rwkv-ops 0.6.1__py3-none-any.whl

rwkv_ops/rwkv6_kernel/jax_rwkv_kernel.py

@@ -0,0 +1,722 @@
+import os
+import pybind11
+import importlib
+import sysconfig
+import subprocess
+from functools import partial, reduce
+
+import jax
+import jax.numpy as jnp
+from jax import core, dtypes
+from jax.core import ShapedArray
+from jax.interpreters import mlir, xla
+from jax.interpreters.mlir import ir
+from jax.lib import xla_client
+from jaxlib.hlo_helpers import custom_call
+
+
+use_rocm = os.environ.get("RWKV_USE_ROCM", "0") == "1"
+if use_rocm:
+    kernel_dir_name = "jax_kernel_hip"
+    cuda_lib_dir = "/opt/rocm/include"
+    kernel_name = "gpu_ops"
+    platform = "ROCM"
+else:
+    kernel_dir_name = "jax_kernel_cuda"
+    cuda_lib_dir = "/usr/local/cuda/include"
+    kernel_name = "gpu_ops"
+    platform = "gpu"
+
+
+def default_layouts(*shapes):
+    return [range(len(shape) - 1, -1, -1) for shape in shapes]
+
+
+class RWKVKernelOperator:
+    def __init__(self, head_size, max_sequence_length):
+        """
+        加载或构建rwkv内核
+        """
+        rwkv_kernel = RWKVKernelOperator._load_or_build_kernel(
+            head_size, max_sequence_length
+        )
+        """
+        向mlir注册C++算子入口
+        """
+        for _name, _value in rwkv_kernel.get_rwkv_registrations().items():
+            xla_client.register_custom_call_target(_name, _value, platform=platform)
+
+        """
+        构建jax前端数据类型到C++后端的映射
+        """
+
+        def element_type_to_descriptor_type_mapping(element_type):
+            _element_type_to_descriptor_type_mapping = {
+                ir.BF16Type.get(): rwkv_kernel.ElementType.BF16,
+                ir.F16Type.get(): rwkv_kernel.ElementType.F16,
+                ir.F32Type.get(): rwkv_kernel.ElementType.F32,
+            }
+            return _element_type_to_descriptor_type_mapping.get(element_type)
+
+        """
+        定义前向过程算子
+        """
+
+        def build_forward_op():
+            _rwkv_fwd_p = core.Primitive("rwkv_fwd")
+            _rwkv_fwd_p.multiple_results = False
+            _rwkv_fwd_p.def_impl(partial(xla.apply_primitive, _rwkv_fwd_p))
+
+            """
+            构建前向过程方法
+            """
+
+            def _rwkv_fwd(r, k, v, w, u):
+                y = _rwkv_fwd_p.bind(r, k, v, w, u)
+                ctx = r, k, v, w, u
+                return y, ctx
+
+            """
+            milr调用上面注册的前向过程算子，milr帮助算子创建输出张量
+            """
+
+            def _rwkv_fwd_cuda_lowering(ctx, r, k, v, w, u):
+                r_type = ir.RankedTensorType(r.type)
+                k_type = ir.RankedTensorType(k.type)
+                v_type = ir.RankedTensorType(v.type)
+                w_type = ir.RankedTensorType(w.type)
+                u_type = ir.RankedTensorType(u.type)
+                assert all(
+                    [
+                        r_type.element_type == xx.element_type
+                        for xx in [k_type, v_type, w_type, u_type]
+                    ]
+                )
+                assert all(
+                    [r_type.shape == xx.shape for xx in [k_type, v_type, w_type]]
+                )
+                assert r_type.element_type in [
+                    ir.F32Type.get(),
+                    ir.BF16Type.get(),
+                    ir.F16Type.get(),
+                ]
+                bz, seq_len, hd_sz = r_type.shape
+
+                assert hd_sz % head_size == 0
+                assert reduce(lambda x, y: x * y, u_type.shape, 1) == hd_sz, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+                input_type = r_type.element_type
+
+                if input_type in [ir.F32Type.get(), ir.BF16Type.get()]:
+                    output_type = input_type
+                else:
+                    output_type = ir.F32Type.get()
+
+                opaque = rwkv_kernel.create_rwkv_descriptor(
+                    bz,
+                    seq_len,
+                    hd_sz,
+                    hd_sz // head_size,
+                    False,  # unused
+                    element_type_to_descriptor_type_mapping(input_type),
+                    element_type_to_descriptor_type_mapping(output_type),
+                )
+
+                out = custom_call(
+                    b"wkv_forward",
+                    result_types=[
+                        ir.RankedTensorType.get(r_type.shape, output_type),
