通义千问 - Code Qwen能力算法赛道季军方案

{
  "task_id": "Python/0",
  "prompt": "from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n    \"\"\" Check if in given list of numbers, are any two numbers closer to each other than\n    given threshold.\n    >>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n    False\n    >>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n    True\n    \"\"\"\n",
  "declaration": "from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n",
  "canonical_solution": "    for idx, elem in enumerate(numbers):\n        for idx2, elem2 in enumerate(numbers):\n            if idx != idx2:\n                distance = abs(elem - elem2)\n                if distance < threshold:\n                    return True\n\n    return False\n",
  "buggy_solution": "    for idx, elem in enumerate(numbers):\n        for idx2, elem2 in enumerate(numbers):\n            if idx != idx2:\n                distance = elem - elem2\n                if distance < threshold:\n                    return True\n\n    return False\n",
  "bug_type": "missing logic",
  "failure_symptoms": "incorrect output",
  "entry_point": "has_close_elements",
  "import": ""
  "test_setup": ""
  "test": "\n\n\n\n\ndef check(has_close_elements):\n    assert has_close_elements([1.0, 2.0, 3.9, 4.0, 5.0, 2.2], 0.3) == True\n    assert has_close_elements([1.0, 2.0, 3.9, 4.0, 5.0, 2.2], 0.05) == False\n    assert has_close_elements([1.0, 2.0, 5.9, 4.0, 5.0], 0.95) == True\n    assert has_close_elements([1.0, 2.0, 5.9, 4.0, 5.0], 0.8) == False\n    assert has_close_elements([1.0, 2.0, 3.0, 4.0, 5.0, 2.0], 0.1) == True\n    assert has_close_elements([1.1, 2.2, 3.1, 4.1, 5.1], 1.0) == True\n    assert has_close_elements([1.1, 2.2, 3.1, 4.1, 5.1], 0.5) == False\n\ncheck(has_close_elements)",
  "example_test": "def check(has_close_elements):\n    assert has_close_elements([1.0, 2.0, 3.0], 0.5) == False\n    assert has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3) == True\ncheck(has_close_elements)\n",
  "signature": "has_close_elements(numbers: List[float], threshold: float) -> bool",
  "docstring": "Check if in given list of numbers, are any two numbers closer to each other than\ngiven threshold.\n>>> has_close_elements([1.0, 2.0, 3.0], 0.5)\nFalse\n>>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\nTrue",
  "instruction": "Write a Python function `has_close_elements(numbers: List[float], threshold: float) -> bool` to solve the following problem:\nCheck if in given list of numbers, are any two numbers closer to each other than\ngiven threshold.\n>>> has_close_elements([1.0, 2.0, 3.0], 0.5)\nFalse\n>>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\nTrue"
}

{
    'task_id': 1,
    'text': 'Write a function to find the minimum cost path to reach (m, n) from (0, 0) for the given cost matrix cost[][] and a position (m, n) in cost[][].',
    'code': 'R = 3\r\nC = 3\r\ndef min_cost(cost, m, n): \r\n\ttc = [[0 for x in range(C)] for x in range(R)] \r\n\ttc[0][0] = cost[0][0] \r\n\tfor i in range(1, m+1): \r\n\t\ttc[i][0] = tc[i-1][0] + cost[i][0] \r\n\tfor j in range(1, n+1): \r\n\t\ttc[0][j] = tc[0][j-1] + cost[0][j] \r\n\tfor i in range(1, m+1): \r\n\t\tfor j in range(1, n+1): \r\n\t\t\ttc[i][j] = min(tc[i-1][j-1], tc[i-1][j], tc[i][j-1]) + cost[i][j] \r\n\treturn tc[m][n]',
    'test_list': [
        'assert min_cost([[1, 2, 3], [4, 8, 2], [1, 5, 3]], 2, 2) == 8',
        'assert min_cost([[2, 3, 4], [5, 9, 3], [2, 6, 4]], 2, 2) == 12',
        'assert min_cost([[3, 4, 5], [6, 10, 4], [3, 7, 5]], 2, 2) == 16'],
    'test_setup_code': '',
    'challenge_test_list': []
}

# Specify the number of GPU cards per node. Please set according to your specific situation!
N_GPU_PER_NODE=1
# Specify the number of nodes (machines) can be used. Please set according to your specific situation!
N_NODE=1

accelerate launch \
  --num_machines $N_NODE \
  --num_processes $(($N_NODE*$N_GPU_PER_NODE)) \
  --use_deepspeed \
  --deepspeed_multinode_launcher 'standard' \
  --zero_stage 2 \
  --offload_optimizer_device 'cpu' \
  --offload_param_device 'none' \
  --gradient_accumulation_steps 1 \
  --gradient_clipping 1.0 \
  --zero3_init_flag false \
  --zero3_save_16bit_model false \
  --main_training_function 'main' \
  --mixed_precision 'bf16' \
  --dynamo_backend 'no' \
  --same_network \
  --machine_rank $RANK \
  --main_process_ip $MASTER_ADDR \
  --main_process_port $MASTER_PORT \
  --rdzv_backend 'static' \
  mft_accelerate.py --train_config configs/qwen_train_config_1_8B.json

