--- datasets: - tiiuae/falcon-refinedweb - HuggingFaceFW/fineweb-edu language: - en license: - other license_name: falcon-mamba-7b-license license_link: https://falconllm.tii.ae/falcon-mamba-7b-terms-and-conditions.html base_model: tiiuae/falcon-mamba-7b --- drawing **Make sure to install `bitsandbytes` and have a GPU compatible with `bitsandbytes` to run this model** # Table of Contents 0. [TL;DR](#TL;DR) 1. [Model Details](#model-details) 2. [Usage](#usage) 3. [Training Details](#training-details) 4. [Evaluation](#evaluation) # TL;DR # Model Details ## Model Description - **Developed by:** [https://www.tii.ae](https://www.tii.ae) - **Model type:** Causal decoder-only - **Architecture:** Mamba - **Language(s) (NLP):** Mainly English - **License:** TII Falcon-Mamba License 2.0 ### Model Source - **Paper:** *coming soon*. # Usage Find below some example scripts on how to use the model in `transformers` (Make sure to have the latest transformers, or the one built from source): ## Using the Pytorch model This checkpoint will only run on a GPU device with `bitsandbytes` installed. See below for more details on how to load it
Click to expand ```python from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-4bit") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-4bit") input_text = "Question: How many hours in one day? Answer: " input_ids = tokenizer(input_text, return_tensors="pt").input_ids outputs = model.generate(input_ids) print(tokenizer.decode(outputs[0])) ```
You can also dequantize the model with `model.dequantize()` method:
Click to expand ```python from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-4bit") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-4bit") model = model.dequantize() input_text = "Question: How many hours in one day? Answer: " input_ids = tokenizer(input_text, return_tensors="pt").input_ids outputs = model.generate(input_ids) print(tokenizer.decode(outputs[0])) ```
# Training Details ## Training Data Falcon-Mamba has been trained with ~ 5,500 GT mainly coming from [Refined-Web](https://huggingface.co/datasets/tiiuae/falcon-refinedweb), a large volume web-only dataset filtered and deduplicated. Similar to the others [Falcon](https://huggingface.co/tiiuae/falcon-11B) suite models, Falcon-Mamba has been trained leveraging a multi-stage training strategy to increase the context-length training from 2,048 up to 8,192. Note that at inference the context-length is not relevant as the Mamba architecture has no limit on long range dependency. At the last training stage, small portion of high-quality curated data was used to further enhance performance. Overall, the data sources included RefinedWeb-English, high quality technical data, code data and conversational data extracted from public sources. In particular, we used samples coming from [Fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu). The data was tokenized with the Falcon-[7B](https://huggingface.co/tiiuae/falcon-7B)/[11B](https://huggingface.co/tiiuae/falcon-11B) tokenizer. ## Training Procedure Falcon-Mamba-7B was trained on 256 H100 80GB GPUs for the majority of the training, using a 3D parallelism strategy (TP=1, PP=1, DP=256) combined with ZeRO. #### Training Hyperparameters | **Hyperparameter** | **Value** | **Comment** | |--------------------|------------|-------------------------------------------| | Precision | `bfloat16` | | | Optimizer | AdamW | | | Max learning rate | 6.4e-4 | Following a WSD (warmup-stable-decay) learning rate schedule | | Weight decay | 1e-1 | | | Batch size | 2048 | | The model was trained AdamW optimizer, WSD (warmup-stable-decay) learning rate schedule, and a batch size rampup from \\(b_{\mathrm{min}}=128\\) to \\(b_{\mathrm{max}}=2048\\) during first 50 GT of training. In the stable phase we used maximal learning rate \\(\eta_{\mathrm{max}}=6.4 \times 10^{-4}\\), and decayed it to