LLM: Refine start script for Pipeline Parallel Serving (#11557)

Refine start script and readme for Pipeline Parallel Serving

Author: xiangyuT

docker/llm/serving/xpu/docker/README.md

@@ -23,7 +23,6 @@ export DOCKER_IMAGE=intelanalytics/ipex-llm-serving-xpu:2.1.0-SNAPSHOT
 sudo docker run -itd \
         --net=host \
         --device=/dev/dri \
-        --memory="32G" \
         --name=CONTAINER_NAME \
         --shm-size="16g" \
         $DOCKER_IMAGE

docker/llm/serving/xpu/docker/start-pp_serving-service.sh

@@ -14,11 +14,15 @@ if [[ $KERNEL_VERSION != *"6.5"* ]]; then
 fi
 export TORCH_LLM_ALLREDUCE=0
 
+export IPEX_LLM_LAST_LM_HEAD=1
 export IPEX_LLM_QUANTIZE_KV_CACHE=1
+export IPEX_LLM_LOW_MEM=1
 export num_gpus=2
 export model_path="/llm/models/Llama-2-7b-chat-hf"
 export low_bit="fp8"
 # max requests = max_num_reqs * rank_num
 export max_num_seqs="4"
+export max_prefilled_seqs="0"
+
 cd /llm/pp_serving
-CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $num_gpus pipeline_serving.py --repo-id-or-model-path $model_path --low-bit $low_bit --max-num-seqs $max_num_seqs
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $num_gpus pipeline_serving.py --repo-id-or-model-path $model_path --low-bit $low_bit --max-num-seqs $max_num_seqs --max-prefilled-seqs $max_prefilled_seqs

python/llm/example/GPU/Pipeline-Parallel-FastAPI/run.sh

@@ -28,9 +28,16 @@ if [[ $KERNEL_VERSION != *"6.5"* ]]; then
     export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
 fi
 export TORCH_LLM_ALLREDUCE=0
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+export IPEX_LLM_LAST_LM_HEAD=1
+export IPEX_LLM_LOW_MEM=1
 
 export MODEL_PATH=YOUR_MODEL_PATH
 export NUM_GPUS=2
-export IPEX_LLM_QUANTIZE_KV_CACHE=1
+export ZE_AFFINITY_MASK=0,1
+export LOW_BIT="fp8"
+# max requests = max_num_reqs * rank_num
+export MAX_NUM_SEQS="4"
+export MAX_PREFILLED_SEQS=0
 
-CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS pipeline_serving.py --repo-id-or-model-path $MODEL_PATH --low-bit fp8 --max-num-seqs 4 --max-prefilled-seqs 0
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS pipeline_serving.py --repo-id-or-model-path $MODEL_PATH --low-bit $LOW_BIT --max-num-seqs $MAX_NUM_SEQS --max-prefilled-seqs $MAX_PREFILLED_SEQS

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

@@ -447,6 +447,7 @@ def load_model(self, model_path, world_size, low_bit='sym_int4'):
             model = AutoModelForCausalLM.from_pretrained(model_path,
                                                          load_in_low_bit=low_bit,
                                                          torch_dtype=self.dtype,
+                                                         cpu_embedding=True,
                                                          optimize_model=True,
                                                          trust_remote_code=True,
                                                          use_cache=True,
@@ -499,7 +500,8 @@ def cat_kv_cache(self, model_type, kv_cache_1, kv_cache_2):
             return tuple(result)
         else:
             # num_layers = self.model.layer_end - self.model.layer_start
-            for layer_idx in range(self.model.num_layers):
+            num_cache = min(len(kv_cache_1.key_cache), self.model.num_layers)
+            for layer_idx in range(num_cache):
                 kv_cache_1.key_cache[layer_idx] = \
                     torch.cat([kv_cache_1.key_cache[layer_idx],
                                kv_cache_2.key_cache[layer_idx]], dim=0)