vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-15 19:39:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	Package.swift
#	flake.lock
#	tests/test-chat-template.cpp
This commit is contained in:
Concedo 2024-04-15 21:39:40 +08:00
commit 02892c5cbd
4 changed files with 2762 additions and 2793 deletions
Download patch file Download diff file Expand all files Collapse all files

5409
examples/server/index.html.hpp
View file

File diff suppressed because it is too large Load diff

22
examples/server/public/index.html
View file

@ -51,26 +51,6 @@
margin-bottom: 0.5em; margin-bottom: 0.5em;
} }
button, input, textarea, .button, a.button, select {
color: #666;
border: 1px solid #ddd;
border-radius: 4px;
line-height: 1.5em;
padding: 0.25em 0.25em;
text-decoration: none;
font-size: 1.1rem;
}
button {
border: 1px solid #2a8aad;
background: #3584e4;
font-weight: normal;
color: #fff;
}
button:disabled {
background: #9cbce5;
}
#write form { #write form {
margin: 1em 0 0 0; margin: 1em 0 0 0;
display: flex; display: flex;
@ -587,7 +567,7 @@
runCompletion(); runCompletion();
} }
return html` return html`
<div class="right"> <div>
<button onclick=${submit} type="button" disabled=${generating.value}>Start</button> <button onclick=${submit} type="button" disabled=${generating.value}>Start</button>
<button onclick=${stop} disabled=${!generating.value}>Stop</button> <button onclick=${stop} disabled=${!generating.value}>Stop</button>
<button onclick=${reset}>Reset</button> <button onclick=${reset}>Reset</button>

102
ggml-sycl.cpp
View file

@ -15996,73 +15996,76 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
const ggml_tensor *src1, const ggml_tensor *src1,
ggml_tensor *dst) try { ggml_tensor *dst) try {
#if 0 GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT &&
ggml_sycl_mul_mat_id_sycl(dst); "mul_mat_id does not support split buffers");
// TODO: mmq/mmv support const ggml_tensor *ids = dst->src[2];
#endif
const int64_t nb11 = src1->nb[1];
const int64_t nb1 = dst->nb[1];
const struct ggml_tensor * ids = src0;
const int32_t id = ((int32_t *) dst->op_params)[0];
const int32_t n_as = ((int32_t *) dst->op_params)[1];
std::vector<char> ids_host(ggml_nbytes(ids));
const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; const dpct::queue_ptr stream = g_syclStreams[g_main_device][0];
if (ids->backend == GGML_BACKEND_TYPE_GPU) { const size_t nb11 = src1->nb[1];
const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device]; const size_t nb1 = dst->nb[1];
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)).wait()));
// SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
} else {
memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
}
const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra; const int32_t id = ((int32_t *)dst->op_params)[0];
const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra; const int32_t n_as = src0->ne[2];
std::vector<char> ids_host(ggml_nbytes(ids));
const char *ids_dev = (const char *)ids->data;
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
const ggml_tensor_extra_gpu *src0_extra =
(const ggml_tensor_extra_gpu *)src0->extra;
const ggml_tensor_extra_gpu *src1_extra =
(const ggml_tensor_extra_gpu *)src1->extra;
const ggml_tensor_extra_gpu *dst_extra =
(const ggml_tensor_extra_gpu *)dst->extra;
ggml_tensor_extra_gpu src0_row_extra;
ggml_tensor_extra_gpu src1_row_extra; ggml_tensor_extra_gpu src1_row_extra;
ggml_tensor_extra_gpu dst_row_extra; ggml_tensor_extra_gpu dst_row_extra;
ggml_tensor src0_row = *src0;
ggml_tensor src1_row = *src1; ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst; ggml_tensor dst_row = *dst;
src1_row.backend = GGML_BACKEND_TYPE_GPU; src1_row.backend = GGML_BACKEND_TYPE_GPU;
dst_row.backend = GGML_BACKEND_TYPE_GPU; dst_row.backend = GGML_BACKEND_TYPE_GPU;
src0_row.extra = &src0_row_extra;
src1_row.extra = &src1_row_extra; src1_row.extra = &src1_row_extra;
dst_row.extra = &dst_row_extra; dst_row.extra = &dst_row_extra;
char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ? char *src0_original = src1->backend == GGML_BACKEND_TYPE_CPU
(char *) src1->data : (char *) src1_extra->data_device[g_main_device]; ? (char *)src0->data
char * dst_original = dst->backend == GGML_BACKEND_TYPE_CPU ? : (char *)src0_extra->data_device[g_main_device];
(char *) dst->data : (char *) dst_extra->data_device[g_main_device]; char *src1_original = src1->backend == GGML_BACKEND_TYPE_CPU
? (char *)src1->data
: (char *)src1_extra->data_device[g_main_device];
char *dst_original = dst->backend == GGML_BACKEND_TYPE_CPU
? (char *)dst->data
: (char *)dst_extra->data_device[g_main_device];
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[3] = src0->nb[2];
if (src1->ne[1] == 1) { if (src1->ne[1] == 1) {
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
//int32_t row_id; const int32_t row_id =
//SYCL_CHECK(syclMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), syclMemcpyDeviceToHost, g_syclStreams[g_main_device][0])); *(const int32_t *)(ids_host.data() + i01 * ids->nb[1] +
//SYCL_CHECK(syclStreamSynchronize(g_syclStreams[g_main_device][0])); id * ids->nb[0]);
const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(row_id >= 0 && row_id < n_as); GGML_ASSERT(row_id >= 0 && row_id < n_as);
const struct ggml_tensor * src0_row = dst->src[row_id + 2]; src0_row_extra.data_device[g_main_device] =
src0_original + row_id * src0->nb[2];
src1_row_extra.data_device[g_main_device] =
src1_original + i01 * src1->nb[1];
dst_row_extra.data_device[g_main_device] =
dst_original + i01 * dst->nb[1];
src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1]; ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set?
dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1];
dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set?
ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row);
} }
} else { } else {
sycl_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1)); sycl_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1));
@ -16072,8 +16075,6 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
dst_row_extra.data_device[g_main_device] = dst_contiguous.get(); dst_row_extra.data_device[g_main_device] = dst_contiguous.get();
for (int32_t row_id = 0; row_id < n_as; ++row_id) { for (int32_t row_id = 0; row_id < n_as; ++row_id) {
const struct ggml_tensor * src0_row = dst->src[row_id + 2];
int64_t num_src1_rows = 0; int64_t num_src1_rows = 0;
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
@ -16086,7 +16087,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
SYCL_CHECK(CHECK_TRY_ERROR( SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11, stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11,
src1_original + i01 * nb11, nb11).wait())); src1_original + i01 * nb11, nb11)));
num_src1_rows++; num_src1_rows++;
} }
@ -16094,6 +16095,9 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
continue; continue;
} }
src0_row_extra.data_device[g_main_device] =
src0_original + row_id * src0->nb[2];
src1_row.ne[1] = num_src1_rows; src1_row.ne[1] = num_src1_rows;
dst_row.ne[1] = num_src1_rows; dst_row.ne[1] = num_src1_rows;
@ -16105,7 +16109,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
dst_row.nb[2] = num_src1_rows*nb1; dst_row.nb[2] = num_src1_rows*nb1;
dst_row.nb[3] = num_src1_rows*nb1; dst_row.nb[3] = num_src1_rows*nb1;
ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row); ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
num_src1_rows = 0; num_src1_rows = 0;
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
@ -16119,7 +16123,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
dst_original + i01 * nb1, dst_original + i01 * nb1,
dst_contiguous.get() + num_src1_rows * nb1, nb1).wait())); dst_contiguous.get() + num_src1_rows * nb1, nb1)));
num_src1_rows++; num_src1_rows++;
} }
} }

