Duplicate from ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF

Co-authored-by: Michael Helcig <helcig@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,67 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz 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
+DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00002-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00003-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+assets/banner.png filter=lfs diff=lfs merge=lfs -text
+assets/plots/DeepSeek-V4.1-Flash-1017-benchmarks_bars.png filter=lfs diff=lfs merge=lfs -text
+assets/banner.webp filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00002-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00003-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00004-of-00004.gguf filter=lfs diff=lfs merge=lfs -text
+mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00002-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00003-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00004-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00005-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00002-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00003-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00005-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00006.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00005-of-00006.gguf filter=lfs diff=lfs merge=lfs -text

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e8ddd11441edd4bc19bacaef5dbe38818af2cbdf0820fb70f1d2d851fcc192a
+size 5780704

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00002-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be4ceb1caf1bed28e886860cf6a52adc8f5ca9eb0d2098ade39bb4179d48b323
+size 39340012896

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00003-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:758c66ba65236e8ce0820c21a7b6f9b2947681df8cc509f672a77722e8191b9a
+size 38698805888

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00004-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04a5bea2b38eb576f49efc19c0a677768d6e2a8ea5772d72fb6cb79e1733a212
+size 39707055008

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00005-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c8dbee99890f05dbc420eba6f44393f5228d06248fe61e1ae8ed002449c7549
+size 2969539840

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0182a2a1f42d319a0dbc6e34aa82a063880c069fe3aafbbea86018f386e64e5f
+size 5780704

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c88039fa640460a9fc3465a6dfcd162a4083dd738d55e5b52e025be14555a08
+size 39581533600

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fddebfd05f0a36cfd107ba5dfd2547e64df61adbab39560a5a5f5521ca099a5
+size 39417972640

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c577a43c6e6e615838a1655a0b162541edeac6f36542f08c4a9593a70dcaf500
+size 39079187648

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00005-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:851d8605bf504962054d17ea8a0d9e2e0a228d085d396bd865a4de9f16d90fbd
+size 37141423776

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f356f198561f56df36f7ab92d7238a926646d76de992c34f15bcb939a5f529b9
+size 5780704

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00002-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25af3785ccf925f54137e1ead4c481a8952315dc0b1af6519a99407f67bce8f3
+size 39535781184

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00003-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8168d0998fc494def2e1952d83c131425a108660aae9d29190f5a8670ab62b48
+size 39736712000

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eb9b4de1968f150dfc207eb31c27be9cd9b5666a9a6246bde8a22b184b4bee0
+size 39036908384

DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00005-of-00006.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e35f6b4e6fc46f6c4b152b593d84d683c5ece695ae6ad734f556198c041bec1
+size 19923784832

DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dac4c54893ad7aa89103c0028500eadc78e74112cef3a2930a97e67dc9fb4f3
+size 110595290816

