--- license: mit base_model: karpathy/tinyllamas tags: - llama2 - mqa - gguf - safetensors - transformers - tinyllamas - validation - test-suite --- # TinyStories Llama2 1M MQA (tinymqa1m) GGUF & HF Validation Suite This repository provides an ultra-lightweight Llama2 model variant featuring a **Custom BPE Tokenizer** combined with a strict **MQA (Multi-Query Attention)** structural layout. It is trained on the TinyStories dataset and optimized specifically for compiler, runtime, and hardware kernel validation. --- ## 📊 Comparison: `tinymqa1m` vs Previous Variants To help you choose the correct test asset for your specific engine debugging goals, the architectural differences across the 1M parameter suite are structured below: | Feature / Metric | `tiny1m` (Standard) | `tinybpe1m` (BPE Variant) | `tinymqa1m` (This Repository) | | :--- | :--- | :--- | :--- | | **Attention Mechanism** | **MHA** (Multi-Head Attention) | **MHA** (Multi-Head Attention) | **MQA** (Multi-Query Attention) | | **Attention Heads ($N_{heads} / N_{kv\_heads}$)** | 2 Heads / 2 KV Heads | 2 Heads / 2 KV Heads | **4 Heads / 1 KV Head** (Asymmetric) | | **Tokenizer Type** | Simple Character-level | **SentencePiece BPE** | **SentencePiece BPE** | | **Byte Fallback Support** | No | **Yes** (`byte_fallback=True`) | **Yes** (`byte_fallback=True`) | | **`llama2.c` Compatibility** | **Fully Compatible** (`run.c`) | Incompatible (Corrupts text) | **Incompatible** (Crashes/Corrupts) | | **Primary Debug Target** | Core matrix multiplication & layout | `byte_fallback` decoder loop | **KV-cache alignment & broadcast** | ### Why test with `tinymqa1m`? Modern architectures like Llama 3, Gemma, and Mistral rely on GQA (Grouped-Query Attention) or MQA to optimize memory bandwidth. Implementing these attention patterns in custom inference engines (C/C++, Vulkan, etc.) frequently introduces boundary bugs into KV-cache tensor indexing. This model allows you to thoroughly validate **KV-cache matrix broadcasting logic** under a tight 1M parameter profile without memory overhead. --- ## 📂 Repository Structure & File Descriptions ### 1. GGUF Formats (Root Directory `./`) A complete suite compiled for `llama.cpp` and compatible modern custom runtimes. The structural MQA hyper-parameters and specialized token layouts are fully baked into each GGUF binary: | Filename(s) / Wildcard Pattern | Type | Size | Purpose / Validation Target | | :--- | :--- | :--- | :--- | | **`tinymqa1m.F32.gguf`** | `F32` | ~4.0 MB | **Baseline Test.** Validates GGUF parsing, MQA tensor layout, matrix dimensions, and RoPE indexing without dequantization factors. | | **`tinymqa1m.F16.gguf`**
**`tinymqa1m.BF16.gguf`** | `F16`
`BF16` | ~2.0 MB | **Half-Precision Test.** Validates 16-bit float loading, tensor broadcasting, and structural inference stability. | | **`tinymqa1m.Q8_0.gguf`** | `Q8_0` | ~1.1 MB | **Quantization Level 1.** Validates block-based uniform scaling with 32 elements under MQA dimensions. | | **`tinymqa1m.Q4_0.gguf`**
**`tinymqa1m.Q4_1.gguf`** | `Q4_0`
`Q4_1` | ~0.7 MB | **Quantization Level 2.** Validates classic 4-bit linear quantization and bit-unpacking logic. | | **`tinymqa1m.Q2_K.gguf`** | `Q2_K` | ~0.5 MB | **Standard K-Quant (2-bit).** Validates 2-bit super-block quantization parsing. | | **`tinymqa1m.Q3_K_*.gguf`**
↳ *`tinymqa1m.Q3_K_S.gguf`*
↳ *`tinymqa1m.Q3_K_M.gguf`*
↳ *`tinymqa1m.Q3_K_L.gguf`* | `Q3_K` | ~0.6 MB | **Standard K-Quant (3-bit).** Validates Small, Medium, and Large sub-variants of 3-bit multi-block structures. | | **`tinymqa1m.Q4_K_*.gguf`**
↳ *`tinymqa1m.Q4_K_S.gguf`*
↳ *`tinymqa1m.Q4_K_M.gguf`* | `Q4_K` | ~0.7 MB | **Standard K-Quant (4-bit).** Validates Small and Medium sub-variants of modern 4-bit super-block structural parsing. | | **`tinymqa1m.Q5_K_*.gguf`**
↳ *`tinymqa1m.Q5_K_S.gguf`*
↳ *`tinymqa1m.Q5_K_M.gguf`* | `Q5_K` | ~0.8 MB | **Standard K-Quant (5-bit).** Validates Small and Medium sub-variants of 5-bit mixed precision super-blocks. | | **`tinymqa1m.Q6_K.gguf`** | `Q6_K` | ~0.9 MB | **Standard K-Quant (6-bit).** Validates 6-bit high-fidelity super-block quantization. | | **`tinymqa1m.IQ3_*.gguf`**
