initial support of IPEX_LLM_PERFORMANCE_MODE (#11754)

* add perf mode

* update

* fix style

Author: rnwang04

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

@@ -53,8 +53,8 @@ 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=3, do_sample=False, max_matching_ngram_size=2, max_new_tokens=out_len,
-                                    min_new_tokens=out_len, num_beams=num_beams)
+            output_ids = model.generate(input_ids, lookahead=2, 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,
                                         min_new_tokens=out_len, num_beams=num_beams)
@@ -67,8 +67,8 @@ def run_model_in_thread(model, in_out, tokenizer, result, warm_up, num_beams, in
         torch.xpu.empty_cache()
         actual_out_len = output_ids.shape[1] - actual_in_len
         if i >= warm_up:
-            if lookahead:
-                result[in_out].append([model.first_token_time, (end - st - model.first_token_time)/model.n_token_generated, 0,
+            if lookahead or os.environ.get("IPEX_LLM_PERFORMANCE_MODE", None) == "1":
+                result[in_out].append([model.first_token_time, (end - st - model.first_token_time)/(model.n_token_generated - 1), 0,
                                        actual_in_len, actual_out_len, load_time, 0])
             else:
                 result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time,
@@ -510,7 +510,7 @@ def run_transformer_int4_gpu(repo_id,
     load_time = end - st
     print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
 
-    if not lookahead:
+    if not lookahead and os.environ.get("IPEX_LLM_PERFORMANCE_MODE", None) != "1":
         model = BenchmarkWrapper(model)
 
     result = {}

python/llm/src/ipex_llm/transformers/lookup.py

@@ -21,6 +21,7 @@
 #
 
 from typing import Callable, List, Optional, Tuple
+import os
 import torch
 import time
 import copy
@@ -54,6 +55,9 @@ def generate(
     **kwargs,
 ):
     lookahead = kwargs.pop("lookahead", None)
+    perf_mode = os.environ.get("IPEX_LLM_PERFORMANCE_MODE", None)
+    if perf_mode == "1" and lookahead is None:
+        lookahead = 2  # default to 2 now
     if lookahead:
         from ipex_llm.transformers.convert import get_enable_ipex
         _enable_ipex = get_enable_ipex()