#ifndef KERNELS_H #define KERNELS_H #include #include void add_inplace(float *a, const float *b, int n); /* y = x @ W^T + b: x[seq,in], W[out,in], b[out], y[seq,out]. */ void linear( float *y, const float *x, const float *W, const float *b, int seq_len, int in_dim, int out_dim); void linear_nobias(float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim); /* Mean and population variance are reduced over the hidden axis per row. */ void layer_norm(float *out, const float *x, const float *gamma, const float *beta, int seq_len, int hidden, float eps); void softmax(float *x, int rows, int cols); void gelu_inplace(float *x, int n); /* Attention is block-diagonal: each token attends only within its offsets span. * Missing scratch uses the online-softmax path instead of tiled BLAS. */ size_t bidirectional_gqa_attention_packed_scratch_bytes(const int *offsets, int batch); void bidirectional_gqa_attention_packed_with_scratch(float *out, const float *Q, const float *K, const float *V, const int *offsets, int batch, int n_heads, int n_kv_heads, int head_dim, float scale, float *scratch, size_t scratch_bytes); #endif /* KERNELS_H */