How to run the code with smaller ckpt like opt-6.7B

[czq693497091]

With limited GPU resources, how to use opt-6.7b to just run the code?

[czq693497091]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

[xbzjsj]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

the predictor also needs to be trained, and are you running with a single GPU?

[czq693497091]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

the predictor also needs to be trained, and are you running with a single GPU?

I trained the predictor with more than one gpus. But when I tried to train the mlp predictor, it shows all zero. I don't know how to solve it.

[kecheng-lab]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

[czq693497091]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

still not. I email to the author but she said that is should not be all zero.

[kecheng-lab]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

still not. I email to the author but she said that is should not be all zero.

sad..(T＿T).. thanks for your response

[kecheng-lab]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

still not. I email to the author but she said that is should not be all zero.

Hey, I have found the reason for the problem. The issue is with fp_label, because the author has allocated a fp_label.mmap file of size [400000 ,(4 * hidden_size)] for storing fp_label, but in reality, fp_label does not contain that much data. When the author was dividing the validation set, they selected the last 0.05 * len(fp_label) of data from the fp_labl file, which resulted in the reading of empty data. As a result, the MLP cannot receive effective training. To solve this problem, you can modify the def get_data(args, l) function in ./DejaVu/sparse_predictor/main_mlp.py.

def get_data(args, l):
    if CONFIG[args.model]['ckt_storage'] == "bylayer":
        #path = f"{DATA[args.model][args.dataset]}/mlp_x_{l}.mmap"
        path = f"{DATA[args.model][args.dataset]}/mlp_sp_x_{l}.mmap"
        print(f"Reading query from {path}")
        query = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d']))[: CONFIG[args.model]['N']])
        path = f"{DATA[args.model][args.dataset]}/mlp_label_{l}.mmap"
        print(f"Reading MLP label from {path}")
        label = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d'] * 4))[: CONFIG[args.model]['N']])
        
        num_valid = (label.sum(-1) > 0).sum()
        print(num_valid)
        return  query[:num_valid], label[:num_valid]
        #return  query, label

This is my solution, I hope it will be helpful to you😀

[czq693497091]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

still not. I email to the author but she said that is should not be all zero.

Hey, I have found the reason for the problem. The issue is with fp_label, because the author has allocated a fp_label.mmap file of size [400000 ,(4 * hidden_size)] for storing fp_label, but in reality, fp_label does not contain that much data. When the author was dividing the validation set, they selected the last 0.05 * len(fp_label) of data from the fp_labl file, which resulted in the reading of empty data. As a result, the MLP cannot receive effective training. To solve this problem, you can modify the def get_data(args, l) function in ./DejaVu/sparse_predictor/main_mlp.py.

def get_data(args, l):
    if CONFIG[args.model]['ckt_storage'] == "bylayer":
        #path = f"{DATA[args.model][args.dataset]}/mlp_x_{l}.mmap"
        path = f"{DATA[args.model][args.dataset]}/mlp_sp_x_{l}.mmap"
        print(f"Reading query from {path}")
        query = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d']))[: CONFIG[args.model]['N']])
        path = f"{DATA[args.model][args.dataset]}/mlp_label_{l}.mmap"
        print(f"Reading MLP label from {path}")
        label = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d'] * 4))[: CONFIG[args.model]['N']])
        
        num_valid = (label.sum(-1) > 0).sum()
        print(num_valid)
        return  query[:num_valid], label[:num_valid]
        #return  query, label

This is my solution, I hope it will be helpful to you😀

cool, thanks for your solution and I will try😀

[susavlsh10]

I successfully run the collect_sp_data.sh with opt-6.7b and opt-66b. But when I apply the sparse_predictor to train the mlp predictor for both opt-6.7b and opt-66b, but both shows all zero y, which means that the mlp_label_0.mmap is all zero and the training process ends. Is the problem widespread, and how can it be solved?

I encountered the same problem, but I don't know how to solve it.Have you solved the problem now?

still not. I email to the author but she said that is should not be all zero.

Hey, I have found the reason for the problem. The issue is with fp_label, because the author has allocated a fp_label.mmap file of size [400000 ,(4 * hidden_size)] for storing fp_label, but in reality, fp_label does not contain that much data. When the author was dividing the validation set, they selected the last 0.05 * len(fp_label) of data from the fp_labl file, which resulted in the reading of empty data. As a result, the MLP cannot receive effective training. To solve this problem, you can modify the def get_data(args, l) function in ./DejaVu/sparse_predictor/main_mlp.py.

def get_data(args, l):
    if CONFIG[args.model]['ckt_storage'] == "bylayer":
        #path = f"{DATA[args.model][args.dataset]}/mlp_x_{l}.mmap"
        path = f"{DATA[args.model][args.dataset]}/mlp_sp_x_{l}.mmap"
        print(f"Reading query from {path}")
        query = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d']))[: CONFIG[args.model]['N']])
        path = f"{DATA[args.model][args.dataset]}/mlp_label_{l}.mmap"
        print(f"Reading MLP label from {path}")
        label = np.array(np.memmap(path, dtype='float16', mode='r', shape=(400000,CONFIG[args.model]['d'] * 4))[: CONFIG[args.model]['N']])
        
        num_valid = (label.sum(-1) > 0).sum()
        print(num_valid)
        return  query[:num_valid], label[:num_valid]
        #return  query, label

This is my solution, I hope it will be helpful to you😀

Thanks!! this works :D

[jason-huang03]

sorry, but I met a problem when running 6.7b model. I got

RuntimeError: The expanded size of the tensor (2048) must match the existing size (2047) at non-singleton dimension 1.  Target sizes: [1, 2048].  Tensor sizes: [2047]

on Decentralized_FM_alpha/pipeline_parallel/dist_pipeline_inference_mask_greedy_token_pipe_sync.py line 485

        self.ret_tokens[
            index * self.micro_batch_size : (index + 1) * self.micro_batch_size,
            : self.i_current_token,
        ] = original_indices

I use a single gpu to run, and the script is like

#!/bin/bash

file=./c4_train.jsonl
    
echo "start running ${file}"

ARGS="--model-name opt-6.7b-converted \
--model-type opt-save \
--seed 42 \
--fp16 \
--num-layers 32 \
--max-layers 96 \
--budget 22800 \
--num-iters 2000 \
--dist-url tcp:https://127.0.0.1:9032 \
--token-micro-batch-size 1 \
--world-size 1 --pipeline-group-size 1 --data-group-size 1 \
--pp-mode pipe_sync_sample_mask_token_pipe \
--infer-data ${file}"

python dist_inference_runner.py $(echo ${ARGS}) --cuda-id 0 --rank 0

does anyone know what's the problem?