vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into gg/bpe-preprocess

Browse source
This commit is contained in:
Georgi Gerganov 2024-04-26 18:08:40 +03:00
commit e3f6dc7409
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
5 changed files with 117 additions and 100 deletions
Download patch file Download diff file Expand all files Collapse all files

16
.github/workflows/server.yml vendored
View file

@ -58,6 +58,7 @@ jobs:
git \
cmake \
python3-pip \
python3-venv \
curl \
wget \
language-pack-en \
@ -100,16 +101,13 @@ jobs:
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
- name: Python setup
id: setup_python
uses: actions/setup-python@v5
with:
python-version: '3.11'
- name: Tests dependencies
id: test_dependencies
- name: Setup python env
id: pipenv
run: |
pip install -r examples/server/tests/requirements.txt
cd examples/server/tests
python3 -m venv venv
. venv/bin/activate
pip install -r requirements.txt
- name: Tests
id: server_integration_tests

23
examples/server/server.cpp
View file

@ -1207,6 +1207,27 @@ struct server_context {
LOG_VERBOSE("eos token found", {});
}
auto n_ctx_train = llama_n_ctx_train(model);
if (slot.params.n_predict < 1 && slot.ga_n == 1
&& slot.n_prompt_tokens + slot.n_decoded >= n_ctx_train) {
LOG_WARNING("n_predict is not set and self-context extend is disabled."
" Limiting generated tokens to n_ctx_train to avoid EOS-less generation infinite loop", {
{ "id_slot", slot.id },
{ "params.n_predict", slot.params.n_predict },
{ "slot.n_prompt_tokens", slot.n_prompt_tokens },
{ "slot.n_decoded", slot.n_decoded },
{ "slot.n_predict", slot.n_predict },
{ "n_slots", params.n_parallel },
{ "slot.n_ctx", slot.n_ctx },
{ "n_ctx", n_ctx },
{ "n_ctx_train", n_ctx_train },
{ "ga_n", slot.ga_n },
});
slot.truncated = true;
slot.stopped_limit = true;
slot.has_next_token = false; // stop prediction
}
LOG_VERBOSE("next token", {
{"id_slot", slot.id},
{"id_task", slot.id_task},
@ -2141,7 +2162,7 @@ struct server_context {
});
// process the created batch of tokens
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
for (auto & slot : slots) {

2
ggml.c
View file

@ -20987,7 +20987,7 @@ void gguf_free(struct gguf_context * ctx) {
GGML_FREE(ctx->infos);
}
GGML_ALIGNED_FREE(ctx);
GGML_FREE(ctx);
}
const char * gguf_type_name(enum gguf_type type) {

