Skip to content

feat: support MOSS split QKV weights and batch host loading #2160

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
leejet merged 1 commit into from
Oct 18, 2026
Merged

feat: support MOSS split QKV weights and batch host loading #2160

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
73 changes: 54 additions & 19 deletions src/model_manager.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1022,28 +1022,60 @@ bool ModelManager::load_tensor_parts(TensorState& state) {

bool ModelManager::load_tensors(const std::vector<TensorState*>& states) {
using ReadGroup = std::pair<ModelLoader::FileId, SDVersion>;
using ReadBatch = std::map<std::string, std::vector<TensorState*>>;
using ReadBatch = std::map<std::string, std::vector<ggml_tensor*>>;
std::map<ReadGroup, std::vector<ReadBatch>> groups;
size_t part_count = 0;
for (const auto* state : states) {
if (state != nullptr && state->sources.size() > 1 && ggml_backend_buffer_is_host(state->tensor->buffer)) {
part_count += state->sources.size();
}
}
auto parts_ctx = std::unique_ptr<ggml_context, decltype(&ggml_free)>(
part_count > 0 ? ggml_init({part_count * ggml_tensor_overhead(), nullptr, true}) : nullptr, ggml_free);
if (part_count > 0 && !parts_ctx) {
return false;
}
auto add_destination = [&](const TensorState& state, const std::string& name, ggml_tensor* tensor) {
auto& batches = groups[{state.source_file, state.source_version}];
// The loader supplies one destination per name; incompatible destinations need another batch.
auto batch = std::find_if(batches.begin(), batches.end(), [&](const ReadBatch& candidate) {
auto found = candidate.find(name);
return found == candidate.end() ||
(found->second.front()->type == tensor->type && ggml_are_same_shape(found->second.front(), tensor));
});
if (batch == batches.end()) {
batches.emplace_back();
batch = std::prev(batches.end());
}
(*batch)[name].push_back(tensor);
};
for (auto* state : states) {
if (state == nullptr)
continue;
if (state->sources.size() > 1) {
if (!load_tensor_parts(*state)) {
if (ggml_backend_buffer_is_host(state->tensor->buffer)) {
const size_t size = ggml_nbytes(state->tensor);
size_t offset = 0;
for (const auto& source : state->sources) {
auto* part = ggml_new_tensor(parts_ctx.get(), state->tensor->type, source.n_dims, source.ne);
const size_t part_size = ggml_nbytes(part);
if (part_size > size - offset) {
return false;
}
part->data = static_cast<uint8_t*>(state->tensor->data) + offset;
part->buffer = state->tensor->buffer;
add_destination(*state, source.name, part);
offset += part_size;
}
if (offset != size) {
return false;
}
} else if (!load_tensor_parts(*state)) {
return false;
}
continue;
}
auto& batches = groups[{state->source_file, state->source_version}];
// The loader supplies one destination per name; only conflicting types need another batch.
auto batch = std::find_if(batches.begin(), batches.end(), [&](const ReadBatch& candidate) {
auto found = candidate.find(state->name);
return found == candidate.end() || found->second.front()->tensor->type == state->tensor->type;
});
if (batch == batches.end()) {
batches.emplace_back();
batch = std::prev(batches.end());
}
(*batch)[state->name].push_back(state);
add_destination(*state, state->name, state->tensor);
}
for (auto& group : groups) {
for (auto& batch : group.second) {
Expand All @@ -1057,7 +1089,7 @@ bool ModelManager::load_tensors(const std::vector<TensorState*>& states) {
auto found = batch.find(source.name);
if (found == batch.end())
return true;
*dst = found->second.front()->tensor;
*dst = found->second.front();
std::lock_guard<std::mutex> lock(mutex);
loaded.insert(source.name);
return true;
Expand All @@ -1068,15 +1100,18 @@ bool ModelManager::load_tensors(const std::vector<TensorState*>& states) {
if (!success || loaded != names)
return false;
for (auto& entry : batch) {
auto* first = entry.second.front()->tensor;
for (auto* state : entry.second) {
if (state->tensor != first)
ggml_backend_tensor_copy(first, state->tensor);
state->loaded_to_params_backend = true;
auto* first = entry.second.front();
for (auto* tensor : entry.second) {
if (tensor != first)
ggml_backend_tensor_copy(first, tensor);
}
}
}
}
for (auto* state : states) {
if (state != nullptr)
state->loaded_to_params_backend = true;
}
return true;
}

Expand Down
12 changes: 12 additions & 0 deletions src/name_conversion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1013,6 +1013,16 @@ static std::string convert_diffusers_dit_to_original_pixart(std::string name) {
return prefix + name;
}

static std::string convert_moss_image_dit_name(std::string name) {
static const std::vector<std::pair<std::string, std::string>> name_map = {
{".proj.q.weight", ".proj.qkv.weight"},
{".proj.k.weight", ".proj.qkv.weight.1"},
{".proj.v.weight", ".proj.qkv.weight.2"},
};
replace_with_name_map(name, name_map);
return name;
}

static std::string convert_ming_image_dit_name(std::string name) {
static const std::vector<std::pair<std::string, std::string>> name_map = {
{"all_x_embedder.2-1.", "x_embedder."},
Expand Down Expand Up @@ -1053,6 +1063,8 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
name = convert_diffusers_dit_to_original_flux(name);
} else if (sd_version_is_hunyuan_video(version)) {
name = convert_hunyuan_video_to_original_flux(name);
} else if (version == VERSION_MOSS_IMAGE) {
name = convert_moss_image_dit_name(name);
} else if (version == VERSION_MING_IMAGE) {
name = convert_ming_image_dit_name(name);
} else if (sd_version_is_z_image(version) || sd_version_is_z_image_l2p(version)) {
Expand Down
Loading