Skip to content

Revert "Add flashinfer rmsnorm + quant fusion support SM90, SM100, SM120" #33455

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
Fridge003 merged 1 commit into from
Aug 4, 2026
Merged

Revert "Add flashinfer rmsnorm + quant fusion support SM90, SM100, SM120" #33455

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
185 changes: 0 additions & 185 deletions benchmark/kernels/bench_fused_rmsnorm_fp8_quant.py
View file
Open in desktop

This file was deleted.

30 changes: 8 additions & 22 deletions python/sglang/kernels/aot/benchmark/bench_fp8_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,9 @@
import triton
from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm

from sglang.kernels.ops.quantization.per_tensor_quant_fp8 import per_tensor_quant_fp8
from sglang.kernels.ops.quantization.per_tensor_quant_fp8 import (
per_tensor_quant_fp8,
)
from sglang.utils import is_in_ci

# Optional vLLM import
Expand Down Expand Up @@ -104,7 +106,7 @@ def sglang_scaled_fp8_quant(
if IS_CI:
batch_sizes = [1] # Single batch size for CI
else:
batch_sizes = [1, 2, 8, 16, 64, 128, 256, 512, 1024, 2048]
batch_sizes = [1, 16, 64, 128, 256, 512, 1024, 2048]

# Filter line_vals based on vLLM availability
if VLLM_AVAILABLE:
Expand All @@ -113,39 +115,24 @@ def sglang_scaled_fp8_quant(
"vllm-fp8-bf16",
"sglang-fp8-fp16",
"sglang-fp8-bf16",
"sglang-scalar-a-fp8-fp16",
"sglang-scalar-a-fp8-bf16",
]
line_names = [
"vllm-fp8-fp16",
"vllm-fp8-bf16",
"sglang-fp8-fp16",
"sglang-fp8-bf16",
"sglang-scalar-a-fp8-fp16",
"sglang-scalar-a-fp8-bf16",
]
styles = [
("green", "-"),
("green", "--"),
("blue", "-"),
("blue", "--"),
("red", "-"),
("red", "--"),
]
styles = [("green", "-"), ("green", "--"), ("blue", "-"), ("blue", "--")]
else:
line_vals = [
"sglang-fp8-fp16",
"sglang-fp8-bf16",
"sglang-scalar-a-fp8-fp16",
"sglang-scalar-a-fp8-bf16",
]
line_names = [
"sglang-fp8-fp16",
"sglang-fp8-bf16",
"sglang-scalar-a-fp8-fp16",
"sglang-scalar-a-fp8-bf16",
]
styles = [("blue", "-"), ("blue", "--"), ("red", "-"), ("red", "--")]
styles = [("blue", "-"), ("blue", "--")]


@triton.testing.perf_report(
Expand Down Expand Up @@ -187,9 +174,8 @@ def benchmark(batch_size, provider, N, K):
lambda: vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype),
quantiles=quantiles,
)
elif "sglang" in provider:
a_scale = scale_a_scalar if "scalar-a" in provider else scale_a
a_fp8, scale_a_fp8 = sglang_scaled_fp8_quant(a, a_scale)
elif "sglang-fp8" in provider:
a_fp8, scale_a_fp8 = sglang_scaled_fp8_quant(a, scale_a)
b_fp8, scale_b_fp8 = sglang_scaled_fp8_quant(b, scale_b)
b_fp8 = b_fp8.t()
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
Expand Down
Loading
Loading