[DSv4] Adding TRTLLM gen attention kernel (#43827)

This commit is contained in:
Yongye Zhu
2026-06-04 07:35:09 -07:00
committed by GitHub
parent 3e77036768
commit b5235fca2e
20 changed files with 2659 additions and 86 deletions
+17
View File
@@ -238,6 +238,23 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
torch::stable::Tensor const& cos_sin_cache, int64_t q_head_padded,
double eps, int64_t cache_block_size);
void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert(
torch::stable::Tensor& q, torch::stable::Tensor const& kv,
torch::stable::Tensor& k_cache, torch::stable::Tensor const& slot_mapping,
torch::stable::Tensor const& position_ids,
torch::stable::Tensor const& cos_sin_cache, double eps,
int64_t cache_block_size);
void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
torch::stable::Tensor const& q, torch::stable::Tensor const& kv,
torch::stable::Tensor& q_fp8, torch::stable::Tensor& k_cache,
torch::stable::Tensor const& slot_mapping,
torch::stable::Tensor const& position_ids,
torch::stable::Tensor const& cos_sin_cache,
torch::stable::Tensor const& fp8_scale,
torch::stable::Tensor const& q_fp8_scale_inv, double eps,
int64_t cache_block_size);
#ifndef USE_ROCM
torch::stable::Tensor minimax_allreduce_rms(
torch::stable::Tensor const& input,