drbh
commited on
Commit
·
4080f9c
1
Parent(s):
b0d3c12
fix: adjust types
Browse files- flash_attn/flash_api.cpp +102 -32
flash_attn/flash_api.cpp
CHANGED
@@ -1507,45 +1507,61 @@ mha_fwd(const at::Tensor &q, // batch_size x seqle
|
|
1507 |
float softcap_float = static_cast<float>(softcap);
|
1508 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1509 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1510 |
-
|
1511 |
return FLASH_NAMESPACE::mha_fwd(const_cast<at::Tensor &>(q), k, v, out, alibi_slopes, p_dropout_float, softmax_scale_float, is_causal, window_size_left_int, window_size_right_int, softcap_float, return_softmax, gen);
|
1512 |
}
|
1513 |
|
1514 |
std::vector<at::Tensor>
|
1515 |
-
mha_varlen_fwd(
|
1516 |
-
const at::Tensor &k,
|
1517 |
-
const at::Tensor &v,
|
1518 |
-
const
|
1519 |
-
const at::Tensor &cu_seqlens_q,
|
1520 |
-
const at::Tensor &cu_seqlens_k,
|
|
|
|
|
|
|
|
|
1521 |
const int64_t max_seqlen_q,
|
1522 |
const int64_t max_seqlen_k,
|
1523 |
const double p_dropout,
|
1524 |
const double softmax_scale,
|
1525 |
-
bool
|
|
|
1526 |
const int64_t window_size_left,
|
1527 |
const int64_t window_size_right,
|
1528 |
const double softcap,
|
1529 |
const bool return_softmax,
|
1530 |
-
const
|
1531 |
-
|
1532 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
1533 |
-
|
1534 |
// Prepare the optional arguments as non-const references.
|
1535 |
std::optional<at::Tensor> out = out_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(out_.value())) : std::nullopt;
|
1536 |
-
|
|
|
|
|
|
|
|
|
1537 |
if (!out.has_value()){
|
1538 |
out = torch::empty_like(q);
|
1539 |
}
|
1540 |
-
|
1541 |
// Convert double to float and int64_t to int.
|
1542 |
float p_dropout_float = static_cast<float>(p_dropout);
|
1543 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1544 |
float softcap_float = static_cast<float>(softcap);
|
|
|
|
|
1545 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1546 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1547 |
-
|
1548 |
-
return FLASH_NAMESPACE::mha_varlen_fwd(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1549 |
}
|
1550 |
|
1551 |
std::vector<at::Tensor>
|
@@ -1570,7 +1586,7 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q
|
|
1570 |
std::optional<at::Tensor> &rng_state) {
|
1571 |
|
1572 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
1573 |
-
|
1574 |
// Prepare the optional arguments as non-const references.
|
1575 |
std::optional<at::Tensor> dq = dq_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dq_.value())) : std::nullopt;
|
1576 |
std::optional<at::Tensor> dk = dk_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dk_.value())) : std::nullopt;
|
@@ -1584,7 +1600,15 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q
|
|
1584 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1585 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1586 |
|
1587 |
-
return FLASH_NAMESPACE::mha_bwd(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1588 |
}
|
1589 |
|
1590 |
|
@@ -1595,12 +1619,17 @@ mha_varlen_bwd(const at::Tensor &dout, // batch_size x seqlen_q
|
|
1595 |
const at::Tensor &v, // batch_size x seqlen_k x num_heads_k x head_size
|
1596 |
const at::Tensor &out, // batch_size x seqlen_q x num_heads x head_size
|
1597 |
const at::Tensor &softmax_lse, // b x h x seqlen_q
|
|
|
|
|
|
|
1598 |
const at::Tensor &cu_seqlens_q, // batch_size + 1
|
1599 |
const at::Tensor &cu_seqlens_k, // batch_size + 1
|
|
|
1600 |
const int64_t max_seqlen_q,
|
1601 |
const int64_t max_seqlen_k,
|
1602 |
const double p_dropout,
|
1603 |
const double softmax_scale,
|
|
|
1604 |
const bool is_causal,
|
1605 |
const int64_t window_size_left,
|
1606 |
const int64_t window_size_right,
|
@@ -1608,17 +1637,36 @@ mha_varlen_bwd(const at::Tensor &dout, // batch_size x seqlen_q
|
|
1608 |
const bool deterministic,
|
1609 |
std::optional<at::Generator> gen_,
|
1610 |
std::optional<at::Tensor> &rng_state) {
|
1611 |
-
|
1612 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
1613 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1614 |
// Convert double to float and int64_t to int.
