Skip to content

perf: replace standard attention with unified flash attention #45

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

Open
Skyrion9 wants to merge 3 commits into
base: main
Choose a base branch
from
+76 −43
Open

perf: replace standard attention with unified flash attention #45

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a83467c
perf: replace standard attention with unified flash attention
Skyrion9 Jul 22, 2026
4fb34e9
perf: Prefer reshape over ggml_cpy for the flash-attn output to avoid…
Skyrion9 Jul 22, 2026
5aac582
perf: tree-reduced codebook sum and cont-free KV cache writes
Skyrion9 Jul 30, 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
119 changes: 76 additions & 43 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 @@ -941,14 +942,34 @@ bool SlowARModel::eval_cached(const std::vector<int32_t> & flat_tokens,
ggml_tensor * x = ggml_get_rows(ctx0, weights_.embeddings, semantic_ids);
if (x->type != GGML_TYPE_F32) x = ggml_cast(ctx0, x, GGML_TYPE_F32);

std::vector<ggml_tensor *> cb_id_tensors(hparams_.num_codebooks);
ggml_tensor * codebook_sum = nullptr;
for (int32_t cb = 0; cb < hparams_.num_codebooks; ++cb) {
const int32_t num_cb = hparams_.num_codebooks;
std::vector<ggml_tensor *> cb_id_tensors(num_cb);
std::vector<ggml_tensor *> cb_embs(num_cb);

for (int32_t cb = 0; cb < num_cb; ++cb) {
ggml_tensor * ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
cb_id_tensors[cb] = ids;
ggml_tensor * emb = ggml_get_rows(ctx0, weights_.codebook_embeddings, ids);
if (emb->type != GGML_TYPE_F32) emb = ggml_cast(ctx0, emb, GGML_TYPE_F32);
codebook_sum = (codebook_sum == nullptr) ? emb : ggml_add(ctx0, codebook_sum, emb);
cb_embs[cb] = emb;
}

ggml_tensor * codebook_sum = nullptr;
if (num_cb > 0) {
std::vector<ggml_tensor *> level(cb_embs.begin(), cb_embs.end());
while (level.size() > 1) {
std::vector<ggml_tensor *> next;
next.reserve((level.size() + 1) / 2);
for (size_t i = 0; i < level.size(); i += 2) {
if (i + 1 < level.size()) {
next.push_back(ggml_add(ctx0, level[i], level[i + 1]));
} else {
next.push_back(level[i]);
}
}
level = std::move(next);
}
codebook_sum = level[0];
}

if (codebook_sum != nullptr) {
Expand All @@ -961,6 +982,24 @@ bool SlowARModel::eval_cached(const std::vector<int32_t> & flat_tokens,
x = ggml_mul(ctx0, x, ggml_repeat(ctx0, token_scale, x));
}

ggml_tensor * fa_mask = nullptr;
std::vector<ggml_fp16_t> mask_data;

if (n_tokens > 1) {
int64_t kv_len = n_past_ + n_tokens;
fa_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kv_len, n_tokens, 1, 1);
mask_data.resize(kv_len * n_tokens);
for (int j = 0; j < n_tokens; ++j) {
for (int i = 0; i < kv_len; ++i) {
if (i > n_past_ + j) {
mask_data[j * kv_len + i] = ggml_fp32_to_fp16(-INFINITY);
} else {
mask_data[j * kv_len + i] = ggml_fp32_to_fp16(0.0f);
}
}
}
}

for (int32_t il = 0; il < hparams_.block_count; ++il) {
const auto & layer = weights_.layers[il];

Expand Down Expand Up @@ -990,53 +1029,43 @@ 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);
}

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

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_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 * 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);
// Permute Q to [head_dim, n_tokens, n_head, 1] for flash_attn_ext
ggml_tensor * Q_fa = ggml_permute(ctx0, q, 0, 2, 1, 3);

int64_t kv_len = n_past_ + n_tokens;
ggml_tensor * k_cache = ggml_view_3d(ctx0, memory_k_,
head_dim, kv_len, 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, kv_len, n_head_kv,
memory_v_->nb[1], memory_v_->nb[2], layer_off_v);

ggml_tensor * attn_fa = ggml_flash_attn_ext(
ctx0, Q_fa, k_cache, v_cache, fa_mask, attn_scale, 0.0f, 0.0f);

// Reshape [head_dim, n_head, n_tokens, 1] → [q_size, n_tokens] (zero-copy view)
ggml_tensor * attn_cur = ggml_reshape_2d(ctx0, attn_fa, q_size, n_tokens);

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 Expand Up @@ -1077,6 +1106,10 @@ bool SlowARModel::eval_cached(const std::vector<int32_t> & flat_tokens,
for (int32_t cb = 0; cb < hparams_.num_codebooks; ++cb) {
ggml_backend_tensor_set(cb_id_tensors[cb], cb_vals[cb].data(), 0, n_tokens * sizeof(int32_t));
}

if (fa_mask) {
ggml_backend_tensor_set(fa_mask, mask_data.data(), 0, mask_data.size() * sizeof(ggml_fp16_t));
}

if (ggml_backend_sched_graph_compute(sched_, gf) != GGML_STATUS_SUCCESS) {
std::fprintf(stderr, "[eval_cached] sched compute failed\n");
Expand Down