mirror of
https://github.com/hiyouga/LLaMA-Factory.git
synced 2025-11-04 18:02:19 +08:00
Update cal_mfu.py
Former-commit-id: 0c391b2e59943b0ca9dd4e8561398e7c856a4b29
This commit is contained in:
hoshi-hiyouga
GitHub
parent
9f2427907e
commit
3021b31cf3
@ -11,116 +11,139 @@
|
||||
# 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
|
||||
# limitations under the License.
|
||||
|
||||
import json
|
||||
import os
|
||||
|
||||
import fire
|
||||
import torch
|
||||
from transformers import AutoConfig
|
||||
import fire
|
||||
def model_flops_counter(
|
||||
|
||||
from llamafactory.train.tuner import run_exp
|
||||
|
||||
|
||||
BASE = 2 # gemm (add + mul)
|
||||
|
||||
|
||||
def compute_model_flops(
|
||||
model_name_or_path: str,
|
||||
batch_size: int,
|
||||
seqlen: int,
|
||||
model_config: dict,
|
||||
is_backward: bool = True,
|
||||
is_recompute: bool = False,
|
||||
is_flashattn: bool = False,
|
||||
) -> float:
|
||||
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.
|
||||
"""
|
||||
calculate the FLOPs of model per iteration
|
||||
"""
|
||||
hidden_size = model_config.hidden_size
|
||||
num_attention_heads = model_config.num_attention_heads
|
||||
num_key_value_heads = model_config.num_key_value_heads
|
||||
vocab_size = model_config.vocab_size
|
||||
intermediate_size = model_config.intermediate_size
|
||||
num_hidden_layers = model_config.num_hidden_layers
|
||||
"""
|
||||
B: batch_size
|
||||
S: seqlen
|
||||
L: num_hidden_layers
|
||||
H: hidden_size
|
||||
V: vocab_size
|
||||
I: intermediate_size
|
||||
"""
|
||||
### MLP calculation
|
||||
per_mlp_calculation = 2 * hidden_size * intermediate_size
|
||||
mlp_calculation_per_layer = per_mlp_calculation * 3
|
||||
mlp_calculation = batch_size * seqlen * mlp_calculation_per_layer * num_hidden_layers
|
||||
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)
|
||||
|
||||
### Attention calculation
|
||||
Q_calculation = 2 * hidden_size * hidden_size
|
||||
O_calculation = 2 * hidden_size * hidden_size
|
||||
K_calculation = 2 * hidden_size * hidden_size * num_key_value_heads / num_attention_heads
|
||||
V_calculation = 2 * hidden_size * hidden_size * num_key_value_heads / num_attention_heads
|
||||
|
||||
QKVO_calculation = Q_calculation + O_calculation + K_calculation + V_calculation # 8H^2 / coe
|
||||
self_attn_calculation = seqlen * hidden_size * 2 * 2 # (4 * S * H)
|
||||
attention_calculation = batch_size * seqlen * num_hidden_layers * (QKVO_calculation + self_attn_calculation) # BSL(8H^2/coe + 4S * H)
|
||||
|
||||
#Embedding and LMhead calculation
|
||||
embedding_calculation = hidden_size * vocab_size
|
||||
lmhead_calculation = hidden_size * vocab_size
|
||||
IO_calculation = 3 * batch_size * seqlen * (embedding_calculation + lmhead_calculation) # 2 *(1+2)BSHV
|
||||
E = attention_calculation + mlp_calculation
|
||||
coefficient = 3
|
||||
fix_term = 0
|
||||
if(is_recompute):
|
||||
coefficient = 4
|
||||
if(is_flashattn):
|
||||
fix_term = batch_size *seqlen * self_attn_calculation
|
||||
|
||||
total_calculation = coefficient * E + IO_calculation + fix_term
|
||||
|
||||
return total_calculation
|
||||
# mlp module
|
||||
mlp_flops_per_token = 3 * BASE * hidden_size * intermediate_size # up, gate, down
|
||||
mlp_flops = 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 = 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 = batch_size * num_hidden_layers * sdpa_flops_per_layer
|
||||
|
||||
# embedding module
|
||||
embedding_flops_per_token = hidden_size * vocab_size
|
||||