the minimal value \\(\eta_{\mathrm{min}}=\frac{\eta_{\mathrm{max}}}{256}\\) with exponential schedule over 500 GT. Also, we applied *BatchScaling* during the rampup — rescaling learning rate \\(\eta\\) so that the Adam noise temperature \\(T_{\mathrm{noise}}\equiv\frac{\eta}{\sqrt{b}}\\) is kept constant. #### Speeds, Sizes, Times The model training took roughly two months. # Evaluation ## Benchmarks We evaluate our model on all benchmarks of the leaderboard's version 2 using the `lm-evaluation-harness` package, and we evaluate it on the benchmarks of version 1 using `lighteval`. | `model name` |`IFEval`| `BBH` |`MATH LvL5`| `GPQA`| `MUSR`|`MMLU-PRO`|`Average`| |:--------------------------|:------:|:-----:|:---------:|:-----:|:-----:|:--------:|:-------:| | ***Pure SSM models*** | | | | | | | | | `Falcon-Mamba-7B` | 33.36 | 19.88 | 3.63 | 8.05 | 10.86 | 14.47 |**15.04**| | `TRI-ML/mamba-7b-rw` | 22.46 | 6.71 | 0.45 | 1.12 | 5.51 | 1.69 | 6.25 | |***Hybrid SSM-attention models*** | | | | | | | | `Zamba-7B-v1` | 24.06 | 21.12 | 3.32 | 3.03 | 7.74 | 16.02 | 12.55 | |`recurrentgemma-9b` | 30.76 | 14.80 | 4.83 | 4.70 | 6.60 | 17.88 | 13.20 | |***Transformer models*** | | | | | | | | | `Falcon2-11B` | 32.61 | 21.94 | 2.34 | 2.80 | 7.53 | 15.44 | 13.78 | | `Meta-Llama-3-8B` | 14.55 | 24.50 | 3.25 | 7.38 | 6.24 | 24.55 | 13.41 | | `gemma-7B` | 26.59 | 21.12 | 6.42 | 4.92 | 10.98 | 21.64 |**15.28**| | `Mistral-7B-v0.1` | 23.86 | 22.02 | 2.49 | 5.59 | 10.68 | 22.36 | 14.50 | | `Mistral-Nemo-Base` | 16.83 | 29.37 | 4.98 | 5.82 | 6.52 | 27.46 | 15.08 | | `model name` |`ARC`|`HellaSwag` |`MMLU` |`Winogrande`|`TruthfulQA`|`GSM8K`|`Average` | |:-----------------------------|:------:|:---------:|:-----:|:----------:|:----------:|:-----:|:----------------:| | ***Pure SSM models*** | | | | | | | | | `Falcon-Mamba-7B` |62.03 | 80.82 | 62.11 | 73.64 | 53.42 | 52.54 | **64.09** | | `TRI-ML/mamba-7b-rw` | 46.48 | 80.24 | 57.72 | 76.40 | - | 4.70 | - | |***Hybrid SSM-attention models***| | | | | | | | | `recurrentgemma-9b` |52.00 | 80.40 | 60.50 | 73.60 | 38.60 | 42.60 | 57.95 | | `Zyphra/Zamba-7B-v1` | 46.48 | 80.24 | 57.72 | 76.40 | - | 30.78 | - | |***Transformer models*** | | | | | | | | | `Falcon2-11B` | 59.73 | 82.91 | 58.37 | 78.30 | 52.56 | 53.83 | **64.28** | | `Meta-Llama-3-8B` | 60.24 | 82.23 | 66.70 | 78.45 | 42.93 | 45.19 | 62.62 | | `gemma-7B` | 61.09 | 82.20 | 64.56 | 79.01 | 44.79 | 50.87 | 63.75 | | `Mistral-7B-v0.1` | 59.98 | 83.31 | 64.16 | 78.37 | 42.15 | 37.83 | 60.97 | ## Throughput This model can achieve comparable throughput and performance compared to other transformer based models that use optimized kernels such as Flash Attention 2. Make sure to install the optimized Mamba kernels with the following commands: ```bash pip install "causal-conv1d>=1.4.0" mamba-ssm ``` Refer to our [FalconMamba blogpost](https://huggingface.co/blog/falconmamba) for more details about performance evaluation. # Technical Specifications ## Model Architecture and Objective Falcon-Mamba-7B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token). The model is based on the Mamba architecture ([Gu et al., 2023](https://arxiv.org/abs/2312.00752)). | **Hyperparameter** | **Value** | **Comment** | |--------------------|-----------|----------------------------------------| | Layers | 64 | Number of layers | | `d_model` | 4096 | Hidden dimension | | `d_state` | 16 | The SSM state dimension | | Vocabulary | 65024 | Vocabulary Size | | Sequence length | 8192 | During stages 4 and LR Decay stage | ## Compute Infrastructure ### Hardware Falcon-Mamba-7B was trained on AWS SageMaker, using on average 256 H100 80GB GPUs in 32 p5 instances. ### Software Falcon-Mamba-7B was trained an internal distributed training codebase, Gigatron. It uses a 3D parallelism approach combined with ZeRO, high-performance Triton kernels. # Citation *Paper coming soon* 😊.