[JIT sgl-kernel] Jit support per tensor quant (#15709)

This commit is contained in:
Xiaoyu Zhang
2025-12-25 16:24:37 +08:00
committed by GitHub
parent a89e85e739
commit de2f2880b5
11 changed files with 497 additions and 7 deletions
@@ -90,14 +90,24 @@ else:
configs = list(itertools.product(batch_size_range, seq_len_range))
if VLLM_AVAILABLE:
line_vals = ["vllm", "sglang"]
line_names = ["VLLM", "SGL Kernel"]
styles = [("blue", "-"), ("green", "-")]
else:
line_vals = ["sglang"]
line_names = ["SGL Kernel"]
styles = [("green", "-")]
@triton.testing.perf_report(
triton.testing.Benchmark(
x_names=["batch_size", "seq_len"],
x_vals=configs,
line_arg="provider",
line_vals=["vllm", "sglang"],
line_names=["VLLM", "SGL Kernel"],
styles=[("blue", "-"), ("green", "-")],
line_vals=line_vals,
line_names=line_names,
styles=styles,
ylabel="us",
plot_name="per-tensor-quant-fp8-performance",
args={},