Fail with a message when MiniMax-H3 is run in img_gen mode - #1863
Merged
Merged
Conversation
H3 is video-only: the denoiser always splits the packed latent into a video and
an audio half, and only generate_video computes the audio length. Reaching
generate_image with an H3 checkpoint therefore hits
GGML_ASSERT(!audio_input_cache.empty()) and core dumps, after the minutes it
takes to load the weights and with nothing in the output naming the cause.
Forgetting --mode vid_gen is easy since it is the one flag not implied by
passing --audio-vae.
Before: SIGABRT, exit 134, a ggml assert and a stack trace.
After: "MiniMax-H3 is a video model and cannot be run in img_gen mode; use
--mode vid_gen", exit 1.
The AnimateDiff branch routes vid_gen back through generate_image, but that is
SD1.5 plus a motion module and never H3, so the guard cannot fire there.
danielhanchen
added a commit
to unslothai/stable-diffusion.cpp
that referenced
this pull request
Aug 8, 2026
The prebuilt pipeline builds leejet's source at an aged release tag, not this fork's master, so the three MiniMax-H3 fixes on master reach nobody: every Studio user installs a binary that still aborts on the default cfg-scale, still aborts on --vae-on-cpu, and still quantizes H3's 1-D norms into an output uncorrelated with its own bf16 reference. Building from master instead would throw away the reason the pipeline is shaped this way, which is that what we publish should be traceable to a specific upstream release. So keep the upstream tag as the base and carry the delta explicitly: - patches/ holds one file per fix, each with its upstream pull request in the header. All three are open on leejet: leejet#1861, leejet#1862, leejet#1863. - resolve applies them to the checked-out tag, after running git apply --check over the whole set so a stale patch stops the run before the tree is half modified. That failure is the signal to delete the patch (upstream merged it) or refresh it (upstream moved the code). - a non-empty set moves the published tag to <upstream tag>-u<id>, where id is the sha256 prefix of the concatenated patches. The tag then says whether a box is stock, and a changed patch set republishes rather than matching an existing release and skipping. - the manifest and the release notes both record the applied list. An empty patches/ leaves the tag and every asset name exactly as they are today. Verified by running the resolve step against master-813-bfbef5b with gh stubbed: all three patches apply, the tag becomes master-813-bfbef5b-u<id>, and the stamped source tarball contains the fixes.
Owner
|
Thank you for your contribution. |
rodrigomatta
added a commit
to romara-labs/stable-diffusion.cpp
that referenced
this pull request
Aug 13, 2026
Upstream commits merged: - feat: support INT8 ConvRot safetensors (leejet#1857) - fix: fail with a message when MiniMax-H3 is run in img_gen mode (leejet#1863) - fix: preserve "token_refiner" token for MiniMax H3 LoRAs (leejet#1864) - fix: replace free_compute_buffer with runner_done in vae (leejet#1872) - sync: update ggml (leejet#1873) - fix(ci): trigger builds for ggml updates No conflicts. The local vae.hpp runner_done fix was byte-identical to leejet#1872. Local ClipProj/MiniMax-H3 support and server SSE streaming preserved.
danielhanchen
added a commit
to unslothai/stable-diffusion.cpp
that referenced
this pull request
Sep 21, 2026
* fix: preserve "token_refiner" token for MiniMax H3 LoRAs (leejet#1864) * fix: fail with a message when MiniMax-H3 is run in img_gen mode (leejet#1863) * feat: support INT8 ConvRot safetensors (leejet#1857) * fix: replace free_compute_buffer with runner_done in vae (leejet#1872) * sync: update ggml (leejet#1873) * fix(ci): trigger builds for ggml updates * feat: add taeh3 support (leejet#1874) * fix: prevent gallocr hash overflow in tiny graph-cut segments (leejet#1880) * fix: re-clamp streaming VRAM budget to currently free memory (leejet#1878) * fix: mark graph cuts with both a prefix and a suffix (leejet#1883) * fix: make max_order of lms sampler configurable (leejet#1885) * fix: guard against missing sampler/scheduler names (leejet#1887) * chore: format code * fix: use sd_get_preview_interval() (leejet#1907) * feat: configurable image / video compression (leejet#1909) * feat: support standard Qwen3-VL weights for MiniMax-H3 (leejet#1910) * feat: load scaled FP8 weights without upfront conversion (leejet#1913) * fix: match exact weights in LLM config detection (leejet#1923) * feat: add LTX-2.5 support (leejet#1893) Co-authored-by: leejet <leejet714@gmail.com> * fix: correct MiniMax H3 reference audio encoding (leejet#1886) * fix: correct MiniMax H3 audio Euler steps (leejet#1908) * feat: use backend-native FP8 matmul when supported (leejet#1916) * sync: update ggml * feat: additional `--preview-interval` values (leejet#1915) Co-authored-by: leejet <leejet714@gmail.com> * feat: support numbering for preview images (leejet#1895) * fix: use carrier sampling for MiniMax H3 audio (leejet#1924) * feat: generalize temporal tiling across video VAEs (leejet#1926) * feat: prefetch streamed layers during compute (leejet#1905) Co-authored-by: leejet <leejet714@gmail.com> * refactor: unify runner lifecycles and weight residency (leejet#1940) * feat: