🥷 Safurai-Csharp-34B

This is a codellama/CodeLlama-34b-hf model fine-tuned using QLoRA (4-bit precision) on 13B tokens of csharp evolved Q&A

We obtained state-of-the-art performance on the MultiPL-E code LLM benchmark for csharp, reaching 56% at pass@1 with n=5.

💻 Quantization

This the AWQ quantized version of Safurai-Csharp-34B, it has been made by using the amazing AutoAWQ library.

🔧 Training

It was trained on 2 x NVIDIA A100 PCIe 80GB in 7h 40m with the following configuration file:

base_model: codellama/CodeLlama-34b-hf
base_model_config: codellama/CodeLlama-34b-hf
model_type: LlamaForCausalLM
tokenizer_type: CodeLlamaTokenizer
is_llama_derived_model: true
hub_model_id: "Safurai/Evol-csharp-v1"

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Safurai/EvolInstruct-csharp-16k-13B-Alpaca
    type: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./qlora-out

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: codellama-csharp
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0003

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 40
eval_steps: 40
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

📉 Training loss curve:

📊 Dataset composition:

💻 Usage for AWQ

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer, TextStreamer

quant_path = "Safurai/Safurai-Csharp-34B-AWQ"

# Load model
model = AutoAWQForCausalLM.from_quantized(quant_path, fuse_layers=True)
tokenizer = AutoTokenizer.from_pretrained(quant_path, trust_remote_code=True)
streamer = TextStreamer(tokenizer, skip_special_tokens=True)

# Convert prompt to tokens
prompt_template = """\
A chat between a developer and an AI assistant. The assistant is an expert csharp programmer that can give useful and complete code responses.

USER: {prompt}
ASSISTANT:"""

tokens = tokenizer(
    prompt_template.format(prompt="How are you today?"), 
    return_tensors='pt'
).input_ids.cuda()

# Generate output
generation_output = model.generate(
    tokens, 
    streamer=streamer,
    max_new_tokens=1024
)