What are the hardware requirements for Granite 20B Code Base?

Granite 20B Code Base requires approximately 40 GB of VRAM. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed Granite 20B Code Base and how large is it?

Granite 20B Code Base is developed by ibm-granite, a model with 20b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

Where can I download or evaluate Granite 20B Code Base?

Granite 20B Code Base is hosted on Hugging Face and linked from this page. LLM Explorer also lists quantized variants and similar alternatives if available.

Granite 20B Code Base by ibm-granite — VRAM 40GB

Name: Granite 20B Code Base
Author: ibm-granite

Granite 20B Code Base is an open-source language model by ibm-granite. Features: 20b LLM, VRAM: 40GB, License: apache-2.0, Code Generating, LLM Explorer Score: 0.14.

Arxiv:2405.04324 Autotrain compatible Code Codegen Dataset:bigcode/starcoderdata Dataset:codeparrot/github-code... Dataset:math-ai/stackmathqa Dataset:open-web-math/open-web... Endpoints compatible Gpt bigcode Granite Model-index Region:us Safetensors Sharded Tensorflow

Model Card on HF 🤗: https://huggingface.co/ibm-granite/granite-20b-code-base

Granite 20B Code Base Parameters and Internals

Model Type

code generation, decoder-only, text-generation

Use Cases

Areas:

enterprise use, software engineering productivity

Applications:

code generation, code explanation, code fixing, generating unit tests, generating documentation, addressing technical debt issues, vulnerability detection, code translation

Limitations:

Risks of problematic outputs, No safety alignment, Increased susceptibility to hallucination

Considerations:

Caution against complete reliance for crucial decisions

Supported Languages

116 programming languages (comprehensive)

Training Details

Data Sources:

Publicly available datasets from GitHub Code Clean, Starcoder data

Data Volume:

3 trillion tokens (Phase 1), 500 billion tokens (Phase 2)

Methodology:

Two-phase training strategy (comprehensive understanding, improved reasoning)

Hardware Used:

IBM's Vela and Blue Vela supercomputing clusters, NVIDIA A100 and H100 GPUs

Model Architecture:

Decoder-only code model

Safety Evaluation

Risk Categories:

malicious utilization, unsafe code generation

Ethical Considerations:

The generated code is not guaranteed to work as intended, risks of malicious use.

Responsible Ai Considerations

Mitigation Strategies:

HAP, PII, Malware Filtering

Release Notes

Date:

May 6th, 2024

Notes:

Model released with decoder-only architecture suited for code generative tasks.

LLM Name	Granite 20B Code Base
Repository 🤗	https://huggingface.co/ibm-granite/granite-20b-code-base
Model Size	20b
Required VRAM	40 GB
Updated	2024-09-01
Maintainer	ibm-granite
Model Type	gpt_bigcode
Model Files	5.0 GB: 1-of-9 4.9 GB: 2-of-9 4.9 GB: 3-of-9 4.9 GB: 4-of-9 4.9 GB: 5-of-9 4.9 GB: 6-of-9 4.9 GB: 7-of-9 4.9 GB: 8-of-9 0.7 GB: 9-of-9
Generates Code	Yes
Model Architecture	GPTBigCodeForCausalLM
License	apache-2.0
Model Max Length	9223372036854775807
Transformers Version	4.38.1
Tokenizer Class	GPT2Tokenizer
Padding Token	<\|endoftext\|>
Vocabulary Size	49152
Torch Data Type	bfloat16
Activation Function	gelu_pytorch_tanh

Rank the Granite 20B Code Base capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Quantized Models of the Granite 20B Code Base

Model	Likes	Downloads	VRAM
Granite 20B Code Base GGUF	0	27	12 GB

Best Alternatives to Granite 20B Code Base

Best Alternatives	Context / RAM	Downloads	Likes
Granite 20B Code Instruct 8K	0K / 40 GB	9754	43
Granite 20B Code Instruct	0K / 40 GB	10209	30
Granite 20B Functioncalling	0K / 40 GB	210	36
Granite 20B Code Base 8K	0K / 40 GB	1221	14
Granite 20B Code Instruct R1.1	0K / 40 GB	37	1
Granite 20B Code Base R1.1	0K / 40 GB	23	2
Granite 20B Code Base FP8	0K / 20.4 GB	7	0
Granite 20B Code Base GGUF	0K / 12.8 GB	27	0

Note: green Score (e.g. "73.2") means that the model is better than ibm-granite/granite-20b-code-base.