mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-09-30 09:57:38 -05:00
649dcb1036
* Rebase GLM-Next support onto master, and migrate to llama-memory-hybrid-idx * Add initial MTP support * Merge branch optimizations. Reduce allocated compute buffer size, speed up long context decode, fla, and slight MTP improvements. * Review driven changes, remove env vars, protect tensors * Strip MTP for initial PR * Clean up after mtp strip * Clean up after mtp strip * Update speculative.cpp * Update llama-context.h * Clean up after mtp strip * Fix tokenizer ignore merges * Improve quantization protection selection * Refactor mhc helpers, graph base * Lint Fixes * Apply suggestions from code review Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> * Skip glm5-next in model saver, fix CRLF * Skip glm5-next in sweep * Remove T4 fallback * Review cleanup * Review suggestions * Defer separate MTP gguf handling to MTP PR, drop filter * Repad n_head_kv * kpool init apply * Order by descending score * Drop guard * read kpool from hparams, clarify kpool cache flags, remove kpool_build_state(nullptr) * Add glm5-next support to model saver and add arch test fixture * Review cleanup * Kpool pooled caching clarify * Add multi stream support * Finish Rebase * Sparse FA fir DSA prefill * Const * Update llama-model.cpp to fix rebase error * gguf-py : merge tensor map entries for HC tensors * model : use build_gdn_l2_norm in GLM5_NEXT implementation * chore : remove trailing whitespace * model : use new OP precision setting API in GLM5_NEXT implementation * mtmd : use ggml_swiglu_clamp in GLM5V and apply the image token limit The two clamps around swiglu_split are what ggml_swiglu_clamp already does, so the clamp bounds collapse back to one value. GLM5V also never called set_limit_image_tokens(), so --image-max-tokens had no effect. Assisted-by: Claude Opus 5 (cherry picked from commit 46d18e12d422be4cc04a70e4a9a9e0168bb3d5b7) * llama : keep the GLM5-Next k-pool layout across ubatches The layout was rebuilt from a full cell scan on every ubatch. Pools are fixed by the positions relative to the sequence's first one, so the layout now lives on the memory and a ubatch only appends to it. A sequence edit no longer stales every pooled key either, only the ones at or after the edited position, which makes a tail seq_rm free. The pooling subgraph is built unconditionally so the graph shape no longer changes every kpool tokens, and the pool axis is folded into rows before soft_max, which otherwise exceeds the CUDA gridDim.y limit past n_kv 262144. Assisted-by: Claude Opus 5 (cherry picked from commit 5d1c40b93e17fddbf73b785efe43e0d02ccb3977) * model : write the GLM5-Next recurrent rollback checkpoints The conv state and the delta net state were only written to the live row, so a rollback restored whatever the checkpoint rows happened to hold. Take the same route as kimi-k3: build_recurrent_attn for the state, and write all K_rs conv groups. That also drops a state view that assumed contiguous rows. Enroll the arch in test-recurrent-state-rollback, which catches this under its garbage-filled cache pass. Assisted-by: Claude Opus 5 (cherry picked from commit 5ace37e86d5d448e83ef5dde5632c748185b18cd) * llama: fix PR #27773 test-save-load-state restore failure Clear the attention and indexer cache data after a failed hybrid state restore so restored NaNs cannot affect a later sequence. Assisted-by: Codex * llama: fix PR #27773 gpu-rocm graph reallocation Reserve the full GLM5-Next pool capacity and dirty pool count. The gpu-rocm Test step aborts when n_new grows while the graph node count stays fixed; CUDA, Vulkan, Metal, and WebGPU checks report the same error. Assisted-by: Codex * llama : fix GLM5-Next k-pool layout staleness after edits and shared teardown Two defects in the cross-ubatch k-pool layout added by the k-pool commit: 1. Wrong results. An edited sequence only rebuilt its pool layout when its cell count changed, so if the first ubatch after an edit added back exactly as many cells as were removed, the stale position-to-cell list survived. With a unified cache and more than one sequence, where another sequence takes the freed cells, the reused layout points at the wrong cells (CPU: large logit drift, CUDA: NaN). Rebuild whenever the sequence is stale, not only on a size mismatch. 2. Slowdown. "shared" mode was assumed to end only with an edit that forces a rebuild, but sharing also ends when the other sequence is removed. The survivor kept shared = true, pinning cache_safe off and re-pooling every pool on every ubatch (server trigger: n>1 completions with -kvu, via the seq_cp in copy_state_to). In seq_rm, if the layout has shared cells, stale every sequence so one rebuild re-derives sharing and cache_safe returns to 1. Assisted-by: Claude Opus 5 * llama : fix build_attn_mha stream stride for non-contiguous q build_attn_mha split the batch into streams with a stream stride of q->nb[3]/n_stream. That only equals one stream's span, (ne[2]/n_stream)*nb[2], when q is contiguous. GLM5-Next is nope-only, so it does not concat a rope part and passes the permuted q_absorbed straight in, where nb[3] != ne[2]*nb[2]; the stride was then n_head times too large and every stream