aboutsummaryrefslogtreecommitdiff
path: root/pplx_embed.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-05-24 01:02:17 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:18:50 +0200
commit9bfcc72ee6b2568d0f9fc05a28f9ccc3da176713 (patch)
tree724a5c293a7c6bf7fe028936c38e85e9919534b9 /pplx_embed.c
parent4f0dd572471121135213e939b675fdc1cac32786 (diff)
downloadffwd-9bfcc72ee6b2568d0f9fc05a28f9ccc3da176713.tar.gz
ffwd-9bfcc72ee6b2568d0f9fc05a28f9ccc3da176713.zip
kernels,cpu: fuse tiny BF16 gate up projections
Diffstat (limited to 'pplx_embed.c')
-rw-r--r--pplx_embed.c29
1 files changed, 26 insertions, 3 deletions
diff --git a/pplx_embed.c b/pplx_embed.c
index 5ee7051..ecc7fa0 100644
--- a/pplx_embed.c
+++ b/pplx_embed.c
@@ -65,6 +65,9 @@ struct pplx_workspace {
int pplx_verbose = 0;
int qwen_verbose = 0;
+#define PPLX_BF16_QKV_FUSE_MAX_SEQ 16
+#define PPLX_BF16_PAIR_FUSE_MAX_SEQ 4
+
/* ========================================================================
* Size helpers
* ======================================================================== */
@@ -396,7 +399,8 @@ static void linear_qkv_weight(float *q, float *k, float *v, const float *x,
int seq_len, int in_dim, int q_dim, int kv_dim)
{
if (wq->dtype == DTYPE_BF16 && wk->dtype == DTYPE_BF16 &&
- wv->dtype == DTYPE_BF16 && seq_len <= 16) {
+ wv->dtype == DTYPE_BF16 &&
+ seq_len <= PPLX_BF16_QKV_FUSE_MAX_SEQ) {
qwen_linear_nobias_bf16_qkv(q, k, v, x,
(const uint16_t *)wq->data,
(const uint16_t *)wk->data,
@@ -410,6 +414,24 @@ static void linear_qkv_weight(float *q, float *k, float *v, const float *x,
linear_nobias_weight(v, x, wv, seq_len, in_dim, kv_dim);
}
+static void linear_pair_weight(float *a, float *b, const float *x,
+ const pplx_weight_ref_t *wa,
+ const pplx_weight_ref_t *wb,
+ int seq_len, int in_dim, int a_dim, int b_dim)
+{
+ if (wa->dtype == DTYPE_BF16 && wb->dtype == DTYPE_BF16 &&
+ seq_len <= PPLX_BF16_PAIR_FUSE_MAX_SEQ) {
+ qwen_linear_nobias_bf16_pair(a, b, x,
+ (const uint16_t *)wa->data,
+ (const uint16_t *)wb->data,
+ seq_len, in_dim, a_dim, b_dim);
+ return;
+ }
+
+ linear_nobias_weight(a, x, wa, seq_len, in_dim, a_dim);
+ linear_nobias_weight(b, x, wb, seq_len, in_dim, b_dim);
+}
+
/* ========================================================================
* Working buffer management
* ======================================================================== */
@@ -750,8 +772,9 @@ static int forward_packed_inplace(const pplx_model_t *model,
qwen_rms_norm(x_norm, x, l->post_attn_norm, total_seq, hidden, eps);
- linear_nobias_weight(ffn_gate, x_norm, &l->gate_proj, total_seq, hidden, inter);
- linear_nobias_weight(ffn_up, x_norm, &l->up_proj, total_seq, hidden, inter);
+ linear_pair_weight(ffn_gate, ffn_up, x_norm,
+ &l->gate_proj, &l->up_proj,
+ total_seq, hidden, inter, inter);
qwen_silu_mul_inplace(ffn_gate, ffn_up, total_seq * inter);
linear_nobias_weight(proj_out, ffn_gate, &l->down_proj, total_seq, inter, hidden);
qwen_add_inplace(x, proj_out, total_seq * hidden);