tinygrad/extra/thunder/amd/fa.py

import math, pathlib, functools, struct

from tinygrad import Device, Tensor
from tinygrad.dtype import DTypeLike, dtypes
from tinygrad.helpers import DEBUG
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.empty(*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.empty(*shape, dtype=dtype, device=ref.device).uop.multi(axis), dtype=dtype, device=ref.device)

def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)

@functools.cache
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):
  def grad(dou:UOp, ker:UOp) -> tuple[None, None, UOp, UOp, UOp]:
    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
    dk_partial = _sharded_empty((B * GROUP_SIZE, N, H_KV, D), xk, axis=shard_axis)
    dv_partial = _sharded_empty((B * GROUP_SIZE, 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))[:3]

    # unshuffle dq: atomic_pk_add_bf16_with_warpid creates a shuffled layout within each 16x128 tile
    # decompose each tile into (j=4, a=2, b=2, d=4, e=4, k=4, c=2) and permute to (e, k, j, a, d, b, c) = standard row-major
    dq = dq.reshape(B, H, N//16, 4, 2, 2, 4, 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, N, H_KV, D).sum(1)
    dv = dv_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)

    return None, None, dq.uop, dk.uop, dv.uop
  return grad

def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False):
  assert attn_mask is None, "attn_mask not supported"
  assert is_causal, "only causal attention supported"

  xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)

  B, N, H, D = xq.shape
  H_KV = xk.shape[2]
  assert D == 128, "only D=128 supported"

  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.arch

  attn = _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)

  attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, 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), grad_fxn=grad)[:2]

  return attn.transpose(1, 2)

@functools.cache
def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  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}"]

  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)
  sink = UOp.sink(o.base, l_vec.base, q.base, k.base, v.base,
                  threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z,
                  arg=KernelInfo(name="custom_fa_forward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
  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('<I', lib, rodata_off, 160000)
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  code = (pathlib.Path(__file__).parent / "fa_bwd_pre.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}"]

  DOT_SLICE_QO = 16
  NUM_WARPS = 4
  NUM_THREADS = 64 * NUM_WARPS
  gsz = (B, H, N // (DOT_SLICE_QO * NUM_WARPS))
  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 = o.dtype.itemsize
  mem = 3*B*H*N*D * el + B*H*N * delta_vec.dtype.itemsize
  estimates = Estimates(ops=2*B*H*N*D, lds=mem, mem=mem)
  sink = UOp.sink(delta_vec.base, dq.base, o.base, do.base,
                  threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z,
                  arg=KernelInfo(name="custom_fa_backward_pre", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
  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('<I', lib, rodata_off, 160000)
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  code = (pathlib.Path(__file__).parent / "fa_bwd_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}"]

  BLOCK_SIZE_KV = 256
  NUM_WARPS = 4
  NUM_THREADS = 64 * NUM_WARPS
  gsz = (H, N // BLOCK_SIZE_KV, 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 = (3*B*H*N*D + 4*B*H_KV*N*D) * el + 2*B*H*N * l_vec.dtype.itemsize
  estimates = Estimates(ops=5*B*H*N*N*D, lds=mem, mem=mem)
  sink = UOp.sink(dq.base, dk.base, dv.base, do.base, q.base, k.base, v.base, l_vec.base, delta_vec.base,
                  threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z,
                  arg=KernelInfo(name="custom_fa_backward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
  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('<I', lib, rodata_off, 160000)
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  code = (pathlib.Path(__file__).parent / "fa_bwd_post.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}"]

  DOT_SLICE_QO = 16
  NUM_WARPS = 4
  NUM_THREADS = 64 * NUM_WARPS
  gsz = (B, H, N // (DOT_SLICE_QO * NUM_WARPS))
  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 = dq_out.dtype.itemsize
  mem = 2*B*H*N*D * el
  estimates = Estimates(lds=mem, mem=mem)
  sink = UOp.sink(dq_out.base, dq_in.base,
                  threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z,
                  arg=KernelInfo(name="custom_fa_backward_post", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
  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('<I', lib, rodata_off, 160000)
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))