Skip to content

Commit 2d75e56

Browse files
ngxsonngxson
committed
Merge branch 'master' into mascguy-errno
2 parents c146c30 + ef6dada commit 2d75e56

File tree

7 files changed

+210
-110
lines changed

7 files changed

+210
-110
lines changed

‎.github/workflows/server.yml‎

Lines changed: 22 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -112,9 +112,9 @@ jobs:
112112
-DGGML_OPENMP=OFF ;
113113
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
114114
115-
- name: Build
116-
id: cmake_build
117-
if: ${{ matrix.sanitizer != 'THREAD' }}
115+
- name: Build (sanitizers)
116+
id: cmake_build_sanitizers
117+
if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
118118
run: |
119119
cmake -B build \
120120
-DGGML_NATIVE=OFF \
@@ -124,12 +124,31 @@ jobs:
124124
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
125125
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
126126
127+
- name: Build (sanitizers)
128+
id: cmake_build
129+
if: ${{ matrix.sanitizer == '' }}
130+
run: |
131+
cmake -B build \
132+
-DGGML_NATIVE=OFF \
133+
-DLLAMA_BUILD_SERVER=ON \
134+
-DLLAMA_CURL=ON \
135+
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
136+
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
137+
127138
- name: Tests
128139
id: server_integration_tests
140+
if: ${{ matrix.sanitizer == '' }}
129141
run: |
130142
cd examples/server/tests
131143
./tests.sh
132144
145+
- name: Tests (sanitizers)
146+
id: server_integration_tests_sanitizers
147+
if: ${{ matrix.sanitizer != '' }}
148+
run: |
149+
cd examples/server/tests
150+
LLAMA_SANITIZE=1 ./tests.sh
151+
133152
- name: Slow tests
134153
id: server_integration_tests_slow
135154
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}

‎examples/server/tests/utils.py‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,9 @@
2626
import wget
2727

2828

29+
DEFAULT_HTTP_TIMEOUT = 10 if "LLAMA_SANITIZE" not in os.environ else 30
30+
31+
2932
class ServerResponse:
3033
headers: dict
3134
status_code: int
@@ -88,7 +91,7 @@ def __init__(self):
8891
if "PORT" in os.environ:
8992
self.server_port = int(os.environ["PORT"])
9093

91-
def start(self, timeout_seconds: int = 10) -> None:
94+
def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
9295
if "LLAMA_SERVER_BIN_PATH" in os.environ:
9396
server_path = os.environ["LLAMA_SERVER_BIN_PATH"]
9497
elif os.name == "nt":

‎examples/simple-chat/simple-chat.cpp‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -95,11 +95,11 @@ int main(int argc, char ** argv) {
9595
llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
9696

9797
// helper function to evaluate a prompt and generate a response
98-
auto generate = [&](const std::string & prompt) {
98+
auto generate = [&](const std::string & prompt, bool is_first) {
9999
std::string response;
100100

101101
// tokenize the prompt
102-
const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
102+
const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
103103
std::vector<llama_token> prompt_tokens(n_prompt_tokens);
104104
if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
105105
GGML_ABORT("failed to tokenize the prompt\n");
@@ -180,7 +180,7 @@ int main(int argc, char ** argv) {
180180

181181
// generate a response
182182
printf("\033[33m");
183-
std::string response = generate(prompt);
183+
std::string response = generate(prompt, prev_len == 0);
184184
printf("\n\033[0m");
185185

186186
// add the response to the messages

‎ggml/src/ggml-sycl/common.hpp‎

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -333,8 +333,12 @@ struct ggml_backend_sycl_context {
333333
// pool
334334
std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
335335

336+
std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];
337+
336338
static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);
337339

340+
static std::unique_ptr<ggml_sycl_pool> new_pool_for_host(queue_ptr qptr, int device);
341+
338342
ggml_sycl_pool & pool(int device) {
339343
if (pools[device] == nullptr) {
340344
pools[device] = new_pool_for_device(stream(device,0), device);
@@ -345,6 +349,15 @@ struct ggml_backend_sycl_context {
345349
ggml_sycl_pool & pool() {
346350
return pool(device);
347351
}
352+
353+
ggml_sycl_pool & host_pool(int device) {
354+
if (host_pools[device] == nullptr) {
355+
host_pools[device] = new_pool_for_host(stream(device, 0), device);
356+
}
357+
return *host_pools[device];
358+
}
359+
360+
ggml_sycl_pool & host_pool() { return host_pool(device); }
348361
};
349362

350363
// common device functions

‎ggml/src/ggml-sycl/dpct/helper.hpp‎

Lines changed: 40 additions & 95 deletions
Original file line numberDiff line numberDiff line change
@@ -82,6 +82,14 @@ inline std::string get_device_backend_and_type(const sycl::device &device) {
8282
return device_type.str();
8383
}
8484

