Functionfused_addcmul_fwd_kernel(
hidden,
delta,
ixr, ixw, ixk, ixv, ixa, ixg,
oxr, oxw, oxk, oxv, oxa, oxg,
use_xg: tl.co
fla/ops/rwkv7/fused_addcmul.py:55
Functionfused_beta_sigmoid_bwd_kernel(
x,
dy,
dx,
scale,
n_elements,
BLOCK_SIZE: tl.constexpr,
)
fla/ops/common/gate.py:34
Functionfused_chunk_based_fwd_kernel(
q,
k,
v,
o,
z,
scale, # K ** -0.5
T,
B: tl.constexpr,
H: tl.constexpr,
fla/ops/based/fused_chunk.py:16
Functionfused_chunk_bwd_kernel(
q,
k,
v,
g,
g_gamma,
do,
dq,
dk,
dv,
dg,
h0,
dht,
dh0,
fla/ops/common/fused_chunk.py:180
Functionfused_chunk_fwd_kernel(
q,
k,
v,
g,
g_gamma,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B: tl.
fla/ops/common/fused_chunk.py:46
Functionfused_chunk_ttt_linear_bwd_kernel_dh(
q,
k,
v,
v2,
x,
y,
r,
w,
b,
eta,
h,
dht,
dhbt,
dh0,
fla/ops/ttt/fused_chunk.py:277
Functionfused_chunk_ttt_linear_bwd_kernel_h(
k,
v,
v2,
x,
y,
r,
w,
b,
eta,
h0,
hb0,
h,
do,
dq,
fla/ops/ttt/fused_chunk.py:157
Functionfused_chunk_ttt_linear_fwd_kernel(
q,
k,
v,
eta,
w,
b,
o,
scale,
eps,
h0,
hb0,
ht,
hbt,
fla/ops/ttt/fused_chunk.py:34
Methodfused_grpo_loss(
self,
logits,
ref_logp,
input_ids,
advantages,
beta=0.1,
fla/modules/backends/triton_ascend/__init__.py:254
Methodfused_kl_div_forward(
self,
x,
target_x,
weight,
target_weight,
reduction='batchme
fla/modules/backends/triton_ascend/__init__.py:175
Methodfused_linear_cross_entropy_forward(
self,
x,
target,
weight,
bias=None,
ignore_index=-100,
fla/modules/backends/triton_ascend/__init__.py:92
Functionfused_recurrent_bwd_kernel(
q,
k,
v,
g,
g_gamma,
gk,
gv,
o,
h0,
do,
dq,
dk,
dv,
fla/ops/common/fused_recurrent.py:163
Functionfused_recurrent_comba_fwd_kernel(
q,
k,
p,
v,
g,
beta,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B:
fla/ops/comba/fused_recurrent.py:22
Functionfused_recurrent_delta_rule_bwd_kernel(
q,
k,
v,
beta,
h0,
dh0,
dht,
do,
dq,
dk,
dv,
db,
cu_seql
fla/ops/delta_rule/fused_recurrent.py:109
Functionfused_recurrent_delta_rule_fwd_kernel(
q,
k,
v,
u,
beta,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B: tl.con
fla/ops/delta_rule/fused_recurrent.py:22
Functionfused_recurrent_dplr_delta_rule_fwd_kernel(
q,
k,
v,
a,
b,
gk,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B: t
fla/ops/generalized_delta_rule/dplr/fused_recurrent.py:33
Functionfused_recurrent_fwd_kernel(
q, # query [B, H, L, K]
k, # key [B, H, L, V]
v, # value [B, H, L, V].
