Doc: Update readme.md (#2083)

Signed-off-by: fengding <feng1.ding@intel.com>

Author: feng-intel

README.md

@@ -32,55 +32,33 @@ support AMD CPU, ARM CPU, and NVidia GPU through ONNX Runtime with limited testi
 * [2024/07] Performance optimizations and usability improvements on [client-side](./docs/source/3x/client_quant.md).
 
 ## Installation
+Choose the necessary framework dependencies to install based on your deploy environment.
 ### Install Framework
-#### Install torch for CPU
-```Shell
-pip install torch --index-url https://download.pytorch.org/whl/cpu
+* [Install intel_extension_for_pytorch for CPU](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/)    
+* [Install intel_extension_for_pytorch for XPU](https://intel.github.io/intel-extension-for-pytorch/xpu/latest/)    
+* [Use Docker Image with torch installed for HPU](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#bare-metal-fresh-os-single-click)    
+  **Note**: There is a version mapping between Intel Neural Compressor and Gaudi Software Stack, please refer to this [table](./docs/source/3x/gaudi_version_map.md) and make sure to use a matched combination.    
+* [Install torch for other platform](https://pytorch.org/get-started/locally)    
+* [Install TensorFlow](https://www.tensorflow.org/install)    
+
+### Install Neural Compressor from pypi
 ```
-#### Use Docker Image with torch installed for HPU
-https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#bare-metal-fresh-os-single-click
-
-> **Note**:
-> There is a version mapping between Intel Neural Compressor and Gaudi Software Stack, please refer to this [table](./docs/source/3x/gaudi_version_map.md) and make sure to use a matched combination.
-
-#### Install torch/intel_extension_for_pytorch for Intel GPU
-https://intel.github.io/intel-extension-for-pytorch/index.html#installation
-
-#### Install torch for other platform
-https://pytorch.org/get-started/locally
-
-#### Install tensorflow
-```Shell
-pip install tensorflow
-```
-
-### Install from pypi
-```Shell
 # Install 2.X API + Framework extension API + PyTorch dependency
 pip install neural-compressor[pt]
 # Install 2.X API + Framework extension API + TensorFlow dependency
 pip install neural-compressor[tf]
-```
-> **Note**:
-> Further installation methods can be found under [Installation Guide](./docs/source/installation_guide.md). check out our [FAQ](./docs/source/faq.md) for more details.
+```    
+**Note**: Further installation methods can be found under [Installation Guide](./docs/source/installation_guide.md). check out our [FAQ](./docs/source/faq.md) for more details.
 
 ## Getting Started
+After successfully installing these packages, try your first quantization program. **Following example code demonstrates FP8 Quantization**, it is supported by Intel Gaudi2 AI Accelerator.     
+To try on Intel Gaudi2, docker image with Gaudi Software Stack is recommended, please refer to following script for environment setup. More details can be found in [Gaudi Guide](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#launch-docker-image-that-was-built).    
 
-Setting up the environment:
-```bash
-pip install "neural-compressor>=2.3" "transformers>=4.34.0" torch torchvision
+Run a container with an interactive shell,
 ```
-After successfully installing these packages, try your first quantization program.
-
-### [FP8 Quantization](./docs/source/3x/PT_FP8Quant.md)
-Following example code demonstrates FP8 Quantization, it is supported by Intel Gaudi2 AI Accelerator. 
-
-To try on Intel Gaudi2, docker image with Gaudi Software Stack is recommended, please refer to following script for environment setup. More details can be found in [Gaudi Guide](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#launch-docker-image-that-was-built).
-```bash
-# Run a container with an interactive shell
 docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.19.0/ubuntu24.04/habanalabs/pytorch-installer-2.5.1:latest
 ```
-Run the example:
+Run the example,
 ```python
 from neural_compressor.torch.quantization import (
     FP8Config,
@@ -102,12 +80,10 @@ model = convert(model)
 
 output = model(torch.randn(1, 3, 224, 224).to("hpu")).to("cpu")
 print(output.shape)
-```
-
-### Weight-Only Large Language Model Loading (LLMs)
-
-Following example code demonstrates weight-only large language model loading on Intel Gaudi2 AI Accelerator. 
+```    
+More [FP8 quantization doc](./docs/source/3x/PT_FP8Quant.md).
 
