Skip to content

DeepGEMM vLLM performance benchmarks #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
cleonard530 wants to merge 4 commits into
base: main
Choose a base branch
from
Open

DeepGEMM vLLM performance benchmarks #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
277d098
performance benchmarks for the deep_gemm kernels after migrating to t…
cleonard530 Oct 1, 2026
4ea5a62
new benchmarks using cache update on DeepGEMM for element_size
cleonard530 Oct 2, 2026
f699fc2
updated benchmarks to reflect updates to DeepGEMM with element_size c…
cleonard530 Oct 2, 2026
dd10e5d
used an old file for the _tn stable path. Fixed to use updateded path
cleonard530 Oct 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
157 changes: 157 additions & 0 deletions benchmark_deepgemm_dispatch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,157 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import argparse
import importlib
import json
import time
from collections.abc import Callable

import torch

import vllm.envs as envs
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8,
)
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import per_block_cast_to_fp8


def measure_wall_time(
fn: Callable[[], None], warmups: int, repetitions: int
) -> float:
for _ in range(warmups):
fn()
torch.cuda.synchronize()

start = time.perf_counter()
for _ in range(repetitions):
fn()
torch.cuda.synchronize()
return (time.perf_counter() - start) * 1e3 / repetitions


def main() -> None:
parser = argparse.ArgumentParser(
description=(
"Measure repeated DeepGEMM dispatches in eager mode and CUDA graph "
"replay. The difference exposes per-call host dispatch overhead."
)
)
parser.add_argument("--m", type=int, default=128)
parser.add_argument("--n", type=int, default=4096)
parser.add_argument("--k", type=int, default=7168)
parser.add_argument("--calls-per-forward", type=int, default=200)
parser.add_argument("--warmups", type=int, default=5)
parser.add_argument("--repetitions", type=int, default=20)
parser.add_argument("--skip-cudagraph", action="store_true")
parser.add_argument("--output", type=str)
args = parser.parse_args()

if not torch.cuda.is_available() or not current_platform.support_deep_gemm():
raise RuntimeError("This benchmark requires CUDA with DeepGEMM support")

deep_gemm = importlib.import_module("vllm.third_party.deep_gemm")
binding = importlib.import_module("vllm.third_party.deep_gemm._C")
package_file = str(deep_gemm.__file__)
binding_file = str(binding.__file__)
print(f"DeepGEMM package: {package_file}")
print(f"DeepGEMM binding: {binding_file}")

if current_platform.is_arch_support_pdl():
deep_gemm.set_pdl(True)

if args.m <= 0 or args.n <= 0 or args.k <= 0:
raise ValueError("M, N, and K must be positive")
if args.n % 64 != 0 or args.k % 128 != 0:
raise ValueError("DeepGEMM requires N divisible by 64 and K by 128")
if args.calls_per_forward <= 0:
raise ValueError("--calls-per-forward must be positive")

use_ue8m0 = envs.VLLM_USE_DEEP_GEMM_E8M0
a = torch.randn((args.m, args.k), device="cuda", dtype=torch.bfloat16)
b = torch.randn((args.n, args.k), device="cuda", dtype=torch.bfloat16)
a_fp8, a_scale = per_token_group_quant_fp8(
a,
group_size=128,
column_major_scales=True,
tma_aligned_scales=True,
use_ue8m0=use_ue8m0,
)
b_fp8, b_scale = per_block_cast_to_fp8(
b, block_size=[128, 128], use_ue8m0=use_ue8m0
)
output = torch.empty(
(args.m, args.n), device="cuda", dtype=torch.bfloat16
)

def one_call() -> None:
deep_gemm.fp8_fp4_gemm_tn(
(a_fp8.t(), a_scale.t()),
(b_fp8.t(), b_scale.t()),
output,
disable_ue8m0_cast=not use_ue8m0,
)

def repeated_calls() -> None:
for _ in range(args.calls_per_forward):
one_call()

eager_ms = measure_wall_time(
repeated_calls, args.warmups, args.repetitions
)
results: dict[str, object] = {
"device": torch.cuda.get_device_name(),
"binding": {
"package_file": package_file,
"binding_file": binding_file,
},
"shape": {"m": args.m, "n": args.n, "k": args.k},
"calls_per_forward": args.calls_per_forward,
"warmups": args.warmups,
"repetitions": args.repetitions,
"eager_ms_per_forward": eager_ms,
"eager_ms_per_call": eager_ms / args.calls_per_forward,
}

if not args.skip_cudagraph:
graph = torch.cuda.CUDAGraph()
repeated_calls()
torch.cuda.synchronize()
with torch.cuda.graph(graph):
repeated_calls()

graph_ms = measure_wall_time(
graph.replay, args.warmups, args.repetitions
)
results.update(
{
"cudagraph_ms_per_forward": graph_ms,
"cudagraph_ms_per_call": graph_ms / args.calls_per_forward,
"eager_minus_cudagraph_us_per_call": (
eager_ms - graph_ms
)
* 1000
/ args.calls_per_forward,
}
)

rendered = json.dumps(results, indent=2)
print(rendered)
if args.output:
with open(args.output, "w", encoding="utf-8") as output_file:
output_file.write(rendered + "\n")


if __name__ == "__main__":
main()


# python benchmarks/kernels/benchmark_deepgemm_dispatch.py \
# --m 128 --n 4096 --k 7168 \
# --calls-per-forward 200 \
# --repetitions 20 \
# --output deepgemm_dispatch_results.json


# python benchmarks/kernels/benchmark_deepgemm_dispatch.py --output deepgemm_dispatch_results.json
Loading