+                    ],
+                    operands=[r, k, v, w, u],
+                    backend_config=opaque,
+                    operand_layouts=default_layouts(
+                        r_type.shape,
+                        k_type.shape,
+                        v_type.shape,
+                        w_type.shape,
+                        u_type.shape,
+                    ),
+                    result_layouts=default_layouts(r_type.shape),
+                ).results
+                return out
+
+            """
+            将算子绑定到C++
+            """
+            mlir.register_lowering(
+                _rwkv_fwd_p,
+                _rwkv_fwd_cuda_lowering,
+                platform="gpu",
+            )
+            """
+            定义抽象过程，告知jax输出张量形状与数据类型
+            """
+
+            def _rwkv_fwd_abstract(r, k, v, w, u):
+                assert all([r.shape == xx.shape for xx in [k, v, w]])
+                assert len(r.shape) == 3
+                bz, seq_len, channels = r.shape
+                assert channels % head_size == 0
+                assert seq_len <= max_sequence_length
+                assert reduce(lambda x, y: x * y, u.shape, 1) == channels, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+
+                r_dtype = dtypes.canonicalize_dtype(r.dtype)
+                k_dtype = dtypes.canonicalize_dtype(k.dtype)
+                v_dtype = dtypes.canonicalize_dtype(v.dtype)
+                w_dtype = dtypes.canonicalize_dtype(w.dtype)
+                u_dtype = dtypes.canonicalize_dtype(u.dtype)
+
+                assert all(
+                    [r_dtype == xx for xx in [k_dtype, v_dtype, w_dtype, u_dtype]]
+                )
+                assert r_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
+                if r_dtype in [jnp.float32, jnp.bfloat16]:
+                    output_dtype = r_dtype
+                else:
+                    output_dtype = jnp.float32
+                return ShapedArray(
+                    r.shape, output_dtype, named_shape=r.named_shape
+                )  # output
+
+            _rwkv_fwd_p.def_abstract_eval(_rwkv_fwd_abstract)
+            return _rwkv_fwd, _rwkv_fwd_p
+
+        rwkv_fwd, _ = build_forward_op()
+
+        def build_backward_op():
+            """
+            注册反向传播过程算子
+            """
+            _rwkv_bwd_p = core.Primitive("rwkv_bwd")
+            _rwkv_bwd_p.multiple_results = True
+            _rwkv_bwd_p.def_impl(partial(xla.apply_primitive, _rwkv_bwd_p))
+
+            """
+                反向传播过程
+            """
+
+            def _rwkv_bwd(ctx, gy):
+                r, k, v, w, u = ctx
+                gr, gk, gv, gw, gu = _rwkv_bwd_p.bind(r, k, v, w, u, gy)
+                gu = jnp.sum(gu, axis=0)
+
+                return gr, gk, gv, gw, gu
+
+            """
+                反向传播mlir后端
+            """
+
+            def _rwkv_bwd_cuda_lowering(ctx, r, k, v, w, u, gy):
+                r_type = ir.RankedTensorType(r.type)
+                k_type = ir.RankedTensorType(k.type)
+                v_type = ir.RankedTensorType(v.type)
+                w_type = ir.RankedTensorType(w.type)
+                u_type = ir.RankedTensorType(u.type)
+                gy_type = ir.RankedTensorType(gy.type)
+
+                assert all(
+                    [
+                        r_type.element_type == xx.element_type
+                        for xx in [k_type, v_type, w_type, u_type]
+                    ]
+                )
+                assert all(
+                    [
+                        r_type.shape == xx.shape
+                        for xx in [k_type, v_type, w_type, gy_type]
+                    ]
+                )
+                assert r_type.element_type in [
+                    ir.F32Type.get(),
+                    ir.BF16Type.get(),
+                    ir.F16Type.get(),
+                ]
+                bz, seq_len, hd_sz = r_type.shape
+
+                assert hd_sz % head_size == 0
+                assert reduce(lambda x, y: x * y, u_type.shape, 1) == hd_sz, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+                input_type = r_type.element_type
+
+                if input_type in [ir.F32Type.get(), ir.BF16Type.get()]:
+                    output_type = input_type
+                else:
+                    output_type = ir.F32Type.get()
+
