Add more examples for pipeline parallel inference

python/llm/example/GPU/Pipeline-Parallel-Inference/README.md

@@ -16,7 +16,17 @@ To run this example with IPEX-LLM on Intel GPUs, we have some recommended requir
 - [baichuan-inc/Baichuan2-13B-Chat](./run_baichuan2_arc_2_card.sh)
 - [microsoft/Phi-3-mini-4k-instruct](./run_phi3_arc_2_card.sh)
 - [microsoft/Phi-3-medium-4k-instruct](./run_phi3_arc_2_card.sh)
-
+- [mistralai/Mistral-7B-v0.1](./run_mistral_arc_2_card.sh)
+- [mistralai/Mixtral-8x7B-Instruct-v0.1](./run_mistral_arc_2_card.sh)
+- [01-ai/Yi-6B-Chat](./run_yi_arc_2_card.sh)
+- [01-ai/Yi-34B-Chat](./run_yi_arc_2_card.sh)
+- [codellama/CodeLlama-7b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [codellama/CodeLlama-13b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [codellama/CodeLlama-34b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [upstage/SOLAR-10.7B-Instruct-v1.0](./run_solar_arc_2_card.sh)
+- [lmsys/vicuna-7b-v1.3](./run_vicuna_arc_2_card.sh)
+- [lmsys/vicuna-13b-v1.3](./run_vicuna_arc_2_card.sh)
+- [lmsys/vicuna-33b-v1.3](./run_vicuna_arc_2_card.sh)
 
 ## Example: Run pipeline parallel inference on multiple GPUs
 
@@ -101,6 +111,87 @@ bash run_phi3_arc_2_card.sh
 
 </details>
 
+</details>
+
+<details>
+  <summary> Show Mistral/Mixtral example </summary>
+
+#### Run Mistral-7B-v0.1 / Mixtral-8x7B-Instruct-v0.1 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Mistral / Mixtral to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_mistral_arc_2_card.sh
+```
+
+</details>
+
+</details>
+
+<details>
+  <summary> Show Yi example </summary>
+
+#### Run Yi-6B-Chat / Yi-34B-Chat on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Yi to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_yi_arc_2_card.sh
+```
+
+</details>
+
+</details>
+
+<details>
+  <summary> Show Codellama example </summary>
+
+#### Run CodeLlama-7b-Instruct-hf / CodeLlama-13b-Instruct-hf / CodeLlama-34b-Instruct-hf on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Codellama to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_codellama_arc_2_card.sh
+```
+
+</details>
+
+</details>
+
+<details>
+  <summary> Show Solar example </summary>
+
+#### Run SOLAR-10.7B-Instruct-v1.0 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Solar to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_solar_arc_2_card.sh
+```
+
+</details>
+
+</details>
+
+<details>
+  <summary> Show Vicuna example </summary>
+
+#### Run vicuna-7b-v1.3 / vicuna-13b-v1.3 / vicuna-33b-v1.3 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Vicuna to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_vicuna_arc_2_card.sh
+```
+
+</details>
+
+
 ### 3. Sample Output
 #### [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf)
 ```log

python/llm/example/GPU/Pipeline-Parallel-Inference/run_codellama_arc_2_card.sh

@@ -0,0 +1,25 @@
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run CodeLlama-7b-Instruct-hf
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+    generate.py --repo-id-or-model-path 'codellama/CodeLlama-7b-Instruct-hf' --gpu-num $NUM_GPUS
+
+# To run CodeLlama-13b-Instruct-hf
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path 'codellama/CodeLlama-7b-Instruct-hf' --gpu-num $NUM_GPUS
+
+# To run CodeLlama-34b-Instruct-hf
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path 'codellama/CodeLlama-34b-Instruct-hf' --gpu-num $NUM_GPUS

python/llm/example/GPU/Pipeline-Parallel-Inference/run_mistral_arc_2_card.sh

@@ -0,0 +1,21 @@
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run Mistral-7B-v0.1
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+    generate.py --repo-id-or-model-path 'mistralai/Mistral-7B-v0.1' --gpu-num $NUM_GPUS
+
+# To run Mixtral-8x7B-Instruct-v0.1
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path 'mistralai/Mixtral-8x7B-Instruct-v0.1' --gpu-num $NUM_GPUS

python/llm/example/GPU/Pipeline-Parallel-Inference/run_solar_arc_2_card.sh

@@ -0,0 +1,17 @@
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run SOLAR-10.7B-Instruct-v1.0
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+    generate.py --repo-id-or-model-path 'upstage/SOLAR-10.7B-Instruct-v1.0' --gpu-num $NUM_GPUS

python/llm/example/GPU/Pipeline-Parallel-Inference/run_vicuna_arc_2_card.sh

@@ -0,0 +1,25 @@
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run vicuna-7b-v1.3
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+    generate.py --repo-id-or-model-path 'lmsys/vicuna-7b-v1.3' --gpu-num $NUM_GPUS
+
+# To run vicuna-13b-v1.3
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path 'lmsys/vicuna-13b-v1.3' --gpu-num $NUM_GPUS
+
+# To run vicuna-33b-v1.3
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path 'lmsys/vicuna-33b-v1.3' --gpu-num $NUM_GPUS

python/llm/example/GPU/Pipeline-Parallel-Inference/run_yi_arc_2_card.sh

@@ -0,0 +1,21 @@
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run Yi-6B-Chat
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+    generate.py --repo-id-or-model-path '01-ai/Yi-6B-Chat' --gpu-num $NUM_GPUS
+
+# To run Yi-34B-Chat
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+#     generate.py --repo-id-or-model-path '01-ai/Yi-34B-Chat' --gpu-num $NUM_GPUS

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

@@ -234,7 +234,8 @@ def pipeline_parallel_generate(self,
                                          "make sure that `pad_token_id` is defined.")
             next_ids = next_ids * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)
 
-        if isinstance(outputs.past_key_values, tuple) and local_rank != 0:
+        # Temporarily specify as Baichuan and ChatGLM
+        if self.config.model_type in ["baichuan", "chatglm"] and local_rank != 0:
             value_placeholder = torch.empty_like((outputs.past_key_values)[-1][0])
             past_key_values_placeholder = tuple(
                 (value_placeholder, value_placeholder) for _ in range(layer_start)