Please refrain from using this model yet. It's not any weight at all.
A experts weights of Jamba-v0.1
Required Weights for follow-up research.
The original model is AI21lab's Jamba-v0.1, which requires an >80GB VRAM. Unfortunately, this almonst was not available via Google Colab or cloud computing services. Thus, attempts were made to perform MoE (Mixture of Experts) splitting, using the following resources as a basis:
- Original Model: Jamba-v0.1
- MoE Layer Separation: Consult this script written by @TechxGenusand and use TechxGenus/Jamba-v0.1-9B.
Original Model Card from AI21lab's Jamba-v0.1.
Usage
The code used in AI21lab's Jamba-v0.1.
Presequities
To use Jamba, ensure you have transformers
version 4.40.0 or higher installed (version 4.39.0 or higher is required):
pip install transformers>=4.40.0
For optimized Mamba implementations, install mamba-ssm
and causal-conv1d
:
pip install mamba-ssm causal-conv1d>=1.2.0
Ensure the model is on a CUDA device.
You can run the model without optimized Mamba kernels, but it's not recommended due to significantly lower latencies. To do so, specify use_mamba_kernels=False
when loading the model.
Run the model
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("danielpark/asp-9b-inst-base")
tokenizer = AutoTokenizer.from_pretrained("danielpark/asp-9b-inst-base")
input_ids = tokenizer("In the recent Super Bowl LVIII,", return_tensors='pt').to(model.device)["input_ids"]
outputs = model.generate(input_ids, max_new_tokens=216)
print(tokenizer.batch_decode(outputs))
# ["In the recent Super Bowl LVIII, the Kansas City Chiefs emerged victorious, defeating the San Francisco 49ers in a thrilling overtime showdown. The game was a nail-biter, with both teams showcasing their skills and determination.\n\nThe Chiefs, led by their star quarterback Patrick Mahomes, displayed their offensive prowess, while the 49ers, led by their strong defense, put up a tough fight. The game went into overtime, with the Chiefs ultimately securing the win with a touchdown.\n\nThe victory marked the Chiefs' second Super Bowl win in four years, solidifying their status as one of the top teams in the NFL. The game was a testament to the skill and talent of both teams, and a thrilling end to the NFL season.\n\nThe Super Bowl is not just about the game itself, but also about the halftime show and the commercials. This year's halftime show featured a star-studded lineup, including Usher, Alicia Keys, and Lil Jon. The show was a spectacle of music and dance, with the performers delivering an energetic and entertaining performance.\n"]
When using transformers<4.40.0
, ensure trust_remote_code=True
for running the new Jamba architecture.
Loading the model in half precision
The published checkpoint is saved in BF16. To load it into RAM in BF16/FP16, specify torch_dtype
:
from transformers import AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained("danielpark/asp-9b-inst-base",
torch_dtype=torch.bfloat16) # you can also use torch_dtype=torch.float16
When using half precision, enable the FlashAttention2 implementation of the Attention blocks. To use it, ensure the model is on a CUDA device. Since the model is too big to fit on a single 80GB GPU, parallelize it using accelerate:
from transformers import AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained("danielpark/asp-9b-inst-base",
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map="auto")
Load the model in 8-bit
Using 8-bit precision, up to 140K sequence lengths can fit on a single 80GB GPU. Quantize the model to 8-bit using bitsandbytes. To exclude Mamba blocks from quantization to prevent model quality degradation:
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True,
llm_int8_skip_modules=["mamba"])
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
quantization_config=quantization_config)
Fine-tuning example
Jamba is a base model that can be fine-tuned for custom solutions (including for chat/instruct versions). Fine-tune it using any technique of your choice. Here's an example of fine-tuning with the PEFT library:
from datasets import load_dataset
from trl import SFTTrainer
from peft import LoraConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
tokenizer = AutoTokenizer.from_pretrained("danielpark/asp-9b-inst-base")
model = AutoModelForCausalLM.from_pretrained("danielpark/asp-9b-inst-base", device_map='auto')
dataset = load_dataset("Abirate/english_quotes", split="train")
training_args = TrainingArguments(
output_dir="./results",
num_train_epochs=3,
per_device_train_batch_size=4,
logging_dir='./logs',
logging_steps=10,
learning_rate=2e-3
)
lora_config = LoraConfig(
r=8,
target_modules=["embed_tokens", "x_proj", "in_proj", "out_proj"],
task_type="CAUSAL_LM",
bias="none"
)
trainer = SFTTrainer(
model=model,
tokenizer=tokenizer,
args=training_args,
peft_config=lora_config,
train_dataset=dataset,
dataset_text_field="quote",
)
trainer.train()
Further
Check ai21labs/Jamba-tiny-random, which has 128M parameters (instead of 52B), and is initialized with random weights and did not undergo any training.
- Downloads last month
- 28