[9/n] Migrate attention and cache kernels to torch stable ABI (continued) (#43717)

Signed-off-by: Chris Leonard <chleonar@redhat.com>
Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.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
2026-05-29 04:44:45 +00:00
committed by GitHub
co-authored by Mikayla Gawarecki Shengqi Chen
parent 710f077617
commit 22a58640b4
23 changed files with 909 additions and 720 deletions
+4 -138
View File
@@ -1,4 +1,7 @@
#include "cache.h"
// Provides torch::Tensor for ops.h (previously included transitively via
// cache.h, which is no longer included here after cache ops moved to
// _C_stable_libtorch).
#include <torch/all.h>
#include "cuda_utils.h"
#include "ops.h"
#include "core/registration.h"
@@ -33,35 +36,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
&get_cuda_view_from_cpu_tensor);
// Attention ops
// Compute the attention between an input query and the cached
// keys/values using PagedAttention.
ops.def(
"paged_attention_v1("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v1", torch::kCUDA, &paged_attention_v1);
// PagedAttention V2.
ops.def(
"paged_attention_v2("
" Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
" Tensor! tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);
// Activation ops (quantized only — basic ops moved to _C_stable_libtorch)
ops.def(
"silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
@@ -217,114 +191,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
#endif
}
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
// Cache ops
// Swap in (out) the cache blocks from src to dst.
cache_ops.def(
"swap_blocks(Tensor src, Tensor! dst,"
" int block_size_in_bytes, Tensor block_mapping) -> ()");
cache_ops.impl("swap_blocks", torch::kCUDA, &swap_blocks);
// Batch swap: submit all block copies in a single driver call.
cache_ops.def(
"swap_blocks_batch(Tensor src_ptrs, Tensor dst_ptrs,"
" Tensor sizes,"
" bool is_src_access_order_any=False) -> ()");
cache_ops.impl("swap_blocks_batch", torch::kCPU, &swap_blocks_batch);
// Reshape the key and value tensors and cache them.
cache_ops.def(
"reshape_and_cache(Tensor key, Tensor value,"
" Tensor! key_cache, Tensor! value_cache,"
" Tensor slot_mapping,"
" str kv_cache_dtype,"
" Tensor k_scale, Tensor v_scale) -> ()");
cache_ops.impl("reshape_and_cache", torch::kCUDA, &reshape_and_cache);
// Reshape the key and value tensors and cache them.
cache_ops.def(
"reshape_and_cache_flash(Tensor key, Tensor value,"
" Tensor! key_cache,"
" Tensor! value_cache,"
" Tensor slot_mapping,"
" str kv_cache_dtype,"
" Tensor k_scale, Tensor v_scale) -> ()");
cache_ops.impl("reshape_and_cache_flash", torch::kCUDA,
&reshape_and_cache_flash);
// Concat kv_c and k_pe and cache them.
cache_ops.def(
"concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
" Tensor! kv_cache,"
" Tensor slot_mapping,"
" str kv_cache_dtype,"
" Tensor scale) -> ()");
cache_ops.impl("concat_and_cache_mla", torch::kCUDA, &concat_and_cache_mla);
// Rotate Q and K, then write to kv cache for MLA
cache_ops.def(
"concat_and_cache_mla_rope_fused("
" Tensor positions,"
" Tensor! q_pe,"
" Tensor! k_pe,"
" Tensor kv_c,"
" Tensor cos_sin_cache,"
" bool is_neox,"
" Tensor slot_mapping,"
" Tensor! kv_cache,"
" str kv_cache_dtype,"
" Tensor kv_cache_scale) -> ()");
cache_ops.impl("concat_and_cache_mla_rope_fused", torch::kCUDA,
&concat_and_cache_mla_rope_fused);
// Convert the key and value cache to fp8 data type.
cache_ops.def(
"convert_fp8(Tensor! dst_cache, Tensor src_cache, float scale, "
"str kv_cache_dtype) -> ()");
cache_ops.impl("convert_fp8", torch::kCUDA, &convert_fp8);
// Gather cache blocks from src_cache to dst, dequantizing from
// src_cache's dtype to dst's dtype if necessary.
cache_ops.def(
"gather_and_maybe_dequant_cache(Tensor src_cache, Tensor! dst, "
" Tensor block_table, Tensor cu_seq_lens, "
" Tensor token_to_seq, "
" int num_tokens, "
" str kv_cache_dtype, "
" Tensor scale, Tensor? seq_starts) -> ()");
cache_ops.impl("gather_and_maybe_dequant_cache", torch::kCUDA,
&gather_and_maybe_dequant_cache);
cache_ops.def(
"cp_gather_cache(Tensor src_cache, Tensor! dst, Tensor block_table, "
"Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
cache_ops.impl("cp_gather_cache", torch::kCUDA, &cp_gather_cache);
cache_ops.def(
"cp_gather_and_upconvert_fp8_kv_cache(Tensor src_cache, Tensor! dst, "
"Tensor block_table, Tensor seq_lens, Tensor workspace_starts, int "
"batch_size) -> ()");
cache_ops.impl("cp_gather_and_upconvert_fp8_kv_cache", torch::kCUDA,
&cp_gather_and_upconvert_fp8_kv_cache);
cache_ops.def(
"indexer_k_quant_and_cache(Tensor k, Tensor! kv_cache, Tensor "
"slot_mapping, "
"int quant_block_size, str kv_cache_dtype) -> ()");
cache_ops.impl("indexer_k_quant_and_cache", torch::kCUDA,
&indexer_k_quant_and_cache);
cache_ops.def(
"concat_mla_q(Tensor ql_nope, Tensor q_pe, Tensor! q_out) -> ()");
cache_ops.impl("concat_mla_q", torch::kCUDA, &concat_mla_q);
cache_ops.def(
"cp_gather_indexer_k_quant_cache(Tensor kv_cache, Tensor! dst_k, Tensor! "
"dst_scale, Tensor block_table, Tensor cu_seq_lens) -> ()");
cache_ops.impl("cp_gather_indexer_k_quant_cache", torch::kCUDA,
&cp_gather_indexer_k_quant_cache);
}
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
// Cuda utils