Add SmolLM2-135M model and Gradio app

- app.py: Gradio interface for text generation
- requirements.txt: Dependencies (torch, tiktoken, gradio)
- README.md: Model architecture and training documentation
- smollm2_135m_final.pt: Trained model weights (~135M params)

Model trained from scratch on Suits TV series scripts for 5,050 steps.
Architecture: RMSNorm, RoPE, GQA (9 query, 3 KV heads), SwiGLU MLP.

README.md

@@ -1,12 +1,326 @@
+# SmolLM2-135M Training from Scratch
+
+This repository contains a from-scratch implementation of the SmolLM2-135M model architecture, trained on custom text data.
+
+## Table of Contents
+- [Model Architecture](#model-architecture)
+- [Parameter Calculation](#parameter-calculation)
+- [Training Data](#training-data)
+- [Training Details](#training-details)
+- [Speedups Used](#speedups-used)
+- [Results](#results)
+- [Usage](#usage)
+
+---
+
+## Model Architecture
+
+SmolLM2-135M is a **Llama-based decoder-only transformer** model. Unlike GPT-2, it incorporates modern architectural improvements that have become standard in recent language models.
+
+### Architecture Overview
+
+```
+Input Tokens
+     │
+     ▼
+┌─────────────────┐
+│ Token Embedding │  (No position embeddings - RoPE applied in attention)
+└────────┬────────┘
+         │
+         ▼
+┌─────────────────────────────────────┐
+│         Transformer Block x30       │
+│  ┌─────────────────────────────┐    │
+│  │ RMSNorm                     │    │
+│  │      ↓                      │    │
+│  │ Grouped Query Attention     │    │
+│  │ (9 query heads, 3 KV heads) │    │
+│  │ + RoPE                      │    │
+│  │      ↓                      │    │
+│  │ Residual Connection         │    │
+│  │      ↓                      │    │
+│  │ RMSNorm                     │    │
+│  │      ↓                      │    │
+│  │ SwiGLU MLP                  │    │
+│  │      ↓                      │    │
+│  │ Residual Connection         │    │
+│  └─────────────────────────────┘    │
+└────────────────┬────────────────────┘
+                 │
+                 ▼
+┌─────────────────┐
+│ Final RMSNorm   │
+└────────┬────────┘
+         │
+         ▼
+┌─────────────────┐
+│ LM Head         │  (Tied with token embeddings)
+└────────┬────────┘
+         │
+         ▼
+    Output Logits
+```
+
+### Key Components
+
+#### 1. RMSNorm (Root Mean Square Normalization)
+Unlike LayerNorm, RMSNorm doesn't center activations (no mean subtraction), making it more computationally efficient.
+
+```python
+# Formula: x * weight / sqrt(mean(x²) + eps)
+def forward(self, x):
+    rms = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+    return x * rms * self.weight
+```
+
+#### 2. Rotary Position Embedding (RoPE)
+RoPE encodes position information by rotating query and key vectors. The dot product of rotated vectors naturally encodes relative positions.
+
+- **Advantage**: No learned position embeddings needed
+- **Advantage**: Better extrapolation to longer sequences
+- **theta**: 10,000 (base frequency)
+
+#### 3. Grouped Query Attention (GQA)
+GQA reduces memory bandwidth by sharing key-value heads across multiple query heads.
+
+| Component | Count |
+|-----------|-------|
+| Query Heads | 9 |
+| Key-Value Heads | 3 |
+| KV Groups | 3 (each KV head shared by 3 query heads) |
+| Head Dimension | 64 |
+
+```python
+# Q: (B, T, 9 heads, 64)
+# K, V: (B, T, 3 heads, 64) → repeated to match 9 query heads
+```
+
+#### 4. SwiGLU MLP
+SwiGLU is a gated linear unit variant using SiLU activation, shown to improve model quality.
