Skip to content

Full graph parallel for Qwen3.5 (dense and MoE) #1388

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ikawrakow merged 22 commits into from
Mar 10, 2026
Merged

Full graph parallel for Qwen3.5 (dense and MoE) #1388

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
22 commits
Select commit Hold shift + click to select a range
30d2f2f
WIP
ikawrakow Mar 2, 2026
4e39d16
WIP
ikawrakow Mar 2, 2026
3e57fa2
WIP
ikawrakow Mar 2, 2026
b9f31f5
WIP
ikawrakow Mar 2, 2026
01b7f75
WIP
ikawrakow Mar 2, 2026
eb40760
WIP
ikawrakow Mar 3, 2026
3cec013
WIP
ikawrakow Mar 3, 2026
84f351f
WIP
ikawrakow Mar 3, 2026
6f854c5
WIP
ikawrakow Mar 3, 2026
c04127e
WIP - still wrong
ikawrakow Mar 3, 2026
c8d8c48
Fix after rebase
ikawrakow Mar 6, 2026
49c5ea6
WIP
ikawrakow Mar 6, 2026
22b834b
WIP
ikawrakow Mar 7, 2026
54e4ed0
This seems to be working for dense Qwen3.5!!!
ikawrakow Mar 7, 2026
3187a72
WIP: Qwen3-Next is not quite working
ikawrakow Mar 7, 2026
7c7f6c6
Some cleanup
ikawrakow Mar 7, 2026
38d4cde
Disable Qwen3-Next for now
ikawrakow Mar 8, 2026
891b6ee
Disable graph parallel when mmproj was specified
ikawrakow Mar 9, 2026
b29e613
Read/write split recurrent state
ikawrakow Mar 9, 2026
23d9a7f
That should not crash
ikawrakow Mar 9, 2026
f09d421
Re-enable vision - it works now
ikawrakow Mar 10, 2026
0fe8599
Recurrent layers should now be counted for split cache
ikawrakow Mar 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 91 additions & 34 deletions ggml/src/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -907,40 +907,70 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor([[maybe_unused]]
else if (extra->split_dim == 0) {
int n_interleave = 1;
if (auto it = k_map.find(tensor->type); it != k_map.end()) n_interleave = it->second;
//if (tensor->type >= GGML_TYPE_Q4_0_R8) {
// GGML_ABORT("Dim 0 copy of row-interleaved quants is not supported yet");
//}
auto tt = ggml_internal_get_type_traits(tensor->type);
std::vector<char> host_buffer;
GGML_ASSERT(ggml_is_contiguous(tensor));
int nrows = ggml_nrows(tensor);
auto bs = tt.blck_size;
auto ts = tt.type_size;
auto row_size = ggml_row_size(tensor->type, tensor->ne[0]);
int ne = 0;
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) continue;
GGML_ASSERT(split->ne[1]%n_interleave == 0);
ggml_cuda_set_device(i);
GGML_ASSERT(split->type == tensor->type);
GGML_ASSERT((int)ggml_nrows(split) == nrows);
GGML_ASSERT(split->ne[0] % bs == 0);
auto source_offset = n_interleave*(tt.row_meta_size + (ne / bs) * ts);
auto split_row_size = ggml_row_size(split->type, split->ne[0]);
if (host_buffer.size() < nrows*split_row_size) host_buffer.resize(nrows*split_row_size);
for (int64_t i02 = 0; i02 < split->ne[2]; ++i02) {
void * extra_ptr;
memcpy(&extra_ptr, tensor->op_params, sizeof(extra_ptr));
if (extra_ptr) {
auto & ranges = *(const std::vector<std::vector<std::pair<int,int>>> *)extra_ptr;
GGML_ASSERT(extra->n_device == int(ranges.size()));
GGML_ASSERT(tensor->ne[2]*tensor->ne[3] == 1);
GGML_ASSERT(n_interleave == 1);
GGML_ASSERT(tt.row_meta_size == 0);
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) {
GGML_ASSERT(ranges[i].empty());
continue;
}
GGML_ASSERT(!ranges[i].empty());
GGML_ASSERT((int)ggml_nrows(split) == nrows);
auto split_row_size = ggml_row_size(split->type, split->ne[0]);
if (host_buffer.size() < nrows*split_row_size) host_buffer.resize(nrows*split_row_size);
auto dst = host_buffer.data();
for (int64_t i01 = 0; i01 < split->ne[1]; i01 += n_interleave) {
auto dst = host_buffer.data() + (i02*split->ne[1] + i01)*split_row_size;
auto src = (const char *)data + i02*tensor->nb[2] + i01*tensor->nb[1];
if (tt.row_meta_size > 0) {
memcpy(dst, src, tt.row_meta_size*n_interleave);
for (auto & p : ranges[i]) {
GGML_ASSERT(p.first % bs == 0);
