Functionintra_chunk_preprocess_bwd_kernel(
q, k, w, w2, beta,
AT,
dA_local, dq, dq_new, dk, dk_new, dw, dbeta, dw1, dw2, T,
offsets, in
fla/ops/path_attn/intra_chunk_preprocess_bwd.py:21
Functionintra_chunk_preprocess_fwd_kernel(
q,
k,
v,
w,
beta,
g_cumsum,
o,
A,
L,
M,
w2,
q_new,
k_new
fla/ops/path_attn/intra_chunk_preprocess_fwd.py:20
Functioniplr_chunkwise(q, k, v, alpha, beta, initial_state=None, output_final_state=True, chunk_size=32)
fla/ops/generalized_delta_rule/iplr/naive.py:40
Functioniplr_recurrence(q, k, v, alpha, beta, initial_state=None, output_final_state=True)
fla/ops/generalized_delta_rule/iplr/naive.py:15
Functionk_update_bwd_kernel_long(
grad_out, k, a, ka,
dk, da, dka,
cu_seqlens, chunk_indices,
T, D,
BD: tl.constexpr, BT:
fla/ops/rwkv7/fused_k_update.py:175
Functionk_update_bwd_kernel_short(
grad_out, k, a, ka,
dk, da, dka,
cu_seqlens,
T, D,
BT: tl.constexpr,
BD: tl.constexp
fla/ops/rwkv7/fused_k_update.py:123
Functionk_update_fwd_kernel_long(
k, a, ka, out,
cu_seqlens, chunk_indices,
T, D,
BD: tl.constexpr, BT: tl.constexpr,
IS_V
fla/ops/rwkv7/fused_k_update.py:76
Functionk_update_fwd_kernel_short(
k, a, ka, out,
cu_seqlens,
T, D,
BD: tl.constexpr,
IS_VARLEN: tl.constexpr,
)
fla/ops/rwkv7/fused_k_update.py:33
Functionkda_gate_bwd_kernel(
g,
A_log,
dt_bias,
beta,
dyg,
dyb,
dg,
dA,
dbeta,
lower_bound,
T
fla/ops/kda/gate.py:144
Functionkda_gate_chunk_cumsum_vector_kernel(
s,
A_log,
dt_bias,
o,
scale,
cu_seqlens,
chunk_indices,
lower_bound,
T,
fla/ops/kda/gate.py:363
Functionkda_gate_fwd_kernel(
g,
A_log,
dt_bias,
beta,
yg,
yb,
lower_bound,
T,
H: tl.constexpr,
D:
fla/ops/kda/gate.py:88
Functionkl_div_kernel(
logits,
target_logits,
loss,
s_logits,
s_loss,
reduction: tl.constexpr,
N: tl.co
fla/modules/fused_kl_div.py:27
Functionkl_div_kernel(
logits,
target_logits,
loss,
s_logits,
s_loss,
reduction: tl.constexpr,
N: tl.co
fla/modules/backends/triton_ascend/fused_kl_div.py:22
Functionl2norm_bwd_kernel(
y,
rstd,
dy,
dx,
eps,
T,
D: tl.constexpr,
BD: tl.constexpr,
NB: tl.const
fla/modules/l2norm.py:113
Functionl2norm_fwd_kernel(
x,
y,
rstd,
eps,
T,
D: tl.constexpr,
BD: tl.constexpr,
NB: tl.constexpr,
fla/modules/l2norm.py:83
Methodlayer_norm_bwd(
self,
dy,
x,
weight,
bias,
mean=None,
rstd=None,
fla/modules/backends/triton_ascend/__init__.py:223
Functionlayer_norm_bwd_kernel1(
x,
w,
b,
y,
dy,
dx,
dw,
db,
dres,
dres_in,
mean,
rstd,
T
fla/modules/backends/triton_ascend/layernorm.py:94
Methodlayer_norm_fwd(
self,
x,
weight,
bias,
eps=1e-5,
residual=None,
out_
fla/modules/backends/triton_ascend/__init__.py:198
Functionlayer_norm_fwd_kernel1(
x,
y,
w,
b,
res,
res_out,
mean,
rstd,
eps,
G: tl.constexpr,
D: t
fla/modules/backends/triton_ascend/layernorm.py:26
Functionlogsigmoid_bwd_kernel(
x,
dy,
dx,
stride_x_row,
stride_dy_row,
stride_dx_row,
temperature,
T,
D
fla/modules/activations.py:244
Functionlogsigmoid_bwd_kernel(
x,
dx,
dy,
temperature,
T,
D: tl.constexpr,
stride_x_row,
stride_dx_row,
fla/modules/backends/triton_ascend/activations.py:172
