mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_SPLIT

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+*.gguf.zbst filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,158 @@
 ---
 license: mit
 ---
+# mtp-Qwen3.5-27B
+
+## 🤔 What is this [HuggingFace repository](https://huggingface.co/Thireus/mtp-Qwen3.5-27B-THIREUS-BF16-SPECIAL_SPLIT/) about?
+
+This repository provides **GGUF-quantized tensors** for the [mtp](https://github.com/ggml-org/llama.cpp/blob/master/docs/speculative.md) layer(s) of the Qwen3.5-27B model (official repo: https://huggingface.co/Qwen/Qwen3.5-27B). These GGUF shards are designed to be used with **Thireus’ GGUF Tool Suite** (https://github.com/Thireus/GGUF-Tool-Suite), a collection of tools that automatically finds the perplexity-optimal mix of quantizations for any given a model size target. With this GGUF Tool Suite, you can produce your own Dynamic 3.0 Quants recipes and achieve optimum accuracy & SOTA quantization performance. Give it a try here: https://gguf.thireus.com/quant_assign.html  
+
+- 📖 Documentation: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/docs
+- 🔍 Example of GGUF recipes: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples  
+- 🍳 Cook your own recipe files: https://gguf.thireus.com/quant_assign.html  
+- ☁️ Download GGUF models from recipe files: https://gguf.thireus.com/quant_downloader.html  
+- 📂 Browse available models: https://huggingface.co/Thireus/collections and https://gguf.thireus.com  
+
+*tl;dr: Expand the details section below*
+<details>
+
+```
+cd ~
+
+# Make sure to install all ik_llama.cpp compilation dependencies...
+apt install python3-dev python3-pip python3-venv python3-wheel python3-setuptools git acl netcat-openbsd cmake # pipx
+
+# Obtain ik_llama's Thireus version - Windows/macOS/Linux builds available at https://github.com/Thireus/ik_llama.cpp/releases
+git clone https://github.com/Thireus/ik_llama.cpp
+cd ik_llama.cpp
+git pull
+# Build ik_llama.cpp
+cmake -B build -DGGML_AVX=ON -DGGML_AVX2=ON -DLLAMA_CURL=OFF -DGGML_MAX_CONTEXTS=2048
+cmake --build build --config Release -j16
+cd ..
+
+# Obtain Thireus' GGUF-Tool-Suite
+GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/Thireus/GGUF-Tool-Suite
+
+# Download model quant mix from recipe file - you can also try the web version: https://gguf.thireus.com/quant_downloader.html
+cd GGUF-Tool-Suite
+rm -f download.conf # Make sure to copy the relevant download.conf for the model before running quant_assign.py
+cp -f models/Qwen3.5-27B/download.conf . # Use the download.conf of the chosen model
+mkdir -p kitchen && cd kitchen
+# Obtain a recipe example for the chosen model from ../recipe_examples/
+../quant_downloader.sh ../recipe_examples/ik_llama.cpp_recipes/Qwen3.5-27B.ROOT-3.5993bpw-11.3565ppl.1GB-GGUF_0GB-GPU_0GB-CPU.9888e4b_831ff04.recipe
+
+# Other recipe examples can be found at https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples
+
+# Launch ik_llama's llama-cli:
+ulimit -n 9999 # Lifts "too many open files" limitation on Linux
+~/ik_llama.cpp/build/bin/llama-server \
+  -m Qwen3.5-27B-THIREUS-BF16-SPECIAL_TENSOR-00001-of-*.gguf \
+  -md mtp-Qwen3.5-27B-THIREUS-BF16-SPECIAL_TENSOR-00001-of-*.gguf --spec-type draft-mtp \
+  -fa auto -amb 1024 -ctk q8_0 -c 32768 -ngl 99 \
+  -b 4096 -ub 4096 --warmup-batch --no-mmap --threads 1 \
+  --main-gpu 0
+```
+
+</details>
+
+---
+
+## ❓ Why does this Tool Suite exist?
+
+1. **Compatibility & Speed** – [unsloth](https://huggingface.co/unsloth)’s dynamic quants may not always work optimally with `ik_llama.cpp`.  
+2. **Custom Rig Fit** – No off-the-shelf GGUF model perfectly matched my VRAM/RAM setup, so I built a way to tailor models and leverage extra VRAM/RAM to reduce perplexity.  
+3. **Automated PPL-Optimal Quantization** – To my knowledge, there was no open source flexible, automated method to minimize perplexity for any bits-per-weight (bpw) target—so I created one with excellent results!  
+
+---
+
+## 📊 How does it compare to other GGUFs?
+
+Here’s how Qwen3.5-27B quantized with **Thireus’ GGUF Tool Suite** stacks up against other quantizers (lower perplexity = better at equal or lower bpw):
+
+![PPLs Compared With Others](https://github.com/Thireus/GGUF-Tool-Suite/raw/main/ppl_graphs/Qwen3.5-27B.svg)
+
+> _Note: The `recipe_examples` files illustrate good recipes. The Tool Suite computes the optimal ppl/bpw curve for you — just specify your target RAM, VRAM, and quant types, and `quant_assign.py` finds the best mix._  
+
+More perplexity/bpw graphs for other supported models: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/ppl_graphs  
+
+*Qwen3.5 Thireus' PPL benchmarks are computed with the parameters `-ctk f16 -c 512 -b 512 -ub 512`. Changing any of these parameters will alter the PPL. In particular, reducing `-b 512 -ub 512` increases the PPL, while increasing them decreases the PPL.*
+
+---
+
+## 🚀 How do I get started?
+
