From c9cb32e3106118ba42edd885e775916fafbf4aff Mon Sep 17 00:00:00 2001 From: kangwangamd <100359556+kangwangamd@users.noreply.github.com> Date: Tue, 29 Sep 2026 17:20:42 +0800 Subject: [PATCH] [AMD] fix kda decode flydsl import (#40546) Co-authored-by: Bingxu Chen Co-authored-by: Michael <13900043+michaelzhang-ai@users.noreply.github.com> --- .../kda_flydsl/kernels/kimi_k3_kda_decode.py | 2 +- .../kda_flydsl/kernels/kimi_k3_kda_decode_fb.py | 2 +- .../attention/kda_flydsl/test_kimi_k3_kda_decode.py | 11 ++++------- 3 files changed, 6 insertions(+), 9 deletions(-) diff --git a/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode.py b/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode.py index 81c6165e05..b904aee61d 100644 --- a/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode.py +++ b/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode.py @@ -8,11 +8,11 @@ import math import flydsl.compiler as flyc import flydsl.expr as fx -from aiter.ops.flydsl.kernels import vector from aiter.ops.flydsl.kernels.tensor_shim import GTensor, _to_raw from flydsl._mlir import ir from flydsl._mlir.dialects import gpu as mlir_gpu from flydsl._mlir.dialects import scf +from flydsl._mlir.dialects import vector from flydsl._mlir.dialects import vector as mlir_vector from flydsl.expr import range_constexpr from flydsl.expr.typing import T diff --git a/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode_fb.py b/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode_fb.py index 3e756d35d8..4928dcfeb9 100644 --- a/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode_fb.py +++ b/python/sglang/kernels/ops/attention/kda_flydsl/kernels/kimi_k3_kda_decode_fb.py @@ -8,7 +8,6 @@ import math import flydsl.compiler as flyc import flydsl.expr as fx -from aiter.ops.flydsl.kernels import vector from aiter.ops.flydsl.kernels.tensor_shim import ( AITER_FLYDSL_KERNARG_PRELOAD, AITER_FLYDSL_KERNARG_PRELOAD_COUNT, @@ -18,6 +17,7 @@ from aiter.ops.flydsl.kernels.tensor_shim import ( from flydsl._mlir import ir from flydsl._mlir.dialects import gpu as mlir_gpu from flydsl._mlir.dialects import llvm, scf +from flydsl._mlir.dialects import vector from flydsl._mlir.dialects import vector as mlir_vector from flydsl.expr import arith, const_expr, range_constexpr from flydsl.expr.arith import ArithValue diff --git a/test/registered/kernels/ops/attention/kda_flydsl/test_kimi_k3_kda_decode.py b/test/registered/kernels/ops/attention/kda_flydsl/test_kimi_k3_kda_decode.py index dd33997eaa..6ef7948d58 100644 --- a/test/registered/kernels/ops/attention/kda_flydsl/test_kimi_k3_kda_decode.py +++ b/test/registered/kernels/ops/attention/kda_flydsl/test_kimi_k3_kda_decode.py @@ -5,6 +5,7 @@ from __future__ import annotations +import importlib.util from dataclasses import dataclass import pytest @@ -12,8 +13,7 @@ import torch import torch.nn.functional as F pytest.importorskip("flydsl") -from aiter.jit.utils.chip_info import get_gfx -from aiter.ops.flydsl.utils import is_flydsl_available +pytest.importorskip("aiter") from sglang.kernels.ops.attention.kda_flydsl.kimi_k3_kda_decode import ( _fb_build_options, @@ -27,12 +27,9 @@ register_amd_ci(est_time=120, suite="stage-b-test-1-gpu-small-amd-mi35x") def _gfx950_flydsl_available() -> bool: - if not torch.cuda.is_available() or not is_flydsl_available(): - return False - try: - return get_gfx() == "gfx950" - except (AssertionError, KeyError, RuntimeError): + if importlib.util.find_spec("flydsl") is None: return False + return is_flydsl_kimi_k3_kda_decode_supported() pytestmark = pytest.mark.skipif(