ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-06 12:00:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixing split mode graph with many GPUs (#1152)

Browse Source 
* Attempt to fix the many GPU issue in split mode graph

* WIP: this seems more stable

Still hanging after a while if I try to use all 7 GPUs

* Reenable OpenMP in scheduler async

Seems solid up to 4 GPUs. It did hang with --max-gpu 6.

* printf cleanup
This commit is contained in:
Kawrakow
2026-01-17 08:05:24 +02:00
committed by GitHub
parent cb1063f6cd
commit 709e1a5375
3 changed files with 85 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
ggml/src/ggml-backend.cpp
View File

@@ -2180,10 +2180,20 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
}
int first_reduce = -1;
for (int i = 0; i < sched->n_splits; i++) {
auto split = &sched->splits[i];
if (split->graph.n_nodes == 1 && split->graph.nodes[0]->op == GGML_OP_REDUCE) {
first_reduce = split->backend_id;
break;
}
}
if (!has_cpu_work) {
#pragma omp parallel num_threads(sched->n_backends)
{
int last_reduce = first_reduce;
int ith = omp_get_thread_num();
struct ggml_backend_sched_split * splits = sched->splits;
@@ -2206,9 +2216,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
#pragma omp barrier
}
if (split->n_inputs > 0) {
int copy_thread = last_reduce >= 0 ? last_reduce : 0;
if (ith == copy_thread) {
ggml_backend_sched_copy_inputs(sched, split, sched->needs_sync, ids, unique_ids, last_ids_tensor);
}
#pragma omp barrier
}
if (ith == split_backend_id) {
// copy the input tensors to the split backend
ggml_backend_sched_copy_inputs(sched, split, sched->needs_sync, ids, unique_ids, last_ids_tensor);
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
if (split->n_inputs > 0 && !sched->own_cpy[split_backend_id]) {
sched->needs_sync[split_backend_id] = true;
@@ -2219,10 +2237,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
}
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
}
if (split->graph.nodes[0]->op == GGML_OP_REDUCE) {
last_reduce = split_backend_id;
#pragma omp barrier
}
@@ -2238,8 +2256,9 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
#endif
if (!work_done) {
std::barrier barrier(sched->n_backends, [] () noexcept {});
auto compute = [sched, &barrier] (int ith) {
std::barrier barrier(sched->n_backends);
auto compute = [sched, &barrier, first_reduce] (int ith) {
struct ggml_backend_sched_split * splits = sched->splits;
@@ -2247,6 +2266,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
std::vector<uint32_t> unique_ids;
ggml_tensor * last_ids_tensor = nullptr;
int last_reduce = first_reduce;
for (int i = 0; i < sched->n_splits; i++) {
#if IK_PRINT_TIMING
int64_t tim1 = ggml_time_us();
@@ -2261,10 +2282,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
barrier.arrive_and_wait();
}
if (ith == split_backend_id) {
// copy the input tensors to the split backend
ggml_backend_sched_copy_inputs(sched, split, sched->needs_sync, ids, unique_ids, last_ids_tensor);
if (split->n_inputs > 0) {
int copy_thread = last_reduce >= 0 ? last_reduce : 0;
if (ith == copy_thread) {
ggml_backend_sched_copy_inputs(sched, split, sched->needs_sync, ids, unique_ids, last_ids_tensor);
}
barrier.arrive_and_wait();
}
if (ith == split_backend_id) {
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
if (split->n_inputs > 0 && !sched->own_cpy[split_backend_id]) {
sched->needs_sync[split_backend_id] = true;
} else {
@@ -2274,10 +2302,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
}
sched->statuses[ith] = ggml_backend_sched_eval(sched, split_backend, split);
}
if (split->graph.nodes[0]->op == GGML_OP_REDUCE) {
last_reduce = split_backend_id;
barrier.arrive_and_wait();
}
//if (needs_barrier) {
@@ -2287,6 +2315,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
// record the event of this copy
if (split->n_inputs > 0) {
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
printf("Recording event %d, %d\n", split_backend_id, sched->cur_copy);
ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
}
}

