import math, pathlib, functools, struct from tinygrad import Device, Tensor from tinygrad.dtype import DTypeLike, dtypes from tinygrad.helpers import DEBUG, getenv from tinygrad.renderer import Estimates from tinygrad.runtime.support.compiler_amd import HIPCCCompiler from tinygrad.runtime.support.elf import elf_loader from tinygrad.uop.ops import UOp, Ops, KernelInfo def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None, dtype:DTypeLike|None=None) -> Tensor: dtype = dtype or ref.dtype if not isinstance(ref.device, tuple): return Tensor.invalids(*shape, dtype=dtype, device=ref.device) shard_axis = ref.uop.axis if axis is None else axis shape = tuple(s // len(ref.device) if i == shard_axis else s for i, s in enumerate(shape)) axis = ref.uop.axis if axis is None else axis return Tensor(Tensor.invalids(*shape, dtype=dtype, device=ref.device).uop.unshard(axis), dtype=dtype, device=ref.device) @functools.cache def custom_fused_qkv_rope_forward(q:UOp, k:UOp, v:UOp, xqkv:UOp, freqs_cis:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int): group_size = H // H_KV q, k, v = q.reshape(B, N, H, D), k.reshape(B, N, H_KV, D), v.reshape(B, N, H_KV, D) xqkv = xqkv.reshape(B, N, H_KV, group_size + 2, D) b, n = UOp.range(B, 0), UOp.range(N, 1) pair = UOp.range(D // 2, 2) even = pair * 2 c = freqs_cis[0, n, 0, pair, 0].cast(dtypes.float) s = freqs_cis[0, n, 0, pair, 1].cast(dtypes.float) ordered:UOp|None = None for kvh in range(H_KV): q_out, k_out, v_out = (x.after(ordered) if ordered is not None else x for x in (q, k, v)) x_in = xqkv.after(ordered) if ordered is not None else xqkv stores:list[UOp] = [] for rep in range(group_size): a = x_in[b, n, kvh, rep, even].cast(dtypes.float) bb = x_in[b, n, kvh, rep, even + 1].cast(dtypes.float) h = kvh * group_size + rep stores += [q_out[b, n, h, even].store((a * c - bb * s).cast(q.dtype)), q_out[b, n, h, even + 1].store((a * s + bb * c).cast(q.dtype))] a = x_in[b, n, kvh, group_size, even].cast(dtypes.float) bb = x_in[b, n, kvh, group_size, even + 1].cast(dtypes.float) stores += [k_out[b, n, kvh, even].store((a * c - bb * s).cast(k.dtype)), k_out[b, n, kvh, even + 1].store((a * s + bb * c).cast(k.dtype)), v_out[b, n, kvh, even].store(x_in[b, n, kvh, group_size + 1, even]), v_out[b, n, kvh, even + 1].store(x_in[b, n, kvh, group_size + 1, even + 1])] ordered = UOp.group(*stores) assert ordered is not None return ordered.end(pair, n, b).sink(arg=KernelInfo(name="fused_qkv_rope_forward")) @functools.cache def custom_fused_qkv_rope_backward(dxqkv:UOp, dq:UOp, dk:UOp, dv:UOp, freqs_cis:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int): assert (B, N, H, H_KV, D) == (2, 8192, 32, 8, 128) code = (pathlib.Path(__file__).parent / "fused_qkv_rope_bwd.cpp").read_text() threads = 256 thread_idx = UOp.special(threads, "lidx0") gsz = (B, N // 64, H + 2 * H_KV) block_idx_x, block_idx_y, block_idx_z = (UOp.special(x, f"gidx{i}") for i, x in enumerate(gsz)) sink = UOp.sink(dxqkv.base, dq.base, dk.base, dv.base, freqs_cis.base, thread_idx, block_idx_x, block_idx_y, block_idx_z, arg=KernelInfo(name="fused_qkv_rope_backward")) compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math", f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DTHREADS_PER_BLOCK={threads}"] lib = HIPCCCompiler(arch, compile_args).compile_cached(code) return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib))) def _fa_native_grads(dq:UOp, dk:UOp, dv:UOp) -> tuple[UOp, UOp, UOp]|None: def unwrap_partial(x:UOp) -> UOp|None: expected = (Ops.CAST, Ops.REDUCE, Ops.PERMUTE, Ops.CAST, Ops.RESHAPE, Ops.AFTER) for op in expected: if x.op is not op: return None if op is not Ops.AFTER: x = x.src[0] return x dq_native, dk_partial, dv_partial = dq.base, unwrap_partial(dk), unwrap_partial(dv) if dq_native.op is not Ops.AFTER or dk_partial is None or dv_partial is None: return None