+**Following example code demonstrates weight-only large language model loading** on Intel Gaudi2 AI Accelerator. 
 ```python
 from neural_compressor.torch.quantization import load
 
@@ -119,10 +95,7 @@ model = load(
     torch_dtype=torch.bfloat16,
 )
 ```
-
-**Note:**
-
-Intel Neural Compressor will convert the model format from auto-gptq to hpu format on the first load and save hpu_model.safetensors to the local cache directory for the next load. So it may take a while to load for the first time.
+**Note:** Intel Neural Compressor will convert the model format from auto-gptq to hpu format on the first load and save hpu_model.safetensors to the local cache directory for the next load. So it may take a while to load for the first time.
 
 ## Documentation
 

docs/source/3x/PT_FP8Quant.md

@@ -4,7 +4,8 @@ FP8 Quantization
 1. [Introduction](#introduction)
 2. [Supported Parameters](#supported-parameters)
 3. [Get Start with FP8 Quantization](#get-start-with-fp8-quantization)
-4. [Examples](#examples)  
+4. [Optimum-habana LLM example](#optimum-habana-LLM-example)
+5. [VLLM example](#VLLM-example)
 
 ## Introduction
 
@@ -75,30 +76,192 @@ Intel Neural Compressor provides general quantization APIs to leverage HPU FP8 c
 </tbody></table>
 
 ## Get Start with FP8 Quantization
+[Demo Usage](https://github.com/intel/neural-compressor?tab=readme-ov-file#getting-started)    
+[Computer vision example](../../../examples/3.x_api/pytorch/cv/fp8_quant)
 