|
1615 |
float p_dropout_float = static_cast<float>(p_dropout);
|
1616 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1617 |
float softcap_float = static_cast<float>(softcap);
|
|
|
|
|
1618 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1619 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1620 |
|
1621 |
-
return FLASH_NAMESPACE::mha_varlen_bwd(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1622 |
}
|
1623 |
|
1624 |
std::vector<at::Tensor>
|
@@ -1643,25 +1691,47 @@ mha_fwd_kvcache(const at::Tensor &q, // batch
|
|
1643 |
bool is_rotary_interleaved, // if true, rotary combines indices 0 & 1, else indices 0 & rotary_dim / 2
|
1644 |
const int64_t num_splits
|
1645 |
) {
|
1646 |
-
|
1647 |
-
// Prepare the optional arguments as
|
1648 |
-
std::optional<at::Tensor> k = k_.has_value() ? std::optional<at::Tensor>(
|
1649 |
-
std::optional<at::Tensor> v = v_.has_value() ? std::optional<at::Tensor>(
|
1650 |
-
std::optional<at::Tensor> seqlens_k = seqlens_k_.has_value() ? std::optional<at::Tensor>(
|
1651 |
-
std::optional<at::Tensor> rotary_cos = rotary_cos_.has_value() ? std::optional<at::Tensor>(
|
1652 |
-
std::optional<at::Tensor> rotary_sin = rotary_sin_.has_value() ? std::optional<at::Tensor>(
|
1653 |
-
std::optional<at::Tensor> cache_batch_idx = cache_batch_idx_.has_value() ? std::optional<at::Tensor>(
|
1654 |
-
std::optional<at::Tensor> leftpad_k = leftpad_k_.has_value() ? std::optional<at::Tensor>(
|
|
|
|
|
1655 |
std::optional<at::Tensor> block_table = block_table_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(block_table_.value())) : std::nullopt;
|
1656 |
std::optional<at::Tensor> alibi_slopes = alibi_slopes_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(alibi_slopes_.value())) : std::nullopt;
|
1657 |
std::optional<at::Tensor> out = out_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(out_.value())) : std::nullopt;
|
1658 |
-
|
|
|
|
|
|
|
|
|
1659 |
// Convert double to float and int64_t to int.
|
1660 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1661 |
float softcap_float = static_cast<float>(softcap);
|
1662 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1663 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1664 |
int num_splits_int = static_cast<int>(num_splits);
|
1665 |
-
|
1666 |
-
return FLASH_NAMESPACE::mha_fwd_kvcache(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1667 |
}
|
|
|
1507 |
float softcap_float = static_cast<float>(softcap);
|
1508 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1509 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1510 |
+
|
1511 |
return FLASH_NAMESPACE::mha_fwd(const_cast<at::Tensor &>(q), k, v, out, alibi_slopes, p_dropout_float, softmax_scale_float, is_causal, window_size_left_int, window_size_right_int, softcap_float, return_softmax, gen);
|
1512 |
}
|
1513 |
|
1514 |
std::vector<at::Tensor>
|
1515 |
+
mha_varlen_fwd(at::Tensor &q, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
|
1516 |
+
const at::Tensor &k, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i or num_blocks x page_block_size x num_heads_k x head_>
|
1517 |
+
const at::Tensor &v, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i or num_blocks x page_block_size x num_heads_k x head_>
|
1518 |
+
const std::optional<at::Tensor> &out_, // total_q x num_heads x head_size, total_k := \sum_{i=0}^{b} s_i
|
1519 |
+
const at::Tensor &cu_seqlens_q, // b+1
|
1520 |
+
const at::Tensor &cu_seqlens_k, // b+1
|
1521 |
+
const std::optional<at::Tensor> &seqused_k_, // b. If given, only this many elements of each batch element's keys are used.