20
ggml/src/ggml-cuda/reduce.cu
View File

@@ -270,9 +270,15 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
(const char *)dst->src[peer]->data + ichunk*nelem_per_device*elem_size, info.all_ctx[peer]->device,
this_nelem*elem_size, info.all_ctx[peer]->stream()));
CUDA_CHECK(cudaEventRecord(info.all_ctx[peer]->copy_event, info.all_ctx[peer]->stream()));
//ggml_cuda_set_device(info.all_ctx[i]->device);
//CUDA_CHECK(cudaStreamWaitEvent(info.all_ctx[i]->stream(), info.all_ctx[peer]->copy_event, 0));
ichunk = (ichunk + 1)%nhave;
}
for (int ii = 0; ii < nhave; ++ii) {
int i = idx[ii];
int peer = idx[(ii+1)%nhave];
ggml_cuda_set_device(info.all_ctx[i]->device);
CUDA_CHECK(cudaStreamWaitEvent(info.all_ctx[i]->stream(), info.all_ctx[peer]->copy_event, 0));
ichunk = (ichunk + 1)%nhave;
}
}
ggml_cuda_set_device(ctx.device);
@@ -344,7 +350,7 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
ggml_cuda_set_device(ctx.device);
return;
}
if (dst->ne[1] <= 8 && ctx.p2p_enabled) {
if (dst->ne[1] < 32 && ctx.p2p_enabled) {
for (int ii = 0; ii < nhave; ++ii) {
int i = idx[ii];
GGML_ASSERT(dst->src[i]->type == dst->type);
@@ -357,7 +363,9 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
}
//printf("Recorded events\n");
auto nelem = ggml_nelements(dst);
auto nelem_per_device = (nelem + nhave - 1)/nhave;
auto nelem8 = (nelem + 7)/8;
auto nelem_per_device = 8*((nelem8 + nhave - 1)/nhave);
//auto nelem_per_device = (nelem + nhave - 1)/nhave;
auto elem_size = ggml_element_size(dst);
for (int ii = 0; ii < nhave; ++ii) {
int i = idx[ii];
@@ -471,6 +479,12 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
ggml_cuda_set_device(i);
CUDA_CHECK(cudaStreamWaitEvent(info.all_ctx[i]->stream(), ctx.copy_event, 0));
CUDA_CHECK(cudaMemcpyPeerAsync(dst->src[i]->data, i, dst->data, ctx.device, nbytes, info.all_ctx[i]->stream()));
CUDA_CHECK(cudaEventRecord(info.all_ctx[i]->copy_event, info.all_ctx[i]->stream()));
}
ggml_cuda_set_device(ctx.device);
for (int ii = 0; ii < nhave; ++ii) {
int i = idx[ii];
if (i == ctx.device) continue;
CUDA_CHECK(cudaStreamWaitEvent(ctx.stream(), info.all_ctx[i]->copy_event, 0));
}
}

