aboutsummaryrefslogtreecommitdiff
path: root/liblangdet/kernels.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-09-29 14:38:02 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-09-29 14:38:02 +0200
commit91c6aa07597462705a7f30ee5b97e2c7943b5616 (patch)
treeb16f686279a1f1b0691a6c0d65a544084fcc1c1e /liblangdet/kernels.h
parent6f98fc8cf3f7cc4b9d523d16492a07d3e2aea8c9 (diff)
downloadlangdet-91c6aa07597462705a7f30ee5b97e2c7943b5616.tar.gz
langdet-91c6aa07597462705a7f30ee5b97e2c7943b5616.zip
Remove unused checkpoint, tokenizer, and attention compatibility paths
Diffstat (limited to 'liblangdet/kernels.h')
-rw-r--r--liblangdet/kernels.h26
1 files changed, 12 insertions, 14 deletions
diff --git a/liblangdet/kernels.h b/liblangdet/kernels.h
index 1025602..95f5df0 100644
--- a/liblangdet/kernels.h
+++ b/liblangdet/kernels.h
@@ -23,20 +23,18 @@ void layer_norm(float *out,
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.
+/* Bidirectional attention over one unpadded sequence with equal Q/K/V head counts.
* 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);
+size_t attention_scratch_bytes(int seq);
+void attention_with_scratch(float *out,
+ const float *Q,
+ const float *K,
+ const float *V,
+ int seq,
+ int n_heads,
+ int head_dim,
+ float scale,
+ float *scratch,
+ size_t scratch_bytes);
#endif /* KERNELS_H */