+                assert output_type == gy_type.element_type
+
+                opaque = rwkv_kernel.create_rwkv_descriptor(
+                    bz,
+                    seq_len,
+                    hd_sz,
+                    hd_sz // head_size,
+                    False,  # unused
+                    element_type_to_descriptor_type_mapping(input_type),
+                    element_type_to_descriptor_type_mapping(output_type),
+                )
+
+                gu_shape = (bz,) + u_type.shape
+
+                out = custom_call(
+                    b"wkv_backward",
+                    result_types=[
+                        ir.RankedTensorType.get(r_type.shape, output_type),  # gr
+                        ir.RankedTensorType.get(k_type.shape, output_type),  # gk
+                        ir.RankedTensorType.get(v_type.shape, output_type),  # gw
+                        ir.RankedTensorType.get(w_type.shape, output_type),  # gw
+                        ir.RankedTensorType.get(gu_shape, output_type),  # gu
+                    ],
+                    operands=[r, k, v, w, u, gy],
+                    backend_config=opaque,
+                    operand_layouts=default_layouts(
+                        r_type.shape,
+                        k_type.shape,
+                        v_type.shape,
+                        w_type.shape,
+                        u_type.shape,
+                        gy_type.shape,
+                    ),
+                    result_layouts=default_layouts(
+                        r_type.shape, k_type.shape, v_type.shape, w_type.shape, gu_shape
+                    ),
+                ).results
+                return out
+
+            """
+            注册反向传播算子到mlir
+            """
+            mlir.register_lowering(
+                _rwkv_bwd_p,
+                _rwkv_bwd_cuda_lowering,
+                platform="gpu",
+            )
+            """
+            反向传播的抽象过程
+            """
+
+            def _rwkv_bwd_abstract(r, k, v, w, u, gy):
+                assert all([r.shape == xx.shape for xx in [k, v, w]])
+                assert len(r.shape) == 3
+                bz, seq_len, channels = r.shape
+                assert channels % head_size == 0
+                assert seq_len <= max_sequence_length
+                assert reduce(lambda x, y: x * y, u.shape, 1) == channels, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+
+                r_dtype = dtypes.canonicalize_dtype(r.dtype)
+                k_dtype = dtypes.canonicalize_dtype(k.dtype)
+                v_dtype = dtypes.canonicalize_dtype(v.dtype)
+                w_dtype = dtypes.canonicalize_dtype(w.dtype)
+                u_dtype = dtypes.canonicalize_dtype(u.dtype)
+                gy_dtype = dtypes.canonicalize_dtype(gy.dtype)
+
+                u_shape = (bz,) + u_shape
+                assert all(
+                    [r_dtype == xx for xx in [k_dtype, v_dtype, w_dtype, u_dtype]]
+                )
+                assert r_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
+                if r_dtype in [jnp.float32, jnp.bfloat16]:
+                    output_dtype = r_dtype
+                else:
+                    output_dtype = jnp.float32
+                assert output_dtype == gy_dtype
+
+                outputs = [
+                    ShapedArray(r.shape, output_dtype, named_shape=r.named_shape),
+                    ShapedArray(k.shape, output_dtype, named_shape=k.named_shape),
+                    ShapedArray(v.shape, output_dtype, named_shape=v.named_shape),
+                    ShapedArray(w.shape, output_dtype, named_shape=w.named_shape),
+                    ShapedArray(u_shape, output_dtype, named_shape=u.named_shape),
+                ]  # output
+                return outputs
+
+            _rwkv_bwd_p.def_abstract_eval(_rwkv_bwd_abstract)
+            return _rwkv_bwd, _rwkv_bwd_p
+
+        rwkv_bwd, _ = build_backward_op()
+        """
+        组合算子
+        """
+
+        @jax.custom_vjp
+        def rwkv_normal_operator(r, k, v, w, u):
+            output, _ = rwkv_fwd(r, k, v, w, u)
+            return output
+
+        rwkv_normal_operator.defvjp(rwkv_fwd, rwkv_bwd)
+        self.rwkv_normal_operator = rwkv_normal_operator
+
+        def build_forward_with_state_op():
+            _rwkv_fwd_state_p = core.Primitive("rwkv_fwd_with_state")
+            _rwkv_fwd_state_p.multiple_results = True
