mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-09-04 10:47:38 +02:00
ad1de39e07
* model: add Kimi-K3 text model Hybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, plus five things that architecture does not have: 1. cross-layer residual attention (attn_res_block_size) 2. latent MoE (routed experts run at n_expert_latent) 3. situ activation (replaces SwiGLU everywhere) 4. MLA output gate (sigmoid gate before o_proj) 5. full-rank KDA gate (single ssm_g instead of ssm_g_a/ssm_g_b) K3's text_config reports KimiLinearForCausalLM - the older 48B architecture - so get_model_architecture routes on the top-level name instead. The KDA decay gate has two forms, selected by linear_attn_config's gate_lower_bound. It is not a clamp: when set it swaps the activation entirely (fla/ops/kda/gate.py), from -exp(A_log)*softplus(x) to lower_bound*sigmoid(exp(A_log)*x). K3 sets it to -5.0; kimi-linear leaves it unset, so that path is unchanged. Cross-layer residuals reuse ggml_dsv4_hc_pre for the weighted sum. That op is CPU + CUDA only, so Metal/Vulkan will fall back per-node until those kernels exist. The routed experts ship as compressed-tensors "mxfp4-pack-quantized". That is bit-compatible with ggml's MXFP4 - same E2M1 code assignment, same E8M0 scale byte, only the nibble positions within a block differ - so they are repacked rather than dequantized, losslessly and without a ~5.5 TB bf16 round-trip. The repack is built lazily because gguf_writer holds every added tensor until the final write. DeepSeek-V4 was already doing the identical bit-shuffling, so it now shares the helper. Verified against Moonshot's own code path (transformers + fla's Triton KDA kernels) on a tiny model exercising every K3-specific feature. Final-position logits vs the fp32 reference: 6.7e-05 rel / corr 1.00000000 for both the chunked and the recurrent delta-net path. MXFP4 blocks dequantize to the source weights with 0.0e+00 error. Assisted-By: Claude Opus 5 (1M context) <[email protected]> * model: fix ty errors in the Kimi-K3 converter - `_res_parts` buffers (kind, tensor) pairs, not bare tensors - `get_tensors` must return an Iterator, matching ModelBase - LazyBase's `func` takes one argument, so pass the expert loaders through `args` instead of the closure - borrowing KimiLinearModel.set_vocab from an unrelated TextModel is deliberate and safe, but not expressible in the signature No behaviour change: the MXFP4 repack still dequantizes to the source weights with 0.0e+00 error and end-to-end logits are unchanged (8.386e-03 rel, corr 0.99996630). Assisted-By: Claude Opus 5 (1M context) <[email protected]> * Update conversion/kimi_k3.py Co-authored-by: Boris Dvorkin <[email protected]> * Increase LLAMA_MAX_EXPERTS from 512 to 1024 * tests : support for Kimi K3 in archs test * chat : add Kimi K3 chat format (reasoning, content, typed tool calls) K3's assistant output is an XTML-ish tagged format built by the template's open_tag/close_tag macros. Two properties break generic parsing: 1. The generation prompt ends with open_tag('think'), so the completion starts inside the think section with no opening marker in the output (thinking_forced_open). 2. Only <|open|>/<|close|>/<|sep|>/<|end_of_msg|> are special tokens; tag names ("think", "response", "message") are ordinary text tokens. Adds common_chat_params_init_kimi_k3 (PEG_NATIVE) with detection on the marker trio, reasoning extraction, response unwrapping, and tool-call parsing of the tools/call/argument tag structure with argument types taken from the tool schema. Includes the K3 chat template fixture and 9 test-chat cases derived from real generations of the full 2.8T model. Verified end-to-end against Kimi-K3-Q2_K (GrEarl/Kimi-K3-GGUF) on 8x B200: content, reasoning_content, streaming deltas, and tool_calls all correct; finish_reason stop/tool_calls as appropriate. Co-Authored-By: Claude Fable 5 <[email protected]> * chat : add message_delimiters for Kimi K3 Per-role message-start markers for token-level span splitting. User and assistant messages carry only the role attribute, so their full opener (through <|sep|>) is