README.md

metadata

datasets:
  - tiiuae/falcon-refinedweb
  - HuggingFaceFW/fineweb-edu
language:
  - en
license: apache-2.0

Model Card for Falcon-Mamba-7B

Table of Contents

0. TL;DR
1. Model Details
2. Usage
3. Training Details
4. Evaluation

TL;DR

Model Details

Model Description

• Developed by: 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

Running the model on a CPU

Click to expand

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b")

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]))

Running the model on a GPU

Click to expand

# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b", device_map="auto")

input_text = "Question: How many hours in one day? Answer: "
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))

Running the model on a GPU using different precisions

FP16

Click to expand

# pip install accelerate
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b", device_map="auto", torch_dtype=torch.float16)

input_text = "Question: How many hours in one day? Answer: "
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))

4-bit

Click to expand

# pip install bitsandbytes accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True))

input_text = "Question: How many hours in one day? Answer: "
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))

Training Details

Training Data

Falcon-Mamba has been trained with ~ 6,000 GT mainly coming from Refined-Web, a large volume web-only dataset filtered and deduplicated. Similar to the others Falcon 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.

The data was tokenized with the Falcon-7B/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 $b_{\mathrm{m}\mathrm{i}\mathrm{n}}=128b\_\{\backslash mathrm\{min\}\}=128$ to $b_{\mathrm{m}\mathrm{a}\mathrm{x}}=2048b\_\{\backslash mathrm\{max\}\}=2048$ during first 50 GT of training. In the stable phase we used maximal learning rate ${\eta }_{\mathrm{m}\mathrm{a}\mathrm{x}}=6.4\times 10^{-4}\backslash eta\_\{\backslash mathrm\{max\}\}=6.4\; \backslash times\; 10^\{-4\}$, and decayed it to the minimal value ${\eta }_{\mathrm{m}\mathrm{i}\mathrm{n}}=\frac{{\eta }_{\mathrm{m}\mathrm{a}\mathrm{x}}}{256}\backslash eta\_\{\backslash mathrm\{min\}\}=\backslash frac\{\backslash eta\_\{\backslash mathrm\{max\}\}\}\{256\}$ with exponential schedule over 500 GT. Also, we applied BatchScaling during the rampup — rescaling learning rate $\eta \backslash eta$ so that the Adam noise temperature $T_{\mathrm{n}\mathrm{o}\mathrm{i}\mathrm{s}\mathrm{e}}\equiv \frac{\eta }{\sqrt{b}}T\_\{\backslash mathrm\{noise\}\}\backslash equiv\backslash frac\{\backslash eta\}\{\backslash sqrt\{b\}\}$ 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 L2	ARC	HellaSwag	MMLU	Winogrande	TruthfulQA	GSM8K	Average L1
meta-llama/Meta-Llama-3-8B	14.55	24.50	3.25	7.38	6.24	24.55	13.41	60.24	82.23	66.70	78.45	42.93	45.19	62.62
tiiuae/falcon2-11B	32.61	21.94	2.34	2.8	7.53	15.44	13.78	59.73	82.91	58.37	78.30	52.56	53.83	64.28
mistralai/Mistral-7B-v0.1	23.86	22.02	2.49	5.59	10.68	22.36	14.50	59.98	83.31	64.16	78.37	42.15	37.83	60.97
Zyphra/Zamba-7B-v1	-	-	-	-	-	-	-	46.48	80.24	57.72	76.4	-	-	-
Ours	32.16	21.07	4.08	10.18	6.97	13.43	14.65	61.69	80.63	61.05	74.03	53.60	51.86	63.81

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:

pip install "causal-conv1d>=1.4.0" mamba-ssm

Refer to our technical report 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).

Hyperparameter	Value	Comment
Layers	64
d_model	4096
d_state	16	The SSM state dimension
Vocabulary	65024
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 😊.