|
1522 |
+
const std::optional<const at::Tensor> &leftpad_k_, // batch_size
|
1523 |
+
const std::optional<at::Tensor> &block_table_, // batch_size x max_num_blocks_per_seq
|
1524 |
+
const std::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
|
1525 |
const int64_t max_seqlen_q,
|
1526 |
const int64_t max_seqlen_k,
|
1527 |
const double p_dropout,
|
1528 |
const double softmax_scale,
|
1529 |
+
const bool zero_tensors,
|
1530 |
+
const bool is_causal,
|
1531 |
const int64_t window_size_left,
|
1532 |
const int64_t window_size_right,
|
1533 |
const double softcap,
|
1534 |
const bool return_softmax,
|
1535 |
+
const std::optional<at::Generator> gen_) {
|
|
|
1536 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
|
|
1537 |
// Prepare the optional arguments as non-const references.
|
1538 |
std::optional<at::Tensor> out = out_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(out_.value())) : std::nullopt;
|
1539 |
+
std::optional<at::Tensor> seqused_k = seqused_k_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(seqused_k_.value())) : std::nullopt;
|
1540 |
+
std::optional<const at::Tensor> leftpad_k = leftpad_k_.has_value() ? std::optional<const at::Tensor>(leftpad_k_.value()) : std::nullopt;
|
1541 |
+
std::optional<at::Tensor> block_table = block_table_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(block_table_.value())) : std::nullopt;
|
1542 |
+
std::optional<at::Tensor> alibi_slopes = alibi_slopes_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(alibi_slopes_.value())) : std::nullopt;
|
1543 |
+
|
1544 |
if (!out.has_value()){
|
1545 |
out = torch::empty_like(q);
|
1546 |
}
|
|
|
1547 |
// Convert double to float and int64_t to int.
|
1548 |
float p_dropout_float = static_cast<float>(p_dropout);
|
1549 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1550 |
float softcap_float = static_cast<float>(softcap);
|
1551 |
+
int max_seqlen_q_int = static_cast<int>(max_seqlen_q);
|
1552 |
+
int max_seqlen_k_int = static_cast<int>(max_seqlen_k);
|
1553 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1554 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1555 |
+
|
1556 |
+
return FLASH_NAMESPACE::mha_varlen_fwd(
|
1557 |
+
const_cast<at::Tensor &>(q), k, v, out,
|
1558 |
+
cu_seqlens_q, cu_seqlens_k,
|
1559 |
+
seqused_k, leftpad_k, block_table, alibi_slopes,
|
1560 |
+
max_seqlen_q_int, max_seqlen_k_int,
|
1561 |
+
p_dropout_float, softmax_scale_float,
|
1562 |
+
zero_tensors, is_causal,
|
1563 |
+
window_size_left_int, window_size_right_int,
|
1564 |
+
softcap_float, return_softmax, gen);
|
1565 |
}
|
1566 |
|
1567 |
std::vector<at::Tensor>
|
|
|
1586 |
std::optional<at::Tensor> &rng_state) {
|
1587 |
|
1588 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
1589 |
+
|
1590 |
// Prepare the optional arguments as non-const references.
