Coercing NVIDA bfloat16 LLM models to run on Tesla GPUs that only support float16

Newer LLM models are built around the bfloat16 data type that has different types of precision than the older float16. My Tesla vintage GPU supports the lower precision float16, not the newer bfloat16.  you can coerce the model from bfloat16 to float16. 

Disclaimer: The difference in precision can result in errors. Your mileage and accuracy may vary depending on the model 

Hacking config.json for FP16

The basic steps are 
  1. Download the model either manually or as part of an attempted run.  
  2. Find the location of the model on disk. This typically is in the model cache directory
    1. ~/.cache/nvidia/nvidia-nims/ngc/<some_path>/config.json
    2. ~/.cache/nim/ngc/<some_path>/config.json
    3. ~/.cache/nvidia-nims/ngc/hub/<some-model>/snapshots/config.json
  3. Edit the config.json found for the model.
  4. Find the line "torch_dtype":"bfloat16"
  5. Change the value of torch_dtype to one supported by the card, float16 in my case. "torch_dtype":"float16".
The property is torch.dtype on the command line but is torch_dtype in JSON files.

sample config.json 

You can see "torch_dtype": "bfloat16",  partway down the file list

(base) joe@hp-z820:~$ cat ./.cache/nim/ngc/hub/models--nim--mistralai--mistral-7b-instruct-v03/snapshots/hf/config.json
{
  "architectures": [
    "MistralForCausalLM"
  ],
  "attention_dropout": 0.0,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 32768,
  "model_type": "mistral",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "rms_norm_eps": 1e-05,
  "rope_theta": 1000000.0,
  "sliding_window": null,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.42.0.dev0",
  "use_cache": true,
  "vocab_size": 32768
}


Now we can walk through a few model definition cards to see if they fit in our card and are a candidate for dtype coercion.  Some of the model cards show that they are float16 but they download with bfloat16 configuration files.  We can overcome that with the cache edit.

NIM Anywhere Models

The NIM Anywhere project ships with a sample service that uses a model that expects to run bfloat16. 
I was able to load the LLM model by changing the torch_dtype.  I validated the model configuration with a few simple queries. I didn't do any precision verification testing.

Model meta-llama/Meta-Llama-3-8B-Instruct

  • Model: meta-llama/Meta-Llama-3-8B-Instruct
  • config.json location: ~/.cache/nvidia-nims/ngc/hub/models--nim--meta--llama3-8b-instruct/snapshots/hf/config.json
  • torch_dtype: bfloat16
  • model weight size: 15GB

Modifications to run on Tesla GPUs
  • torch_dtype: float16




Hybrid RAG Ungated Models

These are the six local models supported by the NVIDIA AI Workstation Hybrid Rag example.  Two are immediately available.  The other four are gated, behind a token Hugging Face token wall.

nvidia/Llama3-ChatQA-1.5-8b

  • Model: nvidia/Llama3-ChatQA-1.5-8b
  • Mode Location:  /data/models--nvidia--Llama3-ChatQA-1.5-8B
  • config.json location: /data/models--nvidia--Llama3-ChatQA-1.5-8B/snapshots/3b98162e3f97550d62aeeb19ea50208f968c678a/config.json
  • torch_dtype: float16


Phi-3-mini-128k-instruct

  • Model: Phi-3-mini-128k-instruct
  • ModelLocation: /data/models--microsoft--Phi-3-mini-128k-instruct/
  • config.json location: /data/models--microsoft--Phi-3-mini-128k-instruct/snapshots/d548c233192db00165d842bf8edff054bb3212f8/config.json
  • torch_dtype in config.json: bfloat16




Hybrid Rag Gated Models Supported in AI Workbench

Most of these models are BFloat16.  The first one is FP16 so it would be runnable without modification on a Tesla CPU.  This currently doesn't work in the AI Workbench. I think this is because NVidia erroneously set the torch_dtype to bf16 no matter what the model card says.

Llama-2-7b--chat-hf

* Model: Llama-2-7b--chat-hf
* torch_dtype in config.json: not investigated



Meta-Llama-3-8b-Instruct

* Model: Meta-Llama-3-8b-Instruct
* torch_dtype in config.json: blfoat16


Mistral-7B-Instruct-v0.1

* Model: Mistral-7B-Instruct-v0.1
* torch_dtype in config.json: bfloat16



Mistral-7B-Instruct-v0.2

* Model: Mistral-7B-Instruct-v0.2
* torch_dtype in config.json: bfloat16




Other

Revision History

Create 2024 08

Comments

Post a Comment

Popular posts from this blog

Installing the RNDIS driver on Windows 11 to use USB Raspberry Pi as network attached

Image
I do a lot of my development and configuration via ssh into my Raspberry Pi Zero over the RNDIS connection. Some models of the Raspberry PIs can be configured with  gadget drivers  that let the Raspberry pi emulate different devices when plugged into computers via USB. My favorite gadget  is the network profile that makes a Raspberry Pi look like an RNDIS-attached network device.  All types of network services travel over an RNDIS device without knowing it is a USB hardware connection. A Raspberry Pi shows up as a Remote NDIS (RNDIS) device when you plug the Pi into a PC or Mac via a USB cable. The gadget  in the Windows Device Manager picture shows this RNDIS Gadget connectivity between a Windows machine and a Raspberry Pi. The Problem Windows 11 and Windows 10 no longer auto-installs the RNDIS driver that makes magic happen. Windows recognizes that the Raspberry Pi is some type of generic USB COM device. Manually running W indows Update  or  Update Driver  does  not  install the RNDI
Read more

Understanding your WSL2 RAM and swap - Changing the default 50%-25%

Image
The Windows Subsystem for Linux operates as a virtual machine that can dynamically grow the amount of RAM to a maximum set at startup time.  Microsoft sets a default maximum RAM available to 50% of the physical memory and a swap-space that is 1/4 of the maximum WSL RAM.  You can scale those numbers up or down to allocate more or less RAM to the Linux instance. The first drawing shows the default WSL memory and swap space sizing. The images below show a developer machine that is running a dev environment in WSL2 and Docker Desktop.  Docker Desktop has two of its own WSL modules that need to be accounted for.  You can see that the memory would actually be oversubscribed, 3 x 50% if every VM used its maximum memory.  The actual amount of memory used is significantly smaller allowing every piece to fit. Click to Enlarge The second drawing shows the memory allocation on my 64GB laptop. WSL Linux defaults to a maximum RAM size of 5
Read more

Almost PaaS Document Parsing with Tika and AWS Elastic Beanstalk

Image
The Apache Tika project provides a  library capable of parsing and extracting data and meta data from over 1000 file types.  Tika is available as a single jar file that can be included inside applications or as a deployable jar file that runs Tika as a standalone service. This blog describes deploying the Tika jar as an auto-scale service in Amazon AWS Elastic Beanstalk.  I selected Elastic Beanstalk because it supports jar based deployments without any real Infrastructure configuration. Elastic Beanstalk auto-scale should take care of scaling up and down for for the number of requests you get. Tika parses documents and extracts their text completely in memory. Tika was deployed for this blog using EC2 t2.micro instances available in the AWS free tier. t2.micro VMs are 1GB which means that you are restricted in document complexity and size. You would size your instances appropriately for your largest documents.   Preconditions An AWS account. AWS access id and secret key.
Read more