README.md

@@ -0,0 +1,237 @@
+---
+base_model: deepseek-ai/DeepSeek-V4.1-Flash-1017
+base_model_relation: quantized
+pipeline_tag: text-generation
+library_name: gguf
+license: mit
+tags:
+  - gguf
+  - gsq
+  - rco
+  - quantization
+  - mixed-precision
+  - ist-daslab
+  - moe
+  - deepseek
+---
+
+<div align="center">
+
+<a href="https://github.com/IST-DASLab"><img src="assets/banner.webp" alt="DeepSeek-V4.1-Flash-1017, GSQ-RCO dynamic non-uniform GGUF" width="100%"/></a>
+
+<br/>
+
+# DeepSeek-V4.1-Flash-1017 &middot; GSQ-RCO GGUFs
+
+**Non-uniform GGUF quantizations** of a 763B-parameter, 384-expert MoE, produced with **GSQ** and **RCO**.
+
+[![arXiv: GSQ](https://img.shields.io/badge/arXiv-GSQ_2604.18556-b31b1b?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2604.18556)
+[![arXiv: RCO](https://img.shields.io/badge/arXiv-RCO_2605.00649-b31b1b?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2605.00649)
+[![GSQ code](https://img.shields.io/badge/code-GSQ-181717?logo=github&logoColor=white)](https://github.com/IST-DASLab/GSQ)
+[![RCO code](https://img.shields.io/badge/code-RCO-181717?logo=github&logoColor=white)](https://github.com/IST-DASLab/RCO)
+[![DASLab](https://img.shields.io/badge/DASLab-GitHub-101048?logo=github&logoColor=white)](https://github.com/IST-DASLab)
+[![license](https://img.shields.io/badge/license-MIT-19a34a)](#license)
+
+</div>
+
+![Base model vs GSQ-RCO quants on AIME25, GPQA-Diamond and LiveCodeBench v6](assets/plots/DeepSeek-V4.1-Flash-1017-benchmarks_bars.png)
+
+> **Update, October 27, 2026.** The files now use llama.cpp's own DeepSeek-V4.1 layout and load in llama.cpp [b11870](https://github.com/ggml-org/llama.cpp/releases/tag/b11870) or newer, including images.
+> - **Downloaded the earlier files?** You do not need to download the model again. Run [convert_ds41_upstream.py](./convert_ds41_upstream.py) on your model folder: it rewrites the files locally into exactly the current ones and fetches only 80 KB. See [Updating an existing download](#updating-an-existing-download).
+> - **New download?** The shared Engram file is shard 6 of every build, so link it once per build you use (`<ID>` is `IQ1_M`, `IQ2_XXS` or `IQ2_XS`; see [llama.cpp](#llamacpp)):
+>
+>   ```sh
+>   ln -s DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf DeepSeek-V4.1-Flash-1017-GSQ-RCO-<ID>-00006-of-00006.gguf
+>   ```
+
+---
+
+## Overview
+
+This repository provides GGUF quantizations of **DeepSeek-V4.1-Flash-1017** at three sizes: **IQ1_M (1.75 bits per weight), IQ2_XXS (2.00) and IQ2_XS (2.25)**. In contrast to uniform quantization, which applies a single quantization type to all weight tensors, each model here assigns a separate quantization type to every tensor. The assignment is obtained by a gradient-based search that allocates precision according to per-tensor sensitivity, subject to a total size budget. The resulting files are standard GGUF.
+
+DeepSeek-V4.1-Flash-1017 is a sparse mixture-of-experts model with **763.2B parameters**: **384 routed experts per layer**, of which 6 are active per token, and a 1M-token context window. The routed experts dominate the parameter count, so they also dominate the size budget, and the search spends most of its freedom there, assigning a quantization type **per layer** (not per expert, which GGUF cannot express).
+
+With [Strata](https://github.com/Niko1221/Strata), every build runs on a **single GPU with 12 GB of VRAM**. See [Usage](#usage).
+
+> **Method summary.** GSQ provides accurate low-bit scalar quantization of each tensor at a given quantization type; RCO assigns the per-tensor quantization types under a size budget. Together they yield a non-uniform GGUF at the requested size.
+
+| Method | Description |
+|---|---|
+| **GSQ** (Gumbel-Softmax Quantization, [paper](https://arxiv.org/abs/2604.18556), [code](https://github.com/IST-DASLab/GSQ)) | Post-training scalar quantization that jointly learns the per-coordinate grid assignments and the per-group scales via a Gumbel-Softmax relaxation. GSQ closes most of the gap between scalar and vector quantization at 2 to 3 bits while remaining deployable in standard scalar formats such as GGUF. |
+| **RCO** (Riemannian Constrained Optimization, [paper](https://arxiv.org/abs/2605.00649), [code](https://github.com/IST-DASLab/RCO)) | Assigns one of K quantization types to each of N tensors under a total size budget. The budget constraint is reformulated as a smooth Riemannian manifold in logit space, which permits gradient-based optimization directly on the task loss while enforcing the budget exactly, without constraint-specific hyperparameter tuning. |
+
+Both methods were developed at the [Deep Algorithms and Systems Lab (DASLab)](https://github.com/IST-DASLab), Institute of Science and Technology Austria.
+
+---
+
+## Available files
+
+Each build is a backbone GGUF in five shards, named after its quantization identifier: `DeepSeek-V4.1-Flash-1017-GSQ-RCO-<ID>-0000i-of-00006.gguf`, i = 1 to 5. Shard 1 holds the metadata only, shards 2 to 5 the transformer weights, **551.9B parameters**. The bpw is the average bit-width over these, which is what the search controls.
+
+The **Engram file**, `DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf`, holds the model's two Engram lookup tables (196.6B parameters) at IQ4_NL, 4.50 bpw. It is identical for all builds, so it is stored once and needs to be downloaded only once, also when trying several builds. llama.cpp loads it as shard 6 of a build, so each build needs a link to it named `...-<ID>-00006-of-00006.gguf` (one command, see [llama.cpp](#llamacpp)). The tables are a lookup, not a matmul weight, so they are held at a fixed type in every build and excluded from the search. They are quantized in the evaluated models, not just the shipped ones.
+
+| Build | bpw | Backbone | Engram | Total | Notes |
+|---|---|---|---|---|---|
+| `IQ1_M` | 1.75 | 120.72 GB | 110.60 GB | 231.32 GB | Smallest |
+| `IQ2_XXS` | 2.00 | 138.24 GB | 110.60 GB | 248.84 GB | Recommended; closest to the base model |
+| `IQ2_XS` | 2.25 | 155.23 GB | 110.60 GB | 265.83 GB | |
+
+All three builds run on a single 12 GB GPU with [Strata](https://github.com/Niko1221/Strata).
+
+The Engram column is the same shared file in every row. The vision projector, `mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf` (0.97 GB), also serves all builds. GB is 10<sup>9</sup> bytes.
+
+The IQ2_XXS build is the stronger operating point: it matches the base model on all three benchmarks and so reaches **100.0%** of its task average (93.54). The IQ1_M build trades 1.2 points of task average for 17.5 GB less.
+
+### Memory requirements
+
+<!-- TBD: confirm how Strata places the model on a 12 GB GPU (which tensors stay in VRAM, how much system RAM it needs) and fill in this section. -->
+
+Only the backbone holds transformer weights. The Engram tables are read sparsely, one row per token, so they can stay memory-mapped on disk and do not need to occupy VRAM or RAM. With Strata, the backbone does not need to fit in VRAM either: every build runs on a **12 GB GPU**, with the rest of the model held in system memory. Keep the Engram file on an SSD so the paging stays cheap.
+
+---
+
+## Results
+
+All models are evaluated against the base model **deepseek-ai/DeepSeek-V4.1-Flash-1017** as released (FP8 dense weights, FP4 routed experts). We report three reasoning and generation benchmarks: **AIME25**, **GPQA-Diamond**, and **LiveCodeBench v6**. Task average is the mean of the three; recovery is the task average relative to the base model.
+
+All numbers in this card were measured at **reasoning effort 50**.
+
+| Variant | bpw | AIME25↑ | GPQA-D↑ | LCB v6↑ | Task avg↑ | Recovery |
+|---|---|---|---|---|---|---|
+| Base (FP8 / FP4) | n/a | 100.00 | 90.91 | 89.71 | 93.54 | 100.0% |
+| **GSQ-RCO IQ1_M** | 1.75 | **100.00** | 88.38 | 88.57 | 92.32 | 98.7% |
+| **GSQ-RCO IQ2_XXS** | 2.00 | **100.00** | **90.91** | **89.71** | **93.54** | **100.0%** |
+| **GSQ-RCO IQ2_XS** | 2.25 | **100.00** | 90.40 | **89.71** | 93.37 | 99.8% |
+
+IQ2_XXS (2.00 bpw) keeps **100.0%** of the base model's task average: it matches the base model on AIME25 (100.00), GPQA-Diamond (90.91) and LiveCodeBench v6 (89.71). IQ2_XS (2.25 bpw) matches it on AIME25 and LiveCodeBench v6 and scores 90.40 on GPQA-Diamond, 99.8% overall. The smallest build, IQ1_M (1.75 bpw), also scores 100.00 on AIME25 and retains 98.7% overall.
+
+---