+            _rwkv_fwd_state_p.def_impl(partial(xla.apply_primitive, _rwkv_fwd_state_p))
+
+            """
+            构建带有状态前向过程方法
+            """
+
+            def _rwkv_fwd_with_state(r, k, v, w, u, init_state=None, state_map=None):
+                bz = r.shape[0]
+                if init_state is not None:  # shape=(B,H,D,D)
+                    assert len(init_state.shape) in [
+                        3,
+                        4,
+                    ], "init_state的shape为(Batch_size,num_heads,head_size,head_size)"
+                    if len(init_state.shape) == 3:
+                        state_map = jnp.zeros((bz,), dtype=jnp.int32)
+
+                    if state_map is None:
+                        n_state = jnp.shape(init_state)[0]
+                        B = jnp.shape(r)[0]
+                        # print('ns:',n_state,'B:',B,r.shape,k.shape,v.shape)
+                        assert n_state == 1 or n_state == B, (
+                            "我无法为您推断state_map的形状，请手动指定。"
+                        )
+                        if n_state == 1:
+                            state_map = jnp.array([0] * B, dtype=jnp.int32)
+                        elif n_state == B:
+                            state_map = jnp.array(
+                                [i for i in range(B)], dtype=jnp.int32
+                            )
+                        else:
+                            assert False, "未实现"
+                    else:
+                        # assert state_map is not None,"请传入一个state_map,这是一个int32类型的shape为(bz,)的数组，存放的是int_state到每一维度上的映射关系"
+                        assert state_map.dtype in [
+                            jnp.int64,
+                            jnp.int32,
+                        ], "state_map的数值类型必须为int32"
+                        state_map = jnp.astype(state_map, jnp.int32)
+                        assert jnp.all(state_map >= 0) and jnp.add(state_map < bz), (
+                            f"state_map内为state的映射下标，因此范围为: [0,{bz})"
+                        )
+                assert (init_state is None) == (state_map is None), (
+                    "init_state与state_map必须同时传入"
+                )
+
+                if init_state is None:
+                    y, s = _rwkv_fwd_state_p.bind(r, k, v, w, u)
+                else:
+                    # state_map = jnp.astype(state_map,jnp.int32)
+                    # print('sd',state_map.dtype)
+                    y, s = _rwkv_fwd_state_p.bind(r, k, v, w, u, init_state, state_map)
+                return y, s
+
+            """
+            milr调用上面注册的前向过程算子，milr帮助算子创建输出张量
+            """
+
+            def _rwkv_fwd_with_state_cuda_lowering(
+                ctx, r, k, v, w, u, s=None, s_map=None
+            ):
+                r_type = ir.RankedTensorType(r.type)
+                k_type = ir.RankedTensorType(k.type)
+                v_type = ir.RankedTensorType(v.type)
+                w_type = ir.RankedTensorType(w.type)
+                u_type = ir.RankedTensorType(u.type)
+
+                with_init_state = s is not None
+
+                assert all(
+                    [
+                        r_type.element_type == xx.element_type
+                        for xx in [k_type, v_type, w_type, u_type]
+                    ]
+                )
+                assert all(
+                    [r_type.shape == xx.shape for xx in [k_type, v_type, w_type]]
+                )
+                assert r_type.element_type in [
+                    ir.F32Type.get(),
+                    ir.BF16Type.get(),
+                    ir.F16Type.get(),
+                ]
+                bz, seq_len, hd_sz = r_type.shape
+
+                assert hd_sz % head_size == 0
+                num_heads = hd_sz // head_size
+                assert reduce(lambda x, y: x * y, u_type.shape, 1) == hd_sz, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+                input_type = r_type.element_type
+
+                if input_type in [ir.F32Type.get(), ir.BF16Type.get()]:
+                    output_type = input_type
+                else:
+                    output_type = ir.F32Type.get()
+                state_shape = (bz, num_heads, head_size, head_size)
+
+                if with_init_state:
+                    assert s_map is not None, (
+                        "您必须同时传入init_state与state_map 或者都赋值为None."