B, N, H, D, H_KV = dq.shape[0], dq.shape[1], dq.shape[2], dq.shape[3], dk.shape[2] heads_per_wg = 2 if D == 128 and (H // H_KV) % 2 == 0 else 1 partials = (H // H_KV) // heads_per_wg if dq_native.shape != (B, H, N, D) or dk_partial.shape != (B * partials, N, H_KV, D) or dv_partial.shape != dk_partial.shape: return None return dq_native, dk_partial, dv_partial def _fused_qkv_rope_grad(dq_u:UOp, dk_u:UOp, dv_u:UOp, call:UOp) -> tuple[None, None, None, UOp, None]: dq, dk, dv = Tensor(dq_u, device=dq_u.device), Tensor(dk_u, device=dk_u.device), Tensor(dv_u, device=dv_u.device) xqkv_u, freqs_u = call.src[4], call.src[5] xqkv, freqs_cis = Tensor(xqkv_u, device=xqkv_u.device), Tensor(freqs_u, device=freqs_u.device) B, N, _ = xqkv.shape H, H_KV, D = dq.shape[2], dk.shape[2], dq.shape[3] num_devices = len(xqkv.device) if isinstance(xqkv.device, tuple) else 1 is_dp, is_mp = xqkv.uop.axis == 0, xqkv.uop.axis == 2 B_local = B // num_devices if is_dp else B H_local = H // num_devices if is_mp else H H_KV_local = H_KV // num_devices if is_mp else H_KV single_device = xqkv.device[0] if isinstance(xqkv.device, tuple) else xqkv.device arch = Device[single_device].renderer.target.arch fa_native = _fa_native_grads(dq_u, dk_u, dv_u) assert fa_native is not None, "fused QKV RoPE backward requires native Flash Attention gradients" dq, dk, dv = (Tensor(x, device=x.device) for x in fa_native) dxqkv = _sharded_empty_like(xqkv, axis=xqkv.uop.axis if isinstance(xqkv.device, tuple) else None) fxn = functools.partial(custom_fused_qkv_rope_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D) dxqkv = Tensor.custom_kernel(dxqkv, dq, dk, dv, freqs_cis, fxn=fxn)[0] return None, None, None, dxqkv.uop, None def fused_qkv_rope(xqkv:Tensor, freqs_cis:Tensor, n_heads:int, n_kv_heads:int, head_dim:int) -> tuple[Tensor, Tensor, Tensor]: B, N, packed_dim = xqkv.shape assert packed_dim == n_kv_heads * (n_heads // n_kv_heads + 2) * head_dim assert freqs_cis.dtype == dtypes.bfloat16, f"fused QKV RoPE requires bfloat16 frequencies, got {freqs_cis.dtype}" assert freqs_cis.shape == (1, freqs_cis.shape[1], 1, head_dim // 2, 2) and freqs_cis.shape[1] >= N, \ f"invalid RoPE frequency shape {freqs_cis.shape} for sequence length {N} and head dimension {head_dim}" num_devices = len(xqkv.device) if isinstance(xqkv.device, tuple) else 1 is_dp, is_mp = xqkv.uop.axis == 0, xqkv.uop.axis == 2 B_local = B // num_devices if is_dp else B H_local = n_heads // num_devices if is_mp else n_heads H_KV_local = n_kv_heads // num_devices if is_mp else n_kv_heads assert H_local % H_KV_local == 0 and head_dim % 2 == 0 and head_dim <= 512 single_device = xqkv.device[0] if isinstance(xqkv.device, tuple) else xqkv.device arch = Device[single_device].renderer.target.arch axis = 0 if is_dp else 2 if is_mp else None q = _sharded_empty((B, N, n_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16) k = _sharded_empty((B, N, n_kv_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16) v = _sharded_empty((B, N, n_kv_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16) fxn = functools.partial(custom_fused_qkv_rope_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=head_dim) q, k, v, *_ = Tensor.custom_kernel(q, k, v, xqkv, freqs_cis, fxn=fxn, grad_fxn=_fused_qkv_rope_grad) return q, k, v def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor: return _sharded_empty(ref.shape, ref, axis) def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=0): def grad(dou:UOp, ker:UOp) -> tuple: do = Tensor(dou, device=dou.device) attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device) l_vec = Tensor(ker.src[2].after(ker), device=ker.src[2].device) xq = Tensor(ker.src[3], device=ker.src[3].device) xk = Tensor(ker.src[4], device=ker.src[4].device) xv = Tensor(ker.src[5], device=ker.src[5].device) dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t) GROUP_SIZE = H_local // H_KV_local HEADS_PER_WG = 2 if D == 128 and GROUP_SIZE % 2 == 0 else 1 dk_partial = _sharded_empty((B * GROUP_SIZE // HEADS_PER_WG, N, H_KV, D), xk, axis=shard_axis) dv_partial = _sharded_empty((B * GROUP_SIZE // HEADS_PER_WG, N, H_KV, D), xv, axis=shard_axis) # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach() delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t) delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2] dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, window=window))[:3] if D == 64: dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2) else: dq = dq.reshape(B, H, N//16, 4, 2, 2, D//32, 4, 4, 2).permute(0, 1, 2, 7, 8, 3, 4, 6, 5, 9).reshape(B, H, N, D).transpose(1, 2) # reduce partial dK/dV across GROUP_SIZE query heads dk = dk_partial.reshape(B, GROUP_SIZE // HEADS_PER_WG, N, H_KV, D).sum(1) dv = dv_partial.reshape(B, GROUP_SIZE // HEADS_PER_WG, N, H_KV, D).sum(1) if not has_sink: return None, None, dq.uop, dk.uop, dv.uop sinks = Tensor(ker.src[6], device=ker.src[6].device) p_sink = (sinks.reshape(1, H, 1, 1) - l_vec).exp() dsink = -(delta_vec.float() * p_sink).sum(axis=(0, 2, 3)) return None, None, dq.uop, dk.uop, dv.uop, dsink.uop return grad # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None, window:int=0): assert attn_mask is None, "attn_mask not supported" assert is_causal, "only causal attention supported" B, N, H, D = xq.shape H_KV = xk.shape[2] assert D in (64, 128), "only D=64 or D=128 supported" has_sink = sinks is not None if has_sink: sinks = sinks.float() num_devices = len(xq.device) if isinstance(xq.device, tuple) else 1 is_dp = xq.uop.axis == 0 is_mp = xq.uop.axis == 2 B_local = B // num_devices if is_dp else B H_local = H // num_devices if is_mp else H H_KV_local = H_KV // num_devices if is_mp else H_KV shard_axis = 0 if is_dp else 2 if is_mp else None shard_axis_t = 0 if is_dp else 1 if is_mp else None if DEBUG >= 2: print(f"Flash Attention {B=} {B_local=} {N=} {H=} {H_local=} {H_KV=} {H_KV_local=} {D=} on {num_devices} devices, {'DP' if is_dp else 'MP' if is_mp else 'no sharding'}") single_device = xq.device[0] if isinstance(xq.device, tuple) else xq.device arch = Device[single_device].renderer.target.arch attn = _sharded_empty_like(xq, axis=shard_axis) attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis) l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t) grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=window) fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ()) attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink, window=window), grad_fxn=grad)[:2] return attn, attn, l_vec @functools.cache def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True, window:int=0): code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text() compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math", f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}", f"-DWINDOW={window}"] Q_BLOCK_SIZE = 32 NUM_WARPS = 8 NUM_THREADS = 64 * NUM_WARPS gsz = (H, (math.ceil((N // Q_BLOCK_SIZE) / NUM_WARPS)), B) lsz = (NUM_THREADS, 1, 1) threadIdx_x = UOp.special(lsz[0], "lidx0") blockIdx_x, blockIdx_y, blockIdx_z = UOp.special(gsz[0], "gidx0"), UOp.special(gsz[1], "gidx1"), UOp.special(gsz[2], "gidx2") el = q.dtype.itemsize mem = (2*B*N*H*D + 2*B*N*H_KV*D) * el + B*H*N * l_vec.dtype.itemsize estimates = Estimates(ops=2*B*H*N*N*D, lds=mem, mem=mem) buf_inputs = (o.base, l_vec.base, q.base, k.base, v.base) + ((sinks.base,) if has_sink else ()) sink = UOp.sink(*buf_inputs, threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z, arg=KernelInfo(name="custom_fa_forward", estimates=estimates)) lib = HIPCCCompiler(arch, compile_args).compile_cached(code) if not getenv("NO_HIPCC"): lib = bytearray(lib) rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata") struct.pack_into('