170
sgemm.cpp
View file

@ -50,7 +50,6 @@
#pragma GCC diagnostic ignored "-Wignored-attributes"
#include "sgemm.h"
#include <algorithm>
#include "ggml-impl.h"
#include "ggml-quants.h"
@ -243,23 +242,23 @@ template <> inline __m512 load(const ggml_fp16_t *p) {
template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
class tinyBLAS {
public:
tinyBLAS(int k,
const TA *A, int lda,
const TB *B, int ldb,
TC *C, int ldc,
tinyBLAS(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int m, int n, int task) {
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
mnpack(0, m, 0, n);
}
private:
NOINLINE void mnpack(int m0, int m, int n0, int n) {
int mc, nc, mp, np;
switch ((std::min(m - m0, 5) << 4) | std::min(n - n0, 5)) {
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
#if VECTOR_REGISTERS == 32
case 0x55:
mc = 5;
@ -409,27 +408,27 @@ class tinyBLAS {
}
template <int RM, int RN>
NOINLINE void gemm(int m0, int m, int n0, int n) {
int ytiles = (m - m0) / RM;
int xtiles = (n - n0) / RN;
int tiles = xtiles * ytiles;
int duty = (tiles + nth - 1) / nth;
int start = duty * ith;
int end = start + duty;
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int job = start; job < end; ++job) {
int ii = m0 + job / xtiles * RM;
int jj = n0 + job % xtiles * RN;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
D Cv[RN][RM] = {};
for (int l = 0; l < k; l += KN)
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t l = 0; l < k; l += KN)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
load<V>(B + ldb * (jj + j) + l),
Cv[j][i]);
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
@ -437,10 +436,10 @@ class tinyBLAS {
const TA *const A;
const TB *const B;
TC *const C;
const int k;
const int lda;
const int ldb;
const int ldc;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
@ -452,23 +451,23 @@ class tinyBLAS {
template <typename TA>
class tinyBLAS_Q0_ARM {
public:
tinyBLAS_Q0_ARM(int k,
const TA *A, int lda,
const block_q8_0 *B, int ldb,
float *C, int ldc,
tinyBLAS_Q0_ARM(int64_t k,
const TA *A, int64_t lda,
const block_q8_0 *B, int64_t ldb,
float *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int m, int n, int task) {
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
mnpack(0, m, 0, n);
}
private:
NOINLINE void mnpack(int m0, int m, int n0, int n) {
int mc, nc, mp, np;
switch ((std::min(m - m0, 3) << 4) | std::min(n - n0, 3)) {
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
case 0x33:
mc = 3;
nc = 3;
@ -524,22 +523,22 @@ class tinyBLAS_Q0_ARM {
}
template <int RM, int RN>
NOINLINE void gemm(int m0, int m, int n0, int n) {
int ytiles = (m - m0) / RM;
int xtiles = (n - n0) / RN;
int tiles = xtiles * ytiles;
int duty = (tiles + nth - 1) / nth;
int start = duty * ith;
int end = start + duty;
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int job = start; job < end; ++job) {
int ii = m0 + job / xtiles * RM;
int jj = n0 + job % xtiles * RN;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
float32x4_t Cv[RN][RM] = {};
for (int l = 0; l < k; ++l)
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t l = 0; l < k; ++l)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
Cv[j][i] = vmlaq_n_f32(Cv[j][i],
vcvtq_f32_s32(vdotq_s32(
vdotq_s32(vdupq_n_s32(0),
@ -549,8 +548,8 @@ class tinyBLAS_Q0_ARM {
load_hi(B + ldb * (jj + j) + l))),
unhalf(A[lda * (ii + i) + l].d) *
unhalf(B[ldb * (jj + j) + l].d));
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
@ -577,10 +576,10 @@ class tinyBLAS_Q0_ARM {
const TA *const A;
const block_q8_0 *const B;
float *const C;
const int k;
const int lda;
const int ldb;
const int ldc;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
@ -590,23 +589,23 @@ class tinyBLAS_Q0_ARM {
template <typename TA, typename TB, typename TC>
class tinyBLAS_Q0_AVX2 {
public:
tinyBLAS_Q0_AVX2(int k,
const TA *A, int lda,
const TB *B, int ldb,
TC *C, int ldc,
tinyBLAS_Q0_AVX2(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int m, int n, int task) {
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
mnpack(0, m, 0, n);
}
private:
void mnpack(int m0, int m, int n0, int n) {
int mc, nc, mp, np;
switch ((std::min(m - m0, 4) << 4) | std::min(n - n0, 4)) {
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
#if VECTOR_REGISTERS == 32
case 0x44:
mc = 4;
@ -714,22 +713,22 @@ class tinyBLAS_Q0_AVX2 {
}
template <int RM, int RN>
NOINLINE void gemm(int m0, int m, int n0, int n) {
int ytiles = (m - m0) / RM;
int xtiles = (n - n0) / RN;
int tiles = xtiles * ytiles;
int duty = (tiles + nth - 1) / nth;
int start = duty * ith;
int end = start + duty;
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int job = start; job < end; ++job) {
int ii = m0 + job / xtiles * RM;
int jj = n0 + job % xtiles * RN;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
__m256 Cv[RN][RM] = {};
for (int l = 0; l < k; ++l)
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t l = 0; l < k; ++l)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
unhalf(B[ldb * (jj + j) + l].d)),
updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
@ -737,8 +736,8 @@ class tinyBLAS_Q0_AVX2 {
_mm256_sign_epi8(load(B + ldb * (jj + j) + l),
load(A + lda * (ii + i) + l))),
Cv[j][i]);
for (int j = 0; j < RN; ++j)
for (int i = 0; i < RM; ++i)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
@ -771,10 +770,10 @@ class tinyBLAS_Q0_AVX2 {
const TA *const A;
const TB *const B;
TC *const C;
const int k;
const int lda;
const int ldb;
const int ldc;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
@ -813,8 +812,8 @@ class tinyBLAS_Q0_AVX2 {
* @param Ctype is GGML data type of `C`
* @return true if this function was able to service the matmul request
*/
bool llamafile_sgemm(int m, int n, int k, const void *A, int lda, const void *B, int ldb, void *C,
int ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype) {
bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
int64_t ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype) {
assert(m >= 0);
assert(n >= 0);
@ -824,9 +823,6 @@ bool llamafile_sgemm(int m, int n, int k, const void *A, int lda, const void *B,
assert(ldc >= m);
assert(nth > 0);
assert(ith < nth);
assert(1ll * lda * m <= 0x7fffffff);
assert(1ll * ldb * n <= 0x7fffffff);
assert(1ll * ldc * n <= 0x7fffffff);
if (Ctype != GGML_TYPE_F32)
return false;

6
sgemm.h
View file

@ -1,11 +1,13 @@
#pragma once
#include <stdint.h>
#include <stdbool.h>
#ifdef __cplusplus
extern "C" {
#endif
bool llamafile_sgemm(int, int, int, const void *, int, const void *, int,
void *, int, int, int, int, int, int, int);
bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
const void *, int64_t, void *, int64_t, int, int,
int, int, int, int);
#ifdef __cplusplus
}