+
+```python
+# Formula: down_proj(silu(gate_proj(x)) * up_proj(x))
+def forward(self, x):
+    return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+```
+
+### Configuration
+
+| Parameter | Value |
+|-----------|-------|
+| `vocab_size` | 50,304 (GPT-2 compatible) |
+| `hidden_size` | 576 |
+| `intermediate_size` | 1,536 |
+| `num_hidden_layers` | 30 |
+| `num_attention_heads` | 9 |
+| `num_key_value_heads` | 3 |
+| `max_position_embeddings` | 2,048 |
+| `rms_norm_eps` | 1e-5 |
+| `rope_theta` | 10,000 |
+| `hidden_act` | SiLU |
+| `tie_word_embeddings` | True |
+
+---
+
+## Parameter Calculation
+
+### Component-by-Component Breakdown
+
+#### 1. Token Embeddings
+```
+vocab_size × hidden_size = 50,304 × 576 = 28,975,104 parameters
+```
+
+#### 2. Attention (per layer)
+```
+Q projection:  hidden_size × (num_heads × head_dim)     = 576 × 576 = 331,776
+K projection:  hidden_size × (num_kv_heads × head_dim)  = 576 × 192 = 110,592
+V projection:  hidden_size × (num_kv_heads × head_dim)  = 576 × 192 = 110,592
+O projection:  (num_heads × head_dim) × hidden_size     = 576 × 576 = 331,776
+──────────────────────────────────────────────────────────��──────────────────
+Total per layer: 884,736 parameters
+Total for 30 layers: 884,736 × 30 = 26,542,080 parameters
+```
+
+#### 3. MLP (per layer)
+```
+gate_proj:  hidden_size × intermediate_size = 576 × 1,536 = 884,736
+up_proj:    hidden_size × intermediate_size = 576 × 1,536 = 884,736
+down_proj:  intermediate_size × hidden_size = 1,536 × 576 = 884,736
+─────────────────────────────────────────────────────────────────────
+Total per layer: 2,654,208 parameters
+Total for 30 layers: 2,654,208 × 30 = 79,626,240 parameters
+```
+
+#### 4. RMSNorm (per layer + final)
+```
+input_layernorm:           hidden_size = 576
+post_attention_layernorm:  hidden_size = 576
+─────────────────────────────────────────────
+Total per layer: 1,152 parameters
+Total for 30 layers: 1,152 × 30 = 34,560 parameters
+Final norm: 576 parameters
+Total normalization: 35,136 parameters
+```
+
+#### 5. LM Head
+```
+Tied with token embeddings: 0 additional parameters
+```
+
+### Total Parameter Count
+
+| Component | Parameters | Percentage |
+|-----------|------------|------------|
+| Embedding | 28,975,104 | 21.4% |
+| Attention | 26,542,080 | 19.6% |
+| MLP | 79,626,240 | 58.9% |
+| Normalization | 35,136 | 0.03% |
+| **Total** | **135,178,560** | **100%** |
+
+> **Note**: Our implementation uses vocab_size=50,304 instead of the original 49,152, adding ~663K parameters to the embedding layer. Original SmolLM2-135M has ~134.5M parameters.
+
+---
+
+## Training Data
+
+### Dataset Description
+The model was trained on dialogue scripts from the television series **"Suits"**, a legal drama that follows characters working at a fictional New York law firm.
+
+### Characteristics
+- **Content Type**: Television dialogue scripts
+- **Genre**: Legal drama
+- **Language Style**: Professional legal terminology mixed with casual dialogue
+- **Text Format**: Character names followed by dialogue
+
+### Tokenization
+- **Tokenizer**: GPT-2 BPE tokenizer (tiktoken)
+- **Vocabulary Size**: 50,257 tokens (padded to 50,304 for GPU efficiency)
+
+---
+
+## Training Details
+
+### Hyperparameters
+
+| Parameter | Value |
+|-----------|-------|
+| Total Steps | 5,000 + 50 (resumed) |
+| Batch Size | 16 (effective) |
+| Micro Batch Size | 4 |
+| Gradient Accumulation | 4 steps |
+| Sequence Length | 1,024 tokens |
+| Learning Rate | 6e-4 (max) |
+| LR Schedule | Cosine with warmup |
+| Warmup Steps | 500 |
+| Weight Decay | 0.1 |
+| Gradient Clipping | 1.0 |
+| Optimizer | AdamW (fused) |
+| Precision | bfloat16 |
+
+### Checkpointing
+- Checkpoints saved every **500 steps**
+- Text generation with fixed prompts at each checkpoint
+- Final checkpoint at step 5,050 (after resume demonstration)
+
+### Fixed Evaluation Prompts
+```
+1. "Once upon a time"
+2. "The meaning of life is"
+3. "In a galaxy far away"
+```
+
 ---
-title: SmolLM2 135M From Scratch
-emoji: 🐨
-colorFrom: yellow
-colorTo: yellow
-sdk: gradio
-sdk_version: 6.3.0
-app_file: app.py
-pinned: false
+
+## Speedups Used
+
+| Speedup | Implementation |
+|---------|----------------|
