Remove old methods.
This commit is contained in:
Piotr Wilkin
parent
b3f55ead92
commit
34e1ed9093
|
|
@ -86,356 +86,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
|
|||
ggml_build_forward_expand(gf, cur);
|
||||
}
|
||||
|
||||
// utility to get one slice from the third dimension
|
||||
// input dim: [x, y, c, b]
|
||||
// output dim: [x, y, 1, b]
|
||||
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
|
||||
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
|
||||
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
|
||||
}
|
||||
|
||||
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chunking(
|
||||
ggml_tensor * q,
|
||||
ggml_tensor * k,
|
||||
ggml_tensor * v,
|
||||
ggml_tensor * g,
|
||||
ggml_tensor * beta,
|
||||
ggml_tensor * state,
|
||||
ggml_tensor * causal_mask,
|
||||
ggml_tensor * identity,
|
||||
ggml_tensor * diag_mask,
|
||||
int il) {
|
||||
const int64_t S_k = q->ne[0];
|
||||
const int64_t H_k = q->ne[1];
|
||||
const int64_t n_tokens = q->ne[2];
|
||||
const int64_t n_seqs = q->ne[3];
|
||||
|
||||
const int64_t S_v = v->ne[0];
|
||||
const int64_t H_v = v->ne[1];
|
||||
|
||||
GGML_ASSERT(v->ne[2] == n_tokens);
|
||||
GGML_ASSERT(k->ne[2] == n_tokens);
|
||||
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
|
||||
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
|
||||
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
|
||||
|
||||
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
|
||||
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
|
||||
|
||||
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
|
||||
|
||||
const float eps_norm = hparams.f_norm_rms_eps;
|
||||
|
||||
q = ggml_l2_norm(ctx0, q, eps_norm);
|
||||
k = ggml_l2_norm(ctx0, k, eps_norm);
|
||||
|
||||
const float scale = 1.0f / sqrtf(S_v);
|
||||
|
||||
q = ggml_scale(ctx0, q, scale);
|
||||
|
||||
beta = ggml_sigmoid(ctx0, beta);
|
||||
|
||||
cb(q, "q_in", il);
|
||||
cb(k, "k_in", il);
|
||||
cb(v, "v_in", il);
|
||||
cb(beta, "beta_in", il);
|
||||
cb(g, "g_in", il);
|
||||
|
||||
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
|
||||
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
|
||||
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
|
||||
g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
|
||||
|
||||
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
|
||||
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
|
||||
|
||||
cb(q, "q_perm", il);
|
||||
cb(k, "k_perm", il);
|
||||
cb(v, "v_perm", il);
|
||||
cb(beta, "beta_perm", il);
|
||||
cb(g, "g_perm", il);
|
||||
cb(state, "state_in", il);
|
||||
|
||||
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
|
||||
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
|
||||
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
|
||||
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
|
||||
|
||||
// Do padding
|
||||
const int64_t chunk_size = CHUNK_SIZE;
|
||||
|
||||
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
|
||||
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
|
||||
|
||||
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
|
||||
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
|
||||
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
|
||||
g = ggml_pad(ctx0, g, pad, 0, 0, 0);
|
||||
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
|
||||
|
||||
cb(q, "q_pad", il);
|
||||
cb(k, "k_pad", il);
|
||||
cb(v, "v_pad", il);
|
||||
cb(beta, "beta_pad", il);
|
||||
cb(g, "g_pad", il);
|
||||
|
||||
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
|
||||
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
|
||||
|
||||
cb(v_beta, "v_beta", il);
|
||||
cb(k_beta, "k_beta", il);
|
||||
|
||||
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
|
||||
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
|
||||
k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
|
||||
v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
|
||||
v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
|
||||
|
||||
g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
|
||||
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
|
||||
|
||||
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
|
||||
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
|
||||
ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
|
||||
|
||||
ggml_tensor * gcs_j_broadcast =
|
||||
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
|
||||
|
||||
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
|
||||
cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
|
||||
decay_mask = ggml_exp(ctx0, decay_mask);
|
||||
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
|
||||
|
||||
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
|
||||
|
||||
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
|
||||
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
|
||||
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
|
||||
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
|
||||
|
||||
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
|
||||
attn = ggml_mul(ctx0, lin_solve, causal_mask);
|
||||
attn = ggml_add(ctx0, attn, identity);
|
||||
cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
|
||||
|
||||
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
|
||||
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
|
||||
|
||||
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
|
||||
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * k_cumdecay =
|
||||
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
|
||||
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
|
||||
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
|
||||
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
|
||||
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
|
||||
// vectorized calculation of key_gdiff
|
||||
// improved from the chunked version:
|
||||
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
|
||||
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
|
||||
// key_gdiff = key * g_diff.unsqueeze(-1)
|
||||
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
|
||||
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
|
||||
|
||||
// get last element in g_cumsum along chunk_size dimension (ne0)
|
||||
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
|
||||
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
|
||||
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
|
||||
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
|
||||
g_last = ggml_cont(ctx0, g_last);
|
||||
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
|
||||
cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
|
||||
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
|
||||
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp);
|
||||
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
