ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

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

Browse source 
Refine start script and readme for Pipeline Parallel Serving
This commit is contained in:
Xiangyu Tian 2024-07-11 15:45:27 +08:00 committed by GitHub
parent 1355b2ce06
commit 7f5111a998
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 17 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
docker/llm/serving/xpu/docker/README.md
View file

@ -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

6
docker/llm/serving/xpu/docker/start-pp_serving-service.sh
View file

@ -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

11
python/llm/example/GPU/Pipeline-Parallel-FastAPI/run.sh
View file

@ -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

4
python/llm/src/ipex_llm/transformers/pipeline_parallel.py
View file

@ -447,6 +447,7 @@ class ModelRunner:
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 @@ class ModelRunner:
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)