refactor mixtral moe block.

[lkk12014402]

What does this PR do?

moe block forward regresion brought by #1511 for training

[lkk12014402]

the pr fix the segmentation fault issue that caused by DynamicMoE from this #1511, when train mixtral model

[lkk12014402]

training

DEEPSPEED_HPU_ZERO3_SYNC_MARK_STEP_REQUIRED=1 python ../gaudi_spawn.py --world_size 4 --use_deepspeed sft.py \
    --model_name_or_path mistralai/Mixtral-8x7B-Instruct-v0.1 \
    --dataset_name "philschmid/dolly-15k-oai-style" \
    --subset 'data/' \
    --streaming False \
    --deepspeed ../language-modeling/llama2_ds_zero3_config.json \
    --output_dir="./model_mixtral" \
    --do_train \
    --max_steps=500 \
    --logging_steps=10 \
    --save_steps=100 \
    --per_device_train_batch_size=2 \
    --per_device_eval_batch_size=1 \
    --gradient_accumulation_steps=2 \
    --learning_rate=1e-4 \
    --lr_scheduler_type="cosine" \
    --warmup_steps=100 \
    --weight_decay=0.05 \
    --optim="paged_adamw_32bit" \
    --lora_target_modules "q_proj" "v_proj" \
    --bf16 \
    --remove_unused_columns=False \
    --max_seq_length 512 \
    --run_name="sft_mixtral" \
    --report_to=none \
    --use_habana \
    --use_lazy_mode

[lkk12014402]

inference

QUANT_CONFIG=./quantization_config/maxabs_measure.json python run_generation.py --model_name_or_path "mistralai/Mixtral-8x7B-Instruct-v0.1" --use_hpu_graphs --use_kv_cache --limit_hpu_graphs --bucket_size 128 --max_new_tokens 128 --batch_size 1 --bf16

QUANT_CONFIG=./quantization_config/maxabs_quant_mixtral.json python run_generation.py --model_name_or_path "mistralai/Mixtral-8x7B-Instruct-v0.1" --use_hpu_graphs --use_kv_cache --limit_hpu_graphs --bucket_size 128 --max_new_tokens 128 --batch_size 1 --bf16

python run_generation.py --model_name_or_path "mistralai/Mixtral-8x7B-Instruct-v0.1" --use_hpu_graphs --use_kv_cache --limit_hpu_graphs --bucket_size 128 --max_new_tokens 512 --batch_size 4 --bf16