+| **Flash Attention** | `F.scaled_dot_product_attention(is_causal=True)` |
+| **Mixed Precision** | `torch.autocast(dtype=torch.bfloat16)` |
+| **torch.compile** | JIT compilation for CUDA |
+| **TF32 Precision** | `torch.set_float32_matmul_precision('high')` |
+| **Gradient Accumulation** | 4 micro-batches per step |
+| **Fused AdamW** | `fused=True` for CUDA |
+| **Power-of-2 Vocab** | 50,304 for efficient GPU memory access |
+
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+## Results
+
+### Training Progress
+- **Initial Loss**: ~10.8 (random initialization)
+- **Final Loss**: Significantly reduced after 5,050 steps
+- **Checkpoint Resume**: Successfully demonstrated loading from step 5,000 and continuing training
+
+### Checkpoints Saved
+```
+checkpoints/
+├── checkpoint_step_500.pt
+├── checkpoint_step_1000.pt
+├── checkpoint_step_1500.pt
+├── checkpoint_step_2000.pt
+├── checkpoint_step_2500.pt
+├── checkpoint_step_3000.pt
+├── checkpoint_step_3500.pt
+├── checkpoint_step_4000.pt
+├── checkpoint_step_4500.pt
+├── checkpoint_step_5000.pt
+└── checkpoint_step_5050.pt
+```
+
+---
+
+## Usage
+
+### Requirements
+```bash
+pip install torch tiktoken matplotlib
+```
+
+### Training
+1. Upload `input.txt` (training data) to the working directory
+2. Run the notebook cells sequentially
+3. Checkpoints will be saved to `checkpoints/` directory
+
+### Loading a Checkpoint
+```python
+checkpoint = torch.load('checkpoints/checkpoint_step_5000.pt')
+model.load_state_dict(checkpoint['model_state_dict'])
+optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
+```
+
+### Text Generation
+```python
+generated = generate_text(
+    model,
+    prompt="Once upon a time",
+    max_new_tokens=50,
+    temperature=0.8,
+    top_k=50
+)
+print(generated)
+```
+
+---
+
+## References
+
+- [SmolLM2 - HuggingFace](https://huggingface.co/HuggingFaceTB/SmolLM2-135M)
+- [RoPE: Rotary Position Embedding](https://arxiv.org/abs/2104.09864)
+- [GQA: Grouped Query Attention](https://arxiv.org/abs/2305.13245)
+- [SwiGLU Activation](https://arxiv.org/abs/2002.05202)
+- [RMSNorm](https://arxiv.org/abs/1910.07467)
+
+---
+
+## License
+
+This project is for educational purposes.
+
+---
+
+## Acknowledgments
+
+- HuggingFace for the SmolLM2 model and training recipes
+- Andrej Karpathy for nanoGPT inspiration

app.py

@@ -0,0 +1,401 @@
+"""
+SmolLM2-135M Text Generation - Gradio App
+Trained from scratch on Suits TV series scripts
+"""
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from dataclasses import dataclass
+from typing import Optional, Tuple
+import tiktoken
+import gradio as gr
+
+# ============================================================================
+# Model Architecture
+# ============================================================================
+
+@dataclass
+class SmolLM2Config:
+    """SmolLM2-135M Configuration"""
+    vocab_size: int = 50304
+    hidden_size: int = 576
+    intermediate_size: int = 1536
+    num_hidden_layers: int = 30
+    num_attention_heads: int = 9
+    num_key_value_heads: int = 3
+    max_position_embeddings: int = 2048
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 10000.0
+    hidden_act: str = "silu"
+    initializer_range: float = 0.041666666666666664
+    tie_word_embeddings: bool = True
+    bos_token_id: int = 0
+    eos_token_id: int = 0
+
+    @property
+    def head_dim(self) -> int:
+        return self.hidden_size // self.num_attention_heads
+
+
+class RMSNorm(nn.Module):
+    """Root Mean Square Layer Normalization"""
+    def __init__(self, hidden_size: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        rms = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        return x * rms * self.weight
+
+
+class RotaryEmbedding(nn.Module):
+    """Rotary Position Embedding (RoPE)"""
+    def __init__(self, dim: int, max_position_embeddings: int = 2048, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.theta = theta
+
+        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float32) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._set_cos_sin_cache(max_position_embeddings)
+
