diff --git a/python/sglang/kernels/ops/gemm/smallm_fp8_blockscale_gfx950/__init__.py b/python/sglang/kernels/ops/gemm/smallm_fp8_blockscale_gfx950/__init__.py index d0ef085cfb..aea5192e56 100644 --- a/python/sglang/kernels/ops/gemm/smallm_fp8_blockscale_gfx950/__init__.py +++ b/python/sglang/kernels/ops/gemm/smallm_fp8_blockscale_gfx950/__init__.py @@ -7,7 +7,7 @@ import tempfile import torch -from sglang.kernels.ops.moe.smallm_moe_gfx950 import _check, _hip_lib, _hipcc, _Kernel +from sglang.kernels.jit.utils.hipcc import HipKernel, check_hip, hip_lib, hipcc_path _SRC = os.path.join(os.path.dirname(__file__), "smallm_fp8_blockscale_gemm.hip") # (N, K) -> (max M, N tiles/block, K steps/wave, waves), first match wins. @@ -49,24 +49,26 @@ def try_smallm_fp8_bs_linear(x, WQ, w_scale): return None if _mod is None: co = os.path.join(tempfile.mkdtemp(), "k.co") - cmd = [_hipcc(), "--genco", "--offload-arch=gfx950", "-O3", "-o", co, _SRC] + cmd = [hipcc_path(), "--genco", "--offload-arch=gfx950", "-O3", "-o", co, _SRC] try: subprocess.run(cmd, check=True, capture_output=True) _mod = ctypes.c_void_p() - _check(_hip_lib().hipModuleLoad(ctypes.byref(_mod), co.encode()), "load") + check_hip(hip_lib().hipModuleLoad(ctypes.byref(_mod), co.encode()), "load") except (OSError, subprocess.CalledProcessError, RuntimeError) as e: _mod = False print(f"[smallm_fp8_bs] disabled: {e}", flush=True) return None key = ((M + 15) // 16, *cfg) - hip, fn = _hip_lib(), _fns.get(key) + hip, fn = hip_lib(), _fns.get(key) if fn is None: - fn = _fns[key] = _Kernel(_mod, "smallm_fp8_bs_m{}_n{}_s{}_w{}".format(*key)).fn + fn = _fns[key] = HipKernel( + _mod, "smallm_fp8_bs_m{}_n{}_s{}_w{}".format(*key) + ).fn out = torch.empty(M, N, dtype=torch.bfloat16, device=x.device) args = Args(*(t.data_ptr() for t in (x, WQ, w_scale, out)), M, N, x.stride(0)) params = (ctypes.c_void_p * 1)(ctypes.addressof(args)) stream = torch.cuda.current_stream().cuda_stream grid, block = N // (16 * key[1]), 64 * key[3] rc = hip.hipModuleLaunchKernel(fn, grid, 1, 1, block, 1, 1, 0, stream, params, None) - _check(rc, "launch") + check_hip(rc, "launch") return out