+Check out the [GGUF Tool Suite README](https://github.com/Thireus/GGUF-Tool-Suite) — focus on these sections:
+
+1. ⚠️ **Requirements** – Which `ik_llama.cpp` (or `llama.cpp`) version to use and how to compile.  
+   - Windows binaries (no patching needed) at: https://github.com/Thireus/ik_llama.cpp/releases  
+2. 📥 **Download Model Shards** – Use `quant_downloader.sh` or [quant_downloader.html](https://gguf.thireus.com/quant_downloader.html) to fetch GGUF shards from any recipe.  
+   - Recipe examples: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples  
+3. 🧠 **Run a Downloaded Model** – Sample usage with `llama-cli`.  
+4. 🛠️ **Generate a Custom Recipe** – Produce recipes tailored to your VRAM/RAM target usage for optimum perplexity.  
+
+---
+
+## ✅ Supported Models
+
+Supported models are listed under `models/` in the [Tool Suite Github repo](https://github.com/Thireus/GGUF-Tool-Suite/tree/main/models). Presence of `ppl_results.csv` indicates official support and compatibility with `quant_assign.py`.
+
+---
+
+## 🤷‍♂️ Will I release baked dynamic quant GGUFs?
+
+No, because I believe in **tailored quantization** for each user’s hardware. If you prefer ready-made shards, you are welcome to merge them via `llama-gguf-split --merge`, or request someone to publish them, or rely on generic GGUF dynamic quants such as [unsloth](https://huggingface.co/unsloth)'s.
+
+Instead, I prefer to share examples of recipes so users can see exactly how they were produced (command included inside these recipe files) and tweak them for their own rigs. The `quant_downloader.sh` script or [quant_downloader.html](https://gguf.thireus.com/quant_downloader.html) (web port of this script) handles automatic fetching and verification of each shard. Note that recipes provided by [Ubergarm](https://huggingface.co/ubergarm) on his model cards are also compatible with `quant_downloader.sh` and [quant_downloader.html](https://gguf.thireus.com/quant_downloader.html), providing a "SPECIAL_SPLIT" version of these models exists (see https://gguf.thireus.com/).
+
+Users who don’t trust the GGUF shards on HuggingFace can also quantize their own by passing recipe lines to `llama-quantize --custom-q` ([see example](https://github.com/Thireus/GGUF-Tool-Suite/blob/main/models/DeepSeek-R1-0528/DeepSeek-R1-0528-THIREUS-ANY-SPECIAL.sh#L482-L486)). Run `llama-quantize --help` to list compatible quants for `quant_assign.py`. This approach is especially useful if you prefer `llama.cpp` over `ik_llama.cpp`.  
+
+---
+
+## 📦 What’s in this repository?
+
+- **00001 GGUF header shard** – Contains metadata (tokens, chat template, tensor count, etc.). This metadata can be explored directly from the HuggingFace web interface after clicking on that shard.  
+- **Tensor shards** – Each shard holds one tensor; see `tensors.map` for names, quant types, sizes, SHA-256 hash, shard IDs, etc.  
+- **GPG-signed files** – `tensors.map` and header shard are signed with the key in [trusted-keys.asc](https://github.com/Thireus/GGUF-Tool-Suite/blob/main/trusted-keys.asc) for tamper detection.  
+- **Security note** – Some papers about various ways to attack GGUFs and LLMs are available online, such as https://arxiv.org/abs/2505.23786, and there are also more classic security exploits like CVE-2024-23496 and CVE-2024-25664 through CVE-2024-25668. Only use GGUFs from reputable, trusted authors—or alternatively self-quantize—to avoid potential exploits. 
+
+---
+
+## 💡 Pro Tips
+
+You can easily download the BF16 model version to quantize your own shards:
+
+```
+mkdir kitchen  
+echo '.*=bf16' > kitchen/bf16.recipe  
+cd kitchen
+../quant_downloader.sh bf16.recipe --qtype BF16 
+```
+
+You can also quantize individual BF16 tensors without the need to download every BF16 .gguf shard:
+
+BF16 model shards can also be individually quantized using a special version of ik_llama.cpp's `llama-quantize` utility which comes with the `--individual-tensors` option.
+
+- Source code: https://github.com/Thireus/ik_llama.cpp/tree/th/quantize_individual_tensors
+- Builds (macOS, Windows and Linux): https://github.com/Thireus/ik_llama.cpp/releases/tag/th-quantize_individual_tensors-b4210-7a44805
+
+Usage example:
+```
+./llama-quantize --keep-split --imatrix imatrix_ubergarm.dat --individual-tensors 2,3,1094 Kimi-K2-Thinking-THIREUS-BF16-SPECIAL_TENSOR-00001-of-01097.gguf my_new_shards.gguf iq3_s 12
+```
+
+For more information about how to use it: https://github.com/Thireus/GGUF-Tool-Suite/issues/45
+
+You can produce your own quantized shards from Thireus' special BF16 model using `quantize_model.sh` found on https://github.com/Thireus/GGUF-Tool-Suite, for example:
+
+```
+./quantize_model.sh --model "Qwen3.5-122B-A10B" --qtype iq2_xxs
+```
+
+You can disable reasoning (thinking) when using jinja templates for supported models:
+
+```
+llama-server ... --jinja --chat-template-kwargs '{"enable_thinking": false}'
+```
+
+Enjoy optimized quantization! 🎉

