Skip to content

feat: Flash Attention and modernized KV Cache layout for Slow-AR #40

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
Skyrion9 wants to merge 3 commits into from
Closed

feat: Flash Attention and modernized KV Cache layout for Slow-AR #40

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1dec132
feat: Flash Attention and modernized KV Cache layout for Slow-AR
Skyrion9 Jul 14, 2026
e62f190
perf(eval_cached): remove redundant permute in the single token FA co…
Skyrion9 Jul 15, 2026
e7506ba
style: fix messy indentation introduced earlier in this PR in get_wei…
Skyrion9 Jul 15, 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
113 changes: 72 additions & 41 deletions src/s2_model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -743,8 +743,9 @@ bool SlowARModel::init_kv_cache(int32_t max_seq_len) {
return false;
}

memory_k_ = ggml_new_tensor_4d(ctx_kv_, GGML_TYPE_F16, head_dim, n_head_kv, max_seq_len, n_layer);
memory_v_ = ggml_new_tensor_4d(ctx_kv_, GGML_TYPE_F16, head_dim, n_head_kv, max_seq_len, n_layer);
// Layout for flash attention
memory_k_ = ggml_new_tensor_4d(ctx_kv_, GGML_TYPE_F16, head_dim, max_seq_len, n_head_kv, n_layer);
memory_v_ = ggml_new_tensor_4d(ctx_kv_, GGML_TYPE_F16, head_dim, max_seq_len, n_head_kv, n_layer);

ggml_backend_t kv_backend = (n_gpu_layers_ > 0 && backend_gpu_) ? backend_gpu_ : backend_cpu_;
kv_buf_ = ggml_backend_alloc_ctx_tensors(ctx_kv_, kv_backend);
Expand Down Expand Up @@ -952,7 +953,6 @@ bool SlowARModel::eval_cached(const std::vector<int32_t> & flat_tokens,
}

if (codebook_sum != nullptr) {

codebook_sum = ggml_mul(ctx0, codebook_sum,
ggml_repeat(ctx0, semantic_mask, codebook_sum));
x = ggml_add(ctx0, x, codebook_sum);
Expand Down Expand Up @@ -990,53 +990,84 @@ bool SlowARModel::eval_cached(const std::vector<int32_t> & flat_tokens,

const size_t layer_off_k = static_cast<size_t>(il) * memory_k_->nb[3];
const size_t layer_off_v = static_cast<size_t>(il) * memory_v_->nb[3];
const size_t token_off_k = static_cast<size_t>(n_past_) * memory_k_->nb[2];
const size_t token_off_v = static_cast<size_t>(n_past_) * memory_v_->nb[2];
const size_t token_off_k = static_cast<size_t>(n_past_) * memory_k_->nb[1];
const size_t token_off_v = static_cast<size_t>(n_past_) * memory_v_->nb[1];

ggml_tensor * k_slot = ggml_view_3d(ctx0, memory_k_,
head_dim, n_head_kv, n_tokens,
head_dim, n_tokens, n_head_kv,
memory_k_->nb[1], memory_k_->nb[2],
layer_off_k + token_off_k);
ggml_tensor * v_slot = ggml_view_3d(ctx0, memory_v_,
head_dim, n_head_kv, n_tokens,
head_dim, n_tokens, n_head_kv,
memory_v_->nb[1], memory_v_->nb[2],
layer_off_v + token_off_v);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, k, k_slot));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, v, v_slot));

ggml_tensor * k_mem = k;
ggml_tensor * v_mem = v;
if (n_past_ > 0) {
ggml_tensor * k_past = ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, memory_k_, static_cast<int64_t>(n_past_) * kv_size, layer_off_k),
head_dim, n_head_kv, n_past_);
ggml_tensor * v_past = ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, memory_v_, static_cast<int64_t>(n_past_) * kv_size, layer_off_v),
head_dim, n_head_kv, n_past_);
if (k_past->type != k->type) k_past = ggml_cast(ctx0, k_past, k->type);
if (v_past->type != v->type) v_past = ggml_cast(ctx0, v_past, v->type);
k_mem = ggml_concat(ctx0, k_past, k, 2);
v_mem = ggml_concat(ctx0, v_past, v, 2);
}

if (n_head != n_head_kv && q->type != GGML_TYPE_F32) {
q = ggml_cast(ctx0, q, GGML_TYPE_F32);

// Permute k and v to match the cache layout
ggml_tensor * k_perm = ggml_cont(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3));
ggml_tensor * v_perm = ggml_cont(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3));

ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_perm, k_slot));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_perm, v_slot));

ggml_tensor * attn_cur = nullptr;

if (n_tokens == 1) {
// Flash attention path
ggml_tensor * Q = ggml_permute(ctx0, q, 0, 2, 1, 3);

ggml_tensor * k_cache = ggml_view_3d(ctx0, memory_k_,
head_dim, n_past_ + 1, n_head_kv,
memory_k_->nb[1], memory_k_->nb[2], layer_off_k);

ggml_tensor * v_cache = ggml_view_3d(ctx0, memory_v_,
head_dim, n_past_ + 1, n_head_kv,
memory_v_->nb[1], memory_v_->nb[2], layer_off_v);

ggml_tensor * attn_fa = ggml_flash_attn_ext(
ctx0, Q, k_cache, v_cache, nullptr, attn_scale, 0.0f, 0.0f);

// attn_fa is [head_dim, n_head, n_tokens=1, 1]; with n_tokens==1 no reorder is needed.
attn_cur = ggml_cpy(ctx0, attn_fa,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, q_size, n_tokens));
} else {
// Prefill path (keep concat logic)
ggml_tensor * k_mem = k;
ggml_tensor * v_mem = v;
if (n_past_ > 0) {
ggml_tensor * k_past_view = ggml_view_3d(ctx0, memory_k_,
head_dim, n_past_, n_head_kv,
memory_k_->nb[1], memory_k_->nb[2], layer_off_k);
ggml_tensor * v_past_view = ggml_view_3d(ctx0, memory_v_,
head_dim, n_past_, n_head_kv,
memory_v_->nb[1], memory_v_->nb[2], layer_off_v);

ggml_tensor * k_past = ggml_cont(ctx0, ggml_permute(ctx0, k_past_view, 0, 2, 1, 3));
ggml_tensor * v_past = ggml_cont(ctx0, ggml_permute(ctx0, v_past_view, 0, 2, 1, 3));

if (k_past->type != k->type) k_past = ggml_cast(ctx0, k_past, k->type);
if (v_past->type != v->type) v_past = ggml_cast(ctx0, v_past, v->type);

k_mem = ggml_concat(ctx0, k_past, k, 2);
v_mem = ggml_concat(ctx0, v_past, v, 2);
}
if (n_head != n_head_kv && q->type != GGML_TYPE_F32) {
q = ggml_cast(ctx0, q, GGML_TYPE_F32);
}
ggml_tensor * k_rep = repeat_interleave_heads(ctx0, k_mem, n_head / n_head_kv);
ggml_tensor * v_rep = repeat_interleave_heads(ctx0, v_mem, n_head / n_head_kv);
ggml_tensor * Q = ggml_permute(ctx0, q, 0, 2, 1, 3);
ggml_tensor * K = ggml_permute(ctx0, k_rep, 0, 2, 1, 3);
ggml_tensor * KQ = mul_mat_checked(ctx0, K, Q, "mul_mat:kq");
ggml_tensor * KQs = ggml_scale(ctx0, KQ, attn_scale);
ggml_tensor * KQm = ggml_diag_mask_inf(ctx0, KQs, n_past_);
ggml_tensor * KQf = ggml_soft_max(ctx0, KQm);
ggml_tensor * V = ggml_cont(ctx0, ggml_permute(ctx0, v_rep, 1, 2, 0, 3));
ggml_tensor * KQV = mul_mat_checked(ctx0, V, KQf, "mul_mat:kqv");
ggml_tensor * KQVm = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
attn_cur = ggml_cpy(ctx0, KQVm,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, q_size, n_tokens));
}
ggml_tensor * k_rep = repeat_interleave_heads(ctx0, k_mem, n_head / n_head_kv);
ggml_tensor * v_rep = repeat_interleave_heads(ctx0, v_mem, n_head / n_head_kv);

ggml_tensor * Q = ggml_permute(ctx0, q, 0, 2, 1, 3);
ggml_tensor * K = ggml_permute(ctx0, k_rep, 0, 2, 1, 3);
ggml_tensor * KQ = mul_mat_checked(ctx0, K, Q, "mul_mat:kq");
ggml_tensor * KQs = ggml_scale(ctx0, KQ, attn_scale);
ggml_tensor * KQm = ggml_diag_mask_inf(ctx0, KQs, n_past_);
ggml_tensor * KQf = ggml_soft_max(ctx0, KQm);

ggml_tensor * V = ggml_cont(ctx0, ggml_permute(ctx0, v_rep, 1, 2, 0, 3));
ggml_tensor * KQV = mul_mat_checked(ctx0, V, KQf, "mul_mat:kqv");
ggml_tensor * KQVm = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
ggml_tensor * attn_cur = ggml_cpy(ctx0, KQVm,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, q_size, n_tokens));
ggml_tensor * attn_out = mul_mat_checked(ctx0, layer.wo, attn_cur, "mul_mat:wo");

ggml_tensor * h = ggml_add(ctx0, x, attn_out);
Expand Down