import os
from typing import List, Optional, Tuple
import pytest
import torch
import triton
import triton.language as tl
from mindspeed.ops.triton.utils import prepare_chunk_indices, exp, check_shared_mem, assert_close
from mindspeed.ops.triton.chunk_o import bwd_chunk_dqkwg
@triton.heuristics({
'USE_G': lambda args: args['g'] is not None,
'USE_G_GAMMA': lambda args: args['g_gamma'] is not None,
'USE_DW': lambda args: args['dw'] is not None,
'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@triton.autotune(
configs=[
triton.Config({}, num_warps=num_warps, num_stages=num_stages)
for num_warps in [2, 4, 8]
for num_stages in [2, 3, 4]
],
key=['H', 'K', 'V', 'BT', 'BK', 'BV', 'USE_G', 'USE_G_GAMMA', 'USE_DW'],
)
@triton.jit(do_not_specialize=['T'])
def ref_chunk_bwd_kernel_dqkwg(
q,
k,
v,
h,
g,
g_gamma,
do,
dh,
dq,
dk,
dg,
w,
dv,
dw,
cu_seqlens,
chunk_indices,
scale,
B: tl.constexpr,
T,
H: tl.constexpr,
K: tl.constexpr,
V: tl.constexpr,
BT: tl.constexpr,
BK: tl.constexpr,
BV: tl.constexpr,
USE_G: tl.constexpr,
USE_G_GAMMA: tl.constexpr,
USE_DW: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_k, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
i_b, i_h = i_bh // H, i_bh % H
if IS_VARLEN:
i_tg = i_t
i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
all_T = T
T = eos - bos
NT = tl.cdiv(T, BT)
else:
NT = tl.cdiv(T, BT)
i_tg = i_b * NT + i_t
bos, eos = i_b * T, i_b * T + T
all_T = B * T
v += (bos * H + i_h) * V
do += (bos * H + i_h) * V
h += (i_tg * H + i_h).to(tl.int64) * K * V
dh += (i_tg * H + i_h).to(tl.int64) * K * V
q += (bos * H + i_h) * K
k += (bos * H + i_h) * K
dq += (bos * H + i_h) * K
dk += (bos * H + i_h) * K
if USE_DW:
w += (bos * H + i_h) * K
dw += (bos * H + i_h) * K
dv += (bos * H + i_h) * V
if USE_G:
dg += i_k * all_T * H
b_dg_last = tl.zeros([1, ], dtype=tl.float32) if USE_G else None
if USE_G_GAMMA:
b_gamma = tl.load(g_gamma + i_h)
b_g = b_gamma * (tl.arange(0, BT) + 1)
b_g_last = b_gamma * min(BT, T - i_t * BT)
b_dq = tl.zeros([BT, BK], dtype=tl.float32)
b_dk = tl.zeros([BT, BK], dtype=tl.float32)
b_ds = tl.zeros([BT, BT], dtype=tl.float32)
b_dw = tl.zeros([BT, BK], dtype=tl.float32) if USE_DW else None
for i_v in range(tl.cdiv(V, BV)):
p_v = tl.make_block_ptr(v, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
p_do = tl.make_block_ptr(do, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
p_h = tl.make_block_ptr(h, (V, K), (1, V), (i_v * BV, i_k * BK), (BV, BK), (0, 1))
p_dh = tl.make_block_ptr(dh, (V, K), (1, V), (i_v * BV, i_k * BK), (BV, BK), (0, 1))
b_v = tl.load(p_v, boundary_check=(0, 1))
b_do = tl.load(p_do, boundary_check=(0, 1))
b_h = tl.load(p_h, boundary_check=(0, 1))
b_dh = tl.load(p_dh, boundary_check=(0, 1))
if USE_G:
b_dg_last += (tl.sum(b_h * b_dh))
b_ds += tl.dot(b_do, tl.trans(b_v))
b_dq += tl.dot(b_do, b_h.to(b_do.dtype))
b_dk += tl.dot(b_v, b_dh.to(b_v.dtype))
if USE_DW:
p_dv = tl.make_block_ptr(dv, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