85+
template <typename Ts> struct matrix_info_t {
86+
oneapi::mkl::transpose transpose_info[2];
87+
Ts value_info[2];
88+
std::int64_t size_info[3];
89+
std::int64_t ld_info[3];
90+
std::int64_t groupsize_info;
91+
};
92+
8593
namespace dpct
8694
{
8795
typedef sycl::queue *queue_ptr;
@@ -1727,26 +1735,13 @@ namespace dpct
17271735
};
17281736

17291737
template <class Ta, class Tb, class Tc, class Ts>
1730-
inline void gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
1731-
oneapi::mkl::transpose b_trans, int m, int n, int k,
1732-
const void *alpha, const void **a, int lda,
1733-
const void **b, int ldb, const void *beta, void **c,
1734-
int ldc, int batch_size)
1735-
{
1736-
struct matrix_info_t
1737-
{
1738-
oneapi::mkl::transpose transpose_info[2];
1739-
Ts value_info[2];
1740-
std::int64_t size_info[3];
1741-
std::int64_t ld_info[3];
1742-
std::int64_t groupsize_info;
1743-
};
1744-
1738+
inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans,
1739+
int m, int n, int k, const void * alpha, const void ** a, int lda, const void ** b,
1740+
int ldb, const void * beta, void ** c, int ldc, int batch_size,
1741+
matrix_info_t<float> * matrix_info) {
17451742
Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
17461743
Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
17471744

1748-
matrix_info_t *matrix_info =
1749-
(matrix_info_t *)std::malloc(sizeof(matrix_info_t));
17501745
matrix_info->transpose_info[0] = a_trans;
17511746
matrix_info->transpose_info[1] = b_trans;
17521747
matrix_info->value_info[0] = alpha_value;
@@ -1763,23 +1758,18 @@ namespace dpct
17631758
sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
17641759
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, matrix_info->transpose_info,
17651760
matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1,
1766-
matrix_info->size_info + 2, matrix_info->value_info, reinterpret_cast<const Ta **>(a),
1767-
matrix_info->ld_info, reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
1768-
matrix_info->value_info + 1, reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1,
1769-
&(matrix_info->groupsize_info));
1761+
matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
1762+
reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
1763+
matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
1764+
reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
17701765
#else
17711766
sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
17721767
q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
1773-
matrix_info->size_info + 1, matrix_info->size_info + 2, matrix_info->value_info,
1768+
matrix_info->size_info + 1, matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
17741769
reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
1775-
matrix_info->ld_info + 1, matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
1776-
matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
1770+
matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
1771+
reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
17771772
#endif
1778-
1779-
q.submit([&](sycl::handler &cgh)
1780-
{
1781-
cgh.depends_on(e);
1782-
cgh.host_task([=] { std::free(matrix_info); }); });
17831773
}
17841774