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00001-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a706512573ea935261e3db60d3deab4dd3bb4ee78d3ff9af5b086901bec48512
+size 10942752

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00002-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b65dd1e0eababa38e99a7f3029c88b32cf734c54eb9b0c1583df4dd862a1af7
+size 874086560

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00003-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bd9e159be4578785418e4ab95cf689ff0cfff573b51e36e3ba22f4e4426ac8e
+size 874086592

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00004-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2075f21a5577c178ca6ced7abd262caeddbad43e9697836d2f3c709beed232e7
+size 61276352

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00005-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b44e2c2c8203bb28ddbab5a7464043c8a2f38069a00ed89cedbf6e511c29ca45
+size 61276352

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00006-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:875e7f3f1f8e63ba52e05e9c5405e65e51399f880f169b8f2cd5241a782337a6
+size 61276352

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00007-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4139d0c5111da1f536ec421910de8594f12b39035d93521b1b8dc49045bbae61
+size 36044992

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00008-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f01af629494700fce301b1b7b954e9734d308b0ec073aff483805f9d2b42cfe1
+size 43253952

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00009-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a55582de06eadc3d724e5b771d2d8c9ff1135339554b4d2d5717ddb6abfc8857
+size 21627072

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00010-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4590cfcb507003c61300b1abfd64d9e98e2fa3091d115f4fc5876a11c6851023
+size 20640

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00011-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57424507743ab0f420971c731168c70d6a78436f4d9162abd15bb0ea8f5f6617
+size 20672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00012-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d016a77a9d9258dcd2a2be3b3d2bc9625bb767e59d79914e5306033e3fbe40a8
+size 20672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00013-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce65b7f5b2027ce61069856bd806a44a66669dbf9e32f9d2b8786bba1d5ddb4e
+size 1216

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00014-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37cdff4ecaa9d3f2b6ae6136140688476ffab894f95f978b082b478ed6bc55f3
+size 3604672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00015-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bb5caaa495668e3cd629734f66d0b42870a234ba6728467fc0a21c63a04741d
+size 1216