s >= 1 read another head's queries. Split-KV (-np N without --kv-unified) multi-stream prefill was wrong for every stream past the first. Unified KV and decode were unaffected (n_stream == 1, and decode takes the gather path). Other MLA models concat rope so q is contiguous and the computed value is unchanged for them. Compute the stride from the token dimension, which is identical for a contiguous q. Assisted-by: Claude Opus 5 * llama : re-derive GLM5-Next k-pool sharing on state_read/state_drop The shared-cell teardown added to seq_rm (stale every sequence when the layout has shared cells, so a survivor does not keep shared = true and pin cache_safe off) was missing from the other paths that can free shared cells: state_read and state_drop staled only the one sequence. Apply the same re-derivation there and correct the comment that claimed sharing ends only via an edit or seq_rm. Assisted-by: Claude Opus 5 * quant : drop duplicate GLM5-Next hc_ filter The hc_ name filter was listed twice in the GLM5_NEXT protection block. Assisted-by: Claude Opus 5 * glm5-next: scope K-pool cache access to indexed operations * glm5-next: keep K-pool access in hybrid index memory * glm5-next: keep mHC graph builders model-local * glm5-next: mark only touched pools per ubatch --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> Co-authored-by: Piotr Wilkin <ilintar@gmail.com>
Multimodal Support in llama.cpp
This directory provides multimodal capabilities for llama.cpp. Initially intended as a showcase for running LLaVA models, its scope has expanded significantly over time to include various other vision-capable models. As a result, LLaVA is no longer the only multimodal architecture supported.
Important
Multimodal support can be viewed as a sub-project within
llama.cpp. It is under very heavy development, and breaking changes are expected.
The naming and structure related to multimodal support have evolved, which might cause some confusion. Here's a brief timeline to clarify:
- #3436: Initial support for LLaVA 1.5 was added, introducing
llava.cppandclip.cpp. Thellava-clibinary was created for model interaction. - #4954: Support for MobileVLM was added, becoming the second vision model supported. This built upon the existing
llava.cpp,clip.cpp, andllava-cliinfrastructure. - Expansion & Fragmentation: Many new models were subsequently added (e.g., #7599, #10361, #12344, and others). However,
llava-clilacked support for the increasingly complex chat templates required by these models. This led to the creation of model-specific binaries likeqwen2vl-cli,minicpmv-cli, andgemma3-cli. While functional, this proliferation of command-line tools became confusing for users. - #12849:
libmtmdwas introduced as a replacement forllava.cpp. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs. - #13012:
mtmd-cliwas added, consolidating the various model-specific CLIs into a single tool powered bylibmtmd.
Pre-quantized models
See the list of pre-quantized model here
How it works and what is mmproj?
Multimodal support in llama.cpp works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
This approach keeps the multimodal components distinct from the core libllama library. Separating these allows for faster, independent development cycles. While many modern vision models are based on Vision Transformers (ViTs), their specific pre-processing and projection steps can vary significantly. Integrating this diverse complexity directly into libllama is currently challenging.
Consequently, running a multimodal model typically requires two GGUF files:
- The standard language model file.
- A corresponding multimodal projector (
mmproj) file, which handles the image encoding and projection.
What is libmtmd?
As outlined in the history, libmtmd is the modern library designed to replace the original llava.cpp implementation for handling multimodal inputs.
Built upon clip.cpp (similar to llava.cpp), libmtmd offers several advantages:
- Unified Interface: Aims to consolidate interaction for various multimodal models.
- Improved UX/DX: Features a more intuitive API, inspired by the
Processorclass in the Hugging Facetransformerslibrary. - Flexibility: Designed to support multiple input types (text, audio, images) while respecting the wide variety of chat templates used by different models.
How to obtain mmproj
Multimodal projector (mmproj) files are specific to each model architecture.
For the following models, you can use convert_hf_to_gguf.py with --mmproj flag to get the mmproj file:
- Gemma 3 ; See the guide here - Note: 1B variant does not have vision support
- SmolVLM (from HuggingFaceTB)
- SmolVLM2 (from HuggingFaceTB)
- Pixtral 12B - only works with
transformers-compatible checkpoint - Qwen 2 VL and Qwen 2.5 VL (from Qwen)
- Mistral Small 3.1 24B
- InternVL 2.5 and InternVL 3 from OpenGVLab (note: we don't support conversion of
InternVL3-*-hfmodel, only non-HF version is supported ;InternLM2Modeltext model is not supported) - MiniCPM-V 4.6 ; See the guide here - requires the standard
transformersv5.7.0+ checkpoint
For older models, please refer to the relevant guide for instructions on how to obtain or create them:
NOTE: conversion scripts are located under tools/mtmd/legacy-models