use fa3 in sgl-kernel (#4954)

python/pyproject.toml

@@ -47,7 +47,7 @@ runtime_common = [
 
 srt = [
     "sglang[runtime_common]",
-    "sgl-kernel==0.0.5.post4",
+    "sgl-kernel==0.0.6",
     "flashinfer_python==0.2.3",
     "torch==2.5.1",
     "cuda-python",

python/sglang/srt/layers/attention/flashattention_backend.py

@@ -22,7 +22,7 @@ if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
     from sglang.srt.model_executor.model_runner import ModelRunner
 
-from flash_attn_interface import flash_attn_with_kvcache
+from sgl_kernel.flash_attn import flash_attn_with_kvcache
 
 
 @dataclass

scripts/ci_install_dependency.sh

@@ -15,7 +15,7 @@ pip install -e "python[all]" --find-links https://flashinfer.ai/whl/cu124/torch2
 rm -rf /root/.cache/flashinfer
 # Force reinstall flashinfer and torch_memory_saver
 pip install flashinfer_python==0.2.3 --find-links ${FLASHINFER_REPO} --force-reinstall --no-deps
-pip install sgl-kernel==0.0.5.post4 --force-reinstall
+pip install sgl-kernel==0.0.6 --force-reinstall
 
 pip install torch_memory_saver
 pip install transformers==4.50.0 sentence_transformers accelerate==1.4.0 peft pandas datasets timm torchaudio