embedding_flops = 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 hardware_flops_counter(
|
||||
seconds: float, # seconds used in given iterations
|
||||
num_gpus: int = 1,
|
||||
) -> float:
|
||||
if "A100" in torch.cuda.get_device_name():
|
||||
return 312 * 1e12 * seconds * num_gpus
|
||||
elif "V100" in torch.cuda.get_device_name():
|
||||
return 125 * 1e12 * seconds * num_gpus
|
||||
def compute_device_flops() -> float:
|
||||
device_name = torch.cuda.get_device_name()
|
||||
device_count = torch.cuda.device_count()
|
||||
if "H100" in device_name or "H800" in device_name:
|
||||
return 989 * 1e12 * device_count
|
||||
elif "A100" in device_name or "A800" in device_name:
|
||||
return 312 * 1e12 * device_count
|
||||
elif "V100" in device_name:
|
||||
return 125 * 1e12 * device_count
|
||||
elif "4090" in device_name:
|
||||
return 98 * 1e12 * device_count
|
||||
else:
|
||||
raise NotImplementedError("Device not supported: {}.".format(device_name))
|
||||
|
||||
|
||||
def compute_mfu(
|
||||
model_name_or_path: str,
|
||||
batch_size: int,
|
||||
seqlen: int,
|
||||
model_config: dict,
|
||||
num_iter: int,
|
||||
seconds: float,
|
||||
num_gpus: int = 1,
|
||||
seq_length: int,
|
||||
finetuning_type: str = "lora",
|
||||
flash_attn: str = "auto",
|
||||
deepspeed_stage: int = 0,
|
||||
disable_gc: bool = False,
|
||||
liger_kernel: bool = False,
|
||||
) -> float:
|
||||
r"""
|
||||
Computes 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
|
||||
"""
|
||||
compute MFU given model configuration, training config and training information
|
||||
"""
|
||||
percentage = (num_iter * model_flops_counter(batch_size,seqlen,model_config)) / hardware_flops_counter(seconds, num_gpus)
|
||||
|
||||
print(f"MFU : {percentage* 100:.2f}%")
|
||||
return percentage
|
||||
|
||||
# User input
|
||||
args = {
|
||||
"model_name_or_path": model_name_or_path,
|
||||
"flash_attn": flash_attn,
|
||||
"disable_gradient_checkpointing": disable_gc,
|
||||
"enable_liger_kernel": liger_kernel,
|
||||
"stage": "pt",
|
||||
"do_train": True,
|
||||
"finetuning_type": finetuning_type,
|
||||
"dataset": "c4_demo",
|
||||
"cutoff_len": seq_length,
|
||||
"output_dir": os.path.join("saves", "test_mfu"),
|
||||
"overwrite_output_dir": True,
|
||||
"per_device_train_batch_size": batch_size,
|
||||
"max_steps": 100,
|
||||
"bf16": True,
|
||||
}
|
||||
if deepspeed_stage in [2, 3]:
|
||||
args["deepspeed"] = "examples/deepspeed/ds_z{}_config.json".format(deepspeed_stage)
|
||||
|
||||
### model_name
|
||||
model_name = "meta-llama/Meta-Llama-3-8B-Instruct"
|
||||
run_exp(args)
|
||||
with open(os.path.join("saves", "test_mfu", "all_results.json"), "r", encoding="utf-8") as f:
|
||||
result = json.load(f)
|
||||
|
||||
### training config
|
||||
batch_size = 8
|
||||
seqlen = 1*1024
|
||||
num_gpus = 1
|
||||
|
||||
### training information
|
||||
num_iter = 225
|
||||
seconds = 605 # time used in {num_iter} iterations
|
||||
|
||||
model_config = AutoConfig.from_pretrained(model_name)
|
||||
if __name__ == "__main__":
|
||||
fire.Fire(
|
||||
compute_mfu(
|
||||
batch_size=batch_size,
|
||||
seqlen=seqlen,
|
||||
model_config=model_config,
|
||||
num_iter=num_iter,
|
||||
seconds=seconds,
|
||||
num_gpus=num_gpus
|
||||
)
|
||||
mfu_value = (
|
||||
result["train_samples_per_second"]
|
||||
* compute_model_flops(model_name_or_path, batch_size, seq_length)
|
||||
/ compute_device_flops()
|
||||
)
|
||||
print("MFU: {:.2f}%".format(mfu_value * 100))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
fire.Fire(compute_mfu)
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user