b_dv = tl.load(p_dv, boundary_check=(0, 1))
b_dw += tl.dot(b_dv.to(b_v.dtype), b_h.to(b_v.dtype))
if USE_DW:
p_dw = tl.make_block_ptr(dw, (T, K), (H * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
tl.store(p_dw, -b_dw.to(p_dw.dtype.element_ty), boundary_check=(0, 1))
tl.debug_barrier()
p_q = tl.make_block_ptr(q, (T, K), (H * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
p_k = tl.make_block_ptr(k, (T, K), (H * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
b_q = tl.load(p_q, boundary_check=(0, 1))
b_k = tl.load(p_k, boundary_check=(0, 1))
p_dq = tl.make_block_ptr(dq, (T, K), (H * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
p_dk = tl.make_block_ptr(dk, (T, K), (H * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
o_t = i_t * BT + tl.arange(0, BT)
m_t = o_t < T
m_A = (o_t[:, None] >= o_t[None, :]) & (m_t[:, None] & m_t)
if USE_G:
b_dg = tl.zeros([BT, ], dtype=tl.float32)
g += bos * H + i_h
dg += bos * H + i_h
p_g = tl.make_block_ptr(g, (T,), (H,), (i_t * BT,), (BT,), (0,))
b_g = tl.load(p_g, boundary_check=(0,))
b_g_last = tl.load(g + (min(i_t * BT + BT, T) - 1) * H)
b_dg_last *= exp(b_g_last)
b_dq = b_dq * exp(b_g)[:, None] * scale
b_dg += tl.sum(b_dq * b_q, axis=1)
b_dk = b_dk * tl.where(m_t, exp(-b_g + b_g_last), 0)[:, None]
b_dg -= tl.sum(b_k * b_dk, axis=1)
b_dg_last += tl.sum(b_dk * b_k)
b_ds = tl.where(m_A, b_ds * exp(b_g[:, None] - b_g[None, :]), 0) * scale
b_ds2 = b_ds * tl.dot(b_q, tl.trans(b_k))
b_dg += tl.sum(b_ds2, axis=1)
b_dg -= tl.sum(b_ds2, axis=0)
b_ds = b_ds.to(b_k.dtype)
b_dq += tl.dot(b_ds, b_k)
b_dk += tl.dot(tl.trans(b_ds), b_q)
p_dg = tl.make_block_ptr(dg, (T,), (H,), (i_t * BT,), (BT,), (0,))
b_dg = tl.where(o_t < min(i_t * BT + BT, T) - 1, b_dg, b_dg + b_dg_last)
tl.store(p_dq, b_dq.to(p_dq.dtype.element_ty), boundary_check=(0, 1))
tl.store(p_dk, b_dk.to(p_dk.dtype.element_ty), boundary_check=(0, 1))
tl.store(p_dg, b_dg.to(p_dg.dtype.element_ty), boundary_check=(0,))
elif USE_G_GAMMA:
b_dq = b_dq * exp(b_g)[:, None] * scale
b_dk = b_dk * tl.where(m_t, exp(-b_g + b_g_last), 0)[:, None]
b_ds = tl.where(m_A, b_ds * exp(b_g[:, None] - b_g[None, :]), 0) * scale
b_ds = b_ds.to(b_k.dtype)
b_dq += tl.dot(b_ds, b_k)
b_dk += tl.dot(tl.trans(b_ds), b_q)
tl.store(p_dq, b_dq.to(p_dq.dtype.element_ty), boundary_check=(0, 1))
tl.store(p_dk, b_dk.to(p_dk.dtype.element_ty), boundary_check=(0, 1))
else:
b_ds = tl.where(m_A, b_ds, 0)
b_ds = b_ds.to(b_k.dtype)
b_dq += tl.dot(b_ds, b_k)
b_dk += tl.dot(tl.trans(b_ds), b_q) * scale
b_dq *= scale
tl.store(p_dq, b_dq.to(p_dq.dtype.element_ty), boundary_check=(0, 1))
tl.store(p_dk, b_dk.to(p_dk.dtype.element_ty), boundary_check=(0, 1))
def ref_chunk_bwd_dqkwg(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
do: torch.Tensor,
h: torch.Tensor,
dh: torch.Tensor,
g: Optional[torch.Tensor] = None,
g_gamma: Optional[torch.Tensor] = None,
dv: Optional[torch.Tensor] = None,
