[Refactor] Reducing code duplication across FP8 CUDA quantization kernels (#4163)

2025-03-06 22:58:52 -08:00
parent c7f254468f
commit 95085d65e9
5 changed files with 32 additions and 64 deletions
--- a/sgl-kernel/benchmark/bench_per_token_quant_fp8.py
+++ b/sgl-kernel/benchmark/bench_per_token_quant_fp8.py
@@ -40,9 +40,6 @@ def calculate_diff(batch_size: int, seq_len: int):
    scale_diff = torch.abs(vllm_scale - sglang_scale).mean().item()
    output_diff = torch.abs(vllm_out.float() - sglang_out.float()).mean().item()

-    print(f"Scale difference: {scale_diff}")
-    print(f"Output difference: {output_diff}")
-
    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):