Compare commits

..

4 Commits

Author SHA1 Message Date
Piotr Wilkin (ilintar) 9b24886f78 Fix compile bug (#20203)
* Fix compile bug

* Update common/chat-auto-parser-helpers.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
2026-03-08 17:15:49 +01:00
Piotr Wilkin (ilintar) 62b8143ad2 Fix structured outputs (#20223)
* Fix structured outputs

* Update common/chat-auto-parser-generator.cpp

Co-authored-by: Aldehir Rojas <hello@alde.dev>

---------

Co-authored-by: Aldehir Rojas <hello@alde.dev>
2026-03-08 17:14:43 +01:00
GiantPrince d088d5b74f ggml-vulkan: Add ELU op support (#20183)
* ggml-Vulkan: add ELU support

* ggml-Vulkan: remove extra spaces and variables

* ggml-Vulkan: fix format issue

* ggml-Vulkan: fix format issue

* fix whitespace issue

* Update Vulkan.csv and ops.md
2026-03-08 12:38:17 +01:00
Jeff Bolz cd18a50ea5 vulkan: Fix data races in coopmat1 mul_mat(_id) (#20084)
* vulkan: Fix data races in coopmat1 mul_mat(_id)

Add barriers between coopmat store and regular loads. We sort of got away with
this because it was the same subgroup accessing the values, but it's still a
race and may not work.

* switch to subgroup control barriers
2026-03-08 12:33:48 +01:00
8 changed files with 2065 additions and 7163 deletions
+14 -8
View File
@@ -1,6 +1,7 @@
#include "chat-auto-parser.h"
#include "chat-peg-parser.h"
#include "chat.h"
#include "common.h"
#include "json-schema-to-grammar.h"
#include "nlohmann/json.hpp"
@@ -51,13 +52,15 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
bool has_tools =
autoparser.tools.format.mode != tool_format::NONE && inputs.tools.is_array() && !inputs.tools.empty();
std::string trigger_marker = !autoparser.tools.format.section_start.empty() ? autoparser.tools.format.section_start :
autoparser.tools.format.per_call_start;
bool include_grammar =
has_tools && ((inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO && !trigger_marker.empty()) ||
inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED);
autoparser.tools.format.per_call_start;
bool has_response_format = !inputs.json_schema.empty() && inputs.json_schema.is_object();
bool include_grammar = has_response_format || (has_tools &&
((inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO && !trigger_marker.empty()) ||
inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
if (include_grammar) {
data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
@@ -68,7 +71,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
});
// Set grammar triggers based on tool section markers (fall back to per-call markers)
if (data.grammar_lazy) { // only do triggers on lazy grammar
if (data.grammar_lazy) {
data.grammar_triggers = {
{ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, trigger_marker }
};
@@ -104,8 +107,11 @@ common_peg_arena autoparser::build_parser(const templates_params & inputs) const
bool has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
if (has_response_format) {
return ctx.reasoning_parser + p.space() +
p.content(p.schema(p.json(), "response-format", inputs.json_schema)) + p.end();
auto response_format = p.rule("response-format", p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
return ctx.reasoning_parser + p.space() + p.choice({
p.literal("```json") + p.space() + response_format + p.space() + p.literal("```"),
response_format
}) + p.end();
}
if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && jinja_caps.supports_tool_calls) {
+1 -1
View File
@@ -162,7 +162,7 @@ diff_split calculate_diff_split(const std::string & left, const std::string & ri
right_fully_consumed = true;
}
auto eat_segment = [](std::string & str, segment & seg) -> std::string { return str.append(seg.value); };
auto eat_segment = [](std::string str, const segment & seg) -> std::string { return std::move(str) + seg.value; };
bool can_have_text_suffix = left_end->type == segment_type::TEXT && right_end->type == segment_type::TEXT;
bool can_have_text_prefix = right_start->type == segment_type::TEXT && left_start->type == segment_type::TEXT;
+7 -7
View File
@@ -37,11 +37,11 @@ Legend:
| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
| DIAG | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | ❌ | ❌ | ❌ |
| DIAG | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | ❌ | ❌ | ❌ |
| DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
| DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
| DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
| ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | | ✅ | ❌ | ❌ |
| ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
| FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
@@ -54,7 +54,7 @@ Legend:
| GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
| GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
| GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
| GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
| GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | | 🟡 | ❌ | ❌ |
| GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
| HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -90,7 +90,7 @@ Legend:
| RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
| SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | 🟡 | | ❌ | ❌ | ❌ |
| SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | 🟡 | | ❌ | ❌ | ❌ |
| SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
| SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
| SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -100,7 +100,7 @@ Legend:
| SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
| SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
| SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
| SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | | ❌ | ❌ | ❌ |
| SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -116,5 +116,5 @@ Legend:
| TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
| TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
| XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | | ✅ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | | ❌ | ❌ | ❌ |
| XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | | ✅ | ❌ | ❌ |
+1999 -7147
View File
File diff suppressed because it is too large Load Diff
+9
View File
@@ -744,6 +744,7 @@ struct vk_device_struct {
// [src/dst 0=fp32,1=fp16]
vk_pipeline pipeline_exp[2];
vk_pipeline pipeline_elu[2];
vk_pipeline pipeline_gelu[2];
vk_pipeline pipeline_gelu_erf[2];
vk_pipeline pipeline_gelu_quick[2];
@@ -4373,6 +4374,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
CREATE_UNARY(elu)
CREATE_UNARY(gelu)
CREATE_UNARY(gelu_erf)
CREATE_UNARY(gelu_quick)
@@ -9241,6 +9243,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
switch (ggml_get_unary_op(dst)) {
case GGML_UNARY_OP_EXP:
return ctx->device->pipeline_exp[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_ELU:
return ctx->device->pipeline_elu[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_SILU:
return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_GELU:
@@ -12852,6 +12856,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
}
switch (ggml_get_unary_op(node)) {
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
@@ -14951,6 +14956,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
@@ -16074,6 +16080,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_UNARY_OP_EXP:
tensor_clone = ggml_exp(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_ELU:
tensor_clone = ggml_elu(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_SILU:
tensor_clone = ggml_silu(ggml_ctx, src_clone[0]);
break;
@@ -0,0 +1,27 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
float x = float(data_a[i]);
if (x < 0.0f) {
x = exp(x) - 1;
}
data_d[i] = D_TYPE(x);
}
@@ -377,6 +377,7 @@ void main() {
[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
barrier();
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
const uint row_i = dc + cm_col * TN + col + store_c;
if (row_i >= _ne1) break;
@@ -387,6 +388,7 @@ void main() {
data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
}
barrier();
}
}
#else
@@ -404,18 +406,22 @@ void main() {
// Full coopMat is within bounds, but stride_d is not aligned
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
} else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
// Partial coopMat is within bounds
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
}
controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
}
}
}
@@ -867,6 +867,8 @@ void process_shaders() {
string_to_spv("hardswish_f32", "hardswish.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("abs_f16", "abs.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("abs_f32", "abs.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("elu_f16", "elu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("elu_f32", "elu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("xielu_f16", "xielu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("xielu_f32", "xielu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});