Skip to content

Sync master with upstream release b6337 #231

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 10 commits into from
Sep 1, 2025
Merged

Sync master with upstream release b6337 #231

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
94e82c7
vulkan: clamp matmul and FA results to the max finite value (#15652)
jeffbolznv Aug 31, 2025
b97c9ed
vulkan: Allow fallback to sysmem memory when vidmem is full (#15649)
jeffbolznv Aug 31, 2025
5c16b9c
vulkan : remove unused portability_enumeration_ext variable (#15679)
danbev Aug 31, 2025
c37052a
vulkan: mul_mat_id coopmat2 optimizations (#15546)
jeffbolznv Aug 31, 2025
bbbf5ec
vulkan: handle large sizes for get_rows (#15686)
jeffbolznv Aug 31, 2025
7d3c9f2
ci : explicitly set fa off or on (#15692)
CISC Aug 31, 2025
9777032
llama : separate compute buffer reserve from fattn check (#15696)
slaren Aug 31, 2025
2749662
llama : fix fattn reserve call n_seqs parameter (#15699)
slaren Aug 31, 2025
4efd5a8
metal : fix checks for available FA kernels (#15700)
ggerganov Aug 31, 2025
0d161f0
server : enable /slots by default and make it secure (#15630)
ggerganov Aug 31, 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
20 changes: 10 additions & 10 deletions ci/run.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -386,10 +386,10 @@ function gg_run_open_llama_7b_v2 {

(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

function check_ppl {
qnt="$1"
Expand Down Expand Up @@ -520,8 +520,8 @@ function gg_run_pythia_1_4b {

(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

function check_ppl {
qnt="$1"
Expand Down Expand Up @@ -651,10 +651,10 @@ function gg_run_pythia_2_8b {

(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

function check_ppl {
qnt="$1"
Expand Down
14 changes: 7 additions & 7 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2962,20 +2962,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.endpoint_metrics = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_METRICS"));
add_opt(common_arg(
{"--slots"},
string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
[](common_params & params) {
params.endpoint_slots = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
add_opt(common_arg(
{"--props"},
string_format("enable changing global properties via POST /props (default: %s)", params.endpoint_props ? "enabled" : "disabled"),
[](common_params & params) {
params.endpoint_props = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_PROPS"));
add_opt(common_arg(
{"--slots"},
string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
[](common_params & params) {
params.endpoint_slots = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
add_opt(common_arg(
{"--no-slots"},
"disables slots monitoring endpoint",
Expand Down
2 changes: 1 addition & 1 deletion common/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -444,7 +444,7 @@ struct common_params {

// "advanced" endpoints are disabled by default for better security
bool webui = true;
bool endpoint_slots = false;
bool endpoint_slots = true;
bool endpoint_props = false; // only control POST requests, not GET
bool endpoint_metrics = false;

Expand Down
3 changes: 3 additions & 0 deletions ggml/include/ggml-backend.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -307,6 +307,9 @@ extern "C" {
GGML_API void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
GGML_API ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);

// Split graph without allocating it
GGML_API void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);

// Allocate and compute graph on the backend scheduler
GGML_API bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // returns success
GGML_API enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
Expand Down
4 changes: 3 additions & 1 deletion ggml/src/ggml-backend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -902,7 +902,7 @@ static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, stru
}

// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
// reset splits
sched->n_splits = 0;
sched->n_graph_inputs = 0;
Expand Down Expand Up @@ -1687,6 +1687,8 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
GGML_ASSERT(sched);
GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);

ggml_backend_sched_reset(sched);

ggml_backend_sched_synchronize(sched);

ggml_backend_sched_split_graph(sched, measure_graph);
Expand Down
50 changes: 11 additions & 39 deletions ggml/src/ggml-metal/ggml-metal.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -523,13 +523,6 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
Expand Down Expand Up @@ -1562,13 +1555,6 @@ @implementation GGMLMetalClass
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40, flash_attn_ext_vec_f16_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40, flash_attn_ext_vec_bf16_h40, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40, flash_attn_ext_vec_q4_0_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40, flash_attn_ext_vec_q4_1_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40, flash_attn_ext_vec_q5_0_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40, flash_attn_ext_vec_q5_1_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40, flash_attn_ext_vec_q8_0_h40, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64, flash_attn_ext_vec_f16_h64, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64, flash_attn_ext_vec_bf16_h64, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64, flash_attn_ext_vec_q4_0_h64, has_simdgroup_reduction);
Expand Down Expand Up @@ -1909,9 +1895,15 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_ARANGE:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[0]->ne[0] == 32) {
// head size == 32 (e.g. bert-bge-small)
// TODO: not sure if it is worth adding kernels for this size
// for new head sizes, add checks here
if (op->src[0]->ne[0] != 40 &&
op->src[0]->ne[0] != 64 &&
op->src[0]->ne[0] != 80 &&
op->src[0]->ne[0] != 96 &&
op->src[0]->ne[0] != 112 &&
op->src[0]->ne[0] != 128 &&
op->src[0]->ne[0] != 192 &&
op->src[0]->ne[0] != 256) {
return false;
}
if (op->src[0]->ne[0] == 576) {
Expand Down Expand Up @@ -5138,10 +5130,8 @@ static int ggml_metal_encode_node(

bool use_vec_kernel = false;

// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
// for now avoiding mainly to keep the number of templates/kernels a bit lower
// these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
// use non-vec kernel if the batch size is large or if the vec-kernel is not supported for this head size
if (ne01 >= 20 || (ne00 == 40 || ne00 == 80 || ne00 == 112)) {
switch (src1->type) {
case GGML_TYPE_F16:
{
Expand Down Expand Up @@ -5329,24 +5319,6 @@ static int ggml_metal_encode_node(
use_vec_kernel = true;

switch (ne00) {
case 40:
{
switch (src1->type) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40].pipeline; break;
case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40].pipeline; break;
default:
{
GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
GGML_LOG_ERROR("add template specialization for this type\n");
GGML_ABORT("add template specialization for this type");
}
}
} break;
case 64:
{
switch (src1->type) {
Expand Down
13 changes: 3 additions & 10 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 @@ -4803,6 +4803,9 @@ kernel void kernel_flash_attn_ext_vec(
ushort3 ntg[[threads_per_threadgroup]],
ushort tiisg[[thread_index_in_simdgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
static_assert(DK % 32 == 0, "DK must be divisible by 32");
static_assert(DV % 32 == 0, "DV must be divisible by 32");

const short nsg = ntg.y; // number of simdgroups
const short iwg = tgpig[2]%nwg;

Expand Down Expand Up @@ -5160,16 +5163,6 @@ kernel void kernel_flash_attn_ext_vec(

typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;

template [[host_name("kernel_flash_attn_ext_vec_f16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 40, 40, 8>;
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 40, 40, 8>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h40")]] 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, 40, 40, 8>;
template [[host_name("kernel_flash_attn_ext_vec_q4_1_h40")]] 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, 40, 40, 8>;
template [[host_name("kernel_flash_attn_ext_vec_q5_0_h40")]] 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, 40, 40, 8>;
template [[host_name("kernel_flash_attn_ext_vec_q5_1_h40")]] 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, 40, 40, 8>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h40")]] 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, 40, 40, 8>;

template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 64, 64, 8>;
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 64, 64, 8>;
Expand Down
Loading