+                    )
+
+                    s_type = ir.RankedTensorType(s.type)
+                    sm_type = ir.RankedTensorType(s_map.type)
+                    # print(sm_type, ir.IntegerType.get_signless(64))
+                    assert sm_type.element_type == ir.IntegerType.get_signless(32), (
+                        "state_map的数据类型必须为int32"
+                    )
+                    # print(sm_type.shape,bz)
+                    assert tuple(sm_type.shape) == (bz,), (
+                        "state_map的shape 形状必须为(batch_size,)"
+                    )
+
+                    assert s_type.element_type == output_type
+                    assert tuple(s_type.shape) == state_shape, (
+                        "the shape of init state must be (batch_size,num_heads,head_size,head_size)"
+                    )
+                    # assert s_type.shape[0] == bz and reduce(lambda x,y: x * y, s_type.shape[1:],1) == head_size * hd_sz,"the shape of init state must be (batch_size,num_heads,head_size,head_size)"
+
+                opaque = rwkv_kernel.create_rwkv_descriptor(
+                    bz,
+                    seq_len,
+                    hd_sz,
+                    hd_sz // head_size,
+                    with_init_state,
+                    element_type_to_descriptor_type_mapping(input_type),
+                    element_type_to_descriptor_type_mapping(output_type),
+                )
+
+                if with_init_state:
+                    operands = [s_map, r, k, v, w, u, s]
+                    operand_layouts = default_layouts(
+                        sm_type.shape,
+                        r_type.shape,
+                        k_type.shape,
+                        v_type.shape,
+                        w_type.shape,
+                        u_type.shape,
+                        s_type.shape,
+                    )
+                else:
+                    operands = [r, k, v, w, u]
+                    operand_layouts = default_layouts(
+                        r_type.shape,
+                        k_type.shape,
+                        v_type.shape,
+                        w_type.shape,
+                        u_type.shape,
+                    )
+
+                out = custom_call(
+                    b"wkv_forward_with_state",
+                    result_types=[
+                        ir.RankedTensorType.get(r_type.shape, output_type),
+                        ir.RankedTensorType.get(state_shape, output_type),
+                    ],
+                    operands=operands,
+                    backend_config=opaque,
+                    operand_layouts=operand_layouts,
+                    result_layouts=default_layouts(r_type.shape, state_shape),
+                ).results
+                return out
+
+            """
+            将算子绑定到C++
+            """
+            mlir.register_lowering(
+                _rwkv_fwd_state_p,
+                _rwkv_fwd_with_state_cuda_lowering,
+                platform="gpu",
+            )
+            """
+            定义抽象过程，告知jax输出张量形状与数据类型
+            """
+
+            def _rwkv_fwd_with_state_abstract(r, k, v, w, u, s=None, s_map=None):
+                assert all([r.shape == xx.shape for xx in [k, v, w]])
+                assert len(r.shape) == 3
+                bz, seq_len, channels = r.shape
+                assert channels % head_size == 0
+                assert seq_len <= max_sequence_length
+                assert reduce(lambda x, y: x * y, u.shape, 1) == channels, (
+                    "the elements of u (time first) is not equal to hidden_size"
+                )
+                num_heads = channels // head_size
+                r_dtype = dtypes.canonicalize_dtype(r.dtype)
+                k_dtype = dtypes.canonicalize_dtype(k.dtype)
+                v_dtype = dtypes.canonicalize_dtype(v.dtype)
+                w_dtype = dtypes.canonicalize_dtype(w.dtype)
+                u_dtype = dtypes.canonicalize_dtype(u.dtype)
+
+                assert all(
+                    [r_dtype == xx for xx in [k_dtype, v_dtype, w_dtype, u_dtype]]
+                )
+                assert r_dtype in [jnp.float32, jnp.float16, jnp.bfloat16]
+                if r_dtype in [jnp.float32, jnp.bfloat16]:
+                    output_dtype = r_dtype
+                else:
+                    output_dtype = jnp.float32
+
+                state_shape = (bz, num_heads, head_size, head_size)
+                if s is not None:
+                    s_dtype = dtypes.canonicalize_dtype(s.dtype)
+                    assert s_dtype == output_dtype
+                    assert s.shape == state_shape, (
+                        "the shape of init_state must be (batch_size, seq_len, num_heads, head_size, head_size)"
+                    )
+
+                return [
+                    ShapedArray(
+                        r.shape, output_dtype, named_shape=r.named_shape
+                    ),  # output
+                    ShapedArray(state_shape, output_dtype, named_shape=r.named_shape),
+                ]
+
+            _rwkv_fwd_state_p.def_abstract_eval(_rwkv_fwd_with_state_abstract)
+            return _rwkv_fwd_with_state, _rwkv_fwd_state_p
+
+        self.rwkv_fwd_with_state, _ = build_forward_with_state_op()
+
+    def __call__(
+        self, r, k, v, w, u, with_state=False, init_state=None, state_map=None
+    ):
+        if with_state:
+            return self.rwkv_fwd_with_state(r, k, v, w, u, init_state, state_map)
+        else:
+            return self.rwkv_normal_operator(r, k, v, w, u), None
+
+    @staticmethod
+    def _load_or_build_kernel(head_size, max_sequence_length):
+        assert head_size % 4 == 0, f"head size必须是4的倍数，而{head_size}显然不是."
