+        self.transformer.wte.weight = self.lm_head.weight
+
+        # weight initialization
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            std = 0.02
+            if hasattr(module, 'NANGPT_SCALE_INIT'):
+                std *= (2 * self.config.n_layer) ** -0.5
+            torch.nn.init.normal_(module.weight, mean = 0.0, std = std)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std = 0.02)
+
+
+
+    def forward(self, idx, targets=None):
+        # idx is of shape (B, T)
+        B, T = idx.size()
+        assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
+        # forward the token and posisition embeddings
+        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
+        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
+        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
+        x = tok_emb + pos_emb
+        # forward the blocks of the transformer
+        for block in self.transformer.h:
+            x = block(x)
+        # forward the final layernorm and the classifier
+        x = self.transformer.ln_f(x)
+        logits = self.lm_head(x) # (B, T, vocab_size)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+
+    @classmethod
+    def from_pretrained(cls, model_type):
+        """Loads pretrained GPT-2 model weights from huggingface"""
+        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
+        from transformers import GPT2LMHeadModel
+        print("loading weights from pretrained gpt: %s" % model_type)
+
+        # n_layer, n_head and n_embd are determined from model_type
+        config_args = {
+            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
+            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
+            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
+            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
+        }[model_type]
+        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
+        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
+        # create a from-scratch initialized minGPT model
+        config = GPTConfig(**config_args)
+        model = GPT(config)
+        sd = model.state_dict()
+        sd_keys = sd.keys()
+        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param
+
+        # init a huggingface/transformers model
+        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
+        sd_hf = model_hf.state_dict()
+
+        # copy while ensuring all of the parameters are aligned and match in names and shapes
+        sd_keys_hf = sd_hf.keys()
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
+        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
+        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
+        # this means that we have to transpose these weights when we import them
+        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
+        for k in sd_keys_hf:
+            if any(k.endswith(w) for w in transposed):
+                # special treatment for the Conv1D weights we need to transpose
+                assert sd_hf[k].shape[::-1] == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k].t())
+            else:
+                # vanilla copy over the other parameters
+                assert sd_hf[k].shape == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k])
+
+        return model
+
+# model = GPT.from_pretrained('gpt2')
+
+device = 'cpu'
+if torch.cuda.is_available():
+    device = 'cuda'
+elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    device = "mps"
+print(f"using device: {device}")
+
+# SEED
+torch.manual_seed(1337)
+if torch.cuda.is_available():
+    torch.cuda.manual_seed(1337)
+
+# STOP
+num_return_sequences = 5
+max_length = 30
+
+
+
+import tiktoken
+
+class DataLoaderLite:
+    def __init__(self, B, T):
+        self.B = B
+        self.T = T
+
+        # at init load tokens from disk and store them in memory
+        with open('input.txt', 'r') as f:
+            text = f.read()
+        enc = tiktoken.get_encoding('gpt2') 
+        tokens = enc.encode(text)
+        self.tokens = torch.tensor(tokens)
+        print(f'loaded {len(self.tokens)} tokens')
+        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')
+
+        # state
+        self.current_position = 0
+    
+    def next_batch(self):
+        B, T = self.B, self.T
+        buf = self.tokens[self.current_position: self.current_position + B * T + 1]
+        x = (buf[:-1]).view(B, T) # inputs
+        y = (buf[1:]).view(B, T) # targets
+        # advance the position in the tensor
+        self.current_position += B*T
+        # if loading the next batch would be out of bounds, reset
+        if self.current_position + (B * T + 1) > len(self.tokens):
+            self.current_position = 0
+        return x, y
+
+
+model = GPT(GPTConfig())
+model.to(device)
+
+train_loader = DataLoaderLite(B = 4, T = 32)
+
+# NEW CODE
+optimizer = torch.optim.AdamW(model.parameters(), lr = 3e-4)
+for i in range(50):
+    x, y = train_loader.next_batch()
+    x, y = x.to(device), y.to(device)
+    optimizer.zero_grad()
+    logits, loss = model(x, y)
+    loss.backward()
+    optimizer.step()
+    print(f'step{i}, loss: {loss.item()}')
+
+
+print(loss)
+import sys; sys.exit(0)
+
+torch.manual_seed(42)
+torch.cuda.manual_seed(42)
+while x.size(1) < max_length:
+    # forward the model to get the logits
+    with torch.no_grad():
+        logits = model(x)[0] # (B, T, vocab_size)
+        # take the logits at the last position
+        logits = logits[:, -1, :] # (B, vocab_size)
+        # get the probabilities
+        probs = F.softmax(logits, dim=-1)
+        # do top-k sampling of 50 (huggingface pipeline default)
+        # topk_probs here becomes (5, 50), topk_indices is (5, 50)
+        topk_probs, topk_indices = torch.topk(probs, 50, dim=-1)
+        # select a token from the top-k probabilities
+        # note: multinomial does not demand the input to sum to 1
+        ix = torch.multinomial(topk_probs, 1) # (B, 1)
+        # gather the corresponding indices
+        xcol = torch.gather(topk_indices, -1, ix) # (B, 1)
+        # append to the sequence
+        x = torch.cat((x, xcol), dim=1)
+
+# print the generated text
+for i in range(num_return_sequences):
+    tokens = x[i, :max_length].tolist()
+    decoded = enc.decode(tokens)
+    print(">", decoded)