↳ *`tinymqa1m.IQ3_XXS.gguf`*
↳ *`tinymqa1m.IQ3_S.gguf`* | `I-Quants` | ~0.5 MB | **Importance Quants (3-bit).** Non-linear 3-bit importance quantization targeting lookup table (codebook) decoding logic. | | **`tinymqa1m.IQ4_*.gguf`**
↳ *`tinymqa1m.IQ4_NL.gguf`*
↳ *`tinymqa1m.IQ4_XS.gguf`* | `I-Quants` | ~0.6 MB | **Importance Quants (4-bit).** Non-linear 4-bit importance quantization variants (Non-Linear and Extra Small). | | **`tinymqa1m.TQ1_0.gguf`**
**`tinymqa1m.TQ2_0.gguf`** | `Ternary` | ~0.4 MB | **Experimental.** Ternary (-1, 0, 1) state quantization for cutting-edge engine testing. | ### 2. Hugging Face Native Format (`./hf/`) Standard configurations and weight layer states used by the PyTorch `transformers` library: * **`hf/model.safetensors`**: Unquantized native model parameters using explicit MQA structures. * **`hf/config.json`**: Architectural settings specifying the asymmetrical head layout (`num_attention_heads: 4`, `num_key_value_heads: 1`). * **`hf/generation_config.json`**: Default generation threshold boundaries. * **`hf/tokenizer_config.json`**: Tokenizer behavior configuration enabling automatic `` (BOS) injection and sequence padding boundaries. * **`hf/special_tokens_map.json`**: Token mappings string keys directly to internal special token IDs. * **`hf/tokenizer.model`**: The master 512-vocab SentencePiece tokenizer binary file. --- ## 🚀 Usage Examples ### A. Running GGUF via llama.cpp To verify your local hardware runtime execution or evaluate token generation logic under MQA parameters: ```bash ./llama-cli -m tinymqa1m.Q4_K_M.gguf -p "Tom and Jerry are " -n 64 --temp 0.0 ``` ### B. Loading Hugging Face Formats via Python With the runtime metadata (`tokenizer_config.json` / `special_tokens_map.json`) fully populated, you can instantiate the configuration directly using standard Hugging Face components without custom workflow wrappers. ```python import torch from transformers import AutoTokenizer, AutoModelForCausalLM repo_id = "shibatch/tinymqa1m" print("Loading tokenizer and MQA model configuration...") tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="hf") model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder="hf") device = "cuda" if torch.cuda.is_available() else "cpu" model = model.to(device) model.eval() prompt = "Tom and Jerry are " # Formatting and (BOS) insertion are handled automatically via configuration metadata inputs = tokenizer(prompt, return_tensors="pt").to(device) print("Executing text generation loop (Validating MQA projection tensors)...") with torch.no_grad(): outputs = model.generate( **inputs, max_length=64, do_sample=False ) generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True) print("\n--- Inference Test Result ---") print("Prompt :", prompt) print("Generated:", generated_text) ``` --- ## 📝 Model Specifications The network scales the attention pipeline to map 4 Query channels down to 1 Key-Value pair, verifying structural broadcasting implementations cleanly. * **Architecture:** Llama 2 with **Multi-Query Attention (MQA)** * **Dataset:** TinyStories * **Total Parameters:** ~1M (Exactly 896,256 parameters) * **Vocabulary Size:** 512 (Custom SentencePiece BPE with `byte_fallback` enabled) * **Hidden Size (`hidden_size`):** 128 * **Number of Hidden Layers (`num_hidden_layers`):** 4 * **Number of Attention Heads (`num_heads`):** 4 *(head_dim = 32)* * **Number of Key-Value Heads (`num_kv_heads`):** 1 *(Strict MQA broadcast ratio)* * **Intermediate Size (`intermediate_size`):** 352 * **Max Position Embeddings (`max_position_embeddings`):** 256 ## 📜 Acknowledgments & License * **Original Implementation:** Inspired by Andrej Karpathy's `llama2.c` project. * **Dataset:** TinyStories dataset. * **License:** **MIT License**. You are free to use, modify, and distribute these assets for any purpose.