Skip to content

Tiny fix bench tgv gemm #2277

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

Merged
yzh119 merged 3 commits into from
Jan 5, 2026
Merged

Tiny fix bench tgv gemm #2277

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
46f87f8
tiny fix pre-commit
vincentzed Dec 31, 2025
fff4ca6
this is with cupti
vincentzed Jan 4, 2026
17a478d
basically scan all cupti... too slow
vincentzed Jan 5, 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
78 changes: 22 additions & 56 deletions benchmarks/bench_tgv_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,10 +4,10 @@
Tests the autotuner integration with TGV BF16 GEMM kernels.
"""

import time
import csv
import torch
import torch.nn.functional as F
from flashinfer.testing.utils import bench_gpu_time_with_cudagraph

from flashinfer import tgv_gemm_sm100, autotune

Expand Down Expand Up @@ -80,25 +80,13 @@ def test_tgv_gemm_bf16_sm100_perf():

torch.cuda.synchronize()

cublas_graph = torch.cuda.CUDAGraph()

# Start graph capture
with torch.cuda.graph(cublas_graph):
for _ in range(100):
_ = F.linear(A, B.T, bias)

# Warmup the graph
for _ in range(3):
cublas_graph.replay()

torch.cuda.synchronize()

# Benchmark using CUDA graph
start_time = time.time()
cublas_graph.replay()
torch.cuda.synchronize()
end_time = time.time()
cublas_avg_time = (end_time - start_time) / 100
cublas_times = bench_gpu_time_with_cudagraph(
lambda: F.linear(A, B.T, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)

@gemini-code-assist gemini-code-assist Bot Dec 31, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The previous implementation captured 100 iterations within the CUDA graph to amortize launch overhead. The bench_gpu_time_with_cudagraph function defaults to num_iters_within_graph=10. To maintain consistency with the previous benchmarking methodology and ensure better amortization of kernel launch overhead, it's recommended to explicitly set num_iters_within_graph=100.

Suggested change
cublas_times = bench_gpu_time_with_cudagraph(
lambda: F.linear(A, B.T, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)
cublas_times = bench_gpu_time_with_cudagraph(
lambda: F.linear(A, B.T, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
num_iters_within_graph=100,
)

cublas_avg_time = sum(cublas_times) / len(cublas_times) / 1000
print(
f"CUBLAS average time: {cublas_avg_time * 1000:.6f} ms, {flops / cublas_avg_time:.3f} TFLOPS"
)
Expand All @@ -110,48 +98,26 @@ def test_tgv_gemm_bf16_sm100_perf():

torch.cuda.synchronize()

tgv_graph = torch.cuda.CUDAGraph()

# Start graph capture
with torch.cuda.graph(tgv_graph):
for _ in range(100):
_ = tgv_gemm_sm100(A, B, bias)

# Warmup the graph
tgv_graph.replay()

torch.cuda.synchronize()

# Benchmark using CUDA graph
start_time = time.time()
tgv_graph.replay()
torch.cuda.synchronize()
end_time = time.time()

tgv_avg_time = (end_time - start_time) / 100
tgv_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)

@gemini-code-assist gemini-code-assist Bot Dec 31, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The previous implementation captured 100 iterations within the CUDA graph to amortize launch overhead. The bench_gpu_time_with_cudagraph function defaults to num_iters_within_graph=10. To maintain consistency with the previous benchmarking methodology and ensure better amortization of kernel launch overhead, it's recommended to explicitly set num_iters_within_graph=100.

Suggested change
tgv_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)
tgv_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
num_iters_within_graph=100,
)

tgv_avg_time = sum(tgv_times) / len(tgv_times) / 1000
print(
f"TGV average time: {tgv_avg_time * 1000:.6f} ms, {flops / tgv_avg_time:.3f} TFLOPS, speedup: {cublas_avg_time / tgv_avg_time:.2f}x"
)

# Test with PDL
print("\nTesting with PDL...")
pdl_graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(pdl_graph):
for _ in range(100):
_ = tgv_gemm_sm100(A, B, bias, pdl=True)

# Warmup the graph
pdl_graph.replay()

torch.cuda.synchronize()

# Benchmark using CUDA graph
start_time = time.time()
pdl_graph.replay()
torch.cuda.synchronize()
end_time = time.time()

pdl_avg_time = (end_time - start_time) / 100
pdl_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias, pdl=True),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)

@gemini-code-assist gemini-code-assist Bot Dec 31, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The previous implementation captured 100 iterations within the CUDA graph to amortize launch overhead. The bench_gpu_time_with_cudagraph function defaults to num_iters_within_graph=10. To maintain consistency with the previous benchmarking methodology and ensure better amortization of kernel launch overhead, it's recommended to explicitly set num_iters_within_graph=100.

Suggested change
pdl_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias, pdl=True),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
)
pdl_times = bench_gpu_time_with_cudagraph(
lambda: tgv_gemm_sm100(A, B, bias, pdl=True),
dry_run_time_ms=100,
repeat_time_ms=500,
cold_l2_cache=False,
num_iters_within_graph=100,
)

pdl_avg_time = sum(pdl_times) / len(pdl_times) / 1000
print(
f"PDL average time: {pdl_avg_time * 1000:.6f} ms, {flops / pdl_avg_time:.3f} TFLOPS, speedup: {cublas_avg_time / pdl_avg_time:.2f}x"
)
Expand Down