8000 Optimized DeepSeek V2/V3 implementation (MLA) by fairydreaming · Pull Request #11446 · ggml-org/llama.cpp · GitHub
[go: up one dir, main page]
Skip to content

Optimized DeepSeek V2/V3 implementation (MLA) #11446

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
wants to merge 10 commits into from
+203 −30
Closed

Optimized DeepSeek V2/V3 implementation (MLA) #11446

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
93864cd
llama : experimental DeepSeek2 MLA implementation that caches latent …
sszymczy Jan 22, 2025
f0ce53f
Merge remote-tracking branch 'origin/master' into deepseek2-mla-exp
sszymczy Jan 24, 2025
de538aa
llama : optimize DeepSeek MLA implementation
sszymczy Jan 25, 2025
ce73063
llama : Update tensor names in DeepSeek2 MLA implementation.
sszymczy Jan 26, 2025
202f323
llama : add a second copy of c^KV cache in DeepSeek2 MLA to avoid tra…
sszymczy Jan 26, 2025
93c5937
llama : modified tensor pe 8000 rmutations to multiply larger matrices duri…
sszymczy Jan 26, 2025
1eee98f
llama : removed unnecessary code in DeepSeek V2 implementation
sszymczy Jan 27, 2025
8ff0991
convert : make lint happy
sszymczy Jan 28, 2025
8a887de
llama : prompt processing optimizations in DeepSeek V2
sszymczy Jan 28, 2025
7654331
llama : avoid ggml_cont() is possible in DeepSeek V2 implementation
sszymczy Jan 30, 2025
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
Prev Previous commit
Next Next commit
llama : add a second copy of c^KV cache in DeepSeek2 MLA to avoid tra… 
…nsposing the cache during inference
  • Loading branch information
@sszymczy
sszymczy committed Jan 26, 2025
commit 202f323e66809bb1df192245caddc49471660466
6 changes: 5 additions & 1 deletion src/llama-kv-cache.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ bool llama_kv_cache_init(
auto it = ctx_map.find(buft);
if (it == ctx_map.end()) {
struct ggml_init_params params = {
/*.mem_size =*/ size_t(4u*n_layer*ggml_tensor_overhead()),
/*.mem_size =*/ size_t(5u*n_layer*ggml_tensor_overhead()),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
Expand All @@ -74,6 +74,7 @@ bool llama_kv_cache_init(
// DeepSeek MLA
cache.kr_l.reserve(n_layer);
cache.kv_l.reserve(n_layer);
cache.kvt_l.reserve(n_layer);

for (int i = 0; i < n_layer; i++) {
const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
Expand Down Expand Up @@ -108,10 +109,13 @@ bool llama_kv_cache_init(
LLAMA_LOG_DEBUG("%s: layer %d: n_embd_head_qk_rope = %d, kv_lora_rank = %d\n", __func__, i, n_embd_head_qk_rope, kv_lora_rank);
ggml_tensor * kr = ggml_new_tensor_1d(ctx, cache.type_kr, n_embd_head_qk_rope*kv_size);
ggml_tensor * kv = ggml_new_tensor_1d(ctx, cache.type_kv, kv_lora_rank*kv_size);
ggml_tensor * kvt = ggml_new_tensor_1d(ctx, cache.type_kv, kv_lora_rank*kv_size);
ggml_format_name(kr, "cache_kr_l%d", i);
ggml_format_name(kv, "cache_kv_l%d", i);
ggml_format_name(kvt, "cache_kvt_l%d", i);
cache.kr_l.push_back(kr);
cache.kv_l.push_back(kv);
cache.kvt_l.push_back(kvt);
}

// allocate tensors and initialize the buffers to avoid NaNs in the padding
Expand Down
1 change: 1 addition & 0 deletions src/llama-kv-cache.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,7 @@ struct llama_kv_cache {
// DeepSeek MLA
std::vector<struct ggml_tensor *> kr_l; // per layer
std::vector<struct ggml_tensor *> kv_l;
std::vector<struct ggml_tensor *> kvt_l;

std::vector<ggml_context_ptr> ctxs;
std::vector<ggml_backend_buffer_ptr> bufs;
Expand Down
16 changes: 13 additions & 3 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6476,13 +6476,26 @@ struct llm_build_context {
// note: storing c^KV in the KV cache
ggml_build_forward_expand(gf, ggml_cpy(ctx0, kv_compressed, kv_cache_view));

struct ggml_tensor * kv_cache_trans_view = ggml_view_2d(ctx0, kv_self.kvt_l[il], n_tokens, kv_lora_rank, ggml_row_size(kv_self.kv_l[il]->type, kv_self.size), ggml_row_size(kv_self.kv_l[il]->type, kv_head));
cb(kv_cache_trans_view, "kv_cache_trans_view", il);

// note: storing transposed c^KV in the transposed KV cache
ggml_build_forward_expand(gf, ggml_cpy(ctx0, ggml_transpose(ctx0, kv_compressed), kv_cache_trans_view));

struct ggml_tensor * kv_cache =
ggml_view_2d(ctx0, kv_self.kv_l[il],
kv_lora_rank, n_kv,
ggml_row_size(kv_self.kv_l[il]->type, kv_lora_rank),
0);
cb(kv_cache, "kv_cache", il);

struct ggml_tensor * kv_cache_trans =
ggml_view_2d(ctx0, kv_self.kvt_l[il],
n_kv, kv_lora_rank,
ggml_row_size(kv_self.kv_l[il]->type, kv_self.size),
0);
cb(kv_cache_trans, "kv_cache_trans", il);

q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
q_pe = ggml_rope_ext(
ctx0, q_pe, inp_pos, nullptr,
Expand Down Expand Up @@ -6552,9 +6565,6 @@ struct llm_build_context {
struct ggml_tensor * kq_perm = ggml_permute(ctx0, kq, 0, 2, 3, 1);
cb(kq_perm, "kq_soft_max_ext_perm", il);

struct ggml_tensor * kv_cache_trans = ggml_cont(ctx0, ggml_transpose(ctx0, kv_cache));
cb(kv_cache_trans, "kv_cache_trans", il);

struct ggml_tensor * kqv_compressed = ggml_mul_mat(ctx0, kv_cache_trans, kq_perm);
cb(kqv_compressed, "kqv_compressed", il);

Expand Down
0