58
src/llama-load-tensors.cpp
View File

@@ -244,14 +244,14 @@ static std::vector<int> create_split(int nr, int granularity, const std::vector<
std::vector<int> result(splits.size());
float last_split = 0;
int sum = 0;
if (verbose) printf("--- %s: %d chunks\n", __func__, nchunk);
if (verbose) LLAMA_LOG_INFO("--- %s: %d chunks\n", __func__, nchunk);
for (int i = 0; i < (int)splits.size(); ++i) {
float p = splits[i] - last_split;
float p0 = p;
p += (p - 1.f*mem_used[i]/tot_memory_used);
result[i] = roundf(p*nchunk);
if (result[i] < 0) result[i] = 0;
if (verbose) printf("i = %d, p0 = %g, p = %g, result = %d\n", i, p0, p, result[i]);
if (verbose) LLAMA_LOG_INFO("i = %d, p0 = %g, p = %g, result = %d\n", i, p0, p, result[i]);
sum += result[i];
last_split = splits[i];
}
@@ -317,7 +317,6 @@ ggml_tensor * create_tensors_helper::create_tensor(ggml_context * ctx, const std
if (actual_context) *actual_context = ctx;
auto tensor = ml.create_tensor(ctx, name, ne, flags);
if (tensor && ctx == split_ctx) {
//printf("%s: adding tensor %s to split tensors\n", __func__, tensor->name);
split_tensors.insert(tensor);
}
return tensor;
@@ -1184,7 +1183,6 @@ bool create_tensors_helper::create_mimo2_tensors(const LLM_TN & tn) {
uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
uint32_t n_head = hparams.n_head(i);
//printf("Layer %2d: n_head = %u, n_embd_head_k = %d, n_embd_head_v = %d, n_embd_k_gqa = %d, n_embd_v_gqa = %d\n", i, n_head, (int)n_embd_head_k, (int)n_embd_head_v, n_embd_k_gqa, n_embd_v_gqa);
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
@@ -1800,7 +1798,7 @@ bool create_tensors_helper::create_deepseek2_tensors(const LLM_TN & tn) {
layer.wkv_a_mqa = ml.create_tensor_as_view(ctx_split, layer.wkq_a_mqa, k_name.c_str(), { wk->ne[0], wk->ne[1] }, wq->ne[1]*wq->nb[1]);
merged = true;
use_mmap_buffer = false;
printf("============== Merged %s (%ld x %ld) and %s (%ld x %ld)\n", q_name.c_str(),
LLAMA_LOG_DEBUG("============== Merged %s (%ld x %ld) and %s (%ld x %ld)\n", q_name.c_str(),
wq->ne[0], wq->ne[1], k_name.c_str(), wk->ne[0], wk->ne[1]);
}
}
@@ -2661,7 +2659,7 @@ bool create_tensors_helper::merge_up_gate_exps(const LLM_TN & tn, int i, int bia
auto g_meta = ml.require_tensor_meta(g_name.c_str());
if (u_meta->type != g_meta->type || u_meta->ne[0] != g_meta->ne[0] || u_meta->ne[2] != g_meta->ne[2]) {
printf("%s: not merging because up/fate meta info is different\n", __func__);
LLAMA_LOG_INFO("%s: not merging because up/fate meta info is different\n", __func__);
return false;
}
@@ -2669,16 +2667,16 @@ bool create_tensors_helper::merge_up_gate_exps(const LLM_TN & tn, int i, int bia
auto g_ctx = get_context_for_tensor(ctx_split, g_name);
if (u_ctx != g_ctx) {
printf("%s: not merging because of context\n", __func__);
LLAMA_LOG_INFO("%s: not merging because of context\n", __func__);
return false;
}
if (bias && (u_ctx != ctx_split || g_ctx != ctx_split)) {
printf("%s: not merging because of context\n", __func__);
LLAMA_LOG_INFO("%s: not merging because of context\n", __func__);
return false;
}
printf("%s: merging up/gate in layer %d\n", __func__, i);
LLAMA_LOG_INFO("%s: merging up/gate in layer %d\n", __func__, i);
layer.ffn_up_gate_exps = ggml_new_tensor_3d(u_ctx, u_meta->type, u_meta->ne[0], u_meta->ne[1] + g_meta->ne[1], u_meta->ne[2]);
