aboutsummaryrefslogtreecommitdiff
path: root/qwen_kernels.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-14 15:19:12 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 15:19:12 +0200
commitf463018cd700690336898d109226d489deca036d (patch)
tree4e1d83f55a5e231af519aa4d647141db5de23af7 /qwen_kernels.c
parentd98ee1568c0bd88304a12e59d12104066e01e8d0 (diff)
downloadffwd-f463018cd700690336898d109226d489deca036d.tar.gz
ffwd-f463018cd700690336898d109226d489deca036d.zip
all: add support for Qwen3-Embedding + causal GQA attention
Diffstat (limited to 'qwen_kernels.c')
-rw-r--r--qwen_kernels.c152
1 files changed, 152 insertions, 0 deletions
diff --git a/qwen_kernels.c b/qwen_kernels.c
index 85871d4..dcd724c 100644
--- a/qwen_kernels.c
+++ b/qwen_kernels.c
@@ -1015,6 +1015,158 @@ void qwen_bidirectional_gqa_attention_packed(float *out, const float *Q,
free(scores);
}
+static void qwen_causal_gqa_attention_packed_online_rows(
+ float *out, const float *Q, const float *K, const float *V,
+ int start, int n_heads, int n_kv_heads, int head_dim,
+ float scale, int h, int query_start, int query_end) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+
+ for (int i = query_start; i < query_end; i++) {
+ const float *q_row = Q + (size_t)i * q_hidden + h * head_dim;
+ float *o_row = out + (size_t)i * q_hidden + h * head_dim;
+
+ float max_score = -1e30f;
+ float sum_exp = 0.0f;
+ for (int d = 0; d < head_dim; d++) o_row[d] = 0.0f;
+
+ for (int j = start; j <= i; j++) {
+ const float *k_row = K + (size_t)j * kv_hidden + kv_h * head_dim;
+ const float *v_row = V + (size_t)j * kv_hidden + kv_h * head_dim;
+ float score = qwen_dot_f32_fast(q_row, k_row, head_dim) * scale;
+
+ if (score > max_score) {
+ float correction = expf(max_score - score);
+ sum_exp = sum_exp * correction + 1.0f;
+ qwen_vec_scale_add(o_row, v_row, correction, head_dim);
+ max_score = score;
+ } else {
+ float wt = expf(score - max_score);
+ sum_exp += wt;
+ qwen_vec_axpy_inplace(o_row, v_row, wt, head_dim);
+ }
+ }
+
+ if (sum_exp > 0.0f)
+ qwen_vec_scale_inplace(o_row, 1.0f / sum_exp, head_dim);
+ }
+}
+
+#ifdef USE_BLAS
+static void qwen_causal_gqa_attention_packed_blas_rows(
+ float *out, const float *Q, const float *K, const float *V,
+ int start, int n_heads, int n_kv_heads, int head_dim,
+ float scale, int h, int query_start, int query_end, float *scores) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+ int rows = query_end - query_start;
+ int keys = query_end - start;
+
+ const float *q = Q + (size_t)query_start * q_hidden + h * head_dim;
+ const float *k = K + (size_t)start * kv_hidden + kv_h * head_dim;
+ const float *v = V + (size_t)start * kv_hidden + kv_h * head_dim;
+ float *o = out + (size_t)query_start * q_hidden + h * head_dim;
+
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+ rows, keys, head_dim, scale, q, q_hidden, k, kv_hidden,
+ 0.0f, scores, keys);
+ for (int r = 0; r < rows; r++) {
+ int allowed = query_start + r - start + 1;
+ float *row = scores + (size_t)r * keys;
+ for (int j = allowed; j < keys; j++) row[j] = -INFINITY;
+ }
+ qwen_softmax(scores, rows, keys);
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+ rows, head_dim, keys, 1.0f, scores, keys, v, kv_hidden,
+ 0.0f, o, q_hidden);
+}
+#endif
+
+static void packed_causal_gqa_attn_worker(int tid, int n_threads, void *arg) {
+ packed_gqa_attn_task_t *t = (packed_gqa_attn_task_t *)arg;
+ int task_id = 0;
+#ifdef USE_BLAS
+ float *scores = t->scores
+ ? t->scores + (size_t)tid * QWEN_PACKED_ATTN_QUERY_TILE * t->max_seq
+ : NULL;
+#endif
+
+ for (int h = 0; h < t->n_heads; h++) {
+ for (int b = 0; b < t->batch; b++) {
+ int start = t->offsets[b];
+ int end = t->offsets[b + 1];
+ for (int q0 = start; q0 < end; q0 += QWEN_PACKED_ATTN_QUERY_TILE) {
+ int q1 = q0 + QWEN_PACKED_ATTN_QUERY_TILE;
+ if (q1 > end) q1 = end;
+ if (task_id++ % n_threads != tid) continue;
+#ifdef USE_BLAS
+ if (scores && end - start >= QWEN_PACKED_ATTN_BLAS_MIN_SEQ) {
+ qwen_causal_gqa_attention_packed_blas_rows(
+ t->out, t->Q, t->K, t->V, start, t->n_heads,
+ t->n_kv_heads, t->head_dim, t->scale, h, q0, q1,
+ scores);
+ continue;
+ }
+#endif
+ qwen_causal_gqa_attention_packed_online_rows(
+ t->out, t->Q, t->K, t->V, start, t->n_heads,
+ t->n_kv_heads, t->head_dim, t->scale, h, q0, q1);
+ }
+ }
+ }
+}
+
+void qwen_causal_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) {
+ long long qk_work = 0;
+ int max_seq = 0;
+ for (int b = 0; b < batch; b++) {
+ int len = offsets[b + 1] - offsets[b];
+ qk_work += (long long)len * (len + 1) / 2 * n_heads;
+ if (len > max_seq) max_seq = len;
+ }
+
+ float *scores = NULL;
+#ifdef USE_BLAS
+ size_t required = qwen_bidirectional_gqa_attention_packed_scratch_bytes(
+ offsets, batch);
+ if (required != 0 && scratch && scratch_bytes >= required) scores = scratch;
+#else
+ (void)scratch;
+ (void)scratch_bytes;
+#endif
+ packed_gqa_attn_task_t task = {
+ .out = out, .Q = Q, .K = K, .V = V, .offsets = offsets,
+ .batch = batch, .n_heads = n_heads, .n_kv_heads = n_kv_heads,
+ .head_dim = head_dim, .scale = scale, .scores = scores,
+ .max_seq = max_seq
+ };
+ if (tp.n_threads > 1 && n_heads >= 2 && qk_work >= 4096)
+ parallel_for(packed_causal_gqa_attn_worker, &task);
+ else
+ packed_causal_gqa_attn_worker(0, 1, &task);
+}
+
+void qwen_causal_gqa_attention_packed(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) {
+ size_t scratch_bytes =
+ qwen_bidirectional_gqa_attention_packed_scratch_bytes(offsets, batch);
+ float *scores = scratch_bytes ? malloc(scratch_bytes) : NULL;
+ qwen_causal_gqa_attention_packed_with_scratch(
+ out, Q, K, V, offsets, batch, n_heads, n_kv_heads, head_dim, scale,
+ scores, scratch_bytes);
+ free(scores);
+}
+
/* ========================================================================
* Position Embeddings
* ======================================================================== */