ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : mul_mat_id use the same tensor for all the experts (#6387)

Browse source 
* ggml : update mul_mat_id to use the same tensor for all the experts

* update cuda

* minor

* update metal

* update test-backend-ops

* fix cuda

* Update ggml-metal.m

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update convert.py

* update convert-hf-to-gguf.py

* update convert.py for mixtral hf models

* Update convert-hf-to-gguf.py

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* cuda : support non-pow-2 number of experts

* allow quantize to work for split and merged experts models in the same way

* cleanup + disable mmap automatically with split tensors models

* update imatrix

* test-backend-ops : test qwen argsort

* update grok model loading

* llama : add merged experts tensors to the grok tensor map

* minor

* gguf : bump version

* fix quantizing of merged experts

* convert-hf-to-gguf.py : update grok (untested)

* make linter happy

* cuda/argsort : use shared memory instead of pool memory

* convert : fix grok tensor names

* metal : add support for non-pow-2 argsort

* llama : more loader cleanup, better error checking

* cuda : fix warning

* llama : still use mmap for loading old models, but copy the data to a host buffer

* add review note

* llama : remove ffn tensor counting + add sanity check

ggml-ci

* convert : fix handling of n_experts == None

ggml-ci

* imatrix : fix ncall counters

* llama : produce error if imatrix size does not match

* quantize : terminate on errors + trace logs

ggml-ci

* metal : pad shared memory to 16 bytes

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
slaren 2024-04-03 15:07:05 +02:00 • committed by GitHub
parent 52604860f9
commit 08a0c02060
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
15 changed files with 756 additions and 888 deletions
Download patch file Download diff file Expand all files Collapse all files

57
ggml.c
View file

@ -4573,45 +4573,38 @@ void ggml_mul_mat_set_prec(
// ggml_mul_mat_id
// NOTE: id will be removed in the future and instead all the experts listed in ids will be computed
// this will allow computing all the used experts in a single matrix multiplication
struct ggml_tensor * ggml_mul_mat_id(
struct ggml_context * ctx,
struct ggml_tensor * const as[],
int n_as,
struct ggml_tensor * as,
struct ggml_tensor * ids,
int id,
struct ggml_tensor * b) {
GGML_ASSERT(ids->type == GGML_TYPE_I32);
GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
GGML_ASSERT(ids->ne[1] == b->ne[1]);
GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1); // ids is 2d
GGML_ASSERT(ids->ne[1] == b->ne[1]); // must have an expert per b row
GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
GGML_ASSERT(id >= 0 && id < ids->ne[0]);
GGML_ASSERT(id >= 0 && id < ids->ne[0]); // valid id
GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat
bool is_node = false;
if (as[0]->grad || b->grad) {
if (as->grad || b->grad) {
is_node = true;
}
const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
const int64_t ne[4] = { as->ne[1], b->ne[1], b->ne[2], b->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
ggml_set_op_params_i32(result, 0, id);
ggml_set_op_params_i32(result, 1, n_as);
result->op = GGML_OP_MUL_MAT_ID;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = ids;
result->src[0] = as;
result->src[1] = b;
for (int i = 0; i < n_as; i++) {
struct ggml_tensor * a = as[i];
GGML_ASSERT(ggml_are_same_shape(as[0], a));
GGML_ASSERT(ggml_can_mul_mat(a, b));
GGML_ASSERT(!ggml_is_transposed(a));
result->src[i + 2] = a;
}
result->src[2] = ids;
return result;
}
@ -10948,10 +10941,9 @@ static void ggml_compute_forward_mul_mat_id(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const struct ggml_tensor * ids = dst->src[0];
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
const struct ggml_tensor * ids = dst->src[2];
GGML_TENSOR_BINARY_OP_LOCALS
@ -10981,13 +10973,13 @@ static void ggml_compute_forward_mul_mat_id(
GGML_ASSERT(nb1 <= nb2);
GGML_ASSERT(nb2 <= nb3);
// broadcast factors
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
// broadcast is not supported with mmid
assert(ne12 == 1);
assert(ne13 == 1);
// row groups
const int id = ggml_get_op_params_i32(dst, 0);
const int n_as = ggml_get_op_params_i32(dst, 1);
const int n_as = src0->ne[2];
char * wdata_src1_end = (src1->type == vec_dot_type) ?
(char *) params->wdata :
@ -11047,7 +11039,7 @@ static void ggml_compute_forward_mul_mat_id(
continue;
}
const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
size_t src0_offset = cur_a*src0->nb[2];
const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@ -11082,9 +11074,6 @@ static void ggml_compute_forward_mul_mat_id(
continue;
}
assert(ne12 % ne02 == 0);
assert(ne13 % ne03 == 0);
// block-tiling attempt
const int64_t blck_0 = 16;
const int64_t blck_1 = 16;
@ -11101,14 +11090,14 @@ static void ggml_compute_forward_mul_mat_id(
const int64_t i11 = MMID_MATRIX_ROW(cur_a, _i11);
// broadcast src0 into src1
const int64_t i03 = i13/r3;
const int64_t i02 = i12/r2;
//const int64_t i03 = i13/r3;
//const int64_t i02 = i12/r2;
const int64_t i1 = i11;
const int64_t i2 = i12;
const int64_t i3 = i13;
const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
const char * src0_row = (const char *) src0->data + src0_offset;
// desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
// if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
@ -18464,13 +18453,13 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
case GGML_OP_MUL_MAT_ID:
{
cur = 0;
const struct ggml_tensor * src0 = node->src[2];
const struct ggml_tensor * src0 = node->src[0];
const struct ggml_tensor * src1 = node->src[1];
const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
if (src1->type != vec_dot_type) {
cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
}
const int n_as = ggml_get_op_params_i32(node, 1);
const int n_as = src0->ne[2];
cur += GGML_PAD(cur, sizeof(int64_t)); // align
cur += n_as * sizeof(int64_t); // matrix_row_counts
cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows