Skip to content

【Hackathon 10th Spring No.53】[Feature][Kernel] Optimize AppendAttention for discrete head-wise block_idx [cf] #7718

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
bob-cloudforge wants to merge 11 commits into
base: develop
Choose a base branch
from
Open

【Hackathon 10th Spring No.53】[Feature][Kernel] Optimize AppendAttention for discrete head-wise block_idx [cf] #7718

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
41debee
feat(append_attn): head-wise SWA recycle + discrete-block-idx ABI
bob-cloudforge May 4, 2026
8f421af
fix(t53/pr2): close CUDA-graph capture, swap_states reorder, doc gap
bob-cloudforge May 4, 2026
eeee930
fix(t53-pr2): read kv_num_heads_local from model_config (warmup-safe)
bob-cloudforge May 5, 2026
c8890aa
fix(t53-pr2): hotfix block_tables_headwise OOB at backend boundary
bob-cloudforge May 5, 2026
3c75f27
fix(t53-pr2): remove CUDA-graph-incompatible OOB assert from forward_…
bob-cloudforge May 5, 2026
0a5cb49
fix(t53/pr2): per-head independent block heaps + revert HOTFIX
bob-cloudforge May 5, 2026
4e57aab
bench(t53/pr2/v4): kv_cache_ratio 0.35→0.30 — TINY envelope diagnosti…
bob-cloudforge May 5, 2026
9f2a7ff
fix(t53/pr2): PD_CHECK INT32 dtype on Python-supplied tensors before …
bob-cloudforge May 6, 2026
e1b27e2
fix(t53/pr2): assert GQA divisibility before kv_num_heads_global // t…
bob-cloudforge May 6, 2026
327a43b
fix(kvcache): use float division in available_gpu_resource for partia…
bob-cloudforge May 6, 2026
3a592ac
fix(scheduler): document SWA recycle block-boundary guard
bob-cloudforge May 6, 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
3 changes: 3 additions & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,3 +185,6 @@ custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm_template.h

custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_*.cu
custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_template.h

# Bench output artifacts (T53)
runs/
12 changes: 12 additions & 0 deletions benchmarks/yaml/eb45-21b-a3b-32k-bf16-kv50-512s.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
# T53 bench workload — KV-bound (not slot-bound); gate: FD_HEAD_WISE_KV_CACHE=1
# max_num_seqs raised to 256 so the KV pool, not the slot count, is the bottleneck.
# kv_cache_ratio: 0.30 → ~24GB KV on A800-80GB (TINY envelope diagnostic per opus v2 verdict).
# (0.35 deterministic OOM at 78.99GB / index 3408/3689 weights load — 4 identical failures.
# 0.50 also OOMs. Revert to 0.35 before SMOKE/FULL only after opus comparability decision.)
# Use with: INPUT_LEN=8192 OUTPUT_LEN=4096 REQUEST_RATE=8
#
max_model_len: 32768
max_num_seqs: 256
kv_cache_ratio: 0.30
tensor_parallel_size: 1
max_num_batched_tokens: 32768
35 changes: 33 additions & 2 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 @@ -48,6 +48,7 @@ void AppendAttentionKernel(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_tables,
const paddle::optional<paddle::Tensor>& block_tables_headwise,
const paddle::Tensor& encoder_batch_ids,
const paddle::Tensor& encoder_tile_ids_per_batch,
const paddle::Tensor& encoder_num_blocks,
Expand Down Expand Up @@ -96,6 +97,17 @@ void AppendAttentionKernel(
typedef typename traits_::DataType DataType_;
typedef typename traits_::data_t data_t;

// Dtype guards for Python-supplied INT32 metadata tensors accessed via
// .data<int>() below. Catches accidental INT64/FP dtype before UB.
PD_CHECK(set_max_lengths.dtype() == paddle::DataType::INT32,
"set_max_lengths must be INT32");
PD_CHECK(encoder_num_blocks.dtype() == paddle::DataType::INT32,
"encoder_num_blocks must be INT32");
PD_CHECK(kv_num_blocks.dtype() == paddle::DataType::INT32,
"kv_num_blocks must be INT32");
PD_CHECK(decoder_num_blocks.dtype() == paddle::DataType::INT32,
"decoder_num_blocks must be INT32");

const int max_len_this_time = set_max_lengths.data<int>()[0];
const int max_enc_len_this_time = set_max_lengths.data<int>()[1];
const int max_dec_len_this_time = set_max_lengths.data<int>()[2];
Expand Down Expand Up @@ -155,6 +167,7 @@ void AppendAttentionKernel(
batch_id_per_token,
cu_seqlens_q,
block_tables,
block_tables_headwise,
lambda_batch_ids,
lambda_tile_ids_per_batch,
cache_quant_type_str,
Expand Down Expand Up @@ -488,6 +501,9 @@ std::vector<paddle::Tensor> AppendAttention(
const paddle::optional<paddle::Tensor>& q_norm_weight,
const paddle::optional<paddle::Tensor>& k_norm_weight,
const paddle::optional<paddle::Tensor>& sinks,
const paddle::optional<paddle::Tensor>&
block_tables_headwise, // logical 3D, physical rank-2 [max_num_seqs *
// local_kv_heads, max_blocks_per_head]
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -580,6 +596,8 @@ std::vector<paddle::Tensor> AppendAttention(
}

if (mask_offset) {
PD_CHECK(mask_offset.get().dtype() == paddle::DataType::INT32,
"mask_offset must be INT32");
meta_data.mask_offset = mask_offset.get().data<int>();
}

