[8/n] Migrate merge_attn_states, mamba, sampler to torch stable ABI (continued) (#43361)

Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Signed-off-by: Chris Leonard <chleonar@redhat.com>
Co-authored-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
Chris Leonard
2026-05-27 09:35:24 -07:00
committed by GitHub
co-authored by Mikayla Gawarecki Shengqi Chen
parent 05c50c721e
commit 284e6f543d
13 changed files with 432 additions and 403 deletions
-43
View File
@@ -54,13 +54,6 @@ void paged_attention_v2(
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void merge_attn_states(
torch::Tensor& output, std::optional<torch::Tensor> output_lse,
const torch::Tensor& prefix_output, const torch::Tensor& prefix_lse,
const torch::Tensor& suffix_output, const torch::Tensor& suffix_lse,
const std::optional<int64_t> prefill_tokens_with_context,
const std::optional<torch::Tensor>& output_scale = std::nullopt);
// rms_norm and fused_add_rms_norm declarations also exist in
// csrc/libtorch_stable/ops.h (torch::stable ABI for CUDA). They remain here
// because the CPU build still uses these torch::Tensor declarations.
@@ -76,26 +69,6 @@ torch::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
torch::Tensor const& cos_sin_cache, int64_t q_head_padded, double eps,
int64_t cache_block_size);
void apply_repetition_penalties_(torch::Tensor& logits,
const torch::Tensor& prompt_mask,
const torch::Tensor& output_mask,
const torch::Tensor& repetition_penalties);
void top_k_per_row_prefill(const torch::Tensor& logits,
const torch::Tensor& rowStarts,
const torch::Tensor& rowEnds, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1,
int64_t topK);
void top_k_per_row_decode(const torch::Tensor& logits, int64_t next_n,
const torch::Tensor& seqLens, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1,
int64_t topK);
void persistent_topk(const torch::Tensor& logits, const torch::Tensor& lengths,
torch::Tensor& output, torch::Tensor& workspace, int64_t k,
int64_t max_seq_len);
void silu_and_mul_per_block_quant(torch::Tensor& out,
torch::Tensor const& input,
torch::Tensor& scales, int64_t group_size,
@@ -150,22 +123,6 @@ void dynamic_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
torch::Tensor& scales,
std::optional<torch::Tensor> const& azp);
void selective_scan_fwd(
const torch::Tensor& u, const torch::Tensor& delta, const torch::Tensor& A,
const torch::Tensor& B, const torch::Tensor& C,
const std::optional<torch::Tensor>& D_,
const std::optional<torch::Tensor>& z_,
const std::optional<torch::Tensor>& delta_bias_, bool delta_softplus,
const std::optional<torch::Tensor>& query_start_loc,
const std::optional<torch::Tensor>& cache_indices,
const std::optional<torch::Tensor>& has_initial_state,
const torch::Tensor& ssm_states, int64_t null_block_id, int64_t block_size,
const std::optional<torch::Tensor>& block_idx_first_scheduled_token,
const std::optional<torch::Tensor>& block_idx_last_scheduled_token,
const std::optional<torch::Tensor>& initial_state_idx,
const std::optional<torch::Tensor>& cu_chunk_seqlen,
const std::optional<torch::Tensor>& last_chunk_indices);
torch::Tensor dynamic_4bit_int_moe_cpu(
torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,