w: Optional[torch.Tensor] = None,
cu_seqlens: Optional[torch.LongTensor] = None,
chunk_size: int = 64,
scale: float = 1.0,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
B, T, H, K, V = *k.shape, v.shape[-1]
BT = min(chunk_size, max(16, triton.next_power_of_2(T)))
chunk_indices = prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
CONST_TILING = 64 if check_shared_mem() else 32
BK = min(max(triton.next_power_of_2(K), 16), CONST_TILING)
BV = min(max(triton.next_power_of_2(V), 16), CONST_TILING)
NK = triton.cdiv(K, BK)
dq = torch.empty_like(q)
dk = torch.empty_like(k)
dg = torch.empty(NK, *g.shape, dtype=torch.float32, device=g.device) if g is not None else None
dw = torch.empty_like(w) if w is not None else None
grid = (NK, NT, B * H)
ref_chunk_bwd_kernel_dqkwg[grid](
q=q,
k=k,
v=v,
h=h,
g=g,
g_gamma=g_gamma,
do=do,
dh=dh,
dv=dv,
w=w,
dw=dw,
dq=dq,
dk=dk,
dg=dg,
cu_seqlens=cu_seqlens,
chunk_indices=chunk_indices,
scale=scale,
B=B,
T=T,
H=H,
K=K,
V=V,
BT=BT,
BK=BK,
BV=BV,
)
if dg is not None:
dg = dg.sum(0)
return dq, dk, dw, dg
@pytest.mark.parametrize(
('B', 'T', 'H', 'D', 'hidden_size', 'scale', 'chunk_size', 'cu_seqlens'),
[
pytest.param(*test, id="B{}-T{}-H{}-D{}-hidden_size{}-scale{}-chunk_size{}-cu_seqlens{}".format(*test))
for test in [
(1, 2048, 32, 128, 2048, 0.5, 16, [0, 1024, 1164, 2048]),
(1, 1024, 32, 128, 2048, 0.5, 16, None),
(1, 2048, 32, 128, 2048, 0.5, 16, None),
(2, 2048, 32, 128, 2048, 0.5, 16, None),
]
]
)
def test_chunk_bwd_dqkwg(B, T, H, D, hidden_size, scale, chunk_size, cu_seqlens):
device = "npu:0"
device_dtype = torch.float32
torch.manual_seed(42)
torch.npu.manual_seed(42)
if cu_seqlens is not None:
cu_seqlens = torch.LongTensor(cu_seqlens).to(device)
q = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
k = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
v = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
w = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
g = torch.rand((B, T, H), device=device, dtype=device_dtype)
h = torch.rand((B, hidden_size, H, D, D), device=device, dtype=device_dtype)
dv = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
do = torch.rand((B, T, H, D), device=device, dtype=device_dtype)
dh = torch.rand((B, hidden_size, H, D, D), device=device, dtype=device_dtype)
ref_dq, ref_dk, ref_dw, ref_dg = ref_chunk_bwd_dqkwg(
q=q,
k=k,
v=v,
h=h,
g=g,
do=do,
dh=dh,
dv=dv,
w=w,
cu_seqlens=cu_seqlens,
scale=scale,
chunk_size=chunk_size
)
dq, dk, dw, dg = bwd_chunk_dqkwg(
q=q,
k=k,
v=v,
h=h,
g=g,
do=do,
dh=dh,
dv=dv,
w=w,
cu_seqlens=cu_seqlens,
scale=scale,
chunk_size=chunk_size
)
print("dq diff:", torch.max(torch.abs(ref_dq - dq)))
print("dk diff:", torch.max(torch.abs(ref_dk - dk)))
print("dw diff:", torch.max(torch.abs(ref_dw - dw)))
print("dg diff:", torch.max(torch.abs(ref_dg - dg)))
assert_close('dq', ref_dq, dq, 0.001)
assert_close('dk', ref_dk, dk, 0.001)
assert_close('dw', ref_dw, dw, 0.001)
assert_close('dg', ref_dg, dg, 0.001)