+        assert isinstance(head_size, int), "你是在搞笑吗？ head_size肯定得是int类型的啊"
+        assert isinstance(max_sequence_length, int), (
+            "你是在搞笑吗？ max_sequence_length肯定得是int类型的啊"
+        )
+        assert head_size > 0 and max_sequence_length > 0, (
+            "难绷，head_size与max_sequence_length肯定得是大于0的正整数啊。"
+        )
+        assert os.path.exists(cuda_lib_dir) and len(os.listdir(cuda_lib_dir)) > 0, (
+            f"请检查{cuda_lib_dir}文件夹是否存在，这个文件本质是是您的cuda library的超链接。"
+        )
+        kernel_dir = os.path.abspath(
+            os.path.join(os.path.dirname(__file__), kernel_dir_name)
+        )
+        builds_dir = os.path.join(kernel_dir, "builds")
+        assert os.path.exists(kernel_dir), (
+            f"找不到{kernel_dir_name}文件夹，请问您的文件是完整的吗？"
+        )
+        if not os.path.exists(builds_dir):
+            os.mkdir(builds_dir)
+        target_dir_name = f"_N_{head_size}_T_{max_sequence_length}"
+        target_dir = os.path.join(builds_dir, target_dir_name)
+        if not os.path.exists(target_dir):
+            os.mkdir(target_dir)
+
+        def get_cflags():
+            getvar = sysconfig.get_config_var
+            flags = [
+                "-I" + sysconfig.get_path("include"),
+                "-I" + sysconfig.get_path("platinclude"),
+            ]
+
+            flags.extend(getvar("CFLAGS").split())
+            return " ".join(flags)
+
+        def get_suffix():
+            getvar = sysconfig.get_config_var
+            return getvar("EXT_SUFFIX")
+
+        build_cmds = []
+
+        # first, build cuda kernel
+        if use_rocm:
+            cu_src = os.path.join(kernel_dir, "rwkv_kernels.hip")
+            assert os.path.exists(cu_src)
+            cu_dst = os.path.join(target_dir, "rwkv_kernels.hip.o")
+            kernel_cmd = (
+                "hipcc -O3 --hipstdpar -xhip -fopenmp -ffast-math"
+                + " -munsafe-fp-atomics -enable-vectorize-compares"
+                + f" -I{cuda_lib_dir} -I{pybind11.get_include()}"
+                + " -fPIC -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2"
+                + " --gpu-max-threads-per-block=120"
+                + f" -c {cu_src} -o {cu_dst} -D _N_={head_size} -D _T_={max_sequence_length}"
+            )
+        else:
+            cu_src = os.path.join(kernel_dir, "rwkv_kernels.cu")
+            assert os.path.exists(cu_src)
+            cu_dst = os.path.join(target_dir, "rwkv_kernels.cu.o")
+            kernel_cmd = (
+                "nvcc --threads 4 -Xcompiler -Wall -ldl --expt-relaxed-constexpr -O3 -DNDEBUG -Xcompiler -O3"
+                + " --generate-code=arch=compute_70,code=[compute_70,sm_70] --generate-code=arch=compute_75,code=[compute_75,sm_75] --generate-code=arch=compute_80,code=[compute_80,sm_80] --generate-code=arch=compute_86,code=[compute_86,sm_86]"
+                + f" -Xcompiler=-fPIC -Xcompiler=-fvisibility=hidden -x cu -c {cu_src} -o {cu_dst} -D _N_={head_size} -D _T_={max_sequence_length}"
+            )
+        build_cmds.append(kernel_cmd)
+
+        so_dst = os.path.join(target_dir, f"{kernel_name}{get_suffix()}")
+        if not os.path.exists(so_dst):
+            # second, build C++ code.