+
+## Usage
+
+> **Runtime support.** llama.cpp supports DeepSeek-V4.1, including images, since release [b11870](https://github.com/ggml-org/llama.cpp/releases/tag/b11870) (October 25, 2026), and the files here use its layout since October 27, 2026. Downloads from before that date can be converted locally, see [Updating an existing download](#updating-an-existing-download). The GGUFs are also supported in [Strata](https://github.com/Niko1221/Strata).
+
+<!-- TBD: confirm that Strata reads the six-shard layout of October 27, 2026. -->
+
+### Strata Engine
+
+[Strata](https://github.com/Niko1221/Strata) runs these GGUFs on a **single NVIDIA GPU with 12 GB of VRAM**, with a one-click installer and an OpenAI-compatible endpoint on localhost. The builds were optimized together with Strata for local inference.
+
+<!-- TBD: Strata command or model-selection steps for this repo. -->
+
+### llama.cpp
+
+Requires llama.cpp [b11870](https://github.com/ggml-org/llama.cpp/releases/tag/b11870) or newer.
+
+```bash
+# download one build and the shared Engram file (requires: pip install -U "huggingface_hub[cli]")
+hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF \
+  --include "*IQ2_XXS*" "*Engram*" --local-dir .
+
+# the Engram file is shard 6 of every build
+ln -s DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00006-of-00006.gguf
+
+llama-cli -m DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf \
+  -p "Explain mixed-precision quantization." -ngl 99
+```
+`-m` takes the first shard; llama.cpp opens shards 2 to 6 from the same directory. To try another build, download only its shards (`--include "*IQ1_M*"`) and add its link. On Windows, `mklink /H <link> DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf` makes the link; [convert_ds41_upstream.py](./convert_ds41_upstream.py), run on the folder, also creates missing links.
+
+The Engram table stays memory-mapped on disk and is read one row per token (llama.cpp's default `--lazy-mode auto` does this for tensors over 4 GiB), so it needs neither VRAM nor RAM. See [Memory requirements](#memory-requirements).
+
+### Images
+
+Download the vision projector and pass it with `--mmproj`:
+
+```bash
+hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF \
+  --include "mmproj-*" --local-dir .
+
+llama-server -m DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf \
+  --mmproj mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf -ngl 99
+```
+With `llama-mtmd-cli`, also pass `--jinja` so that the chat template embedded in the GGUF is used.
+
+### Chat template and reasoning effort
+
+The base repository ships no Jinja chat template; its prompt format is defined by DeepSeek's encoder, `encoding/encoding.py`. The GGUFs embed a template adapted from [Unsloth's DeepSeek-V4.1 release](https://huggingface.co/unsloth/DeepSeek-V4.1-Flash-1017-GGUF), checked against that encoder for text chat, tool use, and image formatting. Reasoning effort accepts an integer from 1 to 100 (also a numeric string), or `low`, `high`, `max` (50, 75, 100). The default effort is DeepSeek's `high` (75), and effort applies only in thinking mode. Structured image blocks, including images in tool results, retain their image placeholders. Image inference needs the vision projector, see [Images](#images).
+
+```bash
+llama-server -m DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf -ngl 99 \
+  --reasoning on --reasoning-effort 50
+```
+Effort 50 is the setting all results in this card were measured with. Per request, set `"reasoning_effort": "50"` in the chat completion body (a string, also for numbers). `--reasoning off` disables thinking.
+
+
+### Updating an existing download
+
+Files downloaded before October 27, 2026 (four shards per build and a separate `DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf`) were written for an earlier llama.cpp branch and do not load in llama.cpp. [convert_ds41_upstream.py](./convert_ds41_upstream.py) turns them into the current files of this repository, byte for byte, without downloading the model again. It needs Python 3.8+ and no extra packages, and fetches only two small files of this repository (`chat_template.jinja` and `exp_probs_b_vl.f32`, 80 KB).
+
+```sh
+cd /path/to/downloaded/model-folder
+hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF convert_ds41_upstream.py --local-dir .
+python3 convert_ds41_upstream.py .
+```
+
+It converts every build in the folder and the mmproj, and writes the new backbone shards next to the old ones; every written tensor is read back and compared with its source. The Engram file is converted in place: only its header changes, its 110 GB of tables stay as they are, and the script adds the shard-6 links. The free space needed is the size of the new shards (121 to 155 GB per build). With `--delete-old`, the old shards of a build are deleted as soon as its new ones are verified, so space for one build suffices. The new first shards take the chat template from `chat_template.jinja`, so `patch_chat_template.py` is not needed.
+
+
+---
+
+## Quantization procedure
+
+1. **Per-tensor database.** Each weight tensor is quantized at every candidate GGUF quantization type with GSQ, yielding a searchable database of quantized tensor variants.
+2. **RCO search.** The budget-constrained Riemannian search assigns one quantization type per tensor such that the backbone's average bit-width meets the target.
+3. **Assembly.** The selected per-tensor variants are stitched into a standard GGUF file.
+
+The dense tensors (attention, shared experts, embeddings, output head) and the routed-expert matrices draw on separate quantization ladders, because the routed experts sit at a much lower bit-width than the dense path at any given budget: the dense tensors choose among Q5_K, Q6_K and Q8_0, the routed experts among types from IQ1_S to IQ3_S. The search's teacher is the released checkpoint itself, with the routed experts in their native FP4, and its calibration data was generated by the model. <!-- TBD: per-build expert types and the searched tensor count. -->
+
+Reference implementations: **GSQ** at [IST-DASLab/GSQ](https://github.com/IST-DASLab/GSQ) and **RCO** at [IST-DASLab/RCO](https://github.com/IST-DASLab/RCO).
+
+### Reproducibility artifacts
+
+Each released build ships the files needed to audit how it was built:
+
+| File | Contents |
+|---|---|
+| `tensor-allocation/DeepSeek-V4.1-Flash-1017-GSQ-RCO-<ID>.rco-allocation.txt` | The quantization type assigned to every tensor of that build (all six shards, 1043 tensors), with a quant-type histogram and the target and measured bit-width. This is the RCO search result, so the allocation can be inspected without opening the model. |
+| `convert_ds41_upstream.py`, `chat_template.jinja`, `exp_probs_b_vl.f32` | The converter for downloads from before October 27, 2026 (see [Updating an existing download](#updating-an-existing-download)), the chat template the GGUFs embed, and the vision router bias the converter adds (40 x 384 F32, `ffn.gate.bias_vl` of the original checkpoint). |
+
+---
+
+## Citation
+
+If you use these models or methods, please cite both papers:
+
+```bibtex
+@article{gsq2026,
+  title  = {GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling},
+  author = {Dadgarnia, Alireza and Tabesh, Soroush and Nikdan, Mahdi and Helcig, Michael and Kurtic, Eldar and Kleinegger, Maximilian and Alistarh, Dan},
+  journal= {arXiv preprint arXiv:2604.18556},
+  year   = {2026}
+}
+@article{rco2026,
+  title  = {Model Compression with Exact Budget Constraints via Riemannian Manifolds},
+  author = {Helcig, Michael and Alistarh, Dan},
+  journal= {arXiv preprint arXiv:2605.00649},
+  year   = {2026}
+}
+```
+
+---
+
+## Acknowledgements
+
+<!-- TBD: confirm the compute acknowledgement for this release (copied from the Flash-Next card). -->
+
+We thank [Verda](https://verda.com/) and Scientific Computing at the Institute of Science and Technology Austria for providing the compute resources used to produce these models.
+
+---
+
+## License
+
+These quantized weights inherit the license of the base model (**DeepSeek-V4.1-Flash-1017**, MIT). The GSQ-RCO tooling is released by the Deep Algorithms and Systems Lab under its repository license.
+
+<div align="center">
+<sub>Built with <b>GSQ</b> and <b>RCO</b> at the <a href="https://github.com/IST-DASLab">Deep Algorithms and Systems Lab</a> &middot; Institute of Science and Technology Austria</sub>
+</div>

