Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (#25370)

Browse Source 
* llama : make all KQ masks (except the lightning indexer one) f16 if FA is used and remove zero attention bias in DeepSeek V4

* llama : remove dead code that repeats unified raw_k cache for each stream in DeepSeek V4 - no longer needed as raw_k is always non-unified.

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming
2026-07-10 09:06:58 +02:00
committed by GitHub
co-authored by Stanisław Szymczyk
parent 082b326fc7
commit 2ed3c1abbb
2 changed files with 30 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+26 -12
View File
@@ -646,7 +646,7 @@ static void dsv4_set_kq_mask(
return;
}
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(n_stream > 0);
GGML_ASSERT(n_tokens%n_stream == 0);
GGML_ASSERT(dst->ne[0] == plan.n_kv);
@@ -656,13 +656,27 @@ static void dsv4_set_kq_mask(
GGML_ASSERT((int64_t) plan.n_visible.size() == (int64_t) n_tokens);
GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
float * data = (float *) dst->data;
if (dst->type == GGML_TYPE_F32) {
float * data = (float *) dst->data;
for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
const int32_t n_visible = plan.n_visible[i];
for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
const int32_t n_visible = plan.n_visible[i];
for (int64_t j = 0; j < dst->ne[0]; ++j) {
data[i*dst->ne[0] + j] = j < n_visible ? 0.0f : -INFINITY;
for (int64_t j = 0; j < dst->ne[0]; ++j) {
data[i*dst->ne[0] + j] = j < n_visible ? 0.0f : -INFINITY;
}
}
} else if (dst->type == GGML_TYPE_F16) {
ggml_fp16_t * data = (ggml_fp16_t *) dst->data;
const ggml_fp16_t fp16_ninf = llama_cast<ggml_fp16_t>(-INFINITY);
const ggml_fp16_t fp16_zero = llama_cast<ggml_fp16_t>(0.0f);
for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
const int32_t n_visible = plan.n_visible[i];
for (int64_t j = 0; j < dst->ne[0]; ++j) {
data[i*dst->ne[0] + j] = j < n_visible ? fp16_zero : fp16_ninf;
}
}
}
}
@@ -679,8 +693,7 @@ static ggml_tensor * dsv4_build_raw_kq_mask(
GGML_ASSERT(n_stream > 0);
GGML_ASSERT(n_tokens%n_stream == 0);
const bool use_fattn = cparams.flash_attn && (!cparams.kv_unified || n_stream == 1);
const auto type = use_fattn ? GGML_TYPE_F16 : GGML_TYPE_F32;
const auto type = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
ggml_tensor * res = ggml_new_tensor_4d(ctx, type, n_kv, n_tokens/n_stream, 1, n_stream);
ggml_set_input(res);
@@ -814,6 +827,7 @@ static void dsv4_build_comp_inputs(
llm_graph_input_dsv4::comp_input & inp,
const llama_kv_cache_dsv4_context::comp_plan & plan,
const char * name,
const llama_cparams & cparams,
int64_t n_stream) {
inp.state_pos = dsv4_build_input_1d(ctx, GGML_TYPE_I32, plan.state_pos.size(), std::string("dsv4_") + name + "_state_pos");
inp.state_persist_src_idxs = dsv4_build_input_1d(ctx, GGML_TYPE_I32, plan.state_persist_src_idxs.size(), std::string("dsv4_") + name + "_state_persist_src_idxs");
@@ -828,7 +842,7 @@ static void dsv4_build_comp_inputs(
GGML_ASSERT(n_stream > 0);
GGML_ASSERT(n_tokens%n_stream == 0);
inp.kq_mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
inp.kq_mask = ggml_new_tensor_4d(ctx, cparams.flash_attn && strcmp(name, "lid") != 0 ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
ggml_set_input(inp.kq_mask);
ggml_set_name(inp.kq_mask, (std::string("dsv4_") + name + "_kq_mask").c_str());
}
@@ -3076,9 +3090,9 @@ llm_graph_input_dsv4 * llm_graph_context::build_inp_dsv4() const {
inp_raw->self_k_rot = raw_ctx->build_input_k_rot(ctx0);
auto inp = std::make_unique<llm_graph_input_dsv4>(cparams, std::move(inp_raw), mctx_cur);
dsv4_build_comp_inputs(ctx0, inp->inp_csa, mctx_cur->get_csa_plan(ubatch), "csa", n_stream);
dsv4_build_comp_inputs(ctx0, inp->inp_hca, mctx_cur->get_hca_plan(ubatch), "hca", n_stream);
dsv4_build_comp_inputs(ctx0, inp->inp_lid, mctx_cur->get_lid_plan(ubatch), "lid", n_stream);
dsv4_build_comp_inputs(ctx0, inp->inp_csa, mctx_cur->get_csa_plan(ubatch), "csa", cparams, n_stream);
dsv4_build_comp_inputs(ctx0, inp->inp_hca, mctx_cur->get_hca_plan(ubatch), "hca", cparams, n_stream);
dsv4_build_comp_inputs(ctx0, inp->inp_lid, mctx_cur->get_lid_plan(ubatch), "lid", cparams, n_stream);
inp->inp_csa.k_rot = mctx_cur->get_csa()->build_input_k_rot(ctx0);
inp->inp_hca.k_rot = mctx_cur->get_hca()->build_input_k_rot(ctx0);
inp->inp_lid.k_rot = mctx_cur->get_lid()->build_input_k_rot(ctx0);