ggml-webgpu : extend GDN for K>1 (#23299)
This commit is contained in:
@@ -1234,6 +1234,7 @@ static webgpu_encoded_op ggml_webgpu_gated_delta_net(webgpu_context & ctx,
|
||||
const uint32_t h = (uint32_t) src2->ne[1];
|
||||
const uint32_t n_tokens = (uint32_t) src2->ne[2];
|
||||
const uint32_t n_seqs = (uint32_t) src2->ne[3];
|
||||
const uint32_t K = (uint32_t) src5->ne[1];
|
||||
const float scale = 1.0f / sqrtf((float) s_v);
|
||||
uint32_t scale_u32;
|
||||
memcpy(&scale_u32, &scale, sizeof(scale_u32));
|
||||
@@ -1258,6 +1259,7 @@ static webgpu_encoded_op ggml_webgpu_gated_delta_net(webgpu_context & ctx,
|
||||
|
||||
(uint32_t) src0->ne[1],
|
||||
(uint32_t) (src2->ne[3] / src0->ne[3]),
|
||||
K,
|
||||
scale_u32,
|
||||
};
|
||||
|
||||
|
||||
Reference in New Issue
Block a user