+    def _set_cos_sin_cache(self, seq_len: int):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(seq_len, dtype=torch.float32)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos(), persistent=False)
+        self.register_buffer("sin_cached", emb.sin(), persistent=False)
+
+    def forward(self, x: torch.Tensor, seq_len: int) -> Tuple[torch.Tensor, torch.Tensor]:
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len)
+        return (
+            self.cos_cached[:seq_len].to(x.dtype),
+            self.sin_cached[:seq_len].to(x.dtype)
+        )
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    cos = cos.unsqueeze(0).unsqueeze(0)
+    sin = sin.unsqueeze(0).unsqueeze(0)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class GroupedQueryAttention(nn.Module):
+    """Grouped Query Attention (GQA)"""
+    def __init__(self, config: SmolLM2Config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = config.head_dim
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary_emb = RotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=config.max_position_embeddings,
+            theta=config.rope_theta
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, T, C = x.size()
+
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        q = q.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
+        k = k.view(B, T, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = v.view(B, T, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary_emb(x, T)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        k = k.repeat_interleave(self.num_kv_groups, dim=1)
+        v = v.repeat_interleave(self.num_kv_groups, dim=1)
+
+        y = F.scaled_dot_product_attention(q, k, v, is_causal=True)
+        y = y.transpose(1, 2).contiguous().view(B, T, self.hidden_size)
+        y = self.o_proj(y)
+        return y
+
+
+class SwiGLUMLP(nn.Module):
+    """SwiGLU Feed-Forward Network"""
+    def __init__(self, config: SmolLM2Config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class SmolLM2Block(nn.Module):
+    """SmolLM2 Transformer Block"""
+    def __init__(self, config: SmolLM2Config):
+        super().__init__()
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.self_attn = GroupedQueryAttention(config)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.mlp = SwiGLUMLP(config)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x + self.self_attn(self.input_layernorm(x))
+        x = x + self.mlp(self.post_attention_layernorm(x))
+        return x
+
+
+class SmolLM2(nn.Module):
+    """SmolLM2-135M Model"""
+    def __init__(self, config: SmolLM2Config):
+        super().__init__()
+        self.config = config
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([SmolLM2Block(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+
+    def forward(self, input_ids: torch.Tensor, targets: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        B, T = input_ids.size()
+        x = self.embed_tokens(input_ids)
+
+        for layer in self.layers:
+            x = layer(x)
+
+        x = self.norm(x)
+        logits = self.lm_head(x)
+
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+
+        return logits, loss
+
+
+# ============================================================================
+# Load Model
+# ============================================================================
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"Using device: {device}")
+
+# Load model
+config = SmolLM2Config()
+model = SmolLM2(config)
+
+# Load trained weights
+checkpoint = torch.load("smollm2_135m_final.pt", map_location=device)
+model.load_state_dict(checkpoint['model_state_dict'])
+model.to(device)
+model.eval()
+
+print(f"Model loaded successfully! Parameters: {sum(p.numel() for p in model.parameters()):,}")
+
+# Load tokenizer
+tokenizer = tiktoken.get_encoding('gpt2')
+
+
+# ============================================================================
+# Generation Function
+# ============================================================================
+
+def generate_text(
+    prompt: str,
+    max_new_tokens: int = 100,
+    temperature: float = 0.8,
+    top_k: int = 50,
+    top_p: float = 0.9,
+) -> str:
+    """Generate text from a prompt"""
+    if not prompt.strip():
+        return "Please enter a prompt."
