minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix invalid KVFP4QuantizeUtil references (#28013)

Browse Source 
Signed-off-by: Zach Zhu <zzqshu@126.com>
This commit is contained in:
Zach Zhu
2026-06-12 13:44:11 -07:00
committed by GitHub
parent 3be5a7ec89
commit 627ed3476b
2 changed files with 31 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/quant/test_kvfp4_quant_dequant.py
+5 -5
View File
@@ -6,7 +6,7 @@ import numpy as np
import pytest
import torch
from sglang.srt.layers.quantization.kvfp4_tensor import KVFP4QuantizeUtil
from sglang.srt.layers.quantization.kvfp4_tensor import BlockFP4KVQuantizeUtil
def calculate_accuracy_metrics(
@@ -52,18 +52,18 @@ def run_benchmark(m, n, k, num_runs=100) -> dict[str, dict[str, float]]:
fp8_metrics = calculate_accuracy_metrics(tensor_bf16, tensor_fp8_dequant)
# --- KVFP4 ---
tensor_fp4, scale_factors = KVFP4QuantizeUtil.batched_quantize(tensor_bf16)
_ = KVFP4QuantizeUtil.batched_dequantize(tensor_fp4, scale_factors)
tensor_fp4, scale_factors = BlockFP4KVQuantizeUtil.batched_quantize(tensor_bf16)
_ = BlockFP4KVQuantizeUtil.batched_dequantize(tensor_fp4, scale_factors)
start = time.time()
for _ in range(num_runs):
tensor_fp4, scale_factors = KVFP4QuantizeUtil.batched_quantize(tensor_bf16)
tensor_fp4, scale_factors = BlockFP4KVQuantizeUtil.batched_quantize(tensor_bf16)
torch.cuda.synchronize()
fp4_quant_time = (time.time() - start) / num_runs
start = time.time()
for _ in range(num_runs):
tensor_fp4_dequant = KVFP4QuantizeUtil.batched_dequantize(
tensor_fp4_dequant = BlockFP4KVQuantizeUtil.batched_dequantize(
tensor_fp4, scale_factors
)
torch.cuda.synchronize()