add config and default value (#11344)

* add config and default value

* add config in taml

* remove lookahead and max_matching_ngram_size in config

* remove streaming and use_fp16_torch_dtype in test yaml

* update task in readme

* update commit of task

docs/readthedocs/source/doc/LLM/Quickstart/benchmark_quickstart.md

@@ -41,6 +41,7 @@ test_api:
   - "transformer_int4_gpu"   # on Intel GPU, transformer-like API, (qtype=int4)
 cpu_embedding: False # whether put embedding to CPU
 streaming: False # whether output in streaming way (only avaiable now for gpu win related test_api)
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'
 ```
 
 Some parameters in the yaml file that you can configure:
@@ -61,6 +62,7 @@ Some parameters in the yaml file that you can configure:
 - `streaming`: Whether to output in a streaming way (only available for GPU Windows-related test_api).
 - `use_fp16_torch_dtype`: Whether to use fp16 for the non-linear layer (only available for "pipeline_parallel_gpu" test_api).
 - `n_gpu`: Number of GPUs to use (only available for "pipeline_parallel_gpu" test_api).
+- `task`: There are three tasks: `continuation`, `QA` and `summarize`. `continuation` refers to writing additional content based on prompt. `QA` refers to answering questions based on prompt. `summarize` refers to summarizing the prompt.
 
 
 ```eval_rst

