mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-08-29 18:24:29 +00:00
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) <noreply@anthropic.com>
* 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) <noreply@anthropic.com>
* Update conversion/kimi_k3.py
Co-authored-by: Boris Dvorkin <b_dvorkin@niuitmo.ru>
* 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 <noreply@anthropic.com>
* 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 <noreply@anthropic.com>
* 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 <b_dvorkin@niuitmo.ru>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Deepankar Singh <singh.deepankar39@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Caleb DeLeeuw <caleb.deleeuw@gmail.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
|
||
|---|---|---|
| .. | ||
| Apertus-8B-Instruct.jinja | ||
| Apriel-1.6-15b-Thinker-fixed.jinja | ||
| Bielik-11B-v3.0-Instruct.jinja | ||
| ByteDance-Seed-OSS.jinja | ||
| Cohere2MoE.jinja | ||
| CohereForAI-c4ai-command-r-plus-tool_use.jinja | ||
| CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | ||
| deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | ||
| deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | ||
| deepseek-ai-DeepSeek-V3.1.jinja | ||
| deepseek-ai-DeepSeek-V3.2.jinja | ||
| deepseek-ai-DeepSeek-V4-Flash-0731.jinja | ||
| deepseek-ai-DeepSeek-V4.jinja | ||
| fireworks-ai-llama-3-firefunction-v2.jinja | ||
| GigaChat3-10B-A1.8B.jinja | ||
| GigaChat3.1-10B-A1.8B.jinja | ||
| GLM-4.6.jinja | ||
| GLM-4.7-Flash.jinja | ||
| google-gemma-2-2b-it.jinja | ||
| google-gemma-4-31B-it-interleaved.jinja | ||
| google-gemma-4-31B-it.jinja | ||
| HuggingFaceTB-SmolLM3-3B.jinja | ||
| ibm-granite-granite-3.3-2B-Instruct.jinja | ||
| ibm-granite-granite-4.0.jinja | ||
| ibm-granite-granite-4.1.jinja | ||
| Kimi-K2-Instruct.jinja | ||
| Kimi-K2-Thinking.jinja | ||
| Kimi-K3.jinja | ||
| LFM2-8B-A1B.jinja | ||
| LFM2.5-8B-A1B.jinja | ||
| LFM2.5-Instruct.jinja | ||
| llama-cpp-deepseek-r1.jinja | ||
| llama-cpp-rwkv-world.jinja | ||
| meetkai-functionary-medium-v3.1.jinja | ||
| meetkai-functionary-medium-v3.2.jinja | ||
| meta-llama-Llama-3.1-8B-Instruct.jinja | ||
| meta-llama-Llama-3.2-3B-Instruct.jinja | ||
| meta-llama-Llama-3.3-70B-Instruct.jinja | ||
| microsoft-Phi-3.5-mini-instruct.jinja | ||
| MiMo-VL.jinja | ||
| MiniMax-M1.jinja | ||
| MiniMax-M2.jinja | ||
| MiniMax-M3.jinja | ||
| Mistral-Small-3.2-24B-Instruct-2506.jinja | ||
| mistralai-Ministral-3-14B-Reasoning-2512.jinja | ||
| mistralai-Mistral-Nemo-Instruct-2407.jinja | ||
| moonshotai-Kimi-K2.jinja | ||
| muse-glimmer.jinja | ||
| NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | ||
| NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja | ||
| NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.jinja | ||
| NVIDIA-Nemotron-Nano-v2.jinja | ||
| openai-gpt-oss-120b.jinja | ||
| openbmb-MiniCPM5-1B.jinja | ||
| poolside-Laguna-S-2.1.jinja | ||
| poolside-Laguna-XS-2.1.jinja | ||
| poolside-Laguna-XS.2.jinja | ||
| Qwen-Qwen2.5-7B-Instruct.jinja | ||
| Qwen-Qwen3-0.6B.jinja | ||
| Qwen-QwQ-32B.jinja | ||
| Qwen3-Coder.jinja | ||
| Qwen3.5-4B.jinja | ||
| README.md | ||
| Reka-Edge.jinja | ||
| StepFun3.5-Flash.jinja | ||
| tencent-Hy3.jinja | ||
| unsloth-Apriel-1.5.jinja | ||
| unsloth-mistral-Devstral-Small-2507.jinja | ||
| upstage-Solar-Open-100B.jinja | ||
These templates can be updated with the following commands:
./scripts/get_chat_template.py CohereForAI/c4ai-command-r-plus tool_use > models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja
./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 default > models/templates/CohereForAI-c4ai-command-r7b-12-2024-default.jinja
./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 rag > models/templates/CohereForAI-c4ai-command-r7b-12-2024-rag.jinja
./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use > models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja
./scripts/get_chat_template.py deepseek-ai/DeepSeek-R1-Distill-Llama-8B > models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja
./scripts/get_chat_template.py deepseek-ai/DeepSeek-R1-Distill-Qwen-32B > models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja
./scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 > models/templates/fireworks-ai-llama-3-firefunction-v2.jinja
./scripts/get_chat_template.py google/gemma-2-2b-it > models/templates/google-gemma-2-2b-it.jinja
./scripts/get_chat_template.py meetkai/functionary-medium-v3.1 > models/templates/meetkai-functionary-medium-v3.1.jinja
./scripts/get_chat_template.py meetkai/functionary-medium-v3.2 > models/templates/meetkai-functionary-medium-v3.2.jinja
./scripts/get_chat_template.py meta-llama/Llama-3.1-8B-Instruct > models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja
./scripts/get_chat_template.py meta-llama/Llama-3.2-3B-Instruct > models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja
./scripts/get_chat_template.py meta-llama/Llama-3.3-70B-Instruct > models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja
./scripts/get_chat_template.py microsoft/Phi-3.5-mini-instruct > models/templates/microsoft-Phi-3.5-mini-instruct.jinja
./scripts/get_chat_template.py mistralai/Mistral-Nemo-Instruct-2407 > models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja
./scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use > models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja
./scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use > models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja
./scripts/get_chat_template.py Qwen/Qwen2.5-7B-Instruct > models/templates/Qwen-Qwen2.5-7B-Instruct.jinja
./scripts/get_chat_template.py Qwen/QwQ-32B > models/templates/Qwen-QwQ-32B.jinja
./scripts/get_chat_template.py Qwen/Qwen3-0.6B > models/templates/Qwen-Qwen3-0.6B.jinja
./scripts/get_chat_template.py zai-org/GLM-4.5 > models/templates/zai-org-GLM-4.5.jinja
./scripts/get_chat_template.py deepseek-ai/DeepSeek-V3.1 > models/templates/deepseek-ai-DeepSeek-V3.1.jinja