17851775
template <class Ta, class Tb, class Tc, class Ts>
@@ -2422,25 +2412,11 @@ namespace dpct
24222412
/// \param [in] ldc Leading dimension of C.
24232413
/// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
24242414
/// \param [in] scaling_type Data type of the scaling factors.
2425-
inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans,
2426-
oneapi::mkl::transpose b_trans, int m, int n, int k,
2427-
const void *alpha, const void *a[],
2428-
library_data_t a_type, int lda, const void *b[],
2429-
library_data_t b_type, int ldb, const void *beta,
2430-
void *c[], library_data_t c_type, int ldc,
2431-
int batch_size, library_data_t scaling_type)
2432-
{
2433-
if (scaling_type == library_data_t::real_float &&
2434-
c_type == library_data_t::complex_float)
2435-
{
2436-
scaling_type = library_data_t::complex_float;
2437-
}
2438-
else if (scaling_type == library_data_t::real_double &&
2439-
c_type == library_data_t::complex_double)
2440-
{
2441-
scaling_type = library_data_t::complex_double;
2442-
}
2443-
2415+
inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
2416+
int n, int k, const void * alpha, const void * a[], library_data_t a_type, int lda,
2417+
const void * b[], library_data_t b_type, int ldb, const void * beta, void * c[],
2418+
library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type,
2419+
matrix_info_t<float> * matrix_info) {
24442420
std::uint64_t key =
24452421
detail::get_type_combination_id(a_type, b_type, c_type, scaling_type);
24462422
switch (key)
@@ -2449,68 +2425,41 @@ namespace dpct
24492425
library_data_t::real_float, library_data_t::real_float,
24502426
library_data_t::real_float, library_data_t::real_float):
24512427
{
2452-
detail::gemm_batch_impl<float, float, float, float>(
2453-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2454-
batch_size);
2428+
detail::gemm_batch_impl<float, float, float, float>(q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb,
2429+
beta, c, ldc, batch_size, matrix_info);
24552430
break;
24562431
}
24572432
case detail::get_type_combination_id(
24582433
library_data_t::real_double, library_data_t::real_double,
24592434
library_data_t::real_double, library_data_t::real_double):
24602435
{
2461-
detail::gemm_batch_impl<double, double, double, double>(
2462-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2463-
batch_size);
2464-
break;
2465-
}
2466-
case detail::get_type_combination_id(
2467-
library_data_t::complex_float, library_data_t::complex_float,
2468-
library_data_t::complex_float, library_data_t::complex_float):
2469-
{
2470-
detail::gemm_batch_impl<std::complex<float>, std::complex<float>,
2471-
std::complex<float>, std::complex<float>>(
2472-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2473-
batch_size);
2474-
break;
2475-
}
2476-
case detail::get_type_combination_id(
2477-
library_data_t::complex_double, library_data_t::complex_double,
2478-
library_data_t::complex_double, library_data_t::complex_double):
2479-
{
2480-
detail::gemm_batch_impl<std::complex<double>, std::complex<double>,
2481-
std::complex<double>, std::complex<double>>(
2482-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2483-
batch_size);
2436+
detail::gemm_batch_impl<double, double, double, double>(q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb,
2437+
beta, c, ldc, batch_size, matrix_info);
24842438
break;
24852439
}
24862440
case detail::get_type_combination_id(
24872441
library_data_t::real_half, library_data_t::real_half,
24882442
library_data_t::real_half, library_data_t::real_half):
24892443
{
2490-
detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half,
2491-
sycl::half>(q, a_trans, b_trans, m, n, k, alpha,
2492-
a, lda, b, ldb, beta, c, ldc,
2493-
batch_size);
2444+
detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half, sycl::half>(
2445+
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
24942446
break;
24952447
}
24962448
#ifdef __INTEL_MKL__
24972449
case detail::get_type_combination_id(
24982450
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
24992451
library_data_t::real_bfloat16, library_data_t::real_float):
25002452
{
2501-
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
2502-
oneapi::mkl::bfloat16, float>(
2503-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2504-
batch_size);
2453+
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
2454+
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
25052455
break;
25062456
}
25072457
case detail::get_type_combination_id(
25082458
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
25092459
library_data_t::real_float, library_data_t::real_float):
25102460
{
2511-
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
2512-
float>(q, a_trans, b_trans, m, n, k, alpha, a, lda,
2513-
b, ldb, beta, c, ldc, batch_size);
2461+
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
2462+
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
25142463
break;
25152464
}
25162465
#endif
@@ -2522,28 +2471,25 @@ namespace dpct
25222471
dpct::get_value(reinterpret_cast<const std::int32_t *>(alpha), q);
25232472
float beta_float =
25242473
dpct::get_value(reinterpret_cast<const std::int32_t *>(beta), q);
2525-
detail::gemm_batch_impl<std::int8_t, std::int8_t, std::int32_t,
2526-
float>(q, a_trans, b_trans, m, n, k, &alpha_float,
2527-
a, lda, b, ldb, &beta_float, c, ldc,
2528-
batch_size);
2474+
detail::gemm_batch_impl<std::int8_t, std::int8_t, std::int32_t, float>(
2475+
q, a_trans, b_trans, m, n, k, &alpha_float, a, lda, b, ldb, &beta_float, c, ldc, batch_size,
2476+
matrix_info);
25292477
break;
25302478
}
25312479
case detail::get_type_combination_id(
25322480
library_data_t::real_int8, library_data_t::real_int8,
25332481
library_data_t::real_float, library_data_t::real_float):
25342482
{
25352483
detail::gemm_batch_impl<std::int8_t, std::int8_t, float, float>(
2536-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2537-
batch_size);
2484+
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
25382485
break;
25392486
}
25402487
case detail::get_type_combination_id(
25412488
library_data_t::real_half, library_data_t::real_half,
25422489
library_data_t::real_float, library_data_t::real_float):
25432490
{
25442491
detail::gemm_batch_impl<sycl::half, sycl::half, float, float>(
2545-
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
2546-
batch_size);
2492+
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
25472493
break;
25482494
}
25492495
case detail::get_type_combination_id(
@@ -2557,8 +2503,7 @@ namespace dpct
25572503
sycl::half alpha_half(alpha_value);
25582504
sycl::half beta_half(beta_value);
25592505
detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half, sycl::half>(
2560-
q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, b, ldb, &beta_half, c, ldc,
2561-
batch_size);
2506+
q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, b, ldb, &beta_half, c, ldc, batch_size, matrix_info);
25622507
break;
25632508
}
25642509
default:

0 commit comments

Comments
 (0)