snprintf(layer.ffn_up_gate_exps->name, GGML_MAX_NAME, "blk.%d.ffn_up_gate_exps.weight", i);
@@ -2803,7 +2801,7 @@ bool create_tensors_helper::merge_qkv(const LLM_TN & tn, int i, int bias, bool i
layer.wq = ml.create_tensor_as_view(ctx_split, layer.wqk, wq_name.c_str(), { wq->ne[0], wq->ne[1] }, 0);
layer.wk = ml.create_tensor_as_view(ctx_split, layer.wqk, wk_name.c_str(), { wk->ne[0], wk->ne[1] }, wq->ne[1]*wq->nb[1]);
layer.wv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
printf("====================== Merged only Q and K in layer %d because V is of different type\n", i);
LLAMA_LOG_INFO("====================== Merged only Q and K in layer %d because V is of different type\n", i);
fused_qkv = true;
if (bias) {
auto bq_name = tn(LLM_TENSOR_ATTN_Q, "bias", i);
@@ -2832,7 +2830,7 @@ bool create_tensors_helper::merge_qkv(const LLM_TN & tn, int i, int bias, bool i
if (!fused_qkv) {
if (ml.merge_qkv) {
printf("%s: did not merge Q, K, V in layer %d because %d, %d, %d\n", __func__, i,
LLAMA_LOG_INFO("%s: did not merge Q, K, V in layer %d because %d, %d, %d\n", __func__, i,
wq->type == wk->type, wq->type == wv->type, (ignore_attn_scale || hparams.f_attention_scale == 0.0f));
}
layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head});
@@ -3066,7 +3064,7 @@ bool create_tensors_helper::create_tensors() {
}
if (model.split_mode == LLAMA_SPLIT_MODE_GRAPH || model.split_mode == LLAMA_SPLIT_MODE_ATTN) {
const int n_layer = model.layers.size() - model.hparams.nextn_predict_layers;
printf("================================ max_gpu = %d\n", model.max_gpu);
LLAMA_LOG_INFO("================================ max_gpu = %d\n", model.max_gpu);
std::vector<size_t> mem_used(model.splits.size(), 0);
const auto & hparams = model.hparams;
auto cur_splits = model.splits;
@@ -3097,17 +3095,17 @@ bool create_tensors_helper::create_tensors() {
if (model.max_gpu > 0 && model.max_gpu < int(model.splits.size()) && il % adjust_step == 0) {
cur_splits = model.splits;
adjust_split(cur_splits, mem_used, model.max_gpu);
printf("Adjusted split at layer %2d:", il);
LLAMA_LOG_INFO("Adjusted split at layer %2d:", il);
float last_split = 0;
for (auto & p : cur_splits) {
printf(" %g", p - last_split);
LLAMA_LOG_INFO(" %g", p - last_split);
last_split = p;
}
printf("\n");
LLAMA_LOG_INFO("\n");
}
//printf("=== Layer %2d. Mem used so far:", il);
//for (auto mem : mem_used) printf(" %g", mem/1024./1024.);
//printf("\n");
LLAMA_LOG_DEBUG("=== Layer %2d. Mem used so far:", il);
for ([[maybe_unused]] auto mem : mem_used) LLAMA_LOG_DEBUG(" %g", mem/1024./1024.);
LLAMA_LOG_DEBUG("\n");
auto & layer = model.layers[il];
auto ctx_split = ctx_for_layer_split(il);
if (layer.attn_norm) {
@@ -3128,10 +3126,10 @@ bool create_tensors_helper::create_tensors() {
}
auto split_vo = create_split(layer.wo->ne[0], granularity_vo, cur_splits, mem_used); //, true);
auto split_kq = create_split(layer.wq->ne[1], granularity_kq, cur_splits, mem_used); //, true);
//printf(" split_vo:"); for (auto s : split_vo) printf(" %d", s);
//printf("\n");
//printf(" split_kq:"); for (auto s : split_kq) printf(" %d", s);
//printf("\n");
LLAMA_LOG_DEBUG(" split_vo:"); for ([[maybe_unused]] auto s : split_vo) LLAMA_LOG_DEBUG(" %d", s);
LLAMA_LOG_DEBUG("\n");
