Skip to content

graph : make FA compatible with MLA + add initial Metal kernels #12953

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
merged 5 commits into from
Apr 17, 2025
Merged

graph : make FA compatible with MLA + add initial Metal kernels #12953

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e330856
graph : make mla compatible with FA
ggerganov Apr 13, 2025
9b64dcc
metal : add exp FA kernels for DeepSeek models
ggerganov Apr 13, 2025
9cc85dd
llama : minor naming updates
ggerganov Apr 15, 2025
43c762b
ggml : disable FA for DS head sizes
ggerganov Apr 15, 2025
facdf87
tests : add FA tests for MLA shapes
ggerganov Apr 15, 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
4 changes: 4 additions & 0 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3237,6 +3237,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
if (op->src[0]->ne[0] == 192) {
return false;
}
if (op->src[0]->ne[0] == 576) {
// DeepSeek MLA
return false;
}
if (op->src[0]->ne[3] != 1) {
return false;
}
Expand Down
83 changes: 77 additions & 6 deletions ggml/src/ggml-metal/ggml-metal.m
View file
Open in desktop

Large diffs are not rendered by default.

17 changes: 17 additions & 0 deletions ggml/src/ggml-metal/ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3546,6 +3546,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_f16_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 192, 192>;
template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 192, 128>;
template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 256, 256>;
template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 576, 512>;

#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 64, 64>;
Expand All @@ -3556,6 +3557,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_bf16_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 192, 192>;
template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 192, 128>;
template [[host_name("kernel_flash_attn_ext_bf16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 256, 256>;
template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 576, 512>;
#endif

template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64, 64>;
Expand All @@ -3566,6 +3568,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q4_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q4_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q4_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;

template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64, 64>;
template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80, 80>;
Expand All @@ -3575,6 +3578,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q4_1_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q4_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q4_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;

template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64, 64>;
template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80, 80>;
Expand All @@ -3584,6 +3588,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q5_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q5_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q5_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;

template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64, 64>;
template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80, 80>;
Expand All @@ -3593,6 +3598,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q5_1_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q5_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q5_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;

template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64, 64>;
template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80, 80>;
Expand All @@ -3602,6 +3608,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q8_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q8_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;

#undef FA_TYPES

Expand Down Expand Up @@ -4009,6 +4016,16 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_
template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1, 8, dequantize_q5_1_t4, 256, 256, 4>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0, 8, dequantize_q8_0_t4, 256, 256, 4>;

template [[host_name("kernel_flash_attn_ext_vec_f16_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 576, 512, 2>;
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 576, 512, 2>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0, 8, dequantize_q4_0_t4, 576, 512, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1, 8, dequantize_q4_1_t4, 576, 512, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0, 8, dequantize_q5_0_t4, 576, 512, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1, 8, dequantize_q5_1_t4, 576, 512, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0, 8, dequantize_q8_0_t4, 576, 512, 2>;

#undef FA_TYPES

template<typename T>
Expand Down
1 change: 1 addition & 0 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9261,6 +9261,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case 112:
case 128:
case 256:
case 575: // DeepSeek MLA
break;
default:
return false;
Expand Down
11 changes: 3 additions & 8 deletions src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -484,7 +484,7 @@ ggml_tensor * llama_context::build_rope_shift(

// See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float yarn_attn_factor_scaled = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;

ggml_tensor * tmp;

Expand All @@ -504,14 +504,14 @@ ggml_tensor * llama_context::build_rope_shift(

tmp = ggml_rope_ext_inplace(ctx0, tmp,
shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
yarn_ext_factor, yarn_attn_factor_scaled, yarn_beta_fast, yarn_beta_slow);
yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);

tmp = ggml_cpy(ctx0, tmp, cur);
} else {
// we rotate only the first n_rot dimensions
tmp = ggml_rope_ext_inplace(ctx0, cur,
shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
yarn_ext_factor, yarn_attn_factor_scaled, yarn_beta_fast, yarn_beta_slow);
yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
}

return tmp;
Expand Down Expand Up @@ -2278,11 +2278,6 @@ llama_context * llama_init_from_model(
params.flash_attn = false;
}

if (params.flash_attn && model->arch == LLM_ARCH_DEEPSEEK2) {
LLAMA_LOG_WARN("%s: flash_attn is not compatible with Deepseek2 - forcing off\n", __func__);
params.flash_attn = false;
}

if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
return nullptr;
Expand Down
14 changes: 8 additions & 6 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1200,9 +1200,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
//const auto & n_embd_head_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v;

// note: for MLA with the absorption optimization, the final embedding size will be changed via v_mla
const auto n_embd_head_v = v_mla == nullptr ? v_trans ? v->ne[1] : v->ne[0] : v_mla->ne[1];

const auto n_tokens = q->ne[1];
const auto n_head = q->ne[2];
const auto n_kv = k->ne[1];
Expand Down Expand Up @@ -1231,7 +1228,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(

ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);

cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
if (v_mla) {
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
cur = ggml_mul_mat(ctx0, v_mla, cur);
}

cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
} else {
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);

Expand Down Expand Up @@ -1274,9 +1276,9 @@ ggml_tensor * llm_graph_context::build_attn_mha(
kqv = ggml_mul_mat(ctx0, v_mla, kqv);
}

ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);

cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);

if (!cparams.offload_kqv) {
// all nodes between the KV store and the attention output are run on the CPU
Expand Down
6 changes: 3 additions & 3 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10050,7 +10050,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));

ggml_tensor * cur;
ggml_tensor * inpL;
Expand Down Expand Up @@ -10127,13 +10127,13 @@ struct llm_build_deepseek2 : public llm_graph_context {

q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor_scaled, beta_fast, beta_slow
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(q_pe, "q_pe", il);

k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor_scaled, beta_fast, beta_slow
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(k_pe, "k_pe", il);

Expand Down
7 changes: 4 additions & 3 deletions tests/test-backend-ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4428,10 +4428,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_timestep_embedding());
test_cases.emplace_back(new test_leaky_relu());

for (int hsk : { 64, 80, 128, 192, 256, }) {
for (int hsv : { 64, 80, 128, 192, 256, }) {
if (hsk != 192 && hsk != hsv) continue;
for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA

for (bool mask : { true, false } ) {
for (float max_bias : { 0.0f, 8.0f }) {
Expand Down
Loading