[V1] [Hybrid] Mamba1 Automatic Prefix Caching (#26377)

Signed-off-by: asafg <39553475+Josephasafg@users.noreply.github.com>
This commit is contained in:
Asaf Joseph Gardin
2025-11-02 04:16:23 -08:00
committed by GitHub
parent 73444b7b56
commit 00b31a36a2
16 changed files with 442 additions and 153 deletions
+5 -1
View File
@@ -611,7 +611,11 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
"Tensor? cache_indices,"
"Tensor? has_initial_state,"
"Tensor! ssm_states,"
"int pad_slot_id) -> ()");
"int pad_slot_id,"
"int block_size,"
"Tensor? block_idx_first_scheduled_token,"
"Tensor? block_idx_last_scheduled_token,"
"Tensor? initial_state_idx) -> ()");
ops.impl("selective_scan_fwd", torch::kCUDA, &selective_scan_fwd);
// Hadamard transforms