mirror of
https://github.com/linkedin/Liger-Kernel.git
synced 2026-10-12 02:17:47 +00:00
lowdy1
ace042ec55
## Summary This PR is a descendant of [[PR #1195](https://github.com/linkedin/Liger-Kernel/pull/1195)] and continues the benchmark refactor effort by migrating linear-space-complexity benchmark scripts to the new reusable `setup` helper infrastructure. ### Benefits * Significantly reduces benchmark script complexity * Makes new benchmark additions easier and more consistent * Centralizes probing and sweep logic * Improves maintainability of benchmark infrastructure ### Notes This PR mainly focuses on kernels with approximately linear memory complexity and migrates them to the new benchmark helper APIs introduced in the parent refactor. - [ ] run `make test` to ensure correctness - [x] run `make checkstyle` to ensure code style - [ ] run `make test-convergence` to ensure convergence
111 lines
3.5 KiB
Python
111 lines
3.5 KiB
Python
import torch
|
|
|
|
from benchmark_model_configs import MODEL_REGISTRY
|
|
from benchmark_model_configs import build_model_config_sweep
|
|
from benchmark_model_configs import build_token_length_sweep
|
|
from benchmark_model_configs import get_benchmark_model_config
|
|
from transformers.models.llama.configuration_llama import LlamaConfig
|
|
from transformers.models.llama.modeling_llama import LlamaMLP
|
|
from utils import SingleBenchmarkRunInput
|
|
from utils import build_memory_bench_fn
|
|
from utils import build_speed_bench_fn
|
|
from utils import parse_benchmark_script_args
|
|
from utils import run_benchmarks
|
|
|
|
from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
|
|
from liger_kernel.utils import infer_device
|
|
|
|
device = infer_device()
|
|
|
|
|
|
def setup_swiglu(input: SingleBenchmarkRunInput):
|
|
"""Create input tensor and SwiGLU layer from benchmark config."""
|
|
cfg = input.extra_benchmark_config
|
|
if isinstance(input.x, str):
|
|
model_cfg = MODEL_REGISTRY[input.x]
|
|
seq_len = cfg["seq_len"]
|
|
hidden_size = model_cfg.hidden_size
|
|
intermediate_size = model_cfg.intermediate_size
|
|
dtype = model_cfg.dtype
|
|
else:
|
|
seq_len = input.x
|
|
hidden_size = cfg["hidden_size"]
|
|
intermediate_size = cfg["intermediate_size"]
|
|
dtype = cfg["dtype"]
|
|
|
|
llama_config = LlamaConfig(
|
|
hidden_size=hidden_size,
|
|
intermediate_size=intermediate_size,
|
|
hidden_act=cfg["hidden_act"],
|
|
)
|
|
x = torch.randn(
|
|
cfg["bsz"],
|
|
seq_len,
|
|
hidden_size,
|
|
device=device,
|
|
dtype=dtype,
|
|
requires_grad=True,
|
|
)
|
|
if input.kernel_provider == "liger":
|
|
layer = LigerSwiGLUMLP(config=llama_config).to(device).to(dtype)
|
|
elif input.kernel_provider == "huggingface":
|
|
layer = LlamaMLP(config=llama_config).to(device).to(dtype)
|
|
else:
|
|
raise ValueError(f"Invalid provider: {input.kernel_provider} for SwiGLU")
|
|
return x, layer
|
|
|
|
|
|
if __name__ == "__main__":
|
|
args = parse_benchmark_script_args()
|
|
|
|
if args.sweep_mode == "model_config":
|
|
common_configs = build_model_config_sweep(
|
|
kernel_name="swiglu",
|
|
setup_fn=setup_swiglu,
|
|
model_keys=["hidden_size", "intermediate_size", "dtype"],
|
|
probe_provider="huggingface",
|
|
extra_configs={
|
|
"bsz": 1,
|
|
"hidden_act": "silu",
|
|
},
|
|
probe_dim="T",
|
|
bt=args.bt,
|
|
overwrite=args.overwrite,
|
|
)
|
|
else:
|
|
model = get_benchmark_model_config(args.model)
|
|
probe_seq_len = 1024
|
|
|
|
common_configs = build_token_length_sweep(
|
|
kernel_name="swiglu",
|
|
probe_x=probe_seq_len,
|
|
model=model,
|
|
setup_fn=setup_swiglu,
|
|
model_keys=["hidden_size", "intermediate_size", "dtype"],
|
|
extra_configs={
|
|
"bsz": 1,
|
|
"hidden_act": "silu",
|
|
},
|
|
scale_dim="T",
|
|
x_label="total tokens",
|
|
probe_provider="huggingface",
|
|
overwrite=args.overwrite,
|
|
)
|
|
|
|
common_configs["kernel_providers"] = ["huggingface", "liger"]
|
|
|
|
run_benchmarks(
|
|
bench_test_fn=build_speed_bench_fn(setup_swiglu),
|
|
kernel_operation_modes=["forward", "backward", "full"],
|
|
metric_name="speed",
|
|
metric_unit="ms",
|
|
**common_configs,
|
|
)
|
|
run_benchmarks(
|
|
bench_test_fn=build_memory_bench_fn(setup_swiglu),
|
|
kernel_operation_modes=["full", "forward", "backward"],
|
|
metric_name="memory",
|
|
metric_unit="MB",
|
|
**common_configs,
|
|
)
|