[Attention][DSA] Enable W4A16 DSA (#51724)

Signed-off-by: Shiyang Chen <shiychen@nvidia.com>
This commit is contained in:
sychen52
2026-09-01 12:37:47 +08:00
committed by GitHub
parent ce2e343be1
commit 7c5dc571cb
17 changed files with 824 additions and 22 deletions
+6
View File
@@ -996,6 +996,10 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C_cache_ops, ops) {
"Tensor block_table, Tensor workspace_starts, int batch_size, Tensor? "
"seq_starts) -> ()");
ops.def(
"cp_gather_and_upconvert_nvfp4_kv_cache(Tensor src_cache, Tensor! dst, "
"Tensor block_table, Tensor workspace_starts, int batch_size) -> ()");
ops.def(
"indexer_k_quant_and_cache(Tensor k, Tensor! kv_cache, Tensor "
"slot_mapping, "
@@ -1066,6 +1070,8 @@ STABLE_TORCH_LIBRARY_IMPL(_C_cache_ops, CUDA, ops) {
ops.impl("cp_gather_cache", TORCH_BOX(&cp_gather_cache));
ops.impl("cp_gather_and_upconvert_fp8_kv_cache",
TORCH_BOX(&cp_gather_and_upconvert_fp8_kv_cache));
ops.impl("cp_gather_and_upconvert_nvfp4_kv_cache",
TORCH_BOX(&cp_gather_and_upconvert_nvfp4_kv_cache));
ops.impl("indexer_k_quant_and_cache", TORCH_BOX(&indexer_k_quant_and_cache));
ops.impl("concat_mla_q", TORCH_BOX(&concat_mla_q));
ops.impl("cp_gather_indexer_k_quant_cache",