scripts/cal_mfu.py

# coding=utf-8
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import json
import os

import fire
import torch
import torch.distributed as dist
from transformers import AutoConfig

from llamafactory.train.tuner import run_exp


BASE = 2  # gemm (add + mul)


def compute_model_flops(
    model_name_or_path: str,
    total_batch_size: int,
    seq_length: int,
    include_backward: bool = True,
    include_recompute: bool = False,
    include_flashattn: bool = False,
) -> int:
    r"""
    Calculates the FLOPs of model per forward/backward pass.
    """
    config = AutoConfig.from_pretrained(model_name_or_path)
    hidden_size = getattr(config, "hidden_size", None)
    vocab_size = getattr(config, "vocab_size", None)
    intermediate_size = getattr(config, "intermediate_size", None)
    num_attention_heads = getattr(config, "num_attention_heads", None)
    num_key_value_heads = getattr(config, "num_key_value_heads", None)
    num_hidden_layers = getattr(config, "num_hidden_layers", None)
    tie_word_embeddings = getattr(config, "tie_word_embeddings", False)

    # mlp module
    mlp_flops_per_token = 3 * BASE * hidden_size * intermediate_size  # up, gate, down
    mlp_flops = total_batch_size * seq_length * num_hidden_layers * mlp_flops_per_token

    # attn projector module
    q_flops_per_token = BASE * hidden_size * hidden_size
    o_flops_per_token = BASE * hidden_size * hidden_size
    k_flops_per_token = BASE * hidden_size * hidden_size * num_key_value_heads // num_attention_heads
    v_flops_per_token = BASE * hidden_size * hidden_size * num_key_value_heads // num_attention_heads
    attn_proj_flops_per_token = q_flops_per_token + o_flops_per_token + k_flops_per_token + v_flops_per_token
    attn_proj_flops = total_batch_size * seq_length * num_hidden_layers * attn_proj_flops_per_token

    # attn sdpa module
    sdpa_flops_per_layer = 2 * BASE * hidden_size * seq_length * seq_length  # (q * k^T) * v
    sdpa_flops = total_batch_size * num_hidden_layers * sdpa_flops_per_layer

    # embedding module
    embedding_flops_per_token = hidden_size * vocab_size
    embedding_flops = total_batch_size * seq_length * embedding_flops_per_token
    if tie_word_embeddings is False:
        embedding_flops *= 2

    non_embedding_flops = mlp_flops + attn_proj_flops + sdpa_flops
    non_embedding_coeff, embedding_coeff = 1, 1
    if include_backward:
        non_embedding_coeff += 2
        embedding_coeff += 2

    if include_recompute:
        non_embedding_coeff += 1

    total_flops = non_embedding_coeff * non_embedding_flops + embedding_coeff * embedding_flops

    if include_flashattn:
        total_flops += sdpa_flops

    return total_flops


def compute_device_flops(world_size: int) -> float:
    r"""
    Calculates the FLOPs of the device capability per second.
    """
    device_name = torch.cuda.get_device_name()
    if "H100" in device_name or "H800" in device_name:
        return 989 * 1e12 * world_size
    elif "A100" in device_name or "A800" in device_name:
        return 312 * 1e12 * world_size
    elif "V100" in device_name:
        return 125 * 1e12 * world_size
    elif "4090" in device_name:
        return 98 * 1e12 * world_size
    else:
        raise NotImplementedError("Device not supported: {}.".format(device_name))


def calculate_mfu(
    model_name_or_path: str,
    batch_size: int = 1,
    seq_length: int = 1024,
    num_steps: int = 100,
    finetuning_type: str = "lora",
    flash_attn: str = "auto",
    deepspeed_stage: int = 0,
    disable_gc: bool = False,
    liger_kernel: bool = False,
    unsloth_gc: bool = False,
) -> float:
    r"""
    Calculates MFU for given model and hyper-params.
    Usage: python cal_mfu.py --model_name_or_path path_to_model --batch_size 1 --seq_length 1024
    """
    args = {
        "model_name_or_path": model_name_or_path,
        "flash_attn": flash_attn,
        "disable_gradient_checkpointing": disable_gc,
        "enable_liger_kernel": liger_kernel,
        "use_unsloth_gc": unsloth_gc,
        "stage": "pt",
        "do_train": True,
        "finetuning_type": finetuning_type,
        "dataset": "c4_demo",
        "cutoff_len": seq_length,
        "output_dir": os.path.join("saves", "test_mfu"),
        "logging_strategy": "no",
        "save_strategy": "no",
        "save_only_model": True,
        "overwrite_output_dir": True,
        "per_device_train_batch_size": batch_size,
        "max_steps": num_steps,
        "bf16": True,
    }
    if deepspeed_stage in [2, 3]:
        args["deepspeed"] = "examples/deepspeed/ds_z{}_config.json".format(deepspeed_stage)

    run_exp(args)
    with open(os.path.join("saves", "test_mfu", "all_results.json"), "r", encoding="utf-8") as f:
        result = json.load(f)

    if dist.is_initialized():
        world_size = dist.get_world_size()
    else:
        world_size = 1

    total_batch_size = batch_size * world_size
    mfu_value = (
        result["train_steps_per_second"]
        * compute_model_flops(model_name_or_path, total_batch_size, seq_length)
        / compute_device_flops(world_size)
    )
    print("MFU: {:.2f}%".format(mfu_value * 100))


if __name__ == "__main__":
    fire.Fire(calculate_mfu)