Skip to content

feat: add split qkv support to moss-image main blocks #2153

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

Closed
wbruna wants to merge 1 commit into from
+30 −8
Closed

feat: add split qkv support to moss-image main blocks #2153

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
38 changes: 30 additions & 8 deletions src/model/diffusion/moss_image.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@ namespace MOSSImage {
int64_t head_dim = 128;
int64_t depth = 42;
int64_t freq_dim = 256;
bool split_qkv = false;

static MOSSImageConfig detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
MOSSImageConfig config;
Expand All @@ -40,6 +41,9 @@ namespace MOSSImage {
} else if (key == "time_embedding.0.weight") {
config.freq_dim = tensor.ne[0];
}
if (key == "layers.0.proj.q.weight") {
config.split_qkv = true;
}
if (starts_with(key, "layers.")) {
depth = std::max<int64_t>(depth, std::stoi(key.substr(7)) + 1);
}
Expand Down Expand Up @@ -107,7 +111,13 @@ namespace MOSSImage {
blocks["norm2"] = std::make_shared<RMSNorm>(config.dim, 1e-6f);
blocks["q_norm"] = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
blocks["k_norm"] = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
blocks["proj.qkv"] = std::make_shared<Linear>(config.dim, config.dim * 3, false);
if (config.split_qkv) {
blocks["proj.q"] = std::make_shared<Linear>(config.dim, config.dim, false);
blocks["proj.k"] = std::make_shared<Linear>(config.dim, config.dim, false);
blocks["proj.v"] = std::make_shared<Linear>(config.dim, config.dim, false);
} else {
blocks["proj.qkv"] = std::make_shared<Linear>(config.dim, config.dim * 3, false);
}
blocks["proj.o"] = std::make_shared<Linear>(config.dim, config.dim, false);
blocks["proj.ffn_up_gate"] = std::make_shared<Linear>(config.dim, config.ffn_dim * 2, false);
blocks["proj.ffn_down"] = std::make_shared<Linear>(config.ffn_dim, config.dim, false);
Expand All @@ -120,13 +130,25 @@ namespace MOSSImage {
auto part = [&](int i) { return ggml_ext_slice(c, mod, 0, i * config.dim, (i + 1) * config.dim); };
auto y = norm("norm1", x);
y = ggml_add(c, ggml_add(c, y, ggml_mul(c, y, part(1))), part(0));
auto qkv = linear("proj.qkv", y);
auto split = [&](int i) {
return ggml_reshape_4d(c, ggml_ext_slice(c, qkv, 0, i * config.dim, (i + 1) * config.dim), config.head_dim, config.dim / config.head_dim, x->ne[1], x->ne[2]);
};
auto q = norm("q_norm", split(0));
auto k = norm("k_norm", split(1));
auto v = split(2);
ggml_tensor *q, *k, *v;
if (config.split_qkv) {
auto reshape = [&](ggml_tensor* t) {
return ggml_reshape_4d(c, t, config.head_dim, config.dim / config.head_dim, x->ne[1], x->ne[2]);
};
q = reshape(linear("proj.q", y));
k = reshape(linear("proj.k", y));
v = reshape(linear("proj.v", y));
} else {
auto qkv = linear("proj.qkv", y);
auto split = [&](int i) {
return ggml_reshape_4d(c, ggml_ext_slice(c, qkv, 0, i * config.dim, (i + 1) * config.dim), config.head_dim, config.dim / config.head_dim, x->ne[1], x->ne[2]);
};
q = split(0);
k = split(1);
v = split(2);
}
q = norm("q_norm", q);
k = norm("k_norm", k);
y = Rope::attention(ctx, q, k, v, pe, nullptr, 1.f, false);
y = linear("proj.o", y);
y = ggml_mul(c, y, part(2));
Expand Down
Loading