[CPU][Spec Decode] Enable DFlash SD for CPU (#44029)

Signed-off-by: guybd <guy.boudoukh@intel.com>
Signed-off-by: Guy Boudoukh <guy.boudoukh@intel.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
guybd
2026-06-25 15:32:48 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 9b215ae60b
commit 2396d91e93
8 changed files with 266 additions and 8 deletions
+23
View File
@@ -237,6 +237,16 @@ void copy_and_expand_eagle_inputs_kernel_impl(
const int64_t padding_token_id, const int64_t parallel_drafting_token_id,
const int64_t total_input_tokens,
const int64_t num_padding_slots_per_request, const bool shift_input_ids);
void copy_and_expand_dflash_inputs_kernel_impl(
const torch::Tensor& next_token_ids, const torch::Tensor& target_positions,
torch::Tensor& out_input_ids, torch::Tensor& out_context_positions,
torch::Tensor& out_query_positions, torch::Tensor& out_context_slot_mapping,
torch::Tensor& out_query_slot_mapping, torch::Tensor& out_token_indices,
const torch::Tensor& block_table, const torch::Tensor& query_start_loc,
const std::optional<torch::Tensor>& num_rejected_tokens,
const int64_t parallel_drafting_token_id, const int64_t block_size,
const int64_t num_query_per_req, const int64_t num_speculative_tokens,
const int64_t total_input_tokens, const bool has_num_rejected);
void rejection_greedy_sample_kernel_impl(
torch::Tensor& output_token_ids, const torch::Tensor& cu_num_draft_tokens,
const torch::Tensor& draft_token_ids, const torch::Tensor& target_argmax,
@@ -599,6 +609,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
"SymInt total_input_tokens, SymInt num_padding_slots_per_request, "
"bool shift_input_ids) -> ()",
&cpu_utils::copy_and_expand_eagle_inputs_kernel_impl);
ops.def(
"copy_and_expand_dflash_inputs_kernel_impl("
"Tensor next_token_ids, Tensor target_positions, "
"Tensor(a2!) out_input_ids, Tensor(a3!) out_context_positions, "
"Tensor(a4!) out_query_positions, "
"Tensor(a5!) out_context_slot_mapping, "
"Tensor(a6!) out_query_slot_mapping, "
"Tensor(a7!) out_token_indices, Tensor block_table, "
"Tensor query_start_loc, Tensor? num_rejected_tokens, "
"SymInt parallel_drafting_token_id, SymInt block_size, "
"SymInt num_query_per_req, SymInt num_speculative_tokens, "
"SymInt total_input_tokens, bool has_num_rejected) -> ()",
&cpu_utils::copy_and_expand_dflash_inputs_kernel_impl);
ops.def(
"rejection_greedy_sample_kernel_impl("
"Tensor(a0!) output_token_ids, Tensor cu_num_draft_tokens, "