|
||||
// state to be updated per chunk
|
||||
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
|
||||
cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
|
||||
|
||||
// shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
|
||||
ggml_tensor * core_attn_out = nullptr;
|
||||
|
||||
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
|
||||
// shape: (S_k, chunk_size, 1, H_k * n_seqs)
|
||||
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
|
||||
|
||||
// shape: (S_v, chunk_size, 1, H_v * n_seqs)
|
||||
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
|
||||
|
||||
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
|
||||
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
|
||||
|
||||
// shape: (chunk_size, 1, H_v * n_seqs)
|
||||
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
|
||||
|
||||
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
|
||||
// replaced by precomputed attn_kq
|
||||
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
|
||||
cb(attn_chunk, "attn_chunk", il);
|
||||
|
||||
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
|
||||
|
||||
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
|
||||
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
|
||||
cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
|
||||
|
||||
// v_new = v_i - v_prime
|
||||
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
|
||||
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
|
||||
cb(v_new, "v_new_chunk", il);
|
||||
|
||||
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
|
||||
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
|
||||
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
|
||||
cb(attn_inter, "attn_inter_chunk", il);
|
||||
|
||||
// core_attn_out[:, :, i] = attn_inter + attn @ v_new
|
||||
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
|
||||
cb(v_attn, "v_attn_chunk", il);
|
||||
|
||||
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
|
||||
cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
|
||||
|
||||
core_attn_out = core_attn_out == nullptr
|
||||
? core_attn_out_chunk
|
||||
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
|
||||
|
||||
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
|
||||
ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk));
|
||||
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
|
||||
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff)));
|
||||
|
||||
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
|
||||
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
|
||||
new_state = ggml_add(ctx0,
|
||||
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
|
||||
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
|
||||
}
|
||||
|
||||
// truncate padded tokens
|
||||
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
|
||||
S_v, n_tokens, H_v, n_seqs,
|
||||
ggml_row_size(core_attn_out->type, S_v),
|
||||
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
|
||||
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
|
||||
output_tokens = ggml_cont(ctx0, output_tokens);
|
||||
cb(output_tokens, "output_tokens", il);
|
||||
|
||||
// permute back to (S_v, H_v, n_tokens, n_seqs)
|
||||
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
|
||||
output_tokens = ggml_cont(ctx0, output_tokens);
|
||||
|
||||
return {output_tokens, new_state};
|
||||
}
|
||||
|
||||
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
|
||||
ggml_tensor * q,
|
||||
ggml_tensor * k,
|
||||
ggml_tensor * v,
|
||||
ggml_tensor * g,
|
||||
ggml_tensor * beta,
|
||||
ggml_tensor * state,
|
||||
int il) {
|
||||
const int64_t S_k = q->ne[0];
|
||||
const int64_t H_k = q->ne[1];
|
||||
const int64_t n_tokens = q->ne[2];
|
||||
const int64_t n_seqs = q->ne[3];
|
||||
|
||||
const int64_t S_v = v->ne[0];
|
||||
const int64_t H_v = v->ne[1];
|
||||
|
||||
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
|
||||
GGML_ASSERT(v->ne[2] == n_tokens);
|
||||
GGML_ASSERT(k->ne[2] == n_tokens);
|
||||
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
|
||||
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
|
||||
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
|
||||
|
||||
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
|
||||
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
|
||||
|
||||
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
|
||||
|
||||
const float eps_norm = hparams.f_norm_rms_eps;
|
||||
|
||||
q = ggml_l2_norm(ctx0, q, eps_norm);
|
||||
k = ggml_l2_norm(ctx0, k, eps_norm);
|
||||
|
||||
const float scale = 1.0f / sqrtf(S_v);
|
||||
|
||||
q = ggml_scale(ctx0, q, scale);
|
||||
beta = ggml_sigmoid(ctx0, beta);
|
||||
|
||||
cb(q, "q_in", il);
|
||||
cb(k, "k_in", il);
|
||||
cb(v, "v_in", il);
|
||||
cb(beta, "beta_in", il);
|
||||
cb(g, "g_in", il);
|
||||
|
||||
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
|
||||
|
||||
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
|
||||
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
|
||||
|
||||
// Apply exponential to g_t
|
||||
g_t = ggml_exp(ctx0, g_t);
|
||||
|
||||
// Apply the gated delta rule for the single timestep
|
||||
// last_recurrent_state = last_recurrent_state * g_t
|
||||
state = ggml_mul(ctx0, state, g_t);
|
||||
|
||||
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
|
||||
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
|
||||
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
|
||||
// we need to sum over dim=-2, so we transpose, sum, then transpose again
|
||||
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
|
||||
|
||||
// v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
|
||||
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
|
||||
// delta = (v_t - kv_mem) * beta_t
|
||||
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
|
||||
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
|
||||
|
||||
// last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
|
||||
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
|
||||
state = ggml_add(ctx0, state, k_t_delta);
|
||||
|
||||
// Compute the attention output
|
||||
// core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
|
||||
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
|
||||
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
|
||||
// again, since it's over dim = -2, transpose, sum, transpose back
|
||||
ggml_tensor * core_attn_out =
|
||||
ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
|
||||
|
||||
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
|
||||
cb(core_attn_out, "output_tokens", il);
|
||||
cb(state, "new_state", il);
|
||||
|
||||
return {core_attn_out, state};
|
||||
}
|
||||
|
||||
ggml_tensor * llm_build_qwen3next::build_norm_gated(
|
||||
ggml_tensor * input,
|
||||
ggml_tensor * weights,
|
||||
|
|
|
|||
Loading…
Reference in New Issue