Skip to content

[BCI] QVAC-17071 feat: add BCI neural signal support (variable conv1 kernel + windowed attention) #9

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
sharmaraju352 wants to merge 2 commits into from
Closed

[BCI] QVAC-17071 feat: add BCI neural signal support (variable conv1 kernel + windowed attention) #9

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
346d7de
feat(bci): add variable conv1 kernel size support
Apr 16, 2026
1e91c27
feat(bci): add windowed self-attention for encoder layers
Apr 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 45 additions & 3 deletions src/whisper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -600,6 +600,9 @@ struct whisper_hparams {
int32_t n_mels = 80;
int32_t ftype = 1;
float eps = 1e-5f;
int32_t n_audio_conv1_kernel = 3;
int32_t n_audio_window_size = 0;
int32_t n_audio_last_window_layer = -1;
};

// audio encoding layer
Expand Down Expand Up @@ -1517,6 +1520,9 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
read_safe(loader, hparams.n_text_layer);
read_safe(loader, hparams.n_mels);
read_safe(loader, hparams.ftype);
read_safe(loader, hparams.n_audio_conv1_kernel);
read_safe(loader, hparams.n_audio_window_size);
read_safe(loader, hparams.n_audio_last_window_layer);

assert(hparams.n_text_state == hparams.n_audio_state);

Expand Down Expand Up @@ -1757,7 +1763,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
// encoder
model.e_pe = create_tensor(ASR_TENSOR_ENC_POS_EMBD, ASR_SYSTEM_ENCODER, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx));

model.e_conv_1_w = create_tensor(ASR_TENSOR_CONV1_WEIGHT, ASR_SYSTEM_ENCODER, ggml_new_tensor_3d(ctx, vtype, 3, n_mels, n_audio_state));
model.e_conv_1_w = create_tensor(ASR_TENSOR_CONV1_WEIGHT, ASR_SYSTEM_ENCODER, ggml_new_tensor_3d(ctx, vtype, hparams.n_audio_conv1_kernel, n_mels, n_audio_state));
model.e_conv_1_b = create_tensor(ASR_TENSOR_CONV1_BIAS, ASR_SYSTEM_ENCODER, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state));

model.e_conv_2_w = create_tensor(ASR_TENSOR_CONV2_WEIGHT, ASR_SYSTEM_ENCODER, ggml_new_tensor_3d(ctx, vtype, 3, n_audio_state, n_audio_state));
Expand Down Expand Up @@ -2095,6 +2101,15 @@ static struct ggml_cgraph * whisper_build_graph_encoder(

struct ggml_tensor * inpL = cur;

struct ggml_tensor * window_mask = nullptr;
const int window_size = hparams.n_audio_window_size;
const int last_window_layer = hparams.n_audio_last_window_layer;
if (window_size > 0 && last_window_layer >= 0) {
window_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_ctx, n_ctx, 1);
ggml_set_name(window_mask, "window_mask");
ggml_set_input(window_mask);
}

for (int il = 0; il < n_layer; ++il) {
const auto & layer = model.layers_encoder[il];

Expand Down Expand Up @@ -2156,7 +2171,9 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
ggml_element_size(kv_pad.v)*n_state_head,
0);

cur = ggml_flash_attn_ext(ctx0, Q, K, V, nullptr, KQscale, 0.0f, 0.0f);
struct ggml_tensor * attn_mask_fa = (window_mask && il <= last_window_layer)
? ggml_cast(ctx0, window_mask, GGML_TYPE_F16) : nullptr;
cur = ggml_flash_attn_ext(ctx0, Q, K, V, attn_mask_fa, KQscale, 0.0f, 0.0f);

cur = ggml_reshape_2d(ctx0, cur, n_state, n_ctx);
} else {
Expand All @@ -2170,7 +2187,8 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
// K * Q
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);

struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, nullptr, KQscale, 0.0f);
struct ggml_tensor * enc_attn_mask = (window_mask && il <= last_window_layer) ? window_mask : nullptr;
struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, enc_attn_mask, KQscale, 0.0f);

struct ggml_tensor * V =
ggml_cast(ctx0,
Expand Down Expand Up @@ -2444,6 +2462,25 @@ static bool whisper_encode_internal(
return false;
}

{
struct ggml_tensor * wmask = ggml_graph_get_tensor(gf, "window_mask");
if (wmask) {
const int n_ctx = wstate.exp_n_audio_ctx > 0
? wstate.exp_n_audio_ctx : wctx.model.hparams.n_audio_ctx;
const int ws = wctx.model.hparams.n_audio_window_size;
const int half_w = ws / 2;
std::vector<float> mask_data(n_ctx * n_ctx);
for (int i = 0; i < n_ctx; ++i) {
for (int j = 0; j < n_ctx; ++j) {
mask_data[i * n_ctx + j] =
(abs(i - j) <= half_w) ? 0.0f : -INFINITY;
}
}
ggml_backend_tensor_set(wmask, mask_data.data(), 0,
n_ctx * n_ctx * sizeof(float));
}
}

if (!ggml_graph_compute_helper(sched, gf, n_threads)) {
return false;
}
Expand Down Expand Up @@ -6958,6 +6995,11 @@ int whisper_full_with_state(
} else {
prompt_init.push_back(whisper_token_transcribe(ctx));
}
} else if (ctx->model.hparams.n_audio_window_size > 0) {
const int lang_id = whisper_lang_id(params.language);
state->lang_id = lang_id;
prompt_init.push_back(whisper_token_lang(ctx, lang_id));
prompt_init.push_back(whisper_token_transcribe(ctx));
}

// first release distilled models require the "no_timestamps" token
Expand Down
Loading