[Refactor] Clean up JIT kernel utilites (#16884)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>

python/sglang/jit_kernel/benchmark/bench_per_tensor_quant_fp8.py

@@ -1,4 +1,3 @@
-import itertools
 import os
 from typing import Optional, Tuple
 
@@ -57,7 +56,7 @@ def sglang_scaled_fp8_quant(
 
 def calculate_diff(batch_size: int, seq_len: int):
     device = torch.device("cuda")
-    x = torch.rand((batch_size, seq_len), dtype=torch.float16, device=device)
+    x = torch.rand((batch_size, seq_len), dtype=torch.bfloat16, device=device)
 
     if not VLLM_AVAILABLE:
         print("vLLM not available, skipping comparison")
@@ -66,25 +65,17 @@ def calculate_diff(batch_size: int, seq_len: int):
     vllm_out, vllm_scale = vllm_scaled_fp8_quant(x)
     sglang_out, sglang_scale = sglang_scaled_fp8_quant(x)
 
-    scale_diff = torch.abs(vllm_scale - sglang_scale).item()
-    output_diff = torch.abs(vllm_out.float() - sglang_out.float()).mean().item()
+    vllm_out = vllm_out.to(torch.float32)
+    sglang_out = sglang_out.to(torch.float32)
 
-    if torch.allclose(
-        vllm_out.to(torch.float32), sglang_out.to(torch.float32), rtol=1e-3, atol=1e-5
-    ) and torch.allclose(vllm_scale, sglang_scale, rtol=1e-3, atol=1e-5):
-        print("All implementations match")
-    else:
-        print("Implementations differ")
+    triton.testing.assert_close(vllm_out, sglang_out, rtol=1e-3, atol=1e-3)
+    triton.testing.assert_close(vllm_scale, sglang_scale, rtol=1e-3, atol=1e-3)
 
 
 if IS_CI:
-    batch_size_range = [16]
-    seq_len_range = [64]
+    element_range = [16384]
 else:
-    batch_size_range = [16, 32, 64, 128]
-    seq_len_range = [64, 128, 256, 512, 1024, 2048]
-
-configs = list(itertools.product(batch_size_range, seq_len_range))
+    element_range = [2**n for n in range(10, 20)]
 
 
 if VLLM_AVAILABLE:
@@ -99,8 +90,8 @@ else:
 
 @triton.testing.perf_report(
     triton.testing.Benchmark(
-        x_names=["batch_size", "seq_len"],
-        x_vals=configs,
+        x_names=["element_count"],
+        x_vals=element_range,
         line_arg="provider",
         line_vals=line_vals,
         line_names=line_names,
@@ -110,11 +101,11 @@ else:
         args={},
     )
 )
-def benchmark(batch_size, seq_len, provider):
+def benchmark(element_count, provider):
     dtype = torch.float16
     device = torch.device("cuda")
 
-    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
+    x = torch.randn(element_count, 4096, device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
 

python/sglang/jit_kernel/benchmark/bench_rmsnorm.py

@@ -0,0 +1,96 @@
+import itertools
+import os
+
+import torch
+import triton
+import triton.testing
+from flashinfer import rmsnorm as fi_rmsnorm
+from sgl_kernel import rmsnorm
+
+from sglang.jit_kernel.norm import rmsnorm as jit_rmsnorm
+
+IS_CI = (
+    os.getenv("CI", "false").lower() == "true"
+    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
+)
+
+
+def sglang_aot_rmsnorm(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+) -> None:
+    rmsnorm(input, weight, out=input)
+
+
+def sglang_jit_rmsnorm(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+) -> None:
+    jit_rmsnorm(input, weight, output=input)
+
+
+def flashinfer_rmsnorm(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+) -> None:
+    fi_rmsnorm(input, weight, out=input)
+
+
+@torch.compile()
+def torch_impl_rmsnorm(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    eps: float = 1e-6,
+) -> None:
+    mean = input.float().pow(2).mean(dim=-1, keepdim=True)
+    norm = (mean + eps).rsqrt()
+    input.copy_(input.float() * norm * weight.float())
+
+
+DTYPE = torch.bfloat16
+DEVICE = "cuda"
+
+if IS_CI:
+    BS_LIST = [16]
+    HIDDEN_SIZE_LIST = [512, 2048]
+else:
+    BS_LIST = [2**n for n in range(0, 14)]
+    HIDDEN_SIZE_LIST = [1536, 3072, 4096, 5120, 8192]
+
+LINE_VALS = ["aot", "jit", "fi", "torch"]
+LINE_NAMES = ["SGL AOT Kernel", "SGL JIT Kernel", "FlashInfer", "PyTorch"]
+STYLES = [("orange", "-"), ("blue", "--"), ("green", "-."), ("red", ":")]
+
+configs = list(itertools.product(HIDDEN_SIZE_LIST, BS_LIST))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["hidden_size", "batch_size"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=LINE_VALS,
+        line_names=LINE_NAMES,
+        styles=STYLES,
+        ylabel="us",
+        plot_name="rmsnorm-performance",
+        args={},
+    )
+)
+def benchmark(hidden_size: int, batch_size: int, provider: str):
+    input = torch.randn((batch_size, hidden_size), dtype=DTYPE, device=DEVICE)
+    weight = torch.randn(hidden_size, dtype=DTYPE, device=DEVICE)
+    FN_MAP = {
+        "aot": sglang_aot_rmsnorm,
+        "jit": sglang_jit_rmsnorm,
+        "fi": flashinfer_rmsnorm,
+        "torch": torch_impl_rmsnorm,
+    }
+    fn = lambda: FN_MAP[provider](input.clone(), weight)
+    quantiles = [0.5, 0.2, 0.8]
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)  # type: ignore
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    benchmark.run(print_data=True)