chat_template.jinja

@@ -0,0 +1,303 @@
+{#- Based on Unsloth's DeepSeek-V4.1 template, extracted from
+    unsloth/DeepSeek-V4.1-Flash-1017-GGUF at revision
+    60ad3651641b5c97679c807e064ab1d7a3ca1455.
+    Adaptations: use DeepSeek's native reasoning-effort aliases and default;
+    accept numeric reasoning_effort strings used by serving APIs;
+    preserve content_blocks on tool messages as in DeepSeek's reference encoder.
+    Image blocks render placeholders; the runtime supplies image embeddings through mmproj.
+-#}
+{%- set IMG = "<｜deepseek_image｜>" -%}
+{%- set DSML = "｜DSML｜" -%}
+
+{%- if thinking_mode is not defined or thinking_mode is none -%}
+    {%- if enable_thinking is defined and enable_thinking is not none -%}
+        {%- set thinking_mode = "thinking" if enable_thinking else "chat" -%}
+    {%- elif thinking is defined and thinking is not none and thinking is not string -%}
+        {%- set thinking_mode = "thinking" if thinking else "chat" -%}
+    {%- else -%}
+        {%- set thinking_mode = "thinking" -%}
+    {%- endif -%}
+{%- endif -%}
+{%- if thinking_mode not in ["chat", "thinking"] -%}
+    {{- raise_exception("Invalid thinking_mode `" ~ thinking_mode ~ "`, should be \"chat\" or \"thinking\"") -}}
+{%- endif -%}
+{%- if drop_thinking is not defined or drop_thinking is none -%}{%- set drop_thinking = true -%}{%- endif -%}
+{%- if add_generation_prompt is not defined -%}{%- set add_generation_prompt = false -%}{%- endif -%}
+{%- if tools is not defined -%}{%- set tools = none -%}{%- endif -%}
+{%- if response_format is not defined -%}{%- set response_format = none -%}{%- endif -%}
+
+{%- if reasoning_effort is not defined or reasoning_effort is none -%}{%- set reasoning_effort = "high" -%}{%- endif -%}
+{%- set effort_map = {"low": 50, "high": 75, "max": 100} -%}
+{%- if reasoning_effort is string and reasoning_effort in effort_map -%}
+    {%- set budget = effort_map[reasoning_effort] -%}
+{%- elif reasoning_effort is integer and reasoning_effort is not boolean and reasoning_effort >= 1 and reasoning_effort <= 100 -%}
+    {%- set budget = reasoning_effort -%}
+{%- elif reasoning_effort is string and (reasoning_effort | int | string) == reasoning_effort and (reasoning_effort | int) >= 1 and (reasoning_effort | int) <= 100 -%}
+    {%- set budget = reasoning_effort | int -%}
+{%- else -%}
+    {{- raise_exception("Invalid reasoning effort: " ~ reasoning_effort ~ ", should be int within [1,100] or one of \"low\", \"high\", \"max\"") -}}
+{%- endif -%}
+
+{#- Splits/validates "ns::name" against an explicit namespace and returns the qualified name. -#}
+{%- macro qualified_name(name, ns) -%}
+    {%- if ns is mapping -%}{%- set ns = ns.name -%}{%- endif -%}
+    {%- set parts = name.split("::", 1) -%}
+    {%- if parts | length == 2 -%}
+        {%- if ns is not none and ns != parts[0] -%}{{- raise_exception("Conflicting tool namespaces: " ~ ns ~ " != " ~ parts[0]) -}}{%- endif -%}
+        {%- set ns = parts[0] -%}
+        {%- set name = parts[1] -%}
+    {%- endif -%}
+    {%- if "::" in name -%}{{- raise_exception("Tool name must not contain '::': " ~ name) -}}{%- endif -%}
+    {%- if ns is not none and "::" in ns -%}{{- raise_exception("Tool namespace must not contain '::': " ~ ns) -}}{%- endif -%}
+    {{- name if ns is none else ns ~ "::" ~ name -}}
+{%- endmacro -%}
+
+{%- macro render_tools(tool_list) -%}
+    {%- set ts = namespace(schemas=[]) -%}
+    {%- for tool in tool_list -%}
+        {%- set fn = tool.function if tool.function is defined else tool -%}
+        {%- set ns = tool.namespace if tool.namespace is defined and tool.namespace is not none else fn.get("namespace") -%}
+        {#- JSON assembled by hand (key order kept, name/description replaced in place): llama.cpp has no ** unpacking. -#}
+        {%- set qname = qualified_name(fn.name, ns) -%}
+        {%- set sch = namespace(parts=[]) -%}
+        {%- for k, v in fn.items() if k != "namespace" -%}
+            {%- if k == "name" -%}{%- set v = qname -%}
+            {%- elif k == "description" and ns is mapping and ns.description -%}{%- set v = ns.description ~ "\n" ~ (fn.description or "") -%}
+            {%- endif -%}
+            {%- set sch.parts = sch.parts + [(k | tojson) ~ ": " ~ (v | tojson)] -%}
+        {%- endfor -%}
+        {%- if "description" not in fn and ns is mapping and ns.description -%}
+            {%- set sch.parts = sch.parts + ['"description": ' ~ ((ns.description ~ "\n") | tojson)] -%}
+        {%- endif -%}
+        {%- set ts.schemas = ts.schemas + ["{" ~ sch.parts | join(", ") ~ "}"] -%}
+    {%- endfor -%}
+    {{- "## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n" -}}
+    {{- ts.schemas | join("\n") -}}
+    {{- "\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n" -}}
+{%- endmacro -%}
+
+{#- 1. Merge tool results (and consecutive user turns) into user messages made of content blocks. -#}
+{%- set mns = namespace(msgs=[]) -%}
+{%- for msg in messages -%}
+    {%- set content = msg.get("content") -%}
+    {%- if msg.role == "tool" and msg.get("content_blocks") and content is not string -%}
+        {%- set content = msg.content_blocks -%}
+    {%- endif -%}
+    {%- set is_blocks = content is iterable and content is not string and content is not mapping -%}
+    {%- if content is string and IMG in content -%}
+        {{- raise_exception("Message content contains image special token " ~ IMG ~ ". Images should be provided as image content blocks.") -}}
+    {%- endif -%}
+    {%- set rc = msg.reasoning_content if msg.get("reasoning_content") is not none else msg.get("thinking") -%}
+    {%- if rc is string and IMG in rc -%}
+        {{- raise_exception("reasoning_content contains image special token " ~ IMG) -}}
+    {%- endif -%}
+    {%- set prev = mns.msgs[-1] if mns.msgs else none -%}
+    {%- if msg.role == "tool" -%}