GGML_ASSERT(p.second % bs == 0);
auto src = (const char *)data + i01*tensor->nb[1] + (p.first/bs)*ts;
auto size = (p.second/bs)*ts;
memcpy(dst, src, size);
dst += size;
}
}
ggml_cuda_set_device(i);
CUDA_CHECK(cudaMemcpyAsync(split->data, host_buffer.data(), nrows*split_row_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
}
} else {
int ne = 0;
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) continue;
GGML_ASSERT(split->ne[1]%n_interleave == 0);
ggml_cuda_set_device(i);
GGML_ASSERT(split->type == tensor->type);
GGML_ASSERT((int)ggml_nrows(split) == nrows);
GGML_ASSERT(split->ne[0] % bs == 0);
auto source_offset = n_interleave*(tt.row_meta_size + (ne / bs) * ts);
auto split_row_size = ggml_row_size(split->type, split->ne[0]);
if (host_buffer.size() < nrows*split_row_size) host_buffer.resize(nrows*split_row_size);
for (int64_t i02 = 0; i02 < split->ne[2]; ++i02) {
for (int64_t i01 = 0; i01 < split->ne[1]; i01 += n_interleave) {
auto dst = host_buffer.data() + (i02*split->ne[1] + i01)*split_row_size;
auto src = (const char *)data + i02*tensor->nb[2] + i01*tensor->nb[1];
if (tt.row_meta_size > 0) {
memcpy(dst, src, tt.row_meta_size*n_interleave);
}
memcpy(dst + tt.row_meta_size*n_interleave, src + source_offset, n_interleave*(split_row_size - tt.row_meta_size));
}
memcpy(dst + tt.row_meta_size*n_interleave, src + source_offset, n_interleave*(split_row_size - tt.row_meta_size));
}
CUDA_CHECK(cudaMemcpyAsync(split->data, host_buffer.data(), nrows*split_row_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
ne += split->ne[0];
}
CUDA_CHECK(cudaMemcpyAsync(split->data, host_buffer.data(), nrows*split_row_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
ne += split->ne[0];
}
}
else if (extra->split_dim == 1) {
Expand All @@ -965,16 +995,43 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor([[maybe_unused]]
} else {
int n_interleave = 1;
if (auto it = k_map.find(tensor->type); it != k_map.end()) n_interleave = it->second;
size_t cur_offset = 0;
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) continue;
GGML_ASSERT(split->ne[1]%n_interleave == 0);
ggml_cuda_set_device(i);
auto size = ggml_nbytes(split);
const char * buf_host = (const char *)data + cur_offset;
CUDA_CHECK(cudaMemcpyAsync(split->data, buf_host, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
cur_offset += size;
void * extra_ptr;
memcpy(&extra_ptr, tensor->op_params, sizeof(extra_ptr));
if (extra_ptr) {
auto & ranges = *(const std::vector<std::vector<std::pair<int,int>>> *)extra_ptr;
GGML_ASSERT(extra->n_device == int(ranges.size()));
GGML_ASSERT(tensor->ne[2]*tensor->ne[3] == 1);
GGML_ASSERT(n_interleave == 1);
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) {
GGML_ASSERT(ranges[i].empty());
continue;
}
GGML_ASSERT(!ranges[i].empty());
ggml_cuda_set_device(i);
auto dst = (char *)split->data;
for (auto & p : ranges[i]) {
GGML_ASSERT(p.first >= 0 && p.first < tensor->ne[1]);
GGML_ASSERT(p.second >= 0 && p.first + p.second <= tensor->ne[1]);
auto src = (const char *)data + p.first*tensor->nb[1];
auto size = p.second*tensor->nb[1];
CUDA_CHECK(cudaMemcpyAsync(dst, src, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
dst += size;
}
}
} else {
size_t cur_offset = 0;
for (int i = 0; i < extra->n_device; ++i) {
auto split = extra->splits[i];
if (!split) continue;
GGML_ASSERT(split->ne[1]%n_interleave == 0);
ggml_cuda_set_device(i);
auto size = ggml_nbytes(split);
const char * buf_host = (const char *)data + cur_offset;
CUDA_CHECK(cudaMemcpyAsync(split->data, buf_host, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
cur_offset += size;
}
}
}
}
Expand Down
1 change: 1 addition & 0 deletions ggml/src/ggml-cuda/delta-net.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -157,6 +157,7 @@ __global__ void delta_net_recurrent_f32(
}