+            cpp_src = os.path.join(kernel_dir, f"{kernel_name}.cpp")
+            cpp_dst = os.path.join(builds_dir, f"{kernel_name}.cpp.o")
+            if not os.path.exists(cpp_dst):
+                if use_rocm:
+                    cpp_cmd = (
+                        f"c++ -I{cuda_lib_dir} -I{pybind11.get_include()} {get_cflags()}"
+                        + " -fPIC -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2"
+                        + " -O3 -DNDEBUG -O3 -fPIC -fvisibility=hidden -flto -fno-fat-lto-objects"
+                        + f" -o {cpp_dst} -c {cpp_src}"
+                    )
+                else:
+                    cpp_cmd = (
+                        f"c++ -I{cuda_lib_dir} -I{pybind11.get_include()} {get_cflags()}"
+                        + " -O3 -DNDEBUG -O3 -fPIC -fvisibility=hidden -flto -fno-fat-lto-objects"
+                        + f" -o {cpp_dst} -c {cpp_src}"
+                    )
+                build_cmds.append(cpp_cmd)
+
+            # third assembly C++ and cuda
+            if use_rocm:
+                assembly_cmd = (
+                    f"c++ -fPIC -O3 -DNDEBUG -O3 -flto -shared  -o {so_dst} {cpp_dst} {cu_dst}"
+                    + f" -fPIC -I{cuda_lib_dir} -I{pybind11.get_include()} {get_cflags()}"
+                    + " -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2"
+                    + " -L/opt/rocm/lib  -lamdhip64 -lpthread -ldl"
+                )
+            else:
+                assembly_cmd = (
+                    f"c++ -fPIC -O3 -DNDEBUG -O3 -flto -shared  -o {so_dst} {cpp_dst} {cu_dst}"
+                    + " -L/usr/local/cuda/lib64  -lcudadevrt -lcudart_static -lrt -lpthread -ldl"
+                )
+            build_cmds.append(assembly_cmd)
+
+            # finally strip the so library
+            strip_cmd = f"strip {so_dst}"
+            build_cmds.append(strip_cmd)
+
+            print("-------------------starting build kernel -------------------")
+            for cmd in build_cmds:
+                print("--------------- execute cmd ---------------")
+                print(cmd)
+                p = subprocess.Popen(cmd, shell=True)
+                p.wait()
+            print("-------------------build kernel finished -------------------")
+        kernel_module_path = f"rwkv_ops.rwkv6_kernel.{kernel_dir_name}.builds.{target_dir_name}.{kernel_name}"
+        print("loading cuda kernel from the path:", kernel_module_path)
+        rwkv_op = importlib.import_module(kernel_module_path)
+        return rwkv_op
+
+
+if __name__ == "__main__":
+    bz, seq_len, hd_sz = 1, 16, 8
+    r = jnp.zeros(shape=(bz, seq_len, hd_sz), dtype="float16") + 2
+    k = jnp.zeros(shape=(bz, seq_len, hd_sz), dtype="float16") + 2
+    v = jnp.zeros(shape=(bz, seq_len, hd_sz), dtype="float16") + 2
+    w = jnp.zeros(shape=(bz, seq_len, hd_sz), dtype="float16") - 2
+    u = jnp.zeros(shape=(hd_sz,), dtype="float16") + 2
+    rwkv_op = RWKVKernelOperator(head_size=hd_sz, max_sequence_length=seq_len)
+    out = rwkv_op(r, k, v, w, u)
+
+    print(out.dtype)
+
+    def ref_loss(r, k, v, w, u):
+        predictions = rwkv_op(r, k, v, w, u)
+        return -jnp.mean(predictions**2)
+
+    ref_out = jax.grad(ref_loss, argnums=(0, 1, 2, 3, 4))(r, k, v, w, u)
+    print(ref_out)