[Kernel]Add norm fused kernel for Step-5 Preview and Muse Spark Open (#62132)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-10-17 01:53:00 -07:00
committed by GitHub
parent eb8de8d1a1
commit e56b1603af
10 changed files with 2461 additions and 0 deletions
+33
View File
@@ -363,6 +363,35 @@ minimax_allreduce_rms_qk(torch::stable::Tensor qkv,
int64_t const nranks, double const eps);
#endif
#ifndef USE_ROCM
namespace vllm {
void fused_qknorm_rope_cache(
torch::stable::Tensor& q_out, torch::stable::Tensor& k_out,
torch::stable::Tensor& v_out, torch::stable::Tensor& qkv,
torch::stable::Tensor& q_weight, torch::stable::Tensor& k_weight,
torch::stable::Tensor& cos, torch::stable::Tensor& sin,
torch::stable::Tensor& positions, torch::stable::Tensor& slot_mapping,
torch::stable::Tensor& k_cache, torch::stable::Tensor& v_cache,
int64_t head_dim, int64_t num_q_heads, int64_t num_kv_heads,
int64_t rotary_pairs, double epsilon, double norm_weight_bias,
const std::string& kv_cache_dtype,
std::optional<torch::stable::Tensor> k_scale,
std::optional<torch::stable::Tensor> v_scale);
} // namespace vllm
#endif
void fused_muse_spark_qknorm_rope_kv_insert(
torch::stable::Tensor& qkv, torch::stable::Tensor const& positions,
int64_t num_heads, int64_t num_kv_heads, int64_t head_dim, double eps,
std::optional<torch::stable::Tensor> cos_sin_cache, int64_t rotary_dim,
std::optional<torch::stable::Tensor> slot_mapping,
std::optional<torch::stable::Tensor> kv_cache, int64_t block_size,
const std::string& kv_cache_dtype,
std::optional<torch::stable::Tensor> k_scale,
std::optional<torch::stable::Tensor> v_scale,
std::optional<torch::stable::Tensor> q_fp8_out,
std::optional<torch::stable::Tensor> q_scale);
// Horizontally-fused MiniMax-M3 QK-norm + partial NeoX RoPE (+ optional KV /
// index-cache insert). Dense layer: norm+RoPE only; sparse layer: also packs
// the index branch and scatters k/v/index_k into their paged caches.
@@ -546,6 +575,10 @@ std::tuple<int64_t, torch::stable::Tensor> allocate_shared_buffer_and_handle(
int64_t open_mem_handle(torch::stable::Tensor& mem_handle);
void free_shared_buffer(int64_t buffer);
void muse_spark_mul_tanh_soft_cap(torch::stable::Tensor& out,
torch::stable::Tensor& input,
double multiplier, double soft_cap_temp);
// Activation kernels (shared CUDA/ROCm)
void silu_and_mul(torch::stable::Tensor& out, torch::stable::Tensor& input);
void silu_and_mul_clamp(torch::stable::Tensor& out,