feat: change oneccl to internal (#12296)

* feat: change oneccl

* fix: restore llama-70b

* fix: remove tab

* fix: remove extra blank

* small fix

* add comments

* fix: add a blank space

Author: cranechu0131

python/llm/example/GPU/Deepspeed-AutoTP/README.md

@@ -20,7 +20,8 @@ conda activate llm
 # below command will install intel_extension_for_pytorch==2.1.10+xpu as default
 pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
 pip install transformers==4.37.0
-pip install oneccl_bind_pt==2.1.100 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+wget https://sourceforge.net/projects/oneccl-wks/files/2024.0.0.5.1-release/oneccl_wks_installer_2024.0.0.5.1.sh 
+bash oneccl_wks_installer_2024.0.0.5.1.sh
 # configures OneAPI environment variables
 source /opt/intel/oneapi/setvars.sh
 pip install git+https://github.com/microsoft/DeepSpeed.git@ed8aed5

python/llm/example/GPU/Deepspeed-AutoTP/deepspeed_autotp.py

@@ -104,7 +104,8 @@ def get_int_from_env(env_keys, default):
     deepspeed.comm.comm.cdb = None
     from deepspeed.comm.comm import init_distributed
     init_distributed()
-
+    from ipex_llm.utils import BenchmarkWrapper
+    model = BenchmarkWrapper(model)
     print(model)
 
     # Load tokenizer
@@ -135,7 +136,7 @@ def get_int_from_env(env_keys, default):
             actual_output_len = output.shape[1] - input_ids.shape[1]
             output_str = tokenizer.decode(output[0], skip_special_tokens=True)
             avg_time = (end - st) / actual_output_len * 1000
-            print(f'Inference time of generating {actual_output_len} tokens: {end-st} s, average token latency is {avg_time} ms/token.')
+            print(f'Inference time of generating {actual_output_len} tokens: {end-st} s, first token cost {model.first_cost} s, rest tokens average cost {model.rest_cost_mean} s')
             print('-'*20, 'Prompt', '-'*20)
             print(prompt)
             print('-'*20, 'Output', '-'*20)

python/llm/example/GPU/Deepspeed-AutoTP/run_llama2_70b_pvc_1550_1_card.sh

@@ -24,7 +24,8 @@ export CCL_ZE_IPC_EXCHANGE=sockets
 export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so:${LD_PRELOAD}
 basekit_root=/opt/intel/oneapi
 source $basekit_root/setvars.sh --force
-source $basekit_root/ccl/latest/env/vars.sh --force
+# source $basekit_root/ccl/latest/env/vars.sh --force   deprecate oneccl_bind_pt and use internal oneccl for better performance
+source /opt/intel/1ccl-wks/setvars.sh
 
 export OMP_NUM_THREADS=$((56/$NUM_GPUS))
 export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=2

python/llm/example/GPU/Deepspeed-AutoTP/run_mistral_7b_instruct_flex_2_card.sh

@@ -22,7 +22,8 @@ export CCL_ZE_IPC_EXCHANGE=sockets
 export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so:${LD_PRELOAD}
 basekit_root=/opt/intel/oneapi
 source $basekit_root/setvars.sh --force
-source $basekit_root/ccl/latest/env/vars.sh --force
+# source $basekit_root/ccl/latest/env/vars.sh --force   deprecate oneccl_bind_pt and use internal oneccl for better performance
+source /opt/intel/1ccl-wks/setvars.sh
 
 NUM_GPUS=2 # number of used GPU
 export USE_XETLA=OFF

python/llm/example/GPU/Deepspeed-AutoTP/run_qwen_14b_arc_2_card.sh

@@ -25,7 +25,8 @@ export CCL_ZE_IPC_EXCHANGE=sockets
 export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so:${LD_PRELOAD}
 basekit_root=/opt/intel/oneapi
 source $basekit_root/setvars.sh --force
-source $basekit_root/ccl/latest/env/vars.sh --force
+# source $basekit_root/ccl/latest/env/vars.sh --force   deprecate oneccl_bind_pt and use internal oneccl for better performance
+source /opt/intel/1ccl-wks/setvars.sh
 
 NUM_GPUS=2 # number of used GPU
 export USE_XETLA=OFF

python/llm/example/GPU/Deepspeed-AutoTP/run_vicuna_33b_arc_2_card.sh

@@ -22,7 +22,8 @@ export CCL_ZE_IPC_EXCHANGE=sockets
 export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so:${LD_PRELOAD}
 basekit_root=/opt/intel/oneapi
 source $basekit_root/setvars.sh --force
-source $basekit_root/ccl/latest/env/vars.sh --force
+# source $basekit_root/ccl/latest/env/vars.sh --force   deprecate oneccl_bind_pt and use internal oneccl for better performance
+source /opt/intel/1ccl-wks/setvars.sh
 
 NUM_GPUS=2 # number of used GPU
 export USE_XETLA=OFF