+        {%- set tblock = {"type": "tool_result", "tool_use_id": msg.get("tool_call_id", ""), "content": content if content is not none else ""} -%}
+        {%- if prev is not none and prev.role == "user" and prev.content_blocks is defined -%}
+            {%- set mns.msgs = mns.msgs[:-1] + [dict(prev, content_blocks=prev.content_blocks + [tblock])] -%}
+        {%- else -%}
+            {%- set mns.msgs = mns.msgs + [{"role": "user", "content_blocks": [tblock]}] -%}
+        {%- endif -%}
+    {%- elif msg.role == "user" -%}
+        {%- set blocks = msg.content_blocks if msg.get("content_blocks") is not none else (content if is_blocks else [{"type": "text", "text": content or ""}]) -%}
+        {%- if prev is not none and prev.role == "user" and prev.content_blocks is defined and prev.get("task") is none -%}
+            {%- set mns.msgs = mns.msgs[:-1] + [dict(prev, content_blocks=prev.content_blocks + blocks)] -%}
+        {%- else -%}
+            {%- set mns.msgs = mns.msgs + [dict(msg, content_blocks=blocks)] -%}
+        {%- endif -%}
+    {%- else -%}
+        {%- set mns.msgs = mns.msgs + [msg] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Top-level tools / response_format attach to a leading system message, or to a new empty one. -#}
+{%- if tools or response_format -%}
+    {%- if not (mns.msgs and mns.msgs[0].role == "system") -%}
+        {%- set mns.msgs = [{"role": "system", "content": ""}] + mns.msgs -%}
+    {%- endif -%}
+    {%- set first = mns.msgs[0] -%}
+    {%- if tools and not first.get("tools") -%}{%- set first = dict(first, tools=tools) -%}{%- endif -%}
+    {%- if response_format and not first.get("response_format") -%}{%- set first = dict(first, response_format=response_format) -%}{%- endif -%}
+    {%- set mns.msgs = [first] + mns.msgs[1:] -%}
+{%- endif -%}
+
+{#- 2. Reorder tool results to match the tool_calls order of the preceding assistant message. -#}
+{%- set sns = namespace(msgs=[], ids=[]) -%}
+{%- for m in mns.msgs -%}
+    {%- if m.role == "assistant" and m.get("tool_calls") -%}
+        {%- set sns.ids = [] -%}
+        {%- for tc in m.tool_calls -%}
+            {%- set sns.ids = sns.ids + [tc.get("id") or (tc.function.get("id", "") if tc.function is defined else "")] -%}
+        {%- endfor -%}
+        {%- set sns.msgs = sns.msgs + [m] -%}
+    {%- elif m.role == "user" and m.content_blocks -%}
+        {%- set tb = namespace(blocks=[]) -%}
+        {%- for b in m.content_blocks if b.type == "tool_result" -%}
+            {%- set order = namespace(idx=0) -%}
+            {%- for tc_id in sns.ids -%}
+                {%- if tc_id and tc_id == b.get("tool_use_id", "") -%}{%- set order.idx = loop.index0 -%}{%- endif -%}
+            {%- endfor -%}
+            {%- set tb.blocks = tb.blocks + [dict(b, _order=order.idx)] -%}
+        {%- endfor -%}
+        {%- if tb.blocks | length > 1 and sns.ids -%}
+            {%- set sorted_blocks = tb.blocks | sort(attribute="_order") -%}
+            {%- set rb = namespace(blocks=[], i=0) -%}
+            {%- for b in m.content_blocks -%}
+                {%- if b.type == "tool_result" -%}
+                    {%- set rb.blocks = rb.blocks + [sorted_blocks[rb.i]] -%}
+                    {%- set rb.i = rb.i + 1 -%}
+                {%- else -%}
+                    {%- set rb.blocks = rb.blocks + [b] -%}
+                {%- endif -%}
+            {%- endfor -%}
+            {%- set sns.msgs = sns.msgs + [dict(m, content_blocks=rb.blocks)] -%}
+        {%- else -%}
+            {%- set sns.msgs = sns.msgs + [m] -%}
+        {%- endif -%}
+    {%- else -%}
+        {%- set sns.msgs = sns.msgs + [m] -%}
+    {%- endif -%}
+{%- endfor -%}
+{%- set msgs = sns.msgs -%}
+
+{#- Last user turn: users and mid-conversation systems count, except user turns made only of tool results.
+    Tools anywhere disable drop_thinking. -#}
+{%- set state = namespace(last_user=-1, has_tools=false) -%}
+{%- for m in msgs -%}
+    {%- if m.role == "user" or (m.role == "system" and loop.index0 > 0) -%}
+        {%- set real = namespace(v=true) -%}
+        {%- if m.get("content_blocks") -%}
+            {%- set real.v = false -%}
+            {%- for b in m.content_blocks if b.type in ["text", "image", "image_url"] -%}{%- set real.v = true -%}{%- endfor -%}
+        {%- endif -%}
+        {%- if real.v -%}{%- set state.last_user = loop.index0 -%}{%- endif -%}
+    {%- endif -%}
+    {%- if m.get("tools") -%}{%- set state.has_tools = true -%}{%- endif -%}
+{%- endfor -%}
+{%- set effective_drop = drop_thinking and not state.has_tools -%}
+{%- set is_thinking = thinking_mode == "thinking" -%}
+{#- Like encoding.py, thinking mode with drop_thinking silently drops messages of other roles before the last user turn. -#}
+{%- if is_thinking and effective_drop -%}
+    {%- set kept = namespace(msgs=[]) -%}
+    {%- for m in msgs -%}
+        {%- if m.role in ["user", "system", "tool", "latest_reminder", "direct_search_results", "assistant"] or loop.index0 >= state.last_user -%}
+            {%- set kept.msgs = kept.msgs + [m] -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {%- if kept.msgs | length != msgs | length -%}
+        {%- set msgs = kept.msgs -%}
+        {%- set state.last_user = -1 -%}
+        {%- for m in msgs -%}
+            {%- if m.role == "user" or (m.role == "system" and loop.index0 > 0) -%}
+                {%- set real = namespace(v=true) -%}
+                {%- if m.get("content_blocks") -%}
+                    {%- set real.v = false -%}
+                    {%- for b in m.content_blocks if b.type in ["text", "image", "image_url"] -%}{%- set real.v = true -%}{%- endfor -%}
+                {%- endif -%}
+                {%- if real.v -%}{%- set state.last_user = loop.index0 -%}{%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- endif -%}
+{%- endif -%}
+
+{{- "<｜begin▁of▁sentence｜>" -}}
+{#- 3. Render each message. -#}
+{%- for m in msgs -%}
+    {%- set idx = loop.index0 -%}