|
1591 |
std::optional<at::Tensor> dq = dq_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dq_.value())) : std::nullopt;
|
1592 |
std::optional<at::Tensor> dk = dk_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dk_.value())) : std::nullopt;
|
|
|
1600 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1601 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1602 |
|
1603 |
+
return FLASH_NAMESPACE::mha_bwd(
|
1604 |
+
const_cast<at::Tensor &>(dout),
|
1605 |
+
q, k, v, out, softmax_lse,
|
1606 |
+
dq, dk, dv, alibi_slopes,
|
1607 |
+
p_dropout_float, softmax_scale_float,
|
1608 |
+
is_causal,
|
1609 |
+
window_size_left_int, window_size_right_int,
|
1610 |
+
softcap_float, deterministic,
|
1611 |
+
gen, rng_state);
|
1612 |
}
|
1613 |
|
1614 |
|
|
|
1619 |
const at::Tensor &v, // batch_size x seqlen_k x num_heads_k x head_size
|
1620 |
const at::Tensor &out, // batch_size x seqlen_q x num_heads x head_size
|
1621 |
const at::Tensor &softmax_lse, // b x h x seqlen_q
|
1622 |
+
const std::optional<at::Tensor> &dq_, // batch_size x seqlen_q x num_heads x head_size
|
1623 |
+
const std::optional<at::Tensor> &dk_, // batch_size x seqlen_k x num_heads_k x head_size
|
1624 |
+
const std::optional<at::Tensor> &dv_, // batch_size x seqlen_k x num_heads_k x head_size
|
1625 |
const at::Tensor &cu_seqlens_q, // batch_size + 1
|
1626 |
const at::Tensor &cu_seqlens_k, // batch_size + 1
|
1627 |
+
const std::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
|
1628 |
const int64_t max_seqlen_q,
|
1629 |
const int64_t max_seqlen_k,
|
1630 |
const double p_dropout,
|
1631 |
const double softmax_scale,
|
1632 |
+
const bool zero_tensors,
|
1633 |
const bool is_causal,
|
1634 |
const int64_t window_size_left,
|
1635 |
const int64_t window_size_right,
|
|
|
1637 |
const bool deterministic,
|
1638 |
std::optional<at::Generator> gen_,
|
1639 |
std::optional<at::Tensor> &rng_state) {
|
1640 |
+
|
1641 |
auto gen = gen_.value_or(at::cuda::detail::getDefaultCUDAGenerator());
|
1642 |
+
|
1643 |
+
// Prepare the optional arguments as non-const references.
|
1644 |
+
std::optional<at::Tensor> dq = dq_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dq_.value())) : std::nullopt;
|
1645 |
+
std::optional<at::Tensor> dk = dk_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dk_.value())) : std::nullopt;
|
1646 |
+
std::optional<at::Tensor> dv = dv_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(dv_.value())) : std::nullopt;
|
1647 |
+
std::optional<at::Tensor> alibi_slopes = alibi_slopes_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(alibi_slopes_.value())) : std::nullopt;
|
1648 |
+
|
1649 |
// Convert double to float and int64_t to int.
