[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
+9
View File
@@ -672,6 +672,15 @@ void cp_gather_and_upconvert_fp8_kv_cache(
int64_t batch_size,
std::optional<torch::stable::Tensor> seq_starts = std::nullopt);
// Gather and upconvert an nvfp4_ds_mla KV cache to a BF16 workspace
void cp_gather_and_upconvert_nvfp4_kv_cache(
torch::stable::Tensor const& src_cache, // [NUM_BLOCKS, BLOCK_SIZE,
// 352]
torch::stable::Tensor const& dst, // [TOT_TOKENS, 576]
torch::stable::Tensor const& block_table, // [BATCH, BLOCK_INDICES]
torch::stable::Tensor const& workspace_starts, // [BATCH]
int64_t batch_size);
// Indexer K quantization and cache function
void indexer_k_quant_and_cache(
torch::stable::Tensor& k, // [num_tokens, head_dim]