+    {%- set content = m.get("content") -%}
+    {%- if content is iterable and content is not string and content is not mapping -%}
+        {%- set sp = namespace(parts=[]) -%}
+        {%- for b in content -%}
+            {%- if b.type == "text" -%}{%- set sp.parts = sp.parts + [b.text or ""] -%}
+            {%- elif b.type in ["image", "image_url"] -%}{%- set sp.parts = sp.parts + [IMG] -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {%- set content = sp.parts | join("\n\n") -%}
+    {%- endif -%}
+
+    {%- if idx == 0 -%}
+        {%- if is_thinking or m.role == "system" -%}{{- "<｜System｜>" -}}{%- endif -%}
+        {%- if is_thinking -%}{{- "Reasoning Effort: " ~ budget ~ " (range 1-100, the higher the value, the more thorough the reasoning)\n\n" -}}{%- endif -%}
+    {%- endif -%}
+
+    {%- if m.role == "system" -%}
+        {%- if idx > 0 -%}{{- "<｜System｜>" -}}{%- endif -%}
+        {{- content or "" -}}
+        {%- if m.get("tools") -%}{{- "\n\n" ~ render_tools(m.tools) -}}{%- endif -%}
+        {%- if m.get("response_format") -%}
+            {{- "\n\n## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n" ~ (m.response_format | tojson) -}}
+        {%- endif -%}
+
+    {%- elif m.role == "user" -%}
+        {{- "<｜User｜>" -}}
+        {%- set up = namespace(parts=[]) -%}
+        {%- for b in m.content_blocks -%}
+            {%- if b.type == "text" -%}
+                {%- if IMG in (b.text or "") -%}{{- raise_exception("Text block contains image placeholder " ~ IMG ~ ". Images should be separate content blocks.") -}}{%- endif -%}
+                {%- set up.parts = up.parts + [b.text or ""] -%}
+            {%- elif b.type in ["image", "image_url"] -%}
+                {%- set up.parts = up.parts + [IMG] -%}
+            {%- elif b.type == "tool_result" -%}
+                {%- set tc = b.get("content", "") -%}
+                {%- if tc is iterable and tc is not string and tc is not mapping -%}
+                    {%- set sub = namespace(parts=[]) -%}
+                    {%- for s in tc -%}
+                        {%- if s.type == "text" -%}
+                            {%- if IMG in (s.text or "") -%}{{- raise_exception("Text block contains image placeholder " ~ IMG ~ ". Images should be separate content blocks.") -}}{%- endif -%}
+                            {%- set sub.parts = sub.parts + [s.get("text", "")] -%}
+                        {%- elif s.type in ["image", "image_url"] -%}{%- set sub.parts = sub.parts + [IMG] -%}
+                        {%- else -%}{%- set sub.parts = sub.parts + ["[Unsupported " ~ s.type ~ "]"] -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    {%- set tc = sub.parts | join("\n\n") -%}
+                {%- endif -%}
+                {%- set up.parts = up.parts + ["<tool_result>" ~ tc ~ "</tool_result>"] -%}
+            {%- else -%}
+                {%- set up.parts = up.parts + ["[Unsupported " ~ b.type ~ "]"] -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {{- up.parts | join("\n\n") -}}
+
+    {%- elif m.role == "latest_reminder" -%}
+        {{- "<｜latest_reminder｜>" ~ (content or "") -}}
+
+    {%- elif m.role == "assistant" -%}
+        {%- set prev_has_task = idx > 0 and msgs[idx - 1].get("task") is not none -%}
+        {%- if is_thinking and not prev_has_task and (not effective_drop or idx > state.last_user) -%}
+            {{- ((m.reasoning_content if m.get("reasoning_content") is not none else m.get("thinking")) or "") ~ "</think>" -}}
+        {%- endif -%}
+        {{- content or "" -}}
+        {%- if m.get("tool_calls") -%}
+            {%- set tcs = namespace(lines=[]) -%}
+            {%- for tc in m.tool_calls -%}
+                {%- set fn = tc.function if tc.function is defined else tc -%}
+                {%- set params = namespace(lines=[]) -%}
+                {%- set args = fn.arguments if fn.arguments is mapping else {"arguments": fn.arguments} -%}
+                {%- for k, v in args.items() -%}
+                    {%- set params.lines = params.lines + ["<" ~ DSML ~ " parameter name=\"" ~ k ~ "\" string=\"" ~ ("true" if v is string else "false") ~ "\">" ~ (v if v is string else v | tojson) ~ "</" ~ DSML ~ " parameter>"] -%}
+                {%- endfor -%}
+                {%- set name = qualified_name(fn.name, tc.get("namespace") or fn.get("namespace")) -%}
+                {%- set tcs.lines = tcs.lines + ["<" ~ DSML ~ " invoke name=\"" ~ name ~ "\">\n" ~ params.lines | join("\n") ~ "\n</" ~ DSML ~ " invoke>"] -%}
+            {%- endfor -%}
+            {{- "\n\n<" ~ DSML ~ " calls>\n" ~ tcs.lines | join("\n") ~ "\n</" ~ DSML ~ " calls>" -}}
+        {%- endif -%}
+        {%- if not m.get("wo_eos", false) -%}{{- "<｜end▁of▁sentence｜>" -}}{%- endif -%}
+
+    {%- else -%}
+        {{- raise_exception("Unknown role: " ~ m.role) -}}
+    {%- endif -%}
+
+    {#- Transition tokens, only when the next message is an assistant/latest_reminder turn or this is the last one. -#}
+    {%- if loop.last or msgs[idx + 1].role in ["assistant", "latest_reminder"] -%}
+        {%- set task = m.get("task") -%}
+        {%- if task is not none -%}
+            {%- set task_tokens = {"action": "<｜action｜>", "query": "<｜query｜>", "authority": "<｜authority｜>", "domain": "<｜domain｜>", "title": "<｜title｜>", "read_url": "<｜read_url｜>"} -%}
+            {%- if task not in task_tokens -%}{{- raise_exception("Invalid task: '" ~ task ~ "'") -}}{%- endif -%}
+            {%- if task == "action" -%}{{- "<｜Assistant｜>" ~ ("<think>" if is_thinking else "</think>") -}}{%- endif -%}
+            {{- task_tokens[task] -}}
+        {%- elif (m.role == "user" or (m.role == "system" and idx > 0)) and (add_generation_prompt or not loop.last) -%}
+            {{- "<｜Assistant｜>" ~ ("<think>" if is_thinking and (not effective_drop or idx >= state.last_user) else "</think>") -}}
+        {%- endif -%}
+    {%- endif -%}
+{%- endfor -%}

