Skip to content

llama: use f16 mask for FA to save VRAM #23764

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

Merged
am17an merged 3 commits into from
May 29, 2026
Merged

llama: use f16 mask for FA to save VRAM #23764

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
483c41e
llama: use f16 mask for FA
am17an May 27, 2026
2f5190c
review: add llama_cast + formatting
am17an May 28, 2026
089e265
simplify
am17an May 29, 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
122 changes: 74 additions & 48 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,10 @@ static ggml_tensor * build_attn_inp_kq_mask(
const auto n_tokens = ubatch.n_tokens;
const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;

ggml_tensor * res = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
// flash attention requires an f16 mask
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);
ggml_set_name(res, "attn_inp_kq_mask");

Expand Down Expand Up @@ -348,7 +351,8 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
}
}

static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
template <typename T>
static void print_mask(const T * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
const char * swa_type_str = "unknown";

Expand All @@ -372,7 +376,12 @@ static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64
for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
LLAMA_LOG_DEBUG(" %2d ", i);
for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
float val = data[i * n_kv + j];
float val;
if constexpr (std::is_same_v<T, ggml_fp16_t>) {
val = ggml_fp16_to_fp32(data[i * n_kv + j]);
} else {
val = data[i * n_kv + j];
}
Comment thread
am17an marked this conversation as resolved.
Outdated
if (val == -INFINITY) {
LLAMA_LOG_DEBUG(" ∞");
} else {
Expand All @@ -387,7 +396,8 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
const int64_t n_kv = ubatch->n_tokens;
const int64_t n_tokens = ubatch->n_tokens;

const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
const auto fill_mask_inner = [&](auto * data, int n_swa, llama_swa_type swa_type) {
using T = std::remove_reference_t<decltype(*data)>;
for (int i1 = 0; i1 < n_tokens; ++i1) {
const llama_seq_id s1 = ubatch->seq_id[i1][0];
const llama_pos p1 = ubatch->pos[i1];
Expand All @@ -413,39 +423,42 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
continue;
}

data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
data[idst + i0] = llama_cast<T>(hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f);
}
}
};

{
GGML_ASSERT(self_kq_mask);
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
const auto fill_mask = [&](ggml_tensor * mask, int n_swa, llama_swa_type swa_type) {
GGML_ASSERT(mask);
GGML_ASSERT(ggml_backend_buffer_is_host(mask->buffer));

float * data = (float *) self_kq_mask->data;
if (mask->type == GGML_TYPE_F16) {
ggml_fp16_t * data = (ggml_fp16_t *) mask->data;

std::fill(data, data + ggml_nelements(self_kq_mask), -INFINITY);
std::fill(data, data + ggml_nelements(mask), llama_cast<ggml_fp16_t>(-INFINITY));

fill_mask(data, 0, LLAMA_SWA_TYPE_NONE);
fill_mask_inner(data, n_swa, swa_type);

if (debug) {
print_mask(data, n_tokens, n_kv, 0, LLAMA_SWA_TYPE_NONE);
}
}
if (debug) {
print_mask(data, n_tokens, n_kv, n_swa, swa_type);
}
} else {
float * data = (float *) mask->data;

if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
GGML_ASSERT(self_kq_mask_swa);
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
std::fill(data, data + ggml_nelements(mask), -INFINITY);

float * data = (float *) self_kq_mask_swa->data;
fill_mask_inner(data, n_swa, swa_type);

std::fill(data, data + ggml_nelements(self_kq_mask_swa), -INFINITY);
if (debug) {
print_mask(data, n_tokens, n_kv, n_swa, swa_type);
}
}
};

fill_mask(data, hparams.n_swa, hparams.swa_type);
fill_mask(self_kq_mask, 0, LLAMA_SWA_TYPE_NONE);

if (debug) {
print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
}
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
fill_mask(self_kq_mask_swa, hparams.n_swa, hparams.swa_type);
}
}

Expand Down Expand Up @@ -568,23 +581,30 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing

float * data = (float *) cross_kq_mask->data;
const auto fill_mask = [&](auto * data) {
using T = std::remove_reference_t<decltype(*data)>;
for (int i = 0; i < n_tokens; ++i) {
GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called first");
for (int j = 0; j < n_enc; ++j) {
float f = -INFINITY;

for (int i = 0; i < n_tokens; ++i) {
GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called first");
for (int j = 0; j < n_enc; ++j) {
float f = -INFINITY;
for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
const llama_seq_id seq_id = ubatch->seq_id[i][s];

for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
const llama_seq_id seq_id = ubatch->seq_id[i][s];

if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
f = 0.0f;
if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
f = 0.0f;
}
}
}

data[i*n_enc + j] = f;
data[i*n_enc + j] = llama_cast<T>(f);
}
}
};

if (cross_kq_mask->type == GGML_TYPE_F16) {
fill_mask((ggml_fp16_t *) cross_kq_mask->data);
} else {
fill_mask((float *) cross_kq_mask->data);
}
}

Expand Down Expand Up @@ -2088,17 +2108,20 @@ ggml_tensor * llm_graph_context::build_attn_mha(
llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const {
auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);

// flash attention requires an f16 mask
const auto type_mask = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;

