[CPU] Support Gemma Diffusion (#45690)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2026-06-16 14:39:56 +08:00
committed by GitHub
parent 8bf374955f
commit a7fdfeef72
7 changed files with 213 additions and 100 deletions
+10 -8
View File
@@ -152,7 +152,8 @@ torch::Tensor get_scheduler_metadata(
const torch::Tensor& seq_lens, at::ScalarType dtype,
const torch::Tensor& query_start_loc, const bool casual,
const int64_t window_size, const std::string& isa_hint,
const bool enable_kv_split);
const bool enable_kv_split,
const std::optional<torch::Tensor>& dynamic_causal);
void cpu_attn_reshape_and_cache(const torch::Tensor& key,
const torch::Tensor& value,
@@ -169,10 +170,10 @@ void cpu_attention_with_kv_cache(
const torch::Tensor& query_start_loc, const torch::Tensor& seq_lens,
const double scale, const bool causal,
const std::optional<torch::Tensor>& alibi_slopes,
const int64_t sliding_window_left, const int64_t sliding_window_right,
const torch::Tensor& block_table, const double softcap,
const torch::Tensor& scheduler_metadata,
const std::optional<torch::Tensor>& s_aux, const double k_scale,
const int64_t sliding_window_left, const torch::Tensor& block_table,
const double softcap, const torch::Tensor& scheduler_metadata,
const std::optional<torch::Tensor>& s_aux,
const std::optional<torch::Tensor>& dynamic_causal, const double k_scale,
const double v_scale, const std::string& kv_cache_dtype);
// Note: just for avoiding importing errors
@@ -500,7 +501,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
"get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
"int head_dim, Tensor seq_lens, ScalarType dtype, Tensor "
"query_start_loc, bool casual, int window_size, str isa_hint, bool "
"enable_kv_split) -> Tensor",
"enable_kv_split, Tensor? dynamic_causal) -> Tensor",
&get_scheduler_metadata);
ops.def(
"cpu_attn_reshape_and_cache(Tensor key, Tensor value, Tensor(a2!) "
@@ -512,8 +513,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
"cpu_attention_with_kv_cache(Tensor query, Tensor key_cache, Tensor "
"value_cache, Tensor(a3!) output, Tensor query_start_loc, Tensor "
"seq_lens, float scale, bool causal, Tensor? alibi_slopes, SymInt "
"sliding_window_left, SymInt sliding_window_right, Tensor block_table, "
"float softcap, Tensor scheduler_metadata, Tensor? s_aux, "
"sliding_window_size, Tensor block_table, "
"float softcap, Tensor scheduler_metadata, Tensor? s_aux, Tensor? "
"dynamic_causal, "
"float k_scale=1.0, float v_scale=1.0, str kv_cache_dtype=\"auto\") -> "
"()",
&cpu_attention_with_kv_cache);