Skip to content

[Feature] Support multimodal RoPE 3D -> 1D in append_attention #7932

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
xiaoxiaohehe001 wants to merge 4 commits into
base: release/online/20260415
Choose a base branch
from
Open

[Feature] Support multimodal RoPE 3D -> 1D in append_attention #7932

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d427953
support mm rope3d -> 1d
xiaoxiaohehe001 May 26, 2026
bcaa76c
fix error
xiaoxiaohehe001 May 29, 2026
12b25e8
fix ci
xiaoxiaohehe001 May 29, 2026
5c642c6
fix ci
xiaoxiaohehe001 May 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions custom_ops/gpu_ops/append_attention.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,6 +72,7 @@ void AppendAttentionKernel(
const paddle::optional<paddle::Tensor>& out_linear_shifts,
const paddle::optional<paddle::Tensor>& out_linear_smooths,
const paddle::optional<paddle::Tensor>& kv_signal_data,
const paddle::optional<paddle::Tensor>& rope_3d_delta,
const paddle::optional<paddle::Tensor>& q_norm_weight,
const paddle::optional<paddle::Tensor>& k_norm_weight,
const paddle::optional<paddle::Tensor>& sinks,
Expand Down Expand Up @@ -213,6 +214,7 @@ void AppendAttentionKernel(
max_input_length,
use_neox_rotary_style,
rope_3d,
rope_3d_delta,
main_stream,
&qkv_out,
const_cast<paddle::Tensor*>(&key_cache),
Expand Down Expand Up @@ -310,6 +312,7 @@ void AppendAttentionKernel(
cache_quant_type_str,
use_neox_rotary_style,
rope_3d,
rope_3d_delta,
max_input_length,
exec_stream,
&qkv_out,
Expand Down Expand Up @@ -337,6 +340,7 @@ void AppendAttentionKernel(
cache_quant_type_str,
use_neox_rotary_style,
rope_3d,
rope_3d_delta,
max_input_length,
exec_stream,
&qkv_out,
Expand Down Expand Up @@ -365,6 +369,7 @@ void AppendAttentionKernel(
cache_quant_type_str,
use_neox_rotary_style,
rope_3d,
rope_3d_delta,
max_input_length,
exec_stream,
&qkv_out,
Expand All @@ -391,6 +396,7 @@ void AppendAttentionKernel(
cache_quant_type_str,
use_neox_rotary_style,
rope_3d,
rope_3d_delta,
max_input_length,
exec_stream,
&qkv_out,
Expand Down Expand Up @@ -485,6 +491,7 @@ std::vector<paddle::Tensor> AppendAttention(
const paddle::optional<paddle::Tensor>& out_linear_smooths,
const paddle::optional<paddle::Tensor>& mask_offset,
const paddle::optional<paddle::Tensor>& kv_signal_data,
const paddle::optional<paddle::Tensor>& rope_3d_delta,
const paddle::optional<paddle::Tensor>& q_norm_weight,
const paddle::optional<paddle::Tensor>& k_norm_weight,
const paddle::optional<paddle::Tensor>& sinks,
Expand Down Expand Up @@ -619,6 +626,7 @@ std::vector<paddle::Tensor> AppendAttention(
out_linear_shifts,
out_linear_smooths,
kv_signal_data,
rope_3d_delta,
q_norm_weight,
k_norm_weight,
sinks,
Expand Down Expand Up @@ -697,6 +705,7 @@ std::vector<paddle::Tensor> AppendAttentionWithOutput(
const paddle::optional<paddle::Tensor>& out_linear_smooths,
const paddle::optional<paddle::Tensor>& mask_offset,
const paddle::optional<paddle::Tensor>& kv_signal_data,
const paddle::optional<paddle::Tensor>& rope_3d_delta,
const paddle::optional<paddle::Tensor>& q_norm_weight,
const paddle::optional<paddle::Tensor>& k_norm_weight,
const paddle::optional<paddle::Tensor>& sinks,
Expand Down Expand Up @@ -777,6 +786,7 @@ std::vector<paddle::Tensor> AppendAttentionWithOutput(
out_linear_shifts,
out_linear_smooths,
kv_signal_data,
rope_3d_delta,
q_norm_weight,
k_norm_weight,
sinks,
Expand Down Expand Up @@ -868,6 +878,7 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
const paddle::optional<std::vector<int64_t>>& rope_3d_delta_shape,
