o
    ߥi                      @   s   d dl Z d dlZd dlZd dlmZmZ d dlmZ d dlm	Z	 d dl
mZmZmZ d dlmZ ddlmZ d	d
 Zdd Zdd Zdd Zdd Zdd ZdS )    N)get_argsget_tokenizer)initialize_distributed)GLM130B)get_model_parallel_groupget_model_parallel_rankget_model_parallel_world_size)load_checkpoint   )quantizec                 C   s0   |  d}|jdddd |jdtddd	 | S )
zArguments for BMInfBMInfz--bminf
store_truez,Use BMInf to support low resource evaluationactionhelpz--bminf-memory-limit   z$Max memory for model per GPU (in GB)typedefaultr   )add_argument_groupadd_argumentintparsergroup r   ]/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/nlp/glm_130b/initialize.pyadd_bminf_args   s   
r   c                 C   s0   |  d}|jdtdd |jdtddd d S )	NQuantizationz--quantization-bit-width   )r   r   z--from-quantized-checkpointTz#Loading from a quantized checkpointr   )r   r   r   boolr   r   r   r   add_quantization_args"   s   

r!   c                 C   s   |  d}|jdddd d S )NInitializationz--sequential-initializationr   zTInitialize sequentially in tensor parallel group (reduce CPU RAM for initialization)r   )r   r   r   r   r   r   add_initialization_args-   s   

r#   c                 C   sL   d| _ d| _d| _d| _d| _d| _d| _d| _d	| _d
| _	d
| _
d| _| S )Nr   F   i 0  i   i L `   i   zicetk-glm-130BpostT	inference)model_parallel_size
num_layershidden_sizeinner_hidden_size
vocab_sizenum_attention_headsmax_sequence_lengthtokenizer_typelayernorm_order	skip_initfp16mode)argsr   r   r   set_up_model_args8   s   r5   c                 C   s   t jdd}t| t| t| t| | | | \}}|g d7 }t|}t	|}t j
di t|t|}d|_t| |S )NF)add_help)z--model-parallel-size4z--moder'   r   )argparseArgumentParserr   r!   r#   r   add_model_specific_argsparse_known_argsr   r5   	Namespacevarsdo_trainr   )extra_args_providerr   known	args_listr4   r   r   r   
initializeH   s   
rB   c                 C   s  t | }tj  t }tt D ]{}t |krt| 	 }| j
r/| jd us)J t|| j}t||  | jd urB| j
sBt|| j}| jr|dd l}tj dkrYtd| j d tj| j |j|d| jd> d}W d    n1 svw   Y  n|| j}| jrtjjt d qtj  tj dkrtdt | d	d
 tj  |  |jj}d|j_t : |tjd| jtj tj dtj!| jtj tj d"ddtj#dd| j| jtj ddk ^}}W d    n1 sw   Y  ||j_tj  ||fS )Nr   z!> BMInf activated, memory limit: z GBF   )quantizationmemory_limit)r   z> Model initialized in z.1fsTr
   )devicedtype)rG   g      ?)$r   torchdistributedbarriertimeranger   r   r   halffrom_quantized_checkpointquantization_bit_widthr   r	   bminfget_rankprintbminf_memory_limitcudarG   wrappertosequential_initializationr   empty_cacheevaltransformerparallel_outputno_gradonesr.   current_deviceint64arangeviewrandn)r4   	tokenizerstartimodelrR   original_parallel_output_r   r   r   initialize_model_and_tokenizerY   s   







rk   )r8   rM   rJ   SwissArmyTransformerr   r   SwissArmyTransformer.argumentsr   SwissArmyTransformer.modelr   SwissArmyTransformer.mpur   r   r   SwissArmyTransformer.trainingr	   rD   r   r   r!   r#   r5   rB   rk   r   r   r   r   <module>   s   