convert_ds41_upstream.py

@@ -0,0 +1,849 @@
+#!/usr/bin/env python3
+"""Convert the ISTA-DASLab DeepSeek-V4.1-Flash-1017 GSQ-RCO GGUFs to llama.cpp's own deepseek41 layout.
+
+The files published before October 27, 2026 were written for ISTA's llama.cpp branch. llama.cpp's DeepSeek-V4.1
+support (ggml-org/llama.cpp#31086, merged October 25, 2026) reads the layout of its converter
+(conversion/deepseek41.py). This script writes that layout from the old files:
+
+  - metadata under the upstream key names; the Engram hash constants (token map, bucket counts, multipliers) move
+    from tensors into metadata, with absolute row offsets into the concatenated table
+  - the Engram tables of all Engram layers become one tensor, per_layer_token_embd. They stay in the shared
+    Engram file, which is converted in place: only its 416-byte header is rewritten, the 110 GB of table data stay
+    where they are. llama.cpp opens it as shard 6 of each build, through a link <build>-00006-of-00006.gguf
+  - blk.N.engram_wkv is renamed blk.N.engram_kv; blk.N.engram_q and blk.N.engram_k are folded into the F32
+    blk.N.engram_gate = q * k (exact: a product of two BF16 values fits in F32)
+  - the vision router bias blk.N.exp_probs_b_vl is added (the repository's exp_probs_b_vl.f32, F32 from the
+    original checkpoint); llama.cpp routes image tokens with it
+  - blk.N.attn_output_a as the 2D matrix of the checkpoint (it was stored as [4096, 1024, 8]; same bytes)
+  - tensors in converter order, split like llama-gguf-split --split-max-size 40G --no-tensor-first-split: shard 1
+    holds only metadata, so later metadata fixes replace a 6 MB file
+  - the chat template of the repository's chat_template.jinja, so downloads that were never patched get the
+    current one too
+  - mmproj: projector type deepseek41v (V4.1 image token layout) instead of deepseek4v
+
+Quantized weights are copied byte for byte. Every written file is read back: each tensor is compared with its
+source, and the file with the SHA-256 of the published one (the Engram tables too, before the header is rewritten),
+so a damaged old download or an unfinished patch_chat_template.py run stops the conversion with an error instead of
+giving wrong files. Standard library only (Python 3.8+). The model is not downloaded again;
+the script only fetches chat_template.jinja and exp_probs_b_vl.f32 (80 KB) from the repository, unless their current
+versions (checked by SHA-256) lie next to it or in the model folder (then it runs offline).
+
+    python3 convert_ds41_upstream.py MODEL_FOLDER [--out NEW_FOLDER]
+
+MODEL_FOLDER holds the old shards (*-0000N-of-00004.gguf), DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf and optionally the
+mmproj. Every build found there is converted: its new shards 1 to 5 are written next to the old ones (or to --out)
+and need as much space as its old shards (121 to 155 GB). With --delete-old, the old shards of a build are deleted as
+soon as its new ones are verified, so free space for one build suffices. On a folder of new builds the script only
+adds missing links to the Engram file.
+"""
+import argparse
+import hashlib
+import os
+import re
+import struct
+import sys
+import urllib.request
+
+REPO = 'ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF'  # chat_template.jinja and exp_probs_b_vl.f32 come from here
+OLD_REVISION = '653a9874aa38c356233c5f3cc8c96b19a43b57eb'  # the repository's last revision with the old files
+BUILD_PREFIX = 'DeepSeek-V4.1-Flash-1017-GSQ-RCO-'
+ARCH = 'deepseek41'
+TABLE = 'per_layer_token_embd.weight'
+ENGRAM_FILE = 'DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf'
+N_SPLIT = 6  # shards per build; the last is the shared Engram file
+SPLIT_MAX_SIZE = 40 * 1000 ** 3  # llama-gguf-split --split-max-size 40G
+ALIGNMENT = 32
+SPLIT_RE = re.compile(r'^(?P<prefix>.+)-(?P<no>\d{5})-of-(?P<count>\d{5})\.gguf$')
+COPY_CHUNK = 64 << 20
+
+# ---------------------------------------------------------------------------------------------------------------
+# GGUF header reading and writing
+
+GGUF_MAGIC = 0x46554747
+T_U16, T_U32, T_I32, T_STRING, T_ARRAY, T_U64 = 2, 4, 5, 8, 9, 10
+SCALAR = {0: '<B', 1: '<b', 2: '<H', 3: '<h', 4: '<I', 5: '<i', 6: '<f', 7: '<?', 10: '<Q', 11: '<q', 12: '<d'}
+GGML_F32, GGML_IQ4_NL, GGML_I32, GGML_I64, GGML_BF16 = 0, 20, 26, 27, 30
+# GGML type -> (elements per block, bytes per block), for the types these files contain
+GGML_BLOCK = {0: (1, 4), 1: (1, 2), 8: (32, 34), 10: (256, 84), 11: (256, 110), 12: (256, 144), 13: (256, 176),
+              14: (256, 210), 16: (256, 66), 17: (256, 74), 18: (256, 98), 19: (256, 50), 20: (32, 18),
+              21: (256, 110), 22: (256, 82), 23: (256, 136), 26: (1, 4), 27: (1, 8), 29: (256, 56), 30: (1, 2),
+              39: (32, 17)}
+
+
+def align(n, a=ALIGNMENT):
+    return (n + a - 1) // a * a
+
+
+def tensor_nbytes(dims, ggml_type):
+    per_block, block_bytes = GGML_BLOCK[ggml_type]
+    n = 1
+    for d in dims:
+        n *= d
+    if dims[0] % per_block:
+        raise ValueError('row of %d elements is not a whole number of type %d blocks' % (dims[0], ggml_type))
+    return n // per_block * block_bytes
+
+
+class _Reader:
+    def __init__(self, buf):
+        self.buf, self.p = buf, 0
+
+    def take(self, n):
+        if self.p + n > len(self.buf):
+            raise EOFError
+        b = self.buf[self.p:self.p + n]
+        self.p += n
+        return b
+
+    def u(self, fmt):
+        return struct.unpack(fmt, self.take(struct.calcsize(fmt)))[0]
+
+    def string(self):
+        return self.take(self.u('<Q')).decode('utf-8')
+
+    def skip_value(self, t):
+        if t == T_STRING:
+            self.take(self.u('<Q'))
+        elif t == T_ARRAY:
+            at, n = self.u('<I'), self.u('<Q')
+            if at in (T_STRING, T_ARRAY):
+                for _ in range(n):
+                    self.skip_value(at)
+            else:
+                self.take(n * struct.calcsize(SCALAR[at]))
+        else:
+            self.take(struct.calcsize(SCALAR[t]))
+
+
+def decode_value(t, raw):
+    r = _Reader(raw)
+    if t == T_STRING:
+        return r.string()
+    if t == T_ARRAY:
+        at, n = r.u('<I'), r.u('<Q')
+        if at == T_STRING:
+            return [r.string() for _ in range(n)]
+        fmt = SCALAR[at]
+        return list(struct.unpack('<%d%s' % (n, fmt[1]), r.take(n * struct.calcsize(fmt))))
+    return struct.unpack(SCALAR[t], raw)[0]
+
+
+def encode_value(t, value, array_type=None):
+    if t == T_STRING:
+        b = value.encode('utf-8')
+        return struct.pack('<Q', len(b)) + b
+    if t == T_ARRAY:
+        head = struct.pack('<IQ', array_type, len(value))
+        if array_type == T_STRING:
+            return head + b''.join(encode_value(T_STRING, v) for v in value)
+        return head + struct.pack('<%d%s' % (len(value), SCALAR[array_type][1]), *value)
+    return struct.pack(SCALAR[t], value)
+
+
+class GGUFFile:
+    """Header of a GGUF file: kv is a list of [key, type, raw value bytes], tensors maps a name to
+    (absolute file offset, dims, ggml type, byte size)."""
+
+    def __init__(self, path):
+        self.path = path
+        size = 1 << 20
+        while True:
+            with open(path, 'rb') as f:
+                buf = f.read(size)
+            try:
+                self._parse(buf)
+                break
+            except EOFError:
+                if len(buf) < size:
+                    raise ValueError('%s: truncated GGUF header' % path)
+                size *= 4
+        self.size = os.path.getsize(path)
+
+    def _parse(self, buf):
+        r = _Reader(buf)