// note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, type_mask, n_tokens, n_tokens, 1, 1);
ggml_set_input(inp->self_kq_mask);

inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
inp->self_kq_mask_cnv = inp->self_kq_mask;

if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, type_mask, n_tokens, n_tokens, 1, 1);
ggml_set_input(inp->self_kq_mask_swa);

inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
inp->self_kq_mask_swa_cnv = inp->self_kq_mask_swa;
} else {
inp->self_kq_mask_swa = nullptr;
inp->self_kq_mask_swa_cnv = nullptr;
Expand Down Expand Up @@ -2175,7 +2198,7 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);

inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur, ubatch, cparams);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
inp->self_kq_mask_cnv = inp->self_kq_mask;
}

inp->self_k_rot = mctx_cur->build_input_k_rot(ctx0);
Expand Down Expand Up @@ -2282,7 +2305,7 @@ static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);

inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur, ubatch, cparams);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
inp->self_kq_mask_cnv = inp->self_kq_mask;
}

return inp;
Expand Down Expand Up @@ -2446,10 +2469,13 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {

const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;

inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, n_tokens, 1, 1);
// flash attention requires an f16 mask
const auto type = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
Comment thread
am17an marked this conversation as resolved.
Outdated

inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, type, n_enc, n_tokens, 1, 1);
ggml_set_input(inp->cross_kq_mask);

inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
inp->cross_kq_mask_cnv = inp->cross_kq_mask;

return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
}
Expand Down Expand Up @@ -2510,7 +2536,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);

inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_cur->get_base(), ubatch, cparams);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
inp->self_kq_mask_cnv = inp->self_kq_mask;
}

{
Expand All @@ -2520,7 +2546,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);

inp->self_kq_mask_swa = build_attn_inp_kq_mask(ctx0, mctx_cur->get_swa(), ubatch, cparams);
inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
inp->self_kq_mask_swa_cnv = inp->self_kq_mask_swa;
}

inp->self_k_rot = mctx_cur->get_base()->build_input_k_rot(ctx0);
Expand Down Expand Up @@ -2689,15 +2715,15 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()
inp_attn->self_v_idxs = attn_ctx->get_base()->build_input_v_idxs(ctx0, ubatch);

inp_attn->self_kq_mask = build_attn_inp_kq_mask(ctx0, attn_ctx->get_base(), ubatch, cparams);
inp_attn->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask, GGML_TYPE_F16) : inp_attn->self_kq_mask;
inp_attn->self_kq_mask_cnv = inp_attn->self_kq_mask;
}

{
inp_attn->self_k_idxs_swa = attn_ctx->get_swa()->build_input_k_idxs(ctx0, ubatch);
inp_attn->self_v_idxs_swa = attn_ctx->get_swa()->build_input_v_idxs(ctx0, ubatch);

inp_attn->self_kq_mask_swa = build_attn_inp_kq_mask(ctx0, attn_ctx->get_swa(), ubatch, cparams);
inp_attn->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask_swa, GGML_TYPE_F16) : inp_attn->self_kq_mask_swa;
inp_attn->self_kq_mask_swa_cnv = inp_attn->self_kq_mask_swa;
}

auto inp = std::make_unique<llm_graph_input_mem_hybrid_iswa>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
Expand Down
28 changes: 14 additions & 14 deletions src/llama-graph.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -274,10 +274,10 @@ class llm_graph_input_attn_no_cache : public llm_graph_input_i {
ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }

// n_tokens == n_batch
ggml_tensor * self_kq_mask = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]

const llama_hparams hparams;
const llama_cparams cparams;
Expand Down Expand Up @@ -307,8 +307,8 @@ class llm_graph_input_attn_kv : public llm_graph_input_i {
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]

ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]

// note: assumes v_rot^2 == I
ggml_tensor * self_k_rot = nullptr;
Expand Down Expand Up @@ -347,8 +347,8 @@ class llm_graph_input_attn_k : public llm_graph_input_i {

ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]

ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]

const llama_hparams hparams;
const llama_cparams cparams;
Expand Down Expand Up @@ -385,10 +385,10 @@ class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]

ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]

ggml_tensor * self_k_rot = nullptr;
ggml_tensor * self_v_rot = nullptr;
Expand All @@ -411,8 +411,8 @@ class llm_graph_input_attn_cross : public llm_graph_input_i {

ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }

ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
ggml_tensor * cross_kq_mask = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]
ggml_tensor * cross_kq_mask_cnv = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]

const llama_cross * cross = nullptr;
};
Expand Down
12 changes: 12 additions & 0 deletions src/llama-impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
#include "ggml.h" // for ggml_log_level

#include <string>
#include <type_traits>
#include <vector>

#ifdef __GNUC__
Expand Down Expand Up @@ -40,6 +41,17 @@ struct no_init {
no_init() = default;
};

template <typename dst_t, typename src_t>
static inline dst_t llama_cast(src_t v) {
if constexpr (std::is_same_v<src_t, dst_t>) {
return v;
} else if (std::is_same_v<src_t, ggml_fp16_t> && std::is_same_v<dst_t, float>) {
return ggml_fp16_to_fp32(v);
} else if (std::is_same_v<src_t, float> && std::is_same_v<dst_t, ggml_fp16_t>) {
return ggml_fp32_to_fp16(v);
}
Comment thread
am17an marked this conversation as resolved.
Outdated
}

struct time_meas {
time_meas(int64_t & t_acc, bool disable = false);
~time_meas();
Expand Down
Loading
Loading