-### Demo Usage
-
-```python
-from neural_compressor.torch.quantization import (
-    FP8Config,
-    prepare,
-    convert,
-)
-import torchvision.models as models
-
-model = models.resnet18()
-qconfig = FP8Config(fp8_config="E4M3")
-model = prepare(model, qconfig)
-# customer defined calibration
-calib_func(model)
-model = convert(model)
+## Optimum-habana LLM example
+### Overview
+[Optimum](https://huggingface.co/docs/optimum) is an extension of Transformers that provides a set of performance optimization tools to train and run models on targeted hardware with maximum efficiency.    
+[Optimum-habana](https://github.com/huggingface/optimum-habana) is the interface between the Transformers, Diffusers libraries and Intel Gaudi AI Accelerators (HPU). It provides higher performance based on modified modeling files, and utilizes Intel Neural Compressor for FP8 quantization internally,  [running-with-fp8](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation#running-with-fp8)    
+![](./imgs/optimum-habana.png)
+### Installation
+Refer to [optimum-habana, install-the-library-and-get-example-scripts](https://github.com/huggingface/optimum-habana?tab=readme-ov-file#install-the-library-and-get-example-scripts)    
+Option to install from source,
+```
+$ git clone https://github.com/huggingface/optimum-habana
+$ cd optimum-habana && git checkout v1.14.0 (change the version)
+$ pip install -e .
+$ pip install git+https://github.com/HabanaAI/DeepSpeed.git@1.18.0
+$ cd examples/text-generation
+$ pip install -r requirements.txt
+$ pip install -r requirements_lm_eval.txt  (Option)
+```
+### Check neural_compressor code
+> optimum-habana/examples/text-generation/utils.py
+>> initialize_model() -> setup_model() -> setup_quantization() -> FP8Config/prepare()/convert() 
+
+### FP8 KV cache
+Introduction: [kv-cache-quantization in huggingface transformers](https://huggingface.co/blog/kv-cache-quantization)    
+
+BF16 KVCache Code -> [Modeling_all_models.py -> KVCache()](https://github.com/huggingface/optimum-habana/blob/main/optimum/habana/transformers/models/modeling_all_models.py)    
+
+FP8 KVCache code trace with neural compressor support, for example Llama models,    
+> optimum-habana/optimum/habana/transformers/models/llama/modeling_llama.py     
+>> GaudiLlamaForCausalLM()  -> self.model()
+>>>    GaudiLlamaModel() -> forward() -> decoder_layer() ->  GaudiLlamaDecoderLayer() forward() -> pre_attn() -> pre_attn_forward() -> self.k_cache.update     
+
+> neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py    
+>> PatchedKVCache() -> update()    
+>> PatchedModuleFusedSDPA()
+
+Models list which support FP8 KV Cache,
+```
+microsoft/Phi-3-mini-4k-instruct
+bigcode/starcoder2-3b
+Qwen/Qwen2.5-7B-Instruct|
+meta-llama/Llama-3.2-3B-Instruct
+tiiuae/falcon-7b-instruct
+mistralai/Mixtral-8x7B-Instruct-v0.1
+EleutherAI/gpt-j-6b
+mistralai/Mistral-Nemo-Instruct-2407
+...
+```
+
+### Running with FP8
+Refer to [here](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation#running-with-fp8).    
+Change "--model_name_or_path" to be your model like    
+"meta-llama/Llama-3.1-8B-Instruct",    
+"Qwen/Qwen2.5-7B-Instruct", or     
+"mistralai/Mixtral-8x7B-Instruct-v0.1" and so on.    
+"--use_kv_cache" is to enable FP8 KV cache.
+
+### Profiling
+Add "--profiling_warmup_steps 5 --profiling_steps 2 --profiling_record_shapes" as args in the end of commandline of run_generation.py.     
+Refer to [torch.profiler.ProfilerActivity.HPU](https://github.com/huggingface/optimum-habana/blob/c9e1c23620618e2f260c92c46dfeb163545ec5ba/optimum/habana/utils.py#L305).    
+
+### FP8 Accuracy 
+"lm_eval.tasks", "lm_eval.evaluator", "lm_eval" are installed from the above requirements_lm_eval.txt. The tasks can be set and the default is ["hellaswag", "lambada_openai", "piqa", "winogrande"], [more info](https://github.com/EleutherAI/lm-evaluation-harness/)    
+
+| `Llama-2-7b-hf`| fp8 & fp8 KVCache| bf16 w/ bf16 KVCache|
+|---------------|---------|--------|
+| hellaswag     | 0.5691097390957977   | 0.5704043019318861    |
+| lambada_openai| 0.7360760721909567   | 0.7372404424607025 | 
+| piqa          | 0.7850924918389554   | 0.7818280739934712 |
+| winogrande    | 0.6929755327545383   | 0.6929755327545383 |
+
+| `Qwen2.5-7B-Instruct`| fp8 & fp8 KVCache| bf16 w/ bf16 KVCache|
+|---------------|---------|--------|
+| hellaswag     |  0.2539334793865764  |   0.2539334793865764    |
+| lambada_openai| 0.0   | 0.0 | 
+| piqa          | 0.5391730141458106   | 0.5391730141458106 |
+| winogrande    | 0.4956590370955012  | 0.4956590370955012 |
+
