Create README.md

README.md

@@ -0,0 +1,126 @@
+---
+tags:
+- fp8
+- vllm
+---
+
+# gemma-2-9b-it-FP8
+
+## Model Overview
+* <h3 style="display: inline;">Model Architecture:</h3> Based on and identical to the gemma-2-9b-it architecture
+* <h3 style="display: inline;">Model Optimizations:</h3> Weights and activations quantized to FP8
+* <h3 style="display: inline;">Release Date:</h3> July 8, 2024
+* <h3 style="display: inline;">Model Developers:</h3> Neural Magic
+
+gemma-2-9b-it quantized to FP8 weights and activations using per-tensor quantization through the [AutoFP8 repository](https://github.com/neuralmagic/AutoFP8), ready for inference with vLLM >= 0.5.0. 
+Calibrated with 1 repeat of each token in the tokenizer in random order to achieve 100% performance recovery on the Open LLM Benchmark evaluations. 
+Reduces space on disk by ~50%.
+Part of the [FP8 LLMs for vLLM collection](https://huggingface.co/collections/neuralmagic/fp8-llms-for-vllm-666742ed2b78b7ac8df13127).
+
+
+## Usage and Creation
+Produced using AutoFP8 with random tokens as calibration, based on [AutoFP8 with calibration samples from ultrachat](https://github.com/neuralmagic/AutoFP8/blob/147fa4d9e1a90ef8a93f96fc7d9c33056ddc017a/example_dataset.py).
+
+```python
+from datasets import load_dataset
+from transformers import AutoTokenizer
+import numpy as np
+import torch
+
+from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig
+
+MODEL_DIR = "google/gemma-2-9b-it"
+final_model_dir = MODEL_DIR.split("/")[-1]
+
+CONTEXT_LENGTH = 4096
+NUM_SAMPLES = 512
+NUM_REPEATS = 1
+
+pretrained_model_dir = MODEL_DIR
+tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True, model_max_length=CONTEXT_LENGTH)
+tokenizer.pad_token = tokenizer.eos_token
+
+tokenizer_num_tokens = len(list(tokenizer.get_vocab().values()))
+total_token_samples = NUM_REPEATS * tokenizer_num_tokens
+num_random_samp = -(-total_token_samples // CONTEXT_LENGTH)
+
+input_ids = np.tile(np.arange(tokenizer_num_tokens), NUM_REPEATS + 1)[:num_random_samp * CONTEXT_LENGTH]
+np.random.shuffle(input_ids)
+input_ids = input_ids.reshape(num_random_samp, CONTEXT_LENGTH)
+input_ids = torch.tensor(input_ids, dtype=torch.int64).to("cuda")
+
+quantize_config = BaseQuantizeConfig(
+    quant_method="fp8",
+    activation_scheme="static",
+)
+
+examples = input_ids
+
+model = AutoFP8ForCausalLM.from_pretrained(pretrained_model_dir, quantize_config=quantize_config)
+
+model.quantize(examples)
+
+quantized_model_dir = f"{final_model_dir}-FP8"
+model.save_quantized(quantized_model_dir)
+```
+
+Evaluated through vLLM>=0.5.1 with the following script:
+
+```bash
+#!/bin/bash
+
+# Example usage:
+# CUDA_VISIBLE_DEVICES=0 ./eval_openllm.sh "neuralmagic/gemma-2-9b-it-FP8" "tensor_parallel_size=1,max_model_len=4096,add_bos_token=True,gpu_memory_utilization=0.7"
+
+export MODEL_DIR=${1}
+export MODEL_ARGS=${2}
+
+declare -A tasks_fewshot=(
+    ["arc_challenge"]=25
+    ["winogrande"]=5
+    ["truthfulqa_mc2"]=0
+    ["hellaswag"]=10
+    ["mmlu"]=5
+    ["gsm8k"]=5
+)
+
+declare -A batch_sizes=(
+    ["arc_challenge"]="auto"
+    ["winogrande"]="auto"
+    ["truthfulqa_mc2"]="auto"
+    ["hellaswag"]="auto"
+    ["mmlu"]=1
+    ["gsm8k"]="auto"
+)
+
+for TASK in "${!tasks_fewshot[@]}"; do
+    NUM_FEWSHOT=${tasks_fewshot[$TASK]}
+    BATCH_SIZE=${batch_sizes[$TASK]}
+    lm_eval --model vllm \
+        --model_args pretrained=$MODEL_DIR,$MODEL_ARGS \
+        --tasks ${TASK} \
+        --num_fewshot ${NUM_FEWSHOT} \
+        --write_out \
+        --show_config \
+        --device cuda \
+        --batch_size ${BATCH_SIZE} \
+        --output_path="results/${TASK}"
+done
+```
+
+
+## Evaluation
+
+Evaluated on the Open LLM Leaderboard evaluations through vLLM.
+
+### Open LLM Leaderboard evaluation scores
+|                      | gemma-2-9b-it       | neuralmagic/gemma-2-9b-it-FP8<br>(this model) |
+| :------------------: | :----------------------: | :------------------------------------------------: |
+| arc-c<br>25-shot     | 71.50                    | 71.50                                              |
+| hellaswag<br>10-shot | 81.91                    | 81.70                                              |
+| mmlu<br>5-shot       | 72.28                    | 71.99                                              |
+| truthfulqa<br>0-shot | 60.32                    | 60.52                                              |
+| winogrande<br>5-shot | 77.11                    | 78.37                                              |
+| gsm8k<br>5-shot      | 76.26                    | 76.87                                              |
+| **Average<br>Accuracy**  | **73.23**                    |              **73.49**                                      |
+| **Recovery**             | **100%**                     |              **100.36%**                                     |