const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
const paddle::optional<std::vector<int64_t>>& sinks_shape,
Expand Down Expand Up @@ -934,6 +945,7 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
const paddle::optional<paddle::DataType>& mask_offset_dtype,
const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
const paddle::optional<paddle::DataType>& rope_3d_delta_dtype,
const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
const paddle::optional<paddle::DataType>& sinks_dtype,
Expand Down Expand Up @@ -1021,6 +1033,7 @@ std::vector<std::vector<int64_t>> AppendAttentionWithOutputInferShape(
const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
const paddle::optional<std::vector<int64_t>>& rope_3d_delta_shape,
const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
const paddle::optional<std::vector<int64_t>>& sinks_shape,
Expand Down Expand Up @@ -1080,6 +1093,7 @@ std::vector<paddle::DataType> AppendAttentionWithOutputInferDtype(
const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
const paddle::optional<paddle::DataType>& mask_offset_dtype,
const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
const paddle::optional<paddle::DataType>& rope_3d_delta_dtype,
const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
const paddle::optional<paddle::DataType>& sinks_dtype,
Expand Down Expand Up @@ -1138,6 +1152,7 @@ PD_BUILD_STATIC_OP(append_attention)
paddle::Optional("out_linear_smooths"),
paddle::Optional("mask_offset"),
paddle::Optional("kv_signal_data"),
paddle::Optional("rope_3d_delta"),
paddle::Optional("q_norm_weight"),
paddle::Optional("k_norm_weight"),
paddle::Optional("sinks")})
Expand Down Expand Up @@ -1201,6 +1216,7 @@ PD_BUILD_STATIC_OP(append_attention_with_output)
paddle::Optional("out_linear_smooths"),
paddle::Optional("mask_offset"),
paddle::Optional("kv_signal_data"),
paddle::Optional("rope_3d_delta"),
paddle::Optional("q_norm_weight"),
paddle::Optional("k_norm_weight"),
paddle::Optional("sinks")})
Expand Down
71 changes: 58 additions & 13 deletions custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_impl.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,7 @@ __global__ void append_decode_cache_T_rope_qk_norm_kernel(
const uint32_t elem_cnt,
const int kv_num_heads,
const bool rope_3d,
const int* rope_3d_delta,
const float* q_norm_weight,
const float* k_norm_weight,
const float rms_norm_eps) {
Expand Down Expand Up @@ -143,8 +144,15 @@ __global__ void append_decode_cache_T_rope_qk_norm_kernel(
if (hi < num_heads + kv_num_heads) {
// q k rope
const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
uint32_t new_emb_idx =
rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
uint32_t new_emb_idx = emb_idx;
if (rope_3d) {
if (rope_3d_delta) {
const int rope_pos = write_seq_id + rope_3d_delta[ori_bi];
new_emb_idx = rope_pos * half_head_size + h_bias / 2;
} else {
new_emb_idx = emb_idx + ori_bi * max_seq_len * head_size;
}
}
Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
}
Expand Down Expand Up @@ -237,7 +245,8 @@ __global__ void append_decode_cache_T_rope_kernel(
const int block_size,
const uint32_t elem_cnt,
const int kv_num_heads,
const bool rope_3d) {
const bool rope_3d,
const int* rope_3d_delta) {
using LoadT = AlignedVector<T, VecSize>;
using LoadBiasT = AlignedVector<T, VecSize>;
using LoadKVT = AlignedVector<T, VecSize>;
Expand Down Expand Up @@ -282,8 +291,15 @@ __global__ void append_decode_cache_T_rope_kernel(
if (hi < num_heads + kv_num_heads) {
// q k rope