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00016-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:609c877597d8b6a5379d64ac880704b88a7b2eac831828c92a0fc4ff830726c2
+size 3604672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00017-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2b8b5abd6e5b248120fae9f23ef73bb6c9959ff6914bcf320ea78f4c9b71ebc
+size 20672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00018-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8fd0bde64470d9653e9846cf0515aef9771d83261961a6e08fdb4bc5b08c385
+size 20672

mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00019-of-00019.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7edce47fd6bcefa830c40964540b2cde273e99dacdfcdb494da75beb01e98a6c
+size 20672

tensors.map

@@ -0,0 +1,18 @@
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00002-of-00019.gguf:4b65dd1e0eababa38e99a7f3029c88b32cf734c54eb9b0c1583df4dd862a1af7:output.weight:shape=(5120, 248320):dtype=q5_k:elements=1271398400:bytes=874086400
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00003-of-00019.gguf:2bd9e159be4578785418e4ab95cf689ff0cfff573b51e36e3ba22f4e4426ac8e:token_embd.weight:shape=(5120, 248320):dtype=q5_k:elements=1271398400:bytes=874086400
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00004-of-00019.gguf:2075f21a5577c178ca6ced7abd262caeddbad43e9697836d2f3c709beed232e7:blk.64.ffn_down.weight:shape=(17408, 5120):dtype=q5_k:elements=89128960:bytes=61276160
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00005-of-00019.gguf:b44e2c2c8203bb28ddbab5a7464043c8a2f38069a00ed89cedbf6e511c29ca45:blk.64.ffn_gate.weight:shape=(5120, 17408):dtype=q5_k:elements=89128960:bytes=61276160
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00006-of-00019.gguf:875e7f3f1f8e63ba52e05e9c5405e65e51399f880f169b8f2cd5241a782337a6:blk.64.ffn_up.weight:shape=(5120, 17408):dtype=q5_k:elements=89128960:bytes=61276160
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00007-of-00019.gguf:4139d0c5111da1f536ec421910de8594f12b39035d93521b1b8dc49045bbae61:blk.64.nextn.eh_proj.weight:shape=(10240, 5120):dtype=q5_k:elements=52428800:bytes=36044800
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00008-of-00019.gguf:f01af629494700fce301b1b7b954e9734d308b0ec073aff483805f9d2b42cfe1:blk.64.attn_q.weight:shape=(5120, 12288):dtype=q5_k:elements=62914560:bytes=43253760
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00009-of-00019.gguf:a55582de06eadc3d724e5b771d2d8c9ff1135339554b4d2d5717ddb6abfc8857:blk.64.attn_output.weight:shape=(6144, 5120):dtype=q5_k:elements=31457280:bytes=21626880
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00010-of-00019.gguf:4590cfcb507003c61300b1abfd64d9e98e2fa3091d115f4fc5876a11c6851023:output_norm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00011-of-00019.gguf:57424507743ab0f420971c731168c70d6a78436f4d9162abd15bb0ea8f5f6617:blk.64.attn_norm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00012-of-00019.gguf:d016a77a9d9258dcd2a2be3b3d2bc9625bb767e59d79914e5306033e3fbe40a8:blk.64.post_attention_norm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00013-of-00019.gguf:ce65b7f5b2027ce61069856bd806a44a66669dbf9e32f9d2b8786bba1d5ddb4e:blk.64.attn_k_norm.weight:shape=(256,):dtype=f32:elements=256:bytes=1024
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00014-of-00019.gguf:37cdff4ecaa9d3f2b6ae6136140688476ffab894f95f978b082b478ed6bc55f3:blk.64.attn_k.weight:shape=(5120, 1024):dtype=q5_k:elements=5242880:bytes=3604480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00015-of-00019.gguf:3bb5caaa495668e3cd629734f66d0b42870a234ba6728467fc0a21c63a04741d:blk.64.attn_q_norm.weight:shape=(256,):dtype=f32:elements=256:bytes=1024
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00016-of-00019.gguf:609c877597d8b6a5379d64ac880704b88a7b2eac831828c92a0fc4ff830726c2:blk.64.attn_v.weight:shape=(5120, 1024):dtype=q5_k:elements=5242880:bytes=3604480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00017-of-00019.gguf:f2b8b5abd6e5b248120fae9f23ef73bb6c9959ff6914bcf320ea78f4c9b71ebc:blk.64.nextn.shared_head_norm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00018-of-00019.gguf:b8fd0bde64470d9653e9846cf0515aef9771d83261961a6e08fdb4bc5b08c385:blk.64.nextn.enorm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480
+mtp-Qwen3.5-27B-THIREUS-Q5_K-SPECIAL_TENSOR-00019-of-00019.gguf:7edce47fd6bcefa830c40964540b2cde273e99dacdfcdb494da75beb01e98a6c:blk.64.nextn.hnorm.weight:shape=(5120,):dtype=f32:elements=5120:bytes=20480