Functionlogsigmoid_fwd_kernel(
x,
y,
stride_x_row,
stride_y_row,
temperature,
T,
D: tl.constexpr,
B: tl.con
fla/modules/activations.py:212
Functionlogsigmoid_fwd_kernel(
x,
y,
temperature,
T,
D: tl.constexpr,
stride_x_row,
stride_y_row,
B: tl.con
fla/modules/backends/triton_ascend/activations.py:137
Methodlogsumexp_fwd(
self,
x,
scale=None,
softcapping=None,
dtype=None,
)
fla/modules/backends/triton_ascend/__init__.py:80
Functionlogsumexp_fwd_kernel(
x,
z,
scale,
D: tl.constexpr,
B: tl.constexpr,
HAS_SCALE: tl.constexpr,
)
fla/ops/utils/logsumexp.py:28
Functionlogsumexp_fwd_kernel(
x,
z,
scale,
softcapping,
D: tl.constexpr,
B: tl.constexpr,
HAS_SCALE: tl.conste
fla/modules/fused_linear_cross_entropy.py:44
Functionlogsumexp_fwd_kernel(
x,
z,
scale,
softcapping: tl.constexpr,
D: tl.constexpr,
B: tl.constexpr,
ROWWIS
fla/modules/backends/triton_ascend/fused_linear_cross_entropy.py:26
Functionmean_pooling_bwd_kernel(
do,
dx,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
D: tl.constexpr,
BT:
fla/ops/utils/pooling.py:76
Functionmean_pooling_fwd_kernel(
x,
o,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
D: tl.constexpr,
BT: tl
fla/ops/utils/pooling.py:29
Functionmerge_16x16_to_32x32_inverse_kernel(
A,
Ai,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BT: tl.constexpr,
USE_
fla/ops/utils/solve_tril.py:105
Functionmerge_16x16_to_64x64_inverse_kernel(
A,
Ai,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BT: tl.constexpr,
USE_
fla/ops/utils/solve_tril.py:194
Functionmesa_net_decoding_one_step_kernel(
q,
k,
v,
g,
o,
lamb,
beta,
prev_h_kk,
prev_h_kv,
curr_h_kk,
curr
fla/ops/mesa_net/decoding_one_step.py:17
Functionnaive_attn_decoding_kernel(
q,
k,
v,
o,
g_cumsum,
sink_bias,
scale,
cu_seqlens,
T,
B: tl.constex
fla/ops/attn/decoding.py:31
Functionnaive_chunk_gated_delta_rule Reference PyTorch implementation of chunk gated delta rule. Args: q: [B, T, H, K] k: [B, T, H, K] v: [B, T, H, V]
fla/ops/gated_delta_rule/naive.py:67
Functionnaive_chunk_rwkv6(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
w: torch.Tensor,
u: torch.Tensor,
fla/ops/rwkv6/chunk_naive.py:12
Functionnaive_mesa_net_CG(q, k, v, g, lamb, beta, chunk_size, max_CG_iteration=30, h_kk_init=None, h_kv_init=None)
fla/ops/mesa_net/naive.py:71
Functionnaive_recurrent_comba Reference PyTorch implementation of recurrent COMBA. Args: q: [B, T, H, K] k: [B, T, H, K] v: [B, T, H, V] p
fla/ops/comba/naive.py:13
Functionnaive_recurrent_rwkv6(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
w: torch.Tensor,
u: torch.Tensor,
fla/ops/rwkv6/recurrent_naive.py:11
Functionnaive_recurrent_rwkv6_bwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
w: torch.Tensor,
u: torch.Tensor,
fla/ops/rwkv6/recurrent_naive.py:47
Functionparallel_attn_bwd_kernel_dkv(
q,
k,
v,
g_cumsum,
lse,
delta,
do,
dk,
dv,
dg_cumsum,
cu_seqlens
fla/ops/attn/parallel.py:353
Functionparallel_attn_bwd_kernel_dq(
q,
k,
v,
lse,
delta,
do,
dq,
dg_cumsum,
g_cumsum,
scale,
cu_seql
fla/ops/attn/parallel.py:208