Expand All @@ -595,6 +613,7 @@ std::vector<paddle::Tensor> AppendAttention(
batch_id_per_token,
cu_seqlens_q,
block_tables,
block_tables_headwise,
encoder_batch_ids,
encoder_tile_ids_per_batch,
encoder_num_blocks,
Expand Down Expand Up @@ -700,6 +719,9 @@ std::vector<paddle::Tensor> AppendAttentionWithOutput(
const paddle::optional<paddle::Tensor>& q_norm_weight,
const paddle::optional<paddle::Tensor>& k_norm_weight,
const paddle::optional<paddle::Tensor>& sinks,
const paddle::optional<paddle::Tensor>&
block_tables_headwise, // logical 3D, physical rank-2 [max_num_seqs *
// local_kv_heads, max_blocks_per_head]
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -738,6 +760,8 @@ std::vector<paddle::Tensor> AppendAttentionWithOutput(
meta_data.batch_size = seq_lens_this_time.dims()[0];

if (mask_offset) {
PD_CHECK(mask_offset.get().dtype() == paddle::DataType::INT32,
"mask_offset must be INT32");
meta_data.mask_offset = mask_offset.get().data<int>();
}

Expand All @@ -753,6 +777,7 @@ std::vector<paddle::Tensor> AppendAttentionWithOutput(
batch_id_per_token,
cu_seqlens_q,
block_tables,
block_tables_headwise,
encoder_batch_ids,
encoder_tile_ids_per_batch,
encoder_num_blocks,
Expand Down Expand Up @@ -871,6 +896,7 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
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,
const paddle::optional<std::vector<int64_t>>& block_tables_headwise_shape,
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -937,6 +963,7 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
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,
const paddle::optional<paddle::DataType>& block_tables_headwise_dtype,
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -1024,6 +1051,7 @@ std::vector<std::vector<int64_t>> AppendAttentionWithOutputInferShape(
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,
const paddle::optional<std::vector<int64_t>>& block_tables_headwise_shape,
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -1083,6 +1111,7 @@ std::vector<paddle::DataType> AppendAttentionWithOutputInferDtype(
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,
const paddle::optional<paddle::DataType>& block_tables_headwise_dtype,
const float rms_norm_eps,
const std::string& compute_dtype,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -1140,7 +1169,8 @@ PD_BUILD_STATIC_OP(append_attention)
paddle::Optional("kv_signal_data"),
paddle::Optional("q_norm_weight"),
paddle::Optional("k_norm_weight"),
paddle::Optional("sinks")})
paddle::Optional("sinks"),
paddle::Optional("block_tables_headwise")})
.Outputs({"fmha_out"})
.Attrs({
"rms_norm_eps: float",
Expand Down Expand Up @@ -1203,7 +1233,8 @@ PD_BUILD_STATIC_OP(append_attention_with_output)
paddle::Optional("kv_signal_data"),
paddle::Optional("q_norm_weight"),
paddle::Optional("k_norm_weight"),
paddle::Optional("sinks")})
paddle::Optional("sinks"),
paddle::Optional("block_tables_headwise")})
.Outputs({"fmha_out_out"})
.SetInplaceMap({{"fmha_out", "fmha_out_out"}})
.Attrs({
Expand Down
8 changes: 8 additions & 0 deletions custom_ops/gpu_ops/append_attn/append_attention_c16_impl.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ void CascadeAppendAttentionC16Kernel(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -109,6 +110,7 @@ void CascadeAppendAttentionC16Kernel(
batch_id_per_token,
cu_seqlens_q,
block_table,
block_table_headwise,
batch_ids,
tile_ids_per_batch,
num_blocks,
Expand Down Expand Up @@ -156,6 +158,7 @@ CascadeAppendAttentionC16Kernel<paddle::bfloat16, paddle::bfloat16>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -204,6 +207,7 @@ CascadeAppendAttentionC16Kernel<paddle::bfloat16, paddle::float8_e4m3fn>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -251,6 +255,7 @@ template void CascadeAppendAttentionC16Kernel<paddle::bfloat16, int8_t>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -298,6 +303,7 @@ template void CascadeAppendAttentionC16Kernel<paddle::float16, paddle::float16>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -346,6 +352,7 @@ CascadeAppendAttentionC16Kernel<paddle::float16, paddle::float8_e4m3fn>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down Expand Up @@ -393,6 +400,7 @@ template void CascadeAppendAttentionC16Kernel<paddle::float16, int8_t>(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
Expand Down
2 changes: 2 additions & 0 deletions custom_ops/gpu_ops/append_attn/append_attention_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@ void CascadeAppendAttentionKernel(
const paddle::Tensor& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::optional<paddle::Tensor>& block_table_headwise,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const std::string& cache_quant_type_str,
Expand Down Expand Up @@ -86,6 +87,7 @@ void CascadeAppendAttentionKernel(
batch_id_per_token,
cu_seqlens_q,
block_table,
block_table_headwise,
batch_ids,
tile_ids_per_batch,
num_blocks,
Expand Down
Loading
Loading