Speculative Ziya on CPU (#10160)

* Speculative Ziya on CPU

* Without part of Accelerate with BIGDL_OPT_IPEX

python/llm/example/CPU/Speculative-Decoding/ziya/README.md

@@ -0,0 +1,57 @@
+# Ziya
+In this directory, you will find examples on how you could run Ziya BF16 inference with self-speculative decoding using BigDL-LLM on [Intel CPUs](../README.md). For illustration purposes,we utilize the [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) as reference Ziya model.
+
+## 0. Requirements
+To run the example with BigDL-LLM on Intel CPUs, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
+
+## Example: Predict Tokens using `generate()` API
+In the example [speculative.py](./speculative.py), we show a basic use case for a Ziya model to predict the next N tokens using `generate()` API, with BigDL-LLM speculative decoding optimizations on Intel CPUs.
+### 1. Install
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9
+conda activate llm
+pip install --pre --upgrade bigdl-llm[all]
+pip install intel_extension_for_pytorch==2.1.0
+pip install transformers==4.35.2
+```
+### 2. Configures high-performing processor environment variables
+```bash
+source bigdl-llm-init -t
+export OMP_NUM_THREADS=48 # you can change 48 here to #cores of one processor socket
+```
+### 3. Run
+
+We recommend to use `numactl` to bind the program to a specified processor socket:
+
+```bash
+numactl -C 0-47 -m 0 python ./speculative.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+
+For example, 0-47 means bind the python program to core list 0-47 for a 48-core socket.
+
+Arguments info:
+
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Ziya model (e.g. `IDEA-CCNL/Ziya-Coding-34B-v1.0`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `IDEA-CCNL/Ziya-Coding-34B-v1.0`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). A default prompt is provided.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `128`.
+
+#### Sample Output
+#### [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0)
+
+```log
+<human>: 
+写一段快速排序
+<bot>: 
+def quick_sort(arr):
+    if len(arr) <= 1:
+        return arr
+    pivot = arr[len(arr) // 2]
+    left = [x for x in arr if x < pivot]
+    middle = [x for x in arr if x == pivot]
+    right = [x for x in arr if x > pivot]
+    return quick_sort(left) + middle + quick_sort(right)
+Tokens generated 100
+E2E Generation time xx.xxxxs
+First token latency xx.xxxxs
+```

python/llm/example/CPU/Speculative-Decoding/ziya/speculative.py

@@ -0,0 +1,87 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import torch
+from bigdl.llm.transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer
+import argparse
+import time
+import numpy as np
+
+
+torch.nn.Linear.reset_parameters = lambda x: None
+seed=42
+torch.manual_seed(seed)
+np.random.seed(seed)
+
+ZIYA_PROMPT_FORMAT = "<human>: \n{prompt}\n<bot>: \n"
+prompt = "写一段快速排序"
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Mistral model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="IDEA-CCNL/Ziya-Coding-34B-v1.0",
+                        help='The huggingface repo id for the Ziya (e.g. `IDEA-CCNL/Ziya-Coding-34B-v1.0`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--prompt', type=str, default=prompt,
+                        help='Prompt to infer')
+    parser.add_argument('--n-predict', type=int, default=128,
+                        help='Max tokens to predict')
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+
+    # Load model in optimized bf16 here.
+    # Set `speculative=True`` to enable speculative decoding,
+    # it only works when load_in_low_bit="fp16" on Intel GPU or load_in_low_bit="bf16" on latest Intel Xeon CPU
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 optimize_model=True,
+                                                 torch_dtype=torch.bfloat16,
+                                                 load_in_low_bit="bf16",
+                                                 speculative=True,
+                                                 torchscript=True,
+                                                 trust_remote_code=True,
+                                                 use_cache=True)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+    with torch.inference_mode():
+        prompt = ZIYA_PROMPT_FORMAT.format(prompt=args.prompt)
+        inputs = tokenizer(prompt, return_tensors='pt')
+        input_ids = inputs.input_ids.to(model.device)
+        actual_in_len = input_ids.shape[1]
+        print("actual input_ids length:" + str(actual_in_len))
+        attention_mask = inputs.attention_mask.to(model.device)
+
+        # warmup
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict,
+                                attention_mask=attention_mask,
+                                do_sample=False)
+        output_str = tokenizer.decode(output[0])
+
+        # speculative decoding
+        st = time.perf_counter()
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict,
+                                attention_mask=attention_mask,
+                                do_sample=False)
+        output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+        end = time.perf_counter()
+
+        print(output_str)
+        print(f"Tokens generated {model.n_token_generated}")
+        print(f"E2E Generation time {(end - st):.4f}s")
+        print(f"First token latency {model.first_token_time:.4f}s")