--- tags: - fp8 - vllm license: other license_name: deepseek-license license_link: https://github.com/deepseek-ai/DeepSeek-Coder-V2/blob/main/LICENSE-MODEL --- # DeepSeek-Coder-V2-Base-FP8 ## Model Overview - **Model Architecture:** DeepSeek-Coder-V2-Base - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** FP8 - **Activation quantization:** FP8 - **Intended Use Cases:** Intended for commercial and research use in English. Similarly to [Meta-Llama-3-7B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-7B-Instruct), this models is intended for assistant-like chat. - **Out-of-scope:** Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English. - **Release Date:** 7/22/2024 - **Version:** 1.0 - **License(s):** [deepseek-license](https://github.com/deepseek-ai/DeepSeek-Coder-V2/blob/main/LICENSE-MODEL) - **Model Developers:** Neural Magic Quantized version of [DeepSeek-Coder-V2-Base](https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Base). It achieves an average score of 80.55 on the [HumanEval+](https://github.com/openai/human-eval?tab=readme-ov-file) benchmark, whereas the unquantized model achieves 79.90. ### Model Optimizations This model was obtained by quantizing the weights and activations of [DeepSeek-Coder-V2-Base](https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Base) to FP8 data type, ready for inference with vLLM >= 0.5.2. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. In particular, this model can now be loaded and evaluated with only 4xH100 GPUs, as opposed to 8. Only the weights and activations of the linear operators within transformers blocks are quantized. Symmetric per-tensor quantization is applied, in which a single linear scaling maps the FP8 representations of the quantized weights and activations. [AutoFP8](https://github.com/neuralmagic/AutoFP8) is used for quantization with 512 sequences of UltraChat. ## Deployment ### Use with vLLM This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below. ```python from transformers import AutoTokenizer from vllm import LLM, SamplingParams max_model_len, tp_size = 4096, 4 model_name = "neuralmagic/DeepSeek-Coder-V2-Base-FP8" tokenizer = AutoTokenizer.from_pretrained(model_name) llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True, enforce_eager=True) sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id]) messages_list = [ [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}], ] prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list] outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params) generated_text = [output.outputs[0].text for output in outputs] print(generated_text) ``` vLLM aslo supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details. ## Creation This model was created by applying [AutoFP8 with calibration samples from ultrachat](https://github.com/neuralmagic/AutoFP8/blob/147fa4d9e1a90ef8a93f96fc7d9c33056ddc017a/example_dataset.py) with expert gates kept at original precision, as presented in the code snipet below. Notably, a custom device map had to be used, as the model was being incorrectly loaded otherwise. Although AutoFP8 was used for this particular model, Neural Magic is transitioning to using [llm-compressor](https://github.com/vllm-project/llm-compressor) which supports several quantization schemes and models not supported by AutoFP8. ```python from datasets import load_dataset from transformers import AutoTokenizer from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig pretrained_model_dir = "deepseek-ai/DeepSeek-Coder-V2-Base" quantized_model_dir = "DeepSeek-Coder-V2-Base-FP8" tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True, model_max_length=4096) tokenizer.pad_token = tokenizer.eos_token ds = load_dataset("mgoin/ultrachat_2k", split="train_sft").select(range(512)) examples = [tokenizer.apply_chat_template(batch["messages"], tokenize=False) for batch in ds] examples = tokenizer(examples, padding=True, truncation=True, return_tensors="pt").to("cuda") quantize_config = BaseQuantizeConfig( quant_method="fp8", activation_scheme="static" ignore_patterns=["re:.*lm_head"], ) device_map = { "model.embed_tokens": 0, "model.layers.0": 0, } for i in range(1, 60): device_map[f"model.layers.{i}"] = i//8 device_map["model.norm"] = 7 device_map["lm_head"] = 7 model = AutoFP8ForCausalLM.from_pretrained( pretrained_model_dir, quantize_config=quantize_config, device_map = device_map ) model.quantize(examples) model.save_quantized(quantized_model_dir) ``` ## Evaluation The model was evaluated on the [HumanEval+](https://github.com/openai/human-eval?tab=readme-ov-file) benchmark with the [Neural Magic fork](https://github.com/neuralmagic/evalplus) of the [EvalPlus implementation of HumanEval+](https://github.com/evalplus/evalplus) and the [vLLM](https://docs.vllm.ai/en/stable/) engine, using the following command: ``` python codegen/generate.py --model neuralmagic/DeepSeek-Coder-V2-Base-FP8 --temperature 0.2 --n_samples 50 --resume --root ~ --dataset humaneval python evalplus/sanitize.py ~/humaneval/neuralmagic--DeepSeek-Coder-V2-Base-FP8_vllm_temp_0.2 evalplus.evaluate --dataset humaneval --samples ~/humaneval/neuralmagic--DeepSeek-Coder-V2-Base-FP8_vllm_temp_0.2-sanitized ``` ### Accuracy #### HumanEval+ evaluation scores
Benchmark DeepSeek-Coder-V2-Base DeepSeek-Coder-V2-Base-FP8(this model) Recovery
base pass@1 78.5 78.5 100.0%
base pass@10 88.4 88.8 100.4%
base+extra pass@1 71.3 72.1 101.1%
base+extra pass@10 81.4 82.8 101.7%
Average 79.90 80.55 100.8%