[Kernel] Triton implementation of causal-conv1d for Mamba-based models (#18218)

Signed-off-by: Tuan M. Hoang-Trong <tmhoangt@us.ibm.com>
Co-authored-by: Tuan M. Hoang-Trong <tmhoangt@us.ibm.com>
Co-authored-by: Tyler Michael Smith <tysmith@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tuan, Hoang-Trong
2025-07-09 12:53:55 -07:00
committed by GitHub
co-authored by Tuan M. Hoang-Trong Tyler Michael Smith Tyler Michael Smith
parent 31b96d1c64
commit 47043eb678
15 changed files with 1117 additions and 1142 deletions
-16
View File
@@ -326,22 +326,6 @@ void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
const std::optional<torch::Tensor>& has_initial_state,
const torch::Tensor& ssm_states, int64_t pad_slot_id);
void causal_conv1d_update(const at::Tensor& x, const at::Tensor& conv_state,
const at::Tensor& weight,
const std::optional<at::Tensor>& bias_,
bool silu_activation,
const std::optional<at::Tensor>& cache_seqlens_,
const std::optional<at::Tensor>& conv_state_indices_,
int64_t pad_slot_id);
void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
const std::optional<at::Tensor>& bias_,
const std::optional<at::Tensor>& conv_states,
const std::optional<at::Tensor>& query_start_loc,
const std::optional<at::Tensor>& cache_indices,
const std::optional<at::Tensor>& has_initial_state,
bool silu_activation, int64_t pad_slot_id);
using fptr_t = int64_t;
fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
torch::Tensor& rank_data, int64_t rank,