ViperEkura/AstrAI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: add out_buf to attn_paged_decode for CUDA graph capture compatibility

Browse Source 
- Pre-allocate decode_out in InferenceWorkspace so attn_paged_decode does not call torch::empty inside graph capture
- Wire decode_out through KVCache, PagePool.bind_tasks, and CudaBackend.fwd_decode
- Run live forward before graph capture to get valid output (graph pool memory is zeroed after capture block exits)
- Greedy generation with graph replay is bit-exact across all batch sizes
- Decode speedups vs no-graph: B=1 2.09x, B=4 1.80x, B=8 1.94x, B=16 1.76x
This commit is contained in:
ViperEkura
2026-08-07 19:45:59 +08:00
parent 6572be4f98
commit af25833fab
7 changed files with 132 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/kernels/attn_paged_decode.cu
+15 -2
View File
@@ -13,7 +13,8 @@ torch::Tensor attn_paged_decode(
int64_t causal_offset,
double scale,
c10::optional<torch::Tensor> o_part_buf,
c10::optional<torch::Tensor> ml_part_buf
c10::optional<torch::Tensor> ml_part_buf,
c10::optional<torch::Tensor> out_buf
) {
const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
auto stream = at::cuda::getCurrentCUDAStream();
@@ -23,7 +24,18 @@ torch::Tensor attn_paged_decode(
req_to_token, req_pool_indices, kv_indptr,
max_seq_len, mask, causal_offset, scale, p);
auto O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
torch::Tensor O;
if (out_buf.has_value() && out_buf->defined()) {
TORCH_CHECK(out_buf->dtype() == q.dtype(), "out_buf dtype must match q");
TORCH_CHECK(out_buf->size(0) >= q.size(0), "out_buf batch too small");
TORCH_CHECK(out_buf->size(1) >= q.size(1), "out_buf heads too small");
TORCH_CHECK(out_buf->size(2) >= q.size(2), "out_buf head_dim too small");
O = out_buf.value().slice(0, 0, q.size(0))
.slice(1, 0, q.size(1))
.slice(2, 0, q.size(2));
} else {
O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
}
p.o = (bf16*)O.data_ptr();
if (o_part_buf.has_value() && ml_part_buf.has_value()
@@ -57,5 +69,6 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
py::arg("scale") = 0.0,
py::arg("o_part_buf") = py::none(),
py::arg("ml_part_buf") = py::none(),
py::arg("out_buf") = py::none(),
"SGLang-style paged decode: flat KV pool + req_to_token + kv_indptr.");
}