N_NODE=1
N_GPU_PER_NODE=1

# If you need to set system prompt, set it here, otherwise you can set it as empty string. Decide whether to add system prompt by yourself.
system="$start_tag"$system$'\n'"$end_tag"$'\n'

for task in "${tasks[@]}"; do

    if [ "$task" == "mbpp" ]; then
        prefix="$system""$start_tag"${user}$'\n'
        suffix="$end_tag"$'\n'"$start_tag"${assistant}
    else
        prefix=""
        suffix=""
    fi

    generations_path=$generation_base_dir/generations_$model_name/generations_$task\_$model_name.json
    if [ ! -d $generation_base_dir/generations_$model_name ]; then
        mkdir $generation_base_dir/generations_$model_name
    fi

    echo "start to launch ...."
    accelerate launch \
            --num_machines $N_NODE \
            --num_processes $(($N_NODE*$N_GPU_PER_NODE)) \
            main.py \
                --model $model \
                --task $task \
                --prompt instruct \
                --n_samples $n_samples \
                --batch_size $batch_size \
                --max_length_generation 2000 \
                --do_sample False \
                --temperature 0.2 \
                --precision bf16 \
                --eos "$eos_token" \
                --seed 999999999 \
                --add_special_tokens True \
                --trust_remote_code \
                --generation_only \
                --save_generations_path $generations_path \
                --prefix "$prefix" \
                --suffix "$suffix"
    
    echo "Task $task done"
done

def loss_func_mft(outputs, labels, task_mask, task_id, weighted_loss_mode, loss_mask=None):
    """
    loss function for MFT loss
    :param outputs:
    :param labels:
    :param task_mask:
    :param task_id:
    :param weighted_loss_mode:
    :param loss_mask:
    :return:
    """
    # task_id shape: [[1], [2], [4], [3], ..., [1]]
    weighted = weighted_loss_mode
    lm_logits = outputs["logits"]
    labels = labels.to(device=lm_logits.device)
    task_mask = task_mask.to(device=lm_logits.device)
    task_id = task_id.to(device=lm_logits.device)
    shift_logits = lm_logits.contiguous()
    labels = labels.contiguous()

    bsz, seq_len = labels.shape
    # loss_mask = None
    if loss_mask is None:
        ineffective_tokens_per_sample = (labels==-100).sum(dim=1)
        effective_tokens_per_sample = - (ineffective_tokens_per_sample - seq_len)
        effective_tokens = bsz * seq_len - ineffective_tokens_per_sample.sum()
        loss_fct = CrossEntropyLoss(reduction='none', ignore_index=-100)
    else:
        loss_mask = loss_mask.to(device=lm_logits.device)
        loss_fct = CrossEntropyLoss(reduction='none')
    losses = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1))  # [B * L, 1]
    losses = losses.contiguous().view(bsz, -1)
    token_losses = losses.clone().detach().float() if loss_mask is None else losses.clone().detach().float() * loss_mask  # [B, L]
    task_mask_trans = torch.transpose(task_mask, 0, 1)
    if weighted_loss_mode == "case3" or weighted_loss_mode == "case4":
        unique_id = torch.unique(task_id)
        loss = 0.0

        for i, w in enumerate(unique_id):
            row_idx = torch.squeeze(task_id) == w.item()
            if weighted_loss_mode == "case3":
                if loss_mask is None:
                    loss += torch.sum(losses[row_idx, :]) / torch.sum(effective_tokens_per_sample[row_idx])
                else:
                    loss += torch.sum((losses * loss_mask)[row_idx, :]) / torch.sum(loss_mask[row_idx, :])
            elif weighted_loss_mode == "case4":
                if loss_mask is None:
                    loss += torch.mean(torch.sum(losses, dim=1)[row_idx] / effective_tokens_per_sample[row_idx])
                else:
                    loss += torch.mean(torch.sum(losses * loss_mask, dim=1)[row_idx] / torch.sum(loss_mask, dim=1)[row_idx])

        loss /= len(unique_id)

    elif weighted_loss_mode == "case2":
        if loss_mask is None:
            loss = torch.mean(torch.sum(losses, dim=1) / effective_tokens_per_sample)
        else:
            loss = torch.mean(torch.sum(losses * loss_mask, dim=1) / torch.sum(loss_mask, dim=1))
    elif weighted_loss_mode == "case1":
        # flatten losses & loss_mask tensor
        if loss_mask is None:
            losses = losses.view(-1)
            loss = torch.sum(losses) / effective_tokens
        else:
            loss_mask = loss_mask.view(-1)
            losses = losses.view(-1)
            loss = torch.sum(losses * loss_mask) / loss_mask.sum()

    # fix task order
    task_loss = torch.zeros(len(ID2TASK)).to(device=task_id.device)
    task_num = torch.zeros(len(ID2TASK)).to(device=task_id.device)
    for i, w in enumerate(unique_id):
        row_idx = torch.squeeze(task_id) == w.item()
        if loss_mask is None:
            task_loss[w] = torch.sum(token_losses[row_idx, :]) / torch.sum(effective_tokens_per_sample[row_idx])
            task_num[w] = len(effective_tokens_per_sample[row_idx])
        else:
            task_loss[w] = torch.sum((losses * loss_mask)[row_idx, :]) / torch.sum(loss_mask[row_idx, :])
    
    return loss, task_loss, task_num