Skip to content

fix(minimax-h3): support ComfyUI GGUF tensor layouts #30

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
aegioscy merged 9 commits into from
Aug 24, 2026
Merged

fix(minimax-h3): support ComfyUI GGUF tensor layouts #30

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
5784620
fix(minimax-h3): support ComfyUI GGUF tensor layouts
aegioscy Aug 20, 2026
47b0ee1
Merge base branch before ComfyUI loader fix
Aug 20, 2026
d83cd77
fix(minimax-h3): harden ComfyUI GGUF loading
Aug 24, 2026
5322d7a
fix(minimax-h3): write restored shapes directly
Aug 24, 2026
6961784
fix(minimax-h3): accept ComfyUI five-dimensional vision shapes
Aug 24, 2026
3c79424
fix(minimax-h3): allow ComfyUI vision shape metadata
Aug 24, 2026
b3cdb0d
fix(minimax-h3): flatten ComfyUI vision shapes
Aug 24, 2026
208d6fa
chore(minimax-h3): update ggml submodule to PR61
aegioscy Aug 24, 2026
90e2a3b
fix(minimax-h3): harden ComfyUI loader boundaries
aegioscy Aug 24, 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
2 changes: 1 addition & 1 deletion ggml
Open in desktop
Submodule ggml updated 1 files
+9 −1 src/ggml-cpu/CMakeLists.txt
29 changes: 23 additions & 6 deletions src/model/diffusion/minimax_h3.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@ namespace MiniMaxH3 {
int64_t time_embed_dim = 2688;
int64_t rope_inv_freq_len = 16;
int64_t adaln_curve_grid = 0;
bool is_comfyui_layout = false;
int patch_t = 1;
int patch_h = 2;
int patch_w = 2;
Expand Down Expand Up @@ -75,6 +76,10 @@ namespace MiniMaxH3 {
}
if (const auto* weight = find("audio_patch_proj.weight")) {
config.audio_latent_channels = weight->ne[0];
// ComfyUI GGUF exports can store video_patch_proj transposed,
// so derive the transformer width from the unambiguous audio
// projection when it is wider than the video-derived value.
config.hidden_size = std::max(config.hidden_size, weight->ne[1]);
}
config.num_layers = count_blocks(tensors, prefix + ".blocks.");
config.token_refiner_num_layers = count_blocks(tensors, prefix + ".token_refiner.blocks.");
Expand Down Expand Up @@ -105,15 +110,19 @@ namespace MiniMaxH3 {
if (const auto* inv_freq = find("rope.inv_freq")) {
config.rope_inv_freq_len = inv_freq->ne[0];
}
const auto* adaln = find("blocks.0.adaln_proj.linear.weight");
config.is_comfyui_layout = adaln != nullptr && adaln->has_comfy_original_shape;

LOG_DEBUG("minimax_h3: layers=%" PRId64 ", hidden=%" PRId64 ", heads=%" PRId64
", head_dim=%" PRId64 ", ffn=%" PRId64 ", adaln_curve=%" PRId64,
", head_dim=%" PRId64 ", ffn=%" PRId64 ", adaln_curve=%" PRId64
", comfyui_layout=%d",
config.num_layers,
config.hidden_size,
config.num_attention_heads,
config.attention_head_dim,
config.ffn_hidden_size,
config.adaln_curve_grid);
config.adaln_curve_grid,
config.is_comfyui_layout);
return config;
}
};
Expand Down Expand Up @@ -286,15 +295,17 @@ namespace MiniMaxH3 {
int expand,
int modalities,
bool apply_silu,
bool force_f32)
bool force_f32,
bool force_prec_f32)
: hidden_size(hidden_size),
expand(expand),
modalities(modalities),
apply_silu(apply_silu) {
blocks["linear"] = std::make_shared<Linear>(time_dim,
hidden_size * expand * modalities,
true,
force_f32);
force_f32,
force_prec_f32);
}

ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t_emb) {
Expand Down Expand Up @@ -403,7 +414,10 @@ namespace MiniMaxH3 {
6,
3,
!config.uses_adaln_curves(),
config.uses_adaln_curves());
config.uses_adaln_curves() &&
Comment thread
aegioscy marked this conversation as resolved.
!config.is_comfyui_layout,
config.uses_adaln_curves() &&
config.is_comfyui_layout);
}

ggml_tensor* forward(GGMLRunnerContext* ctx,
Expand Down Expand Up @@ -465,7 +479,10 @@ namespace MiniMaxH3 {
2,
1,
!config.uses_adaln_curves(),
config.uses_adaln_curves());
config.uses_adaln_curves() &&
!config.is_comfyui_layout,
config.uses_adaln_curves() &&
config.is_comfyui_layout);
blocks["video_out"] = std::make_shared<Linear>(config.hidden_size, video_dim, true, true);
blocks["audio_out"] = std::make_shared<Linear>(config.hidden_size, config.audio_latent_channels, true, true);
}
Expand Down
Loading
Loading