ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-07 15:00:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

FlashMLA-2: reduce compute buffer size (CUDA and CPU) (#260)

Browse Source 
* FlashMLA-2: eliminate intermediate f32 tensors

This works on the CPU. PP performance is ~13% better for 16k tokens
and compute buffer is quite a bit smaller.

* FlashMLA-2: enable fast path only on the CPU for now

I did implement the necessary ops on CUDA, but something is
still wrong there, so for now we only use it when running
CPU-only.

* FlashMLA-2: slightly smaller computer buffer size

* Prepare wk_b when loading DeepSeek models (if wk_b is missing)

* Add some comments

* Fix case where wkv_b is quantized with k- or i-quants.

* Fix CUDA

There is an issue with quantized GEMV on CUDA when the left operand
(the matrix) is not contiguous. So, for now, we also create wv_b
during model loading and use that instead of the 3D view of wkv_b.

* FlashMLA-2: avoid conversions to f32 also on CUDA

* Be able to compute for more than 65535 tokens

On CUDA just a quick hack that allows us to cancatenate tensors
with more than 65535 rows along zroth dimension as needed by
FlashMLA-2. Also needed some care in the perplexity tool to
avoid int overflows when evaluating the computed logits.

* Reduce memory usage for FlashMLA-2

Oh, also fix int overflow in the CUDA concat implementation.

It is funny how the llama.cpp 64-bit police has gone (almost) everywhere
and replaced 32-bit ints with 64-bit ints, needed or not,
but hasn't done it where it is actually needed.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-03-18 07:36:42 +01:00
committed by GitHub
parent 0f19a500a9
commit 9fe2b06f79
5 changed files with 184 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files