used; system and tool messages continue with more attributes (type=/tool=/index=), so those delimiters stop after the role's closing quote. Verified against the K3 tiktoken vocabulary that the closing quote is always a standalone token across all attribute variants, so the token-level prefix match stays exact. Co-Authored-By: Claude Fable 5 <[email protected]> * fix: apply nits from @ngxson and text fixes from @danielhanchen * tests : added missing hyperparameters and tensors for Kimi K3 in test-llama-archs * chore : move overly verbose header file comments to Kimi K3 source file * tests : re-enabled KIMI_K3 in test-llama-archs for WebGPU backend * model-saver : emit kda_gate_lower_bound for Kimi K3 Quick fix. The Kimi K3 loader reads kda_gate_lower_bound and gates a graph branch on it (it scales the KDA gate when the bound is above -INFINITY), but the model saver never wrote the key, so a save->load roundtrip silently dropped it back to the -INFINITY default and changed the model's output. The real K3 config sets gate_lower_bound = -5.0. I propose to emit it from the saver, and set it to -5.0 in the test-llama-archs K3 case so the roundtrip check exercises it (the roundtrip fails without the saver line). * Refactor conditional for model architecture check * tests : re-enabled (again) KIMI_K3 and MINIMAX_M3 in test-llama-archs for WebGPU backend * fix code comments * add template on conversion * move repack_mxfp4_blocks to model base * nits * add_value_length * optimize res_stack construction * nits --------- Co-authored-by: Boris Dvorkin <[email protected]> Co-authored-by: Stanisław Szymczyk <[email protected]> Co-authored-by: Deepankar Singh <[email protected]> Co-authored-by: Claude Fable 5 <[email protected]> Co-authored-by: Caleb DeLeeuw <[email protected]> Co-authored-by: Xuan Son Nguyen <[email protected]>
gguf
This is a Python package for writing binary files in the GGUF (GGML Universal File) format.
See convert_hf_to_gguf.py as an example for its usage.
Installation
pip install gguf
Optionally, you can install gguf with the extra 'gui' to enable the visual GGUF editor.
pip install gguf[gui]
API Examples/Simple Tools
examples/writer.py — Generates example.gguf in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model.
examples/reader.py — Extracts and displays key-value pairs and tensor details from a GGUF file in a readable format.
gguf/scripts/gguf_dump.py — Dumps a GGUF file's metadata to the console.
gguf/scripts/gguf_set_metadata.py — Allows changing simple metadata values in a GGUF file by key.
gguf/scripts/gguf_convert_endian.py — Allows converting the endianness of GGUF files.
gguf/scripts/gguf_new_metadata.py — Copies a GGUF file with added/modified/removed metadata values.
gguf/scripts/gguf_editor_gui.py — Allows for viewing, editing, adding, or removing metadata values within a GGUF file as well as viewing its tensors with a Qt interface.
Development
Maintainers who participate in development of this package are advised to install it in editable mode:
cd /path/to/llama.cpp/gguf-py
pip install --editable .
Note: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires setup.py.
In this case, upgrade Pip to the latest:
pip install --upgrade pip
Automatic publishing with CI
There's a GitHub workflow to make a release automatically upon creation of tags in a specified format.
- Bump the version in
pyproject.toml. - Create a tag named
gguf-vx.x.xwherex.x.xis the semantic version number.
git tag -a gguf-v1.0.0 -m "Version 1.0 release"
- Push the tags.
git push origin --tags
Manual publishing
If you want to publish the package manually for any reason, you need to have twine and build installed:
pip install build twine
Then, follow these steps to release a new version:
- Bump the version in
pyproject.toml. - Build the package:
python -m build
- Upload the generated distribution archives:
python -m twine upload dist/*
Run Unit Tests
From root of this repository you can run this command to run all the unit tests
python -m unittest discover ./gguf-py -v
TODO
- Include conversion scripts as command line entry points in this package.