+        if r.u('<I') != GGUF_MAGIC:
+            raise ValueError('%s is not a GGUF file' % self.path)
+        self.version = r.u('<I')
+        n_tensors, n_kv = r.u('<Q'), r.u('<Q')
+        self.kv = []
+        for _ in range(n_kv):
+            key = r.string()
+            t = r.u('<I')
+            start = r.p
+            r.skip_value(t)
+            self.kv.append([key, t, buf[start:r.p]])
+        infos = []
+        for _ in range(n_tensors):
+            name = r.string()
+            dims = [r.u('<Q') for _ in range(r.u('<I'))]
+            infos.append((name, dims, r.u('<I'), r.u('<Q')))
+        self.header_size = r.p
+        self.data_start = align(r.p, self.get('general.alignment', ALIGNMENT))
+        self.tensors = {}
+        self.order = []
+        for name, dims, gt, off in infos:
+            self.tensors[name] = (self.data_start + off, dims, gt, tensor_nbytes(dims, gt))
+            self.order.append(name)
+
+    def get(self, key, default=None):
+        for k, t, raw in self.kv:
+            if k == key:
+                return decode_value(t, raw)
+        return default
+
+    def read_tensor(self, name):
+        off, _, _, n = self.tensors[name]
+        with open(self.path, 'rb') as f:
+            f.seek(off)
+            data = f.read(n)
+        if len(data) != n:
+            raise ValueError('%s: tensor %s is truncated' % (self.path, name))
+        return data
+
+
+def serialize_header(kv, infos):
+    """kv: [key, type, raw]; infos: (name, dims, ggml type, relative offset). Padded to the data section."""
+    out = [struct.pack('<IIQQ', GGUF_MAGIC, 3, len(infos), len(kv))]
+    for key, t, raw in kv:
+        out += [encode_value(T_STRING, key), struct.pack('<I', t), raw]
+    for name, dims, gt, off in infos:
+        out += [encode_value(T_STRING, name), struct.pack('<I', len(dims)),
+                struct.pack('<%dQ' % len(dims), *dims), struct.pack('<IQ', gt, off)]
+    header = b''.join(out)
+    return header + b'\0' * (align(len(header)) - len(header))
+
+
+# ---------------------------------------------------------------------------------------------------------------
+# conversion
+
+
+class Tensor:
+    """An output tensor: bytes from a file range or computed here."""
+
+    def __init__(self, name, dims, ggml_type, path=None, offset=None, data=None):
+        self.name, self.dims, self.type = name, list(dims), ggml_type
+        self.nbytes = tensor_nbytes(self.dims, ggml_type)
+        self.path, self.offset, self.data = path, offset, data
+        if data is not None and len(data) != self.nbytes:
+            raise ValueError('%s: %d bytes for a tensor of %d' % (name, len(data), self.nbytes))
+
+    def source_hash(self):
+        if self.data is not None:
+            return hashlib.sha256(self.data).hexdigest()
+        return hash_range(self.path, self.offset, self.nbytes)
+
+
+def read_into(f, n, *hashes):
+    """Reads the next n bytes of f into the given hashes."""
+    while n:
+        b = f.read(min(n, COPY_CHUNK))
+        if not b:
+            raise ValueError('%s ends early' % f.name)
+        for h in hashes:
+            h.update(b)
+        n -= len(b)
+
+
+def hash_range(path, offset, n):
+    h = hashlib.sha256()
+    with open(path, 'rb') as f:
+        f.seek(offset)
+        read_into(f, n, h)
+    return h.hexdigest()
+
+
+# the old branch-specific metadata, replaced by the upstream keys
+OLD_KEYS = ['attention.candidate_block_size', 'attention.candidate_topk_blocks', 'attention.candidate_source_layer_id',
+            'kv_source_layer_ids', 'index_source_layer_ids', 'dspark_target_layer_ids', 'expert_scoring_func',
+            'engram.present', 'engram.layer_ids', 'engram.num_embeddings', 'engram_head_dim', 'engram_n_heads',
+            'engram_max_ngram_size', 'engram_vocab_size', 'engram_compressed_vocab_size', 'engram_pad_token_id',
+            'engram.source', 'engram.file']
+CONST_TENSORS = ['engram_token_map.weight', 'engram_primes.weight', 'engram_offsets.weight',
+                 'engram_multipliers.weight']
+
+
+def unpack(data, fmt):
+    return list(struct.unpack('<%d%s' % (len(data) // struct.calcsize(fmt), fmt[1]), data))
+
+
+def convert_metadata(meta, tensor_source, chat_template):
+    """meta: GGUFFile with the old first-shard metadata. Returns the new kv list (without split keys)."""
+    A = ARCH + '.'
+    get = lambda k: meta.get(A + k)
+    if meta.get('general.architecture') != ARCH:
+        raise ValueError('%s: architecture %r, expected %s' % (meta.path, meta.get('general.architecture'), ARCH))
+    missing = [A + k for k in OLD_KEYS if meta.get(A + k) is None and k not in ('dspark_target_layer_ids',
+                                                                              'expert_scoring_func')]
+    if missing:
+        raise ValueError('%s: missing metadata %s' % (meta.path, ', '.join(missing)))
+
+    layers = get('engram.layer_ids')
+    n_gram, n_heads, head_dim = get('engram_max_ngram_size'), get('engram_n_heads'), get('engram_head_dim')
+    n_cols = (n_gram - 1) * n_heads
+    token_map = unpack(tensor_source('engram_token_map.weight'), '<i')
+    primes = unpack(tensor_source('engram_primes.weight'), '<q')
+    local_offsets = unpack(tensor_source('engram_offsets.weight'), '<q')
+    multipliers = unpack(tensor_source('engram_multipliers.weight'), '<q')
+    if len(primes) != len(layers) * n_cols or len(local_offsets) != len(primes) or \
+            len(multipliers) != len(layers) * n_gram:
+        raise ValueError('Engram constants do not match %d layers x %d heads' % (len(layers), n_cols))
+    for e, rows in enumerate(get('engram.num_embeddings')):
+        layer_primes = primes[e * n_cols:(e + 1) * n_cols]
+        if sum(layer_primes) != rows:
+            raise ValueError('Engram layer %d: bucket counts do not add up to its table rows' % layers[e])
+        if local_offsets[e * n_cols:(e + 1) * n_cols] != [sum(layer_primes[:i]) for i in range(n_cols)]:
+            raise ValueError('Engram layer %d: unexpected row offsets' % layers[e])
+    if max(token_map) + 1 != get('engram_compressed_vocab_size') or \
+            len(token_map) != len(meta.get('tokenizer.ggml.tokens')):
+        raise ValueError('Engram token map does not match the vocabulary')
+    if any(m <= 0 or m % 2 == 0 for m in multipliers):
+        raise ValueError('Engram multipliers must be odd and positive')
+
+    new = [
+        ('attention.kv_source_layers', T_ARRAY, get('kv_source_layer_ids'), T_U32),
+        ('attention.indexer.source_layers', T_ARRAY, get('index_source_layer_ids'), T_U32),
+        ('attention.indexer.candidate_source_layer', T_I32, get('attention.candidate_source_layer_id'), None),
+        ('attention.indexer.candidate_top_k_blocks', T_U32, get('attention.candidate_topk_blocks'), None),
+        ('attention.indexer.candidate_block_size', T_U32, get('attention.candidate_block_size'), None),
+        ('ple.layers', T_ARRAY, layers, T_U32),
+        ('ple.ngram_size', T_U32, n_gram, None),
+        ('ple.heads_per_ngram', T_U32, n_heads, None),
+        ('embedding_length_per_layer_input', T_U32, head_dim, None),
+        ('ple.layer_multipliers', T_ARRAY, multipliers, T_U64),
+        ('ple.head_offsets', T_ARRAY, [sum(primes[:i]) for i in range(len(primes))], T_U64),
+        ('ple.head_vocab_sizes', T_ARRAY, primes, T_U64),
+        ('engram.pad_token_id', T_U32, get('engram_pad_token_id'), None),
+        ('engram.token_map', T_ARRAY, token_map, T_I32),
+    ]
+    old = {A + k for k in OLD_KEYS}
+    kv, inserted = [], False
+    for key, t, raw in meta.kv:
+        if key.startswith