mirror of
https://gitee.com/ascend/ModelLink.git
synced 2024-12-02 11:58:26 +08:00
59 lines
1.8 KiB
Bash
59 lines
1.8 KiB
Bash
#!/bin/bash
|
|
|
|
export CUDA_DEVICE_MAX_CONNECTIONS=1
|
|
|
|
# please fill these path configurations
|
|
CHECKPOINT="your model directory path"
|
|
TOKENIZER_PATH="your tokenizer directory path"
|
|
TOKENIZER_MODEL="your tokenizer.model file path"
|
|
|
|
# Change for multinode config
|
|
MASTER_ADDR=localhost
|
|
MASTER_PORT=6001
|
|
NNODES=1
|
|
NODE_RANK=0
|
|
NPUS_PER_NODE=8
|
|
|
|
DISTRIBUTED_ARGS="--nproc_per_node $NPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
|
|
|
|
python -m torch.distributed.launch $DISTRIBUTED_ARGS inference.py \
|
|
--tensor-model-parallel-size 8 \
|
|
--pipeline-model-parallel-size 1 \
|
|
--task chat \
|
|
--hf-chat-template \
|
|
--add-eos-token '<|eot_id|>' \
|
|
--top-p 0.9 \
|
|
--temperature 1 \
|
|
--num-layers 32 \
|
|
--num-layers 80 \
|
|
--hidden-size 8192 \
|
|
--ffn-hidden-size 28672 \
|
|
--position-embedding-type rope \
|
|
--seq-length 4096 \
|
|
--max-new-tokens 256 \
|
|
--micro-batch-size 1 \
|
|
--num-attention-heads 64 \
|
|
--max-position-embeddings 4096 \
|
|
--swiglu \
|
|
--load "${CHECKPOINT}" \
|
|
--tokenizer-type PretrainedFromHF \
|
|
--tokenizer-name-or-path "${TOKENIZER_PATH}" \
|
|
--tokenizer-model "${TOKENIZER_MODEL}" \
|
|
--tokenizer-not-use-fast \
|
|
--fp16 \
|
|
--group-query-attention \
|
|
--num-query-groups 8 \
|
|
--normalization RMSNorm \
|
|
--untie-embeddings-and-output-weights \
|
|
--disable-bias-linear \
|
|
--attention-softmax-in-fp32 \
|
|
--no-load-optim \
|
|
--no-load-rng \
|
|
--no-masked-softmax-fusion \
|
|
--no-gradient-accumulation-fusion \
|
|
--exit-on-missing-checkpoint \
|
|
--make-vocab-size-divisible-by 1 \
|
|
--use-kv-cache \
|
|
--use-flash-attn \
|
|
| tee logs/chat_llama2_70b.log
|