22
llama.cpp
View file

@ -13378,6 +13378,11 @@ struct llama_beam_search_data {
} }
llama_logit_info logit_info(ctx); llama_logit_info logit_info(ctx);
std::vector<llama_token_data> next_tokens = logit_info.top_k(n_beams); std::vector<llama_token_data> next_tokens = logit_info.top_k(n_beams);
// Clear the kv slot so that other beams may try different tokens at this position. The llama_decode()
// call in loop() will conclusively fill in the kv slot once the beams converge at this position.
llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
size_t i=0; size_t i=0;
if (next_beams.size() < n_beams) { if (next_beams.size() < n_beams) {
for (; next_beams.size() < n_beams ; ++i) { for (; next_beams.size() < n_beams ; ++i) {
@ -15778,6 +15783,8 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) {
GGML_ASSERT((uint32_t) id < ctx->cparams.n_batch); GGML_ASSERT((uint32_t) id < ctx->cparams.n_batch);
ctx->output_ids[id] = i; ctx->output_ids[id] = i;
} }
ctx->n_outputs = n_outputs;
} }
} }
@ -16946,6 +16953,21 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) { if (add_ass) {
ss << "### Response:\n"; ss << "### Response:\n";
} }
} else if (tmpl == "command-r" || (tmpl.find("<|START_OF_TURN_TOKEN|>") != std::string::npos && tmpl.find("<|USER_TOKEN|>") != std::string::npos)) {
// CohereForAI/c4ai-command-r-plus
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
} else if (role == "user") {
ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
} else if (role == "assistant") {
ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
}
}
if (add_ass) {
ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
}
} else { } else {
// template not supported // template not supported
return -1; return -1;