From c709fc62ebe0eb2ca596316e5dd22f212766ff19 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Wed, 10 Jun 2026 09:40:35 +0200 Subject: cpu: move BF16 widening scratch into the workspace --- qwen_kernels.h | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) (limited to 'qwen_kernels.h') diff --git a/qwen_kernels.h b/qwen_kernels.h index e92f944..5281361 100644 --- a/qwen_kernels.h +++ b/qwen_kernels.h @@ -35,9 +35,8 @@ void qwen_linear_nobias(float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim); /* bf16 weight variants */ -void qwen_linear_bf16(float *y, const float *x, const uint16_t *W_bf16, - const float *b, int seq_len, int in_dim, int out_dim); +void qwen_bf16_to_f32_buf(float *dst, const uint16_t *src, size_t n); void qwen_linear_nobias_bf16(float *y, const float *x, const uint16_t *W_bf16, int seq_len, int in_dim, int out_dim); @@ -56,8 +55,6 @@ void qwen_linear_nobias_bf16_qkv(float *q, float *k, float *v, const float *x, int seq_len, int in_dim, int q_dim, int kv_dim); -void qwen_matmul_t_bf16(float *C, const float *A, const uint16_t *B_bf16, - int M, int K, int N); /* Dot product using the best available local SIMD implementation. */ float qwen_dot_f32(const float *a, const float *b, int n); -- cgit v1.3.1