+| `Llama-3.1-8B-Instruct`| fp8 & fp8 KVCache| bf16 w/ bf16 KVCache|
+|---------------|---------|--------|
+| hellaswag     | 0.5934076877116112   |   0.5975901214897431    |
+| lambada_openai| 0.7230739375121289   | 0.7255967397632447 | 
+| piqa          | 0.7932535364526659   | 0.8030467899891186 |
+| winogrande    | 0.7434885556432518  | 0.7371744277821626 |
+
+
+| `Mixtral-8x7B-Instruct-v0.1`| fp8 & fp8 KVCache| bf16 w/ bf16 KVCache|
+|---------------|---------|--------|
+| hellaswag     | 0.25323640709022105   |   0.25323640709022105    |
+| lambada_openai| 0.0   | 0.0  |
+| piqa          | 0.528835690968444   | 0.528835690968444  |
+| winogrande    | 0.4956590370955012  | 0.4956590370955012 |
+
+## VLLM example
+### Overview
+![](./imgs/vllm_gaudi.png)
+
+### Installation
+Refer to [Habana vllm-fork](https://github.com/HabanaAI/vllm-fork) to install.    
+Option to install `vllm-hpu-extension`, `neural_compressor` and `vllm` from the source,
+```
+$ git clone https://github.com/HabanaAI/vllm-fork.git
+$ cd vllm-fork
+$ pip install -r requirements-hpu.txt
+$ python setup.py develop --user
+
+## Check
+$ pip list |grep vllm
+vllm                              0.6.3.dev1122+g2f43ebf5.d20241121.gaudi118 /home/fengding/vllm-fork
+vllm-hpu-extension                0.1
+
+## Validation
+$ VLLM_SKIP_WARMUP=true python3 examples/offline_inference.py
+......
+Prompt: 'Hello, my name is', Generated text: ' Kelly and I have a job to do.\nI need someone to come over'
+Prompt: 'The president of the United States is', Generated text: ' facing a sharp criticism of his handling of the coronavirus pandemic, including'
+Prompt: 'The capital of France is', Generated text: ' the capital of the Socialist Party of France (SPF), with its state-'
+Prompt: 'The future of AI is', Generated text: " in what's coming, not what's coming.\nI don't know what"
+```
+
+### Run FP8 calibration
+Refer to [vllm-hpu-extension->calibration](https://github.com/HabanaAI/vllm-hpu-extension/tree/main/calibration)    
+```
+$ git clone https://github.com/HabanaAI/vllm-hpu-extension
+$ cd vllm-hpu-extension/calibration
+
+# For Llama-3.1.8B-Instruct
+$ ./calibrate_model.sh -m meta-llama/Llama-3.1-8B-Instruct -d /home/fengding/processed-data.pkl -o ./output_llama3.1.8b.Instruct -b 128 -t 1 -l 128
+    ## Generate scale factors in ./output_llama3.1.8b.Instruct
+```
+
+### Start vllm server
+```
+$ cd vllm-fork/
+
+$ PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
+PT_HPU_WEIGHT_SHARING=0 \
+VLLM_CONTIGUOUS_PA=true \
+VLLM_SKIP_WARMUP=true \
+QUANT_CONFIG=output_llama3.1.8b.Instruct/maxabs_quant_g2.json \
+python3 -m vllm.entrypoints.openai.api_server \
+--model meta-llama/Llama-3.1-8B-Instruct \
+--port 8080 \
+--gpu-memory-utilization 0.9 \
+--tensor-parallel-size 1 \
+--disable-log-requests \
+--block-size 128 \
+--quantization inc \
+--kv-cache-dtype fp8_inc \
+--device hpu \
+--weights-load-device cpu \
+--dtype bfloat16 \
+--num_scheduler_steps 16 2>&1 > vllm_serving.log &
+```
+Refer to [vllm-fork->README_GAUDI.md](https://github.com/HabanaAI/vllm-fork/blob/habana_main/README_GAUDI.md) for more details.
+
+### Start client to test
+```
+$ curl --noproxy "*" http://localhost:8080/v1/completions -H "Content-Type: application/json" -d '{"model": "meta-llama/Llama-3.1-8B-Instruct", "prompt": "San Francisco is a", "max_tokens": 100}'
+```
+
+### Run benchmark
+```
+python benchmarks/benchmark_serving.py \
+--backend vllm \
+--model meta-llama/Llama-3.1-8B-Instruct  \
+--dataset-name sonnet \
+--dataset-path benchmarks/sonnet.txt \
+--request-rate 128 \
+--num-prompts 128 \
+--port 8080 \
+--sonnet-input-len 128 \
+--sonnet-output-len 128 \
+--sonnet-prefix-len 100
 ```
 
-## Examples
+### FP8 KV cache
+Code trace
+> vllm-fork/vllm/attention/backends/hpu_attn.py
+>> from vllm_hpu_extension.utils import Matmul, Softmax, VLLMKVCache
+>> HPUAttentionImpl() -> self.k_cache() / self.v_cache()    
 
-| Task                 | Example |
-|----------------------|---------|
-| Computer Vision (CV)      |    [Link](../../../examples/3.x_api/pytorch/cv/fp8_quant/)     |
-| Large Language Model (LLM) |    [Link](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation#running-with-fp8)     |
+> neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py
+>> PatchedVLLMKVCache()
 
-> Note: For LLM, Optimum-habana provides higher performance based on modified modeling files, so here the Link of LLM goes to Optimum-habana, which utilize Intel Neural Compressor for FP8 quantization internally.
+> neural_compressor/torch/algorithms/fp8_quant/common.py
+>> "VLLMKVCache": ModuleInfo("kv_cache", PatchedVLLMKVCache)