a, # a [B, H, L,
fla/ops/generalized_delta_rule/iplr/fused_recurrent.py:31
Functionfused_recurrent_fwd_kernel(
q,
k,
v,
g,
g_gamma,
gk,
gv,
o,
h0,
ht,
cu_seqlens,
scale,
fla/ops/common/fused_recurrent.py:30
Functionfused_recurrent_gated_delta_rule_fwd_kernel(
q,
k,
v,
g,
gk,
gv,
beta,
A_log,
dt_bias,
o,
h0,
ht,
cu_
fla/ops/gated_delta_rule/fused_recurrent.py:30
Functionfused_recurrent_gdn2_fwd_kernel(
q,
k,
v,
g,
b,
w,
A_log,
dt_bias,
o,
h0,
ht,
cu_seqlens,
fla/ops/gdn2/fused_recurrent.py:60
Functionfused_recurrent_gsa_inference_kernel(
q,
k,
v,
s,
g,
o,
hk0,
hv0,
hkt,
hvt,
scale,
K: tl.constexpr
fla/ops/gsa/fused_recurrent.py:18
Functionfused_recurrent_hgrn_bwd_kernel(
g,
o,
h0,
dx,
dg,
do,
dht,
dh0,
cu_seqlens,
T,
D: tl.constexpr,
fla/ops/hgrn/fused_recurrent.py:93
Functionfused_recurrent_hgrn_fwd_kernel(
x,
g,
o,
h0,
ht,
cu_seqlens,
T,
D: tl.constexpr,
BD: tl.constexpr,
U
fla/ops/hgrn/fused_recurrent.py:31
Functionfused_recurrent_kda_fwd_kernel(
q,
k,
v,
g,
beta,
A_log,
dt_bias,
o,
h0,
ht,
cu_seqlens,
ssm
fla/ops/kda/fused_recurrent.py:33
Functionfused_recurrent_lightning_attnr""" Args: q (torch.Tensor): queries of shape `[B, T, H, K]`. k (torch.Tensor): keys of shape `[B, T, H, K
fla/ops/lightning_attn/fused_recurrent.py:13
Functionfused_recurrent_oja_fwd_kernel(
q,
k,
v,
gv,
beta,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B: tl.co
fla/ops/gated_oja_rule/fused_recurrent.py:23
Functionfused_recurrent_rwkv4_backward_kernel(
# W
w_ptr,
w_s_c,
# U
u_ptr,
u_s_c,
# K
k_ptr,
k_s_b,
k_s_t,
k_s
fla/ops/rwkv4/fused_recurrent.py:180
Functionfused_recurrent_rwkv4_forward_kernel(
# W
w_ptr,
w_s_c,
# U
u_ptr,
u_s_c,
# K
k_ptr,
k_s_b,
k_s_t,
k_s
fla/ops/rwkv4/fused_recurrent.py:29
Functionfused_recurrent_rwkv6_bwd_kernel_dkv(
q, # query [B, H, T, K]/[B, T, H, K]
k, # key [B, H, T, V]/[B, T, H, V]
v, # value [B, H, T,
fla/ops/rwkv6/fused_recurrent.py:207
Functionfused_recurrent_rwkv6_bwd_kernel_dq(
k, # key [B, H, T, V]/[B, T, H, V]
v, # value [B, H, T, V]/[B, T, H, V]
w, # log gate [B, H,
fla/ops/rwkv6/fused_recurrent.py:117
Functionfused_recurrent_rwkv6_bwd_kernel_dw(
q,
k,
dq,
dk,
dw,
cu_seqlens,
scale,
T,
H: tl.constexpr,
K: tl.const
fla/ops/rwkv6/fused_recurrent.py:305
Functionfused_recurrent_rwkv6_fwd_kernel(
q, # query [B, H, T, K]/[B, T, H, K]
k, # key [B, H, T, K]/[B, T, H, K]
v, # value [B, H, T,
fla/ops/rwkv6/fused_recurrent.py:30
Functionfused_recurrent_rwkv7 Args: r (torch.Tensor): r of shape `[B, T, H, K]`. w (torch.Tensor): log decay of shape `[B, T, H, K]`.
fla/ops/rwkv7/fused_recurrent.py:183
Functionfused_recurrent_rwkv7_fwd_kernel(
r,
w,
k,
v,
kk,
a,
o,
h0,
ht,
cu_seqlens,
scale,
T,
B: t
fla/ops/rwkv7/fused_recurrent.py:34
Methodfwd_inner(q, k, v, z, B, H, T, K, V, BT, BK, BV, NT, normk=False, h0=None, ht=None)
fla/ops/abc/chunk.py:860
Functiongate_output_correction_bwd_kernel(
grad_output_ptr, o_ptr, r_ptr, k_ptr, r_k_ptr, v_ptr, g_ptr,
grad_o_ptr, grad_r_ptr, grad_k_ptr, gra
fla/ops/rwkv7/gate_output_correction.py:126
Functiongate_output_correction_fwd_kernel(
o_ptr, r_ptr, k_ptr, r_k_ptr, v_ptr, g_ptr, output_ptr,
o_b_stride, o_t_stride,
r_b_stride, r_t_
fla/ops/rwkv7/gate_output_correction.py:74
Functiongdn_gate_bwd_kernel(
g,
A_log,
dt_bias,
dyg,
dg,
dA,
T,
H: tl.constexpr,
BT: tl.constexpr,
fla/ops/gated_delta_rule/gate.py:118
Functiongdn_gate_chunk_cumsum_scalar_kernel(
g,
A_log,
dt_bias,
o,
scale,
cu_seqlens,
chunk_indices,
T,
H: tl.constex
fla/ops/gated_delta_rule/gate.py:62
Functiongdn_gate_fwd_kernel(
g,
A_log,
dt_bias,
yg,
T,
H: tl.constexpr,
BT: tl.constexpr,
HAS_BIAS: tl.co
fla/ops/gated_delta_rule/gate.py:237