python/llm/dev/benchmark/all-in-one/README.md

@@ -56,13 +56,10 @@ test_api:
   # - "bigdl_ipex_int8"                     # on Intel CPU, (qtype=int8)
   # - "speculative_cpu"                     # on Intel CPU, inference with self-speculative decoding
   # - "deepspeed_transformer_int4_cpu"      # on Intel CPU, deepspeed autotp inference
-  # - "transformer_int4_fp16_lookahead_gpu" # on Intel GPU, transformer-like API, with lookahead, (qtype=int4), (dtype=fp16)
 cpu_embedding: False # whether put embedding to CPU
 streaming: False # whether output in streaming way (only available now for gpu win related test_api)
 use_fp16_torch_dtype: True # whether use fp16 for non-linear layer (only available now for "pipeline_parallel_gpu" test_api)
-lookahead: 3
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'
-max_matching_ngram_size: 2
-task: 'continuation' # when test_api is "transformer_int4_fp16_lookahead_gpu", task could be 'QA', 'continuation' or 'summarize' 
 
 ```
 

python/llm/dev/benchmark/all-in-one/config.yaml

@@ -35,6 +35,4 @@ test_api:
 cpu_embedding: False # whether put embedding to CPU
 streaming: False # whether output in streaming way (only available now for gpu win related test_api)
 use_fp16_torch_dtype: True # whether use fp16 for non-linear layer (only available now for "pipeline_parallel_gpu" test_api)
-lookahead: 3
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'
-max_matching_ngram_size: 2
-task: 'continuation' # when task is 'continuation', the result is without lookahead. When task is 'QA' or 'summarize', the result is with lookahead

python/llm/dev/benchmark/all-in-one/run.py

@@ -49,7 +49,7 @@ def run_model_in_thread(model, in_out, tokenizer, result, warm_up, num_beams, in
     for i in range(num_trials + warm_up):
         st = time.perf_counter()
         if lookahead:
-            output_ids = model.generate(input_ids, lookahead=conf.lookahead, do_sample=False, max_matching_ngram_size=conf.max_matching_ngram_size, max_new_tokens=out_len,
+            output_ids = model.generate(input_ids, lookahead=3, do_sample=False, max_matching_ngram_size=2, max_new_tokens=out_len,
                                     min_new_tokens=out_len, num_beams=num_beams)
         else:
             output_ids = model.generate(input_ids, do_sample=False, max_new_tokens=out_len,
@@ -106,7 +106,7 @@ def preprocess_prompt(tokenizer, in_len, task):
         input_ids = tokenizer.encode(input_str, return_tensors="pt")    
     return input_ids
 
-def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1, num_trials=3, num_beams=1, low_bit='sym_int4', cpu_embedding=False, batch_size=1, streaming=False, use_fp16_torch_dtype=False, lookahead=False):
+def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1, num_trials=3, num_beams=1, low_bit='sym_int4', cpu_embedding=False, batch_size=1, streaming=False, use_fp16_torch_dtype=False, lookahead=False, task='continuation'):
     # TODO: make a parameter
     result= {}
     if test_api == 'transformer_int4':
@@ -118,7 +118,7 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
     elif test_api == 'transformer_int4_gpu':
         result = run_transformer_int4_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding)
     elif test_api == 'transformer_int4_fp16_gpu':
-        result = run_transformer_int4_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding, fp16=True, lookahead=lookahead)
+        result = run_transformer_int4_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding, fp16=True, lookahead=lookahead, task=task)
     elif test_api == 'optimize_model_gpu':
         result = run_optimize_model_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size)
     elif test_api == 'pytorch_autocast_bf16':
@@ -441,7 +441,8 @@ def run_transformer_int4_gpu(repo_id,
                              batch_size,
                              cpu_embedding,
                              fp16=False,
-                             lookahead=False):
+                             lookahead=False,
+                             task='continuation'):
     from ipex_llm.transformers import AutoModel, AutoModelForCausalLM
     from transformers import AutoTokenizer, GPTJForCausalLM, LlamaTokenizer
     model_path = get_model_path(repo_id, local_model_hub)
@@ -506,7 +507,7 @@ def run_transformer_int4_gpu(repo_id,
             in_out_len = in_out.split("-")
             in_len = int(in_out_len[0])
             out_len = int(in_out_len[1])
-            if conf['task'] == 'continuation':
+            if task == 'continuation':
                 # As different tokenizer has different encodings,
                 # in_len.txt maybe shorter than we need,
                 # use much longer context to make sure input length
@@ -520,8 +521,8 @@ def run_transformer_int4_gpu(repo_id,
                 # slice the input_ids to ensure the prompt length is required length.
                 input_ids = tokenizer.encode(input_str, return_tensors="pt")
                 input_ids = input_ids[:, :in_len]
-            elif conf['task'] == 'summarize' or conf['task'] == 'QA':
+            elif task in ['QA', 'summarize']:
-                input_ids = preprocess_prompt(tokenizer, in_len, conf['task'])
+                input_ids = preprocess_prompt(tokenizer, in_len, task)
             true_str = tokenizer.batch_decode(input_ids)[0]
             input_list = [true_str] * batch_size
             input_ids = tokenizer(input_list, return_tensors="pt").input_ids.to('xpu')
@@ -1824,17 +1825,19 @@ def run_pipeline_parallel_gpu(repo_id,
 
 if __name__ == '__main__':
     from omegaconf import OmegaConf
-    global conf
     conf = OmegaConf.load(f'{current_dir}/config.yaml')
     today = date.today()
     if 'exclude' in conf:
         excludes = conf['exclude']
     streaming = False
     use_fp16_torch_dtype = False
+    task = 'continuation'
     if 'streaming' in conf:
         streaming = conf['streaming']
     if 'use_fp16_torch_dtype' in conf:
         use_fp16_torch_dtype = conf['use_fp16_torch_dtype']
+    if 'task' in conf:
+        task = conf['task']
     lookahead = False
     
     import pandas as pd
@@ -1854,10 +1857,10 @@ if __name__ == '__main__':
                         model_id_input_batch_size = model_id_input + ':' + str(batch_size)
                         if model_id_input in excludes or model_id_input_batch_size in excludes:
                             in_out_pairs.remove(in_out)
-                if conf['task'] in ['QA', 'summarize'] and conf['num_beams'] == 1 and batch_size == 1:
+                if task in ['QA', 'summarize'] and conf['num_beams'] == 1 and batch_size == 1:
                     lookahead = True
                 run_model(model, api, in_out_pairs, conf['local_model_hub'], conf['warm_up'], conf['num_trials'], conf['num_beams'],
-                      conf['low_bit'], conf['cpu_embedding'], batch_size, streaming, use_fp16_torch_dtype, lookahead)
+                      conf['low_bit'], conf['cpu_embedding'], batch_size, streaming, use_fp16_torch_dtype, lookahead, task)
         df = pd.DataFrame(results, columns=['model', '1st token avg latency (ms)', '2+ avg latency (ms/token)', 'encoder time (ms)',
                                             'input/output tokens', 'batch_size', 'actual input/output tokens', 'num_beams', 'low_bit', 'cpu_embedding',
                                             'model loading time (s)', 'peak mem (GB)', 'streaming', 'use_fp16_torch_dtype'])

python/llm/test/benchmark/arc-perf-test-batch2.yaml

@@ -35,4 +35,5 @@ exclude:
   - 'bigcode/starcoder-15.5b-4bit:2048'
   # - 'databricks/dolly-v2-12b:2048'
   - 'baichuan-inc/Baichuan2-13B-Chat-4bit:2048'
-  - 'bigscience/bloomz-7b1:2048'
+  - 'bigscience/bloomz-7b1:2048'
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'

python/llm/test/benchmark/arc-perf-test.yaml

@@ -36,3 +36,4 @@ exclude:
 #  - 'fnlp/moss-moon-003-sft-4bit:2048'
   - 'baichuan-inc/Baichuan2-13B-Chat-4bit:2048'
   - 'bigscience/bloomz-7b1:2048'
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'

python/llm/test/benchmark/arc-perf-transformers-437-batch2.yaml

@@ -17,3 +17,4 @@ in_out_pairs:
 test_api:
   - "transformer_int4_fp16_gpu"  # on Intel GPU
 cpu_embedding: False # whether put embedding to CPU (only avaiable now for gpu win related test_api)
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'

python/llm/test/benchmark/arc-perf-transformers-437.yaml

@@ -17,3 +17,4 @@ in_out_pairs:
 test_api:
   - "transformer_int4_fp16_gpu"  # on Intel GPU
 cpu_embedding: False # whether put embedding to CPU (only avaiable now for gpu win related test_api)
+task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'