+
+    model.eval()
+
+    # Encode prompt
+    tokens = tokenizer.encode(prompt)
+    tokens = torch.tensor(tokens, dtype=torch.long, device=device).unsqueeze(0)
+
+    # Generate tokens
+    with torch.no_grad():
+        for _ in range(max_new_tokens):
+            # Crop to max position embeddings
+            idx_cond = tokens[:, -config.max_position_embeddings:]
+
+            # Get predictions
+            logits, _ = model(idx_cond)
+            logits = logits[:, -1, :] / temperature
+
+            # Top-k filtering
+            if top_k is not None and top_k > 0:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                logits[logits < v[:, [-1]]] = float('-inf')
+
+            # Top-p (nucleus) filtering
+            if top_p is not None and top_p < 1.0:
+                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+                sorted_indices_to_remove[..., 0] = 0
+                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+                logits[indices_to_remove] = float('-inf')
+
+            # Sample
+            probs = F.softmax(logits, dim=-1)
+            next_token = torch.multinomial(probs, num_samples=1)
+            tokens = torch.cat([tokens, next_token], dim=1)
+
+    # Decode
+    generated = tokenizer.decode(tokens[0].tolist())
+    return generated
+
+
+# ============================================================================
+# Gradio Interface
+# ============================================================================
+
+title = "SmolLM2-135M Text Generator"
+description = """
+## About This Model
+
+This is a **SmolLM2-135M** model trained from scratch on dialogue scripts from the TV series "Suits".
+
+### Model Architecture
+- **Type**: Llama-based decoder-only transformer
+- **Parameters**: ~135M
+- **Features**: RMSNorm, RoPE, Grouped Query Attention (GQA), SwiGLU MLP
+
+### Training Details
+- Trained for 5,050 steps
+- Sequence length: 1024 tokens
+- Uses GPT-2 tokenizer
+
+Enter a prompt below and adjust the generation parameters to see what the model generates!
+"""
+
+examples = [
+    ["Harvey walked into the office and said,"],
+    ["The legal case was complicated because"],
+    ["Once upon a time"],
+    ["In a world where lawyers"],
+    ["Mike looked at the contract and noticed"],
+]
+
+# Create interface
+with gr.Blocks(title=title, theme=gr.themes.Soft()) as demo:
+    gr.Markdown(f"# {title}")
+    gr.Markdown(description)
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            prompt_input = gr.Textbox(
+                label="Prompt",
+                placeholder="Enter your prompt here...",
+                lines=3,
+            )
+
+            with gr.Row():
+                max_tokens_slider = gr.Slider(
+                    minimum=10,
+                    maximum=500,
+                    value=100,
+                    step=10,
+                    label="Max New Tokens",
+                )
+                temperature_slider = gr.Slider(
+                    minimum=0.1,
+                    maximum=2.0,
+                    value=0.8,
+                    step=0.1,
+                    label="Temperature",
+                )
+
+            with gr.Row():
+                top_k_slider = gr.Slider(
+                    minimum=1,
+                    maximum=100,
+                    value=50,
+                    step=1,
+                    label="Top-K",
+                )
+                top_p_slider = gr.Slider(
+                    minimum=0.1,
+                    maximum=1.0,
+                    value=0.9,
+                    step=0.05,
+                    label="Top-P (Nucleus)",
+                )
+
+            generate_btn = gr.Button("Generate", variant="primary")
+
+        with gr.Column(scale=2):
+            output_text = gr.Textbox(
+                label="Generated Text",
+                lines=15,
+                show_copy_button=True,
+            )
+
+    gr.Markdown("### Example Prompts")
+    gr.Examples(
+        examples=examples,
+        inputs=prompt_input,
+    )
+
+    # Connect the generate button
+    generate_btn.click(
+        fn=generate_text,
+        inputs=[prompt_input, max_tokens_slider, temperature_slider, top_k_slider, top_p_slider],
+        outputs=output_text,
+    )
+
+    # Also generate on Enter key
+    prompt_input.submit(
+        fn=generate_text,
+        inputs=[prompt_input, max_tokens_slider, temperature_slider, top_k_slider, top_p_slider],
+        outputs=output_text,
+    )
+
+    gr.Markdown("""
+    ---
+    ### Parameter Guide
+    - **Temperature**: Higher = more creative/random, Lower = more focused/deterministic
+    - **Top-K**: Only sample from the top K most likely tokens
+    - **Top-P**: Only sample from tokens whose cumulative probability is below P
+    - **Max New Tokens**: Maximum number of tokens to generate
+
+    ---
+    *Model trained from scratch using PyTorch. Architecture based on SmolLM2-135M (Llama-style).*
+    """)
+
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,3 @@
+torch>=2.0.0
+tiktoken>=0.5.0
+gradio>=4.0.0

smollm2_135m_final.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e2fbf5fa84c5eaf792ef34b906d27ff195684c490f2b8ad572b1f03b3d3b3ee
+size 540826769