const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
uint32_t new_emb_idx =
rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
uint32_t new_emb_idx = emb_idx;
if (rope_3d) {
if (rope_3d_delta) {
const int rope_pos = write_seq_id + rope_3d_delta[ori_bi];
new_emb_idx = rope_pos * half_head_size + h_bias / 2;
} else {
new_emb_idx = emb_idx + ori_bi * max_seq_len * head_size;
}
}
Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
}
Expand Down Expand Up @@ -1221,6 +1237,7 @@ __global__ void append_decode_cache_int8_rope_qk_norm_kernel(
const float min_bound,
const int kv_num_heads,
const bool rope_3d,
const int* rope_3d_delta,
const float rms_norm_eps) {
static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
Expand Down Expand Up @@ -1268,8 +1285,15 @@ __global__ void append_decode_cache_int8_rope_qk_norm_kernel(
Load<T, VecSize>(&qkv_now[bias_idx], &src_vec);
// q rope
const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
const uint32_t new_emb_idx =
rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
uint32_t new_emb_idx = emb_idx;
if (rope_3d) {
if (rope_3d_delta) {
const int rope_pos = write_seq_id + rope_3d_delta[bid];
new_emb_idx = rope_pos * half_head_size + head_bias / 2;
} else {
new_emb_idx = emb_idx + bid * max_seq_len * HeadDim;
}
}
Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
#pragma unroll
Expand Down Expand Up @@ -1363,8 +1387,15 @@ __global__ void append_decode_cache_int8_rope_qk_norm_kernel(
const int v_head_idx = head_idx - num_heads - kv_num_heads;
if (head_idx < num_heads + kv_num_heads) {
const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
const uint32_t new_emb_idx =
rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
uint32_t new_emb_idx = emb_idx;
if (rope_3d) {
if (rope_3d_delta) {
const int rope_pos = write_seq_id + rope_3d_delta[bid];
new_emb_idx = rope_pos * half_head_size + head_bias / 2;
} else {
new_emb_idx = emb_idx + bid * max_seq_len * HeadDim;
}
}
Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
Expand Down Expand Up @@ -1533,7 +1564,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
const float max_bound,
const float min_bound,
const int kv_num_heads,
const bool rope_3d) {
const bool rope_3d,
const int* rope_3d_delta) {
static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
constexpr int NUM_WARPS = 4;
Expand Down Expand Up @@ -1564,7 +1596,13 @@ __global__ void append_decode_cache_int8_rope_kernel(
qkv_out + start_token_idx * hidden_size + head_idx * HeadDim;

uint32_t emb_offset = write_seq_id * half_head_size;
emb_offset += rope_3d ? bid * max_seq_len * HeadDim : 0;
if (rope_3d) {
if (rope_3d_delta) {
emb_offset = (write_seq_id + rope_3d_delta[bid]) * half_head_size;
} else {
emb_offset += bid * max_seq_len * HeadDim;
}
}
apply_rope<T, VecSize, HeadDim, 32, EnforceFmulRN>(qkv_now,
cos_emb + emb_offset,
sin_emb + emb_offset,
Expand Down Expand Up @@ -1634,8 +1672,15 @@ __global__ void append_decode_cache_int8_rope_kernel(
cache_v_scale + v_head_idx * HeadDim + head_bias;
if (head_idx < num_heads + kv_num_heads) {
const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
uint32_t new_emb_idx =
rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
uint32_t new_emb_idx = emb_idx;
if (rope_3d) {
if (rope_3d_delta) {
const int rope_pos = write_seq_id + rope_3d_delta[bid];
new_emb_idx = rope_pos * half_head_size + head_bias / 2;
} else {
new_emb_idx = emb_idx + bid * max_seq_len * HeadDim;
}
}
Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
Expand Down
Loading
Loading