}
__syncthreads();
// Copy the final state to its destination
for (int i = 0; i < HEAD_DIM/num_warps; ++i) {
int col = num_warps*i + col_idx_0;
Expand Down
69 changes: 3 additions & 66 deletions src/llama-build-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4437,7 +4437,6 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
ggml_tensor * cur = nullptr;

for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;

GGML_ASSERT(model.layers[il].attn_norm != nullptr);
GGML_ASSERT(model.layers[il].attn_post_norm != nullptr);
Expand All @@ -4455,27 +4454,7 @@ ggml_cgraph * llm_build_context::build_qwen3next() {


if (hparams.is_recurrent(il)) {
int idx = model.default_layer_device[il];
if (inpL->op == GGML_OP_REDUCE) {
if (kv_self.s_l[il]) {
// This shouldn't be necessary, but just in case.
int idx_s_l = ggml_backend_sched_get_backend_idx(lctx.sched, kv_self.s_l[il]->buffer);
if (idx_s_l >= 0) idx = idx_s_l;
}
if (inpL->src[idx]) {
inpL->view_src = inpL->src[idx];
}
}
auto norm = model.layers[il].attn_norm->extra ? ((ggml_split_tensor_t *)model.layers[il].attn_norm->extra)->splits[idx] : model.layers[il].attn_norm;
cur = llm_build_norm(ctx0, inpL, hparams, norm, nullptr, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
cur = delta.build_layer_attn_linear(ctx0, gf, cur, il, cb);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "attn_residual", il);
cur = delta.build_layer_attn_linear(ctx0, gf, inpL, il == n_layer - 1 ? inp_out_ids : nullptr, il, cb);
} else {
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr, nullptr,
KQ_mask, nullptr, nullptr, KQ_scale, 0.0f, 0, il, true, false, true, false, false);
Expand Down Expand Up @@ -4545,28 +4524,7 @@ ggml_cgraph * llm_build_context::build_qwen35moe() {
for (int il = 0; il < n_layer; ++il) {

if (hparams.is_recurrent(il)) {
ggml_tensor * inpSA = inpL;
int idx = model.default_layer_device[il];
if (inpL->op == GGML_OP_REDUCE) {
if (kv_self.s_l[il]) {
// This shouldn't be necessary, but just in case.
int idx_s_l = ggml_backend_sched_get_backend_idx(lctx.sched, kv_self.s_l[il]->buffer);
if (idx_s_l >= 0) idx = idx_s_l;
}
if (inpL->src[idx]) {
inpL->view_src = inpL->src[idx];
}
}
auto norm = model.layers[il].attn_norm->extra ? ((ggml_split_tensor_t *)model.layers[il].attn_norm->extra)->splits[idx] : model.layers[il].attn_norm;
cur = llm_build_norm(ctx0, inpL, hparams, norm, nullptr, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
cur = delta.build_layer_attn_linear(ctx0, gf, cur, il, cb);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "attn_residual", il);
cur = delta.build_layer_attn_linear(ctx0, gf, inpL, il == n_layer - 1 ? inp_out_ids : nullptr, il, cb);
} else {
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr, nullptr,
KQ_mask, nullptr, nullptr, KQ_scale, 0.0f, 0, il, true, false, true, false, true);
Expand Down Expand Up @@ -4625,28 +4583,7 @@ ggml_cgraph * llm_build_context::build_qwen35() {
for (int il = 0; il < n_layer; ++il) {

if (hparams.is_recurrent(il)) {
ggml_tensor * inpSA = inpL;
int idx = model.default_layer_device[il];
if (inpL->op == GGML_OP_REDUCE) {
if (kv_self.s_l[il]) {
// This shouldn't be necessary, but just in case.
int idx_s_l = ggml_backend_sched_get_backend_idx(lctx.sched, kv_self.s_l[il]->buffer);
if (idx_s_l >= 0) idx = idx_s_l;
}
if (inpL->src[idx]) {
inpL->view_src = inpL->src[idx];
}
}
auto norm = model.layers[il].attn_norm->extra ? ((ggml_split_tensor_t *)model.layers[il].attn_norm->extra)->splits[idx] : model.layers[il].attn_norm;
cur = llm_build_norm(ctx0, inpL, hparams, norm, nullptr, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
cur = delta.build_layer_attn_linear(ctx0, gf, cur, il, cb);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "attn_residual", il);
cur = delta.build_layer_attn_linear(ctx0, gf, inpL, il == n_layer - 1 ? inp_out_ids : nullptr, il, cb);
} else {
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr, nullptr,
KQ_mask, nullptr, nullptr, KQ_scale, 0.0f, 0, il, true, false, true, false, true);
Expand Down
1 change: 1 addition & 0 deletions src/llama-context.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,7 @@ struct llama_kv_cache {

std::vector<llama_split_tensor> split_k_l;
std::vector<llama_split_tensor> split_v_l;
std::vector<llama_split_tensor> split_s_l;

std::vector<struct ggml_context *> ctxs;
std::vector<ggml_backend_buffer_t> bufs;
Expand Down
Loading