add verbose logging and log-level selection (leejet#1941) * feat: enable single-GPU auto-fit with tiered parameter placement (leejet#1942) * fix: reuse graph cut plans across CFG passes (leejet#1943) * refactor: split ggml extensions and move implementations to cpp files (leejet#1945) * fix: preserve K-quantized embedding weights (leejet#1936) * fix: correct SDXL embeddings loading (leejet#1939) * refactor: unify model source and weight lifecycle management (leejet#1956) * docs: reflect GGML_MAX_NAME value change in rpc docs (and in ggml_extend assert) (leejet#1950) * refactor: split generation pipeline out of stable-diffusion.cpp (leejet#1957) * fix: enable VAE decode tiling fallback without auto-fit (leejet#1932) * fix: preserve BF16 embedding weights for get_rows (leejet#1959) * fix: handle invalid option numbers (leejet#1961) * feat: expose the loaded model version name through the public API (leejet#1962) * feat: add SenseNova U1.5 support (leejet#1935) * fix: reuse graph plans when scale parameters change (leejet#1963) * feat: add linear and attention scale overrides (leejet#1964) * feat: preserve explicit backend assignments during auto-fit (leejet#1967) * fix: guard GPU memory capacity and propagate encoding failures (leejet#1958) * fix: bound plain-text runs in parse_prompt_attention regex (leejet#1919) * feat: add Wan2.2 S2V (audio+img-to-video) support (leejet#1925) Co-authored-by: leejet <leejet714@gmail.com> * fix: validate vision projector output dim against LLM hidden size (leejet#1918) * fix: resolve MSVC narrowing conversion warnings (leejet#1969) * feat: Add generation parameters into video metadata (leejet#1901) Co-authored-by: leejet <leejet714@gmail.com> * feat: support external Hugging Face tokenizer JSON files (leejet#1973) * refactor: require external Gemma 2 and GPT-OSS tokenizers (leejet#1974) * feat: support Brownian tree noise in all noise injection samplers (leejet#1899) * fix: use tokenizer-specific pre-tokenization rules (leejet#1975) * fix: remove vision_model. from ununsed tensors (leejet#1983) * perf: eliminate temporary allocations in Philox rounds (leejet#1982) * fix: honor flash attention flag in LLM text encoder attention (leejet#1987) * refactor: remove obsolete unused tensor filtering (leejet#1984) * perf: pad small attention heads to 64 for MMA Flash Attention (leejet#1992) * perf: update ggml for faster direct convolutions (leejet#1993) * perf: accelerate VAE direct 3D convolutions (leejet#1996) * fix: propagate CUDA driver dependency to shared library consumers * fix: prevent clip_preprocess center crop from exceeding the resized image (leejet#1995) * perf: reduce CPU overhead in graph execution and sampling (leejet#1997) * perf: parallelize host tensor elementwise and broadcast ops (leejet#1998) * feat: support building with upstream ggml (leejet#1999) * feat: add Qwen Image 2.1 support (leejet#1994) * feat: restore legacy fp8 handling when building with upstream ggml (leejet#2001) * fix: avoid passing ggml logs as format strings (leejet#2002) * feat: add LLaDA-Image support (leejet#1968) Co-authored-by: leejet <leejet714@gmail.com> * feat: add native CUDA SageAttention support (leejet#2005) * fix: avoid narrowing conversion in SigVQ patch embedding and format code * docs: update CONTRIBUTING.md --------- Co-authored-by: stduhpf <stephduh@live.fr> Co-authored-by: leejet <leejet714@gmail.com> Co-authored-by: LostRuins Concedo <39025047+LostRuins@users.noreply.github.com> Co-authored-by: fszontagh <51741446+fszontagh@users.noreply.github.com> Co-authored-by: Wagner Bruna <wbruna@users.noreply.github.com> Co-authored-by: vmobilis <75476228+vmobilis@users.noreply.github.com> Co-authored-by: Piotr Wilkin (ilintar) <ilintar@gmail.com> Co-authored-by: jk212h20 <101200018+jk212h20@users.noreply.github.com> Co-authored-by: assouan <750048+assouan@users.noreply.github.com> Co-authored-by: nan <zjn32202153@gmail.com> Co-authored-by: Hmission <62598659+Hmission@users.noreply.github.com> Co-authored-by: LED-M <105789115+xledx@users.noreply.github.com> Co-authored-by: Maphist0 <28743569+Maphist0@users.noreply.github.com> Co-authored-by: George <35490284+noctrex@users.noreply.github.com> Co-authored-by: Санька Четвёртый <CAHbKA-IV@mail.ru> Co-authored-by: Lin Xuhao <linxuhao84@gmail.com> Co-authored-by: Fabrice Aneche <akhenakh@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
MiniMax-H3 is video-only. Its denoiser always splits the packed latent into a video half and an audio half, and only `generate_video` computes the audio length, so an H3 checkpoint that reaches `generate_image` is guaranteed to die on `GGML_ASSERT(!audio_input_cache.empty())`.
That happens whenever `--mode vid_gen` is left off, which is easy to do because it is the one flag not implied by passing `--audio-vae`. The failure arrives after the several minutes it takes to load the weights, as a core dump with a raw ggml assert and a stack trace, and nothing in the output points at the missing flag.
Before, on a q4_K H3 denoiser with the Qwen3-VL encoder and both VAEs, no `--mode`:
After, same command:
The AnimateDiff path routes `vid_gen` back through `generate_image`, but that is SD1.5 plus a motion module and never H3, so the guard cannot fire there.