|
1650 |
float p_dropout_float = static_cast<float>(p_dropout);
|
1651 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1652 |
float softcap_float = static_cast<float>(softcap);
|
1653 |
+
int max_seqlen_q_int = static_cast<int>(max_seqlen_q);
|
1654 |
+
int max_seqlen_k_int = static_cast<int>(max_seqlen_k);
|
1655 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1656 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1657 |
|
1658 |
+
return FLASH_NAMESPACE::mha_varlen_bwd(
|
1659 |
+
const_cast<at::Tensor &>(dout),
|
1660 |
+
q, k, v, out, softmax_lse,
|
1661 |
+
dq, dk, dv,
|
1662 |
+
cu_seqlens_q, cu_seqlens_k,
|
1663 |
+
alibi_slopes,
|
1664 |
+
max_seqlen_q_int, max_seqlen_k_int,
|
1665 |
+
p_dropout_float, softmax_scale_float,
|
1666 |
+
zero_tensors, is_causal,
|
1667 |
+
window_size_left_int, window_size_right_int,
|
1668 |
+
softcap_float, deterministic,
|
1669 |
+
gen, rng_state);
|
1670 |
}
|
1671 |
|
1672 |
std::vector<at::Tensor>
|
|
|
1691 |
bool is_rotary_interleaved, // if true, rotary combines indices 0 & 1, else indices 0 & rotary_dim / 2
|
1692 |
const int64_t num_splits
|
1693 |
) {
|
1694 |
+
|
1695 |
+
// Prepare the optional arguments as const references where needed
|
1696 |
+
std::optional<const at::Tensor> k = k_.has_value() ? std::optional<const at::Tensor>(k_.value()) : std::nullopt;
|
1697 |
+
std::optional<const at::Tensor> v = v_.has_value() ? std::optional<const at::Tensor>(v_.value()) : std::nullopt;
|
1698 |
+
std::optional<const at::Tensor> seqlens_k = seqlens_k_.has_value() ? std::optional<const at::Tensor>(seqlens_k_.value()) : std::nullopt;
|
1699 |
+
std::optional<const at::Tensor> rotary_cos = rotary_cos_.has_value() ? std::optional<const at::Tensor>(rotary_cos_.value()) : std::nullopt;
|
1700 |
+
std::optional<const at::Tensor> rotary_sin = rotary_sin_.has_value() ? std::optional<const at::Tensor>(rotary_sin_.value()) : std::nullopt;
|
1701 |
+
std::optional<const at::Tensor> cache_batch_idx = cache_batch_idx_.has_value() ? std::optional<const at::Tensor>(cache_batch_idx_.value()) : std::nullopt;
|
1702 |
+
std::optional<const at::Tensor> leftpad_k = leftpad_k_.has_value() ? std::optional<const at::Tensor>(leftpad_k_.value()) : std::nullopt;
|
1703 |
+
|
1704 |
+
// For non-const tensors
|
1705 |
std::optional<at::Tensor> block_table = block_table_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(block_table_.value())) : std::nullopt;
|
1706 |
std::optional<at::Tensor> alibi_slopes = alibi_slopes_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(alibi_slopes_.value())) : std::nullopt;
|
1707 |
std::optional<at::Tensor> out = out_.has_value() ? std::optional<at::Tensor>(const_cast<at::Tensor &>(out_.value())) : std::nullopt;
|
1708 |
+
|
1709 |
+
if (!out.has_value()){
|
1710 |
+
out = torch::empty_like(q);
|
1711 |
+
}
|
1712 |
+
|
1713 |
// Convert double to float and int64_t to int.
|
1714 |
float softmax_scale_float = static_cast<float>(softmax_scale);
|
1715 |
float softcap_float = static_cast<float>(softcap);
|
1716 |
int window_size_left_int = static_cast<int>(window_size_left);
|
1717 |
int window_size_right_int = static_cast<int>(window_size_right);
|
1718 |
int num_splits_int = static_cast<int>(num_splits);
|
1719 |
+
|
1720 |
+
return FLASH_NAMESPACE::mha_fwd_kvcache(
|
1721 |
+
const_cast<at::Tensor &>(q),
|
1722 |
+
kcache, vcache,
|
1723 |
+
k, v,
|
1724 |
+
seqlens_k,
|
1725 |
+
rotary_cos, rotary_sin,
|
1726 |
+
cache_batch_idx,
|
1727 |
+
leftpad_k,
|
1728 |
+
block_table, alibi_slopes,
|
1729 |
+
out,
|
1730 |
+
softmax_scale_float,
|
1731 |
+
is_causal,
|
1732 |
+
window_size_left_int, window_size_right_int,
|
1733 |
+
softcap_float,
|
1734 |
+
is_rotary_interleaved,
|
1735 |
+
num_splits_int
|
1736 |
+
);
|
1737 |
}
|