LLAMA_LOG_DEBUG(" split_kq:"); for ([[maybe_unused]] auto s : split_kq) LLAMA_LOG_DEBUG(" %d", s);
LLAMA_LOG_DEBUG("\n");
prepare_split_tensors(0, ctx_split, layer.wo, layer.split_wo, split_vo, mem_used);
prepare_split_tensors(1, ctx_split, layer.wq, layer.split_wq, split_kq, mem_used);
if (layer.bo) {
@@ -3183,6 +3181,7 @@ bool create_tensors_helper::create_tensors() {
if (tt.blck_size > ffn_granularity) ffn_granularity = tt.blck_size;
}
auto split = create_split(layer.ffn_down->ne[0], ffn_granularity, cur_splits, mem_used);
LLAMA_LOG_DEBUG(" split_ffn:"); for ([[maybe_unused]] auto s : split) LLAMA_LOG_DEBUG(" %d", s); LLAMA_LOG_DEBUG("\n");
prepare_split_tensors(0, ctx_split, layer.ffn_down, layer.split_ffn_down, split, mem_used);
prepare_split_tensors(1, ctx_split, layer.ffn_up, layer.split_ffn_up, split, mem_used);
prepare_split_tensors(1, ctx_split, layer.ffn_gate, layer.split_ffn_gate, split, mem_used);
@@ -3196,14 +3195,14 @@ bool create_tensors_helper::create_tensors() {
split_tensors.find(layer.ffn_up_exps) != split_tensors.end();
if (use_split) {
//any_ffn_split = true;
int ffn_granularity = 16;
if (ggml_is_quantized(layer.ffn_down_exps->type)) {
auto tt = ggml_internal_get_type_traits(layer.ffn_down_exps->type);
if (tt.blck_size > ffn_granularity) ffn_granularity = tt.blck_size;
}
ffn_split = create_split(layer.ffn_down_exps->ne[0], ffn_granularity, cur_splits, mem_used);
//printf("split(%2d):", il); for (auto & s : split) printf(" %d", s); printf("\n");
LLAMA_LOG_DEBUG(" split_ffn_exps:"); for ([[maybe_unused]] auto s : ffn_split) LLAMA_LOG_DEBUG(" %d", s);
LLAMA_LOG_DEBUG("\n");
prepare_split_tensors(0, ctx_split, layer.ffn_down_exps, layer.split_ffn_down_exps, ffn_split, mem_used);
prepare_split_tensors(1, ctx_split, layer.ffn_up_exps, layer.split_ffn_up_exps, ffn_split, mem_used);
prepare_split_tensors(1, ctx_split, layer.ffn_gate_exps, layer.split_ffn_gate_exps, ffn_split, mem_used);
@@ -3242,9 +3241,9 @@ bool create_tensors_helper::create_tensors() {
}
}
if (!ok) {
printf("=== exp/shexp mismatch in layer %d\n", il);
printf(" experts:"); for (auto& s : ffn_split) printf(" %d", s); printf("\n");
printf(" sh_experts:"); for (auto& s : split ) printf(" %d", s); printf("\n");
LLAMA_LOG_INFO("=== exp/shexp mismatch in layer %d\n", il);
LLAMA_LOG_INFO(" experts:"); for (auto& s : ffn_split) LLAMA_LOG_INFO(" %d", s); LLAMA_LOG_INFO("\n");
LLAMA_LOG_INFO(" sh_experts:"); for (auto& s : split ) LLAMA_LOG_INFO(" %d", s); LLAMA_LOG_INFO("\n");
std::vector<float> aux(ffn_split.size());
float sum = 0;
for (int j = 0; j < int(ffn_split.size()); ++j) {
@@ -3253,7 +3252,10 @@ bool create_tensors_helper::create_tensors() {
}
for (auto& s : aux) s /= sum;
split = create_split(layer.ffn_down_shexp->ne[0], ffn_granularity, aux, mem_used);
printf(" new:"); for (auto& s : split ) printf(" %d", s); printf("\n");
LLAMA_LOG_INFO(" new:"); for (auto& s : split ) LLAMA_LOG_INFO(" %d", s); LLAMA_LOG_INFO("\n");
} else {
LLAMA_LOG_DEBUG(" split_ffn_shexps:"); for ([[maybe_unused]] auto s : split) LLAMA_LOG_DEBUG(" %d", s);
LLAMA_LOG_DEBUG("\n");
}
prepare_split_tensors(0, ctx_split, layer.ffn_down_shexp, layer.split_ffn_down_shexp, split, mem_used);
prepare_split_tensors(1, ctx_split, layer.ffn_up_shexp, layer.split_ffn_up_shexp, split, mem_used);