Skip to content

fix: reduce MiniMax-H3 VRAM spikes and redundant token refinemen #1900

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 5 commits into from
Sep 23, 2026
Merged

fix: reduce MiniMax-H3 VRAM spikes and redundant token refinemen #1900

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
56ba2f1
fix: cache MiniMax-H3 refined context during sampling
assouan Aug 21, 2026
34a3146
fix: harden MiniMax-H3 refined context streaming
assouan Aug 22, 2026
a820acf
Merge branch 'master' into model/minimax-h3/token-refiner-cache
leejet Sep 23, 2026
3587885
fix: use RunnerCache for MiniMax-H3 context caching
leejet Sep 23, 2026
2858438
revert token refiner cache
leejet Sep 23, 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
9 changes: 8 additions & 1 deletion src/model/diffusion/minimax_h3.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -264,6 +264,9 @@ namespace MiniMaxH3 {
for (int64_t i = 0; i < num_layers; ++i) {
auto block = std::dynamic_pointer_cast<TokenRefinerBlock>(blocks["blocks." + std::to_string(i)]);
x = block->forward(ctx, x);
sd::ggml_graph_cut::mark_graph_cut(x,
"minimax_h3.token_refiner.blocks." + std::to_string(i),
"hidden_states");
}
return std::dynamic_pointer_cast<RMSNorm>(blocks["final_norm"])->forward(ctx, x);
}
Expand Down Expand Up @@ -527,7 +530,11 @@ namespace MiniMaxH3 {
GGML_ASSERT(context->ne[0] == config.text_dim);
auto condition_proj = std::dynamic_pointer_cast<Linear>(blocks["condition_proj"]);
auto token_refiner = std::dynamic_pointer_cast<TokenRefiner>(blocks["token_refiner"]);
return token_refiner->forward(ctx, condition_proj->forward(ctx, context));
auto projected = condition_proj->forward(ctx, context);
sd::ggml_graph_cut::mark_graph_cut(projected,
"minimax_h3.condition_proj",
"hidden_states");
return token_refiner->forward(ctx, projected);
}

ggml_tensor* time_embedding(GGMLRunnerContext* ctx,
Expand Down
14 changes: 7 additions & 7 deletions src/model/diffusion/qwen_image_2_1.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -178,13 +178,13 @@ namespace Qwen {
auto h = std::dynamic_pointer_cast<Linear>(blocks[name])->forward(ctx, x);
return ggml_reshape_4d(ctx->ggml_ctx, h, dim_head, heads, x->ne[1], x->ne[2]);
};
auto q = project("to_q");
auto k = project("to_k");
auto v = project("to_v");
q = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"])->forward(ctx, q);
k = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"])->forward(ctx, k);
q = Rope::apply_rope(ctx->ggml_ctx, q, pe);
k = Rope::apply_rope(ctx->ggml_ctx, k, pe);
auto q = project("to_q");
auto k = project("to_k");
auto v = project("to_v");
q = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"])->forward(ctx, q);
k = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"])->forward(ctx, k);
q = Rope::apply_rope(ctx->ggml_ctx, q, pe);
k = Rope::apply_rope(ctx->ggml_ctx, k, pe);
if (cache.mode == QwenImage21PrefixCache::Mode::STORE) {
auto persist = [&](ggml_tensor* tensor, int axis, const char* name) {
auto part = ggml_ext_slice(ctx->ggml_ctx, tensor, axis, 0, cache.prefix_length);
Expand Down
Loading