128
src/llama.cpp
View File

@@ -13755,31 +13755,52 @@ struct llm_build_context {
if (lctx.cparams.mla_attn > 1 && lctx.cparams.flash_attn && (pp_opt || lctx.cparams.mla_attn > 2)) {
auto kv_cache_nope = ggml_view_2d(ctx0, kv_self.kv_l[il], kv_lora_rank, n_kv, kv_self.kv_l[il]->nb[1], 0);
ggml_tensor * k;
ggml_tensor * v;
auto kv_f32_size = model.layers[il].wkv_b->ne[1] * kv_cache_nope->ne[1] * sizeof(float) / (1024*1024);
int n_max_head = n_head;
if (cparams.attn_max_batch > 0 && kv_f32_size > cparams.attn_max_batch) {
while (n_max_head%2 == 0 && kv_f32_size > cparams.attn_max_batch) {
n_max_head /= 2; kv_f32_size /= 2;
}
}
GGML_ASSERT(n_head % n_max_head == 0);
// For now this only works in the CPU implementation, so we only use it if there is just the CPU backend.
// If the code was compiled with CUDA (and/or Metal, Vulkan, whatever) support, this branch will not
// be taken even if no layers were offloaded to the GPU.
if (lctx.backends.size() == 1 && lctx.backends.front() == lctx.backend_cpu) {
auto n_per_head = model.layers[il].wkv_b->ne[1] / n_head;
auto kv_cache_nope = ggml_view_2d(ctx0, kv_self.kv_l[il], kv_lora_rank, n_kv, kv_self.kv_l[il]->nb[1], 0);
auto kv_cache_rope = ggml_view_3d(ctx0, kv_self.kv_l[il], n_embd_head_qk_rope, n_kv, 1,
kv_self.kv_l[il]->nb[1], kv_self.kv_l[il]->nb[2], ggml_row_size(kv_self.kv_l[il]->type, kv_lora_rank));
auto kv_f32 = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cache_nope);
ggml_tensor repeater;
repeater.ne[0] = n_embd_head_qk_rope; repeater.ne[1] = n_kv; repeater.ne[2] = n_max_head; repeater.ne[3] = 1;
auto k_rope = ggml_repeat(ctx0, kv_cache_rope, &repeater);
cb(k_rope, "k_rope", il);
auto q = ggml_concat(ctx0, q_nope, q_rope, 0);
q = ggml_permute(ctx0, q, 0, 2, 1, 3);
cb(q, "q_concat", il);
ggml_build_forward_expand(gf, q);
for (int iter = 0; iter < n_head/n_max_head; ++iter) {
auto wkv_b = ggml_view_2d(ctx0, model.layers[il].wkv_b, model.layers[il].wkv_b->ne[0], n_per_head*n_max_head,
model.layers[il].wkv_b->nb[1], model.layers[il].wkv_b->nb[1]*n_per_head*n_max_head*iter);
auto kv_f32 = ggml_mul_mat(ctx0, wkv_b, kv_cache_nope);
cb(kv_f32, "kv_f32", il);
auto v_f32 = ggml_view_3d(ctx0, kv_f32, hparams.n_embd_head_v, n_kv, n_head,
ggml_row_size(kv_f32->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
auto v_f32 = ggml_view_3d(ctx0, kv_f32, hparams.n_embd_head_v, n_kv, n_max_head,
ggml_row_size(kv_f32->type, n_max_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope));
cb(v_f32, "v_f32", il);
v = ggml_cast(ctx0, v_f32, kv_self.kv_l[il]->type);
auto v = ggml_cast(ctx0, v_f32, kv_self.kv_l[il]->type);
cb(v, "v", il);
auto k_nope_f32 = ggml_view_3d(ctx0, kv_f32, n_embd_head_qk_nope, n_kv, n_head,
ggml_row_size(kv_f32->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
auto k_nope_f32 = ggml_view_3d(ctx0, kv_f32, n_embd_head_qk_nope, n_kv, n_max_head,
ggml_row_size(kv_f32->type, n_max_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope + hparams.n_embd_head_v), 0);
cb(k_nope_f32, "k_nope_f32", il);
@@ -13789,74 +13810,27 @@ struct llm_build_context {
ggml_build_forward_expand(gf, k_nope);
ggml_build_forward_expand(gf, v);
auto kv_cache_rope = ggml_view_3d(ctx0, kv_self.kv_l[il], n_embd_head_qk_rope, n_kv, 1,
kv_self.kv_l[il]->nb[1], kv_self.kv_l[il]->nb[2], ggml_row_size(kv_self.kv_l[il]->type, kv_lora_rank));
ggml_tensor repeater;
repeater.ne[0] = n_embd_head_qk_rope; repeater.ne[1] = n_kv; repeater.ne[2] = n_head; repeater.ne[3] = 1;
auto k_rope = ggml_repeat(ctx0, kv_cache_rope, &repeater);
cb(k_rope, "k_rope", il);
k = ggml_concat(ctx0, k_nope, k_rope, 0);
auto k = ggml_concat(ctx0, k_nope, k_rope, 0);
cb(k, "k", il);
ggml_build_forward_expand(gf, k);
auto q_iter = ggml_view_3d(ctx0, q, q->ne[0], q->ne[1], n_max_head,
q->nb[1], q->nb[2], q->nb[2]*n_max_head*iter);
kqv = ggml_flash_attn_ext(ctx0, q_iter, k, v, KQ_mask, kq_scale, hparams.f_max_alibi_bias, 0.f);
if (q->ne[1] <= 8) {
ggml_flash_attn_ext_set_prec(kqv, GGML_PREC_F32);
}
cb(kqv, "kqv", il);
if (iter == 0) {
cur = ggml_reshape_2d(ctx0, kqv, n_embd_head_v*n_max_head, n_tokens);
} else {
cur = ggml_concat(ctx0, cur, ggml_reshape_2d(ctx0, kqv, n_embd_head_v*n_max_head, n_tokens), 0);
}
}
else {
// Hahaha, we need to convert the KV cache for this layer to f32 because the general purpose ML library ggml does not
// provide ops on (almost) anything other than f32. In this case, the cache will be the second operand to a matrix
// multiplication, which *must* be f32.
auto kv_cache_view = ggml_view_2d(ctx0, kv_self.kv_l[il], kv_self.kv_l[il]->ne[0], n_kv, kv_self.kv_l[il]->nb[1], 0);
auto kv_cache_view_f32 = ggml_cast(ctx0, kv_cache_view, GGML_TYPE_F32);
cb(kv_cache_view_f32, "kv_cache_view_f32", il);
// The no- and rotational position encoding portions of the KV cache
auto kv_cache_nope = ggml_view_2d(ctx0, kv_cache_view_f32, kv_lora_rank, n_kv, kv_cache_view_f32->nb[1], 0);
auto kv_cache_rope = ggml_view_3d(ctx0, kv_cache_view_f32, n_embd_head_qk_rope, 1, n_kv,
kv_cache_view_f32->nb[1], kv_cache_view_f32->nb[1], ggml_row_size(kv_cache_view_f32->type, kv_lora_rank));
auto kv_f32 = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cache_nope);
cb(kv_f32, "kv_f32", il);
auto k_nope_f32 = ggml_view_3d(ctx0, kv_f32, n_embd_head_qk_nope, n_kv, n_head,
ggml_row_size(kv_f32->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope + hparams.n_embd_head_v), 0);
cb(k_nope_f32, "k_nope_f32", il);
ggml_tensor repeater;
repeater.ne[0] = n_embd_head_qk_rope; repeater.ne[1] = n_head; repeater.ne[2] = n_kv; repeater.ne[3] = 1;
auto k_rope_f32 = ggml_permute(ctx0, ggml_repeat(ctx0, kv_cache_rope, &repeater), 0, 2, 1, 3);
cb(k_rope_f32, "k_rope_f32", il);
auto k_f32 = ggml_concat(ctx0, k_nope_f32, k_rope_f32, 0);
cb(k_f32, "k_f32", il);
k = ggml_cast(ctx0, k_f32, kv_self.kv_l[il]->type);
cb(k, "k", il);
auto v_f32 = ggml_view_3d(ctx0, kv_f32, hparams.n_embd_head_v, n_kv, n_head,
ggml_row_size(kv_f32->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
ggml_row_size(kv_f32->type, n_embd_head_qk_nope));
cb(v_f32, "v_f32", il);
v = ggml_cast(ctx0, v_f32, kv_self.kv_l[il]->type);
cb(v, "v", il);
}
auto q = ggml_concat(ctx0, q_nope, q_rope, 0);
q = ggml_permute(ctx0, q, 0, 2, 1, 3);
cb(q, "q_concat", il);
ggml_build_forward_expand(gf, q);
kqv = ggml_flash_attn_ext(ctx0, q, k, v, KQ_mask, kq_scale, hparams.f_max_alibi_bias, 0.f);
if (q->ne[1] <= 8) {
ggml_flash_attn_ext_set_prec(kqv, GGML_PREC_F32);
}
cb(kqv, "kqv", il);
cur = ggml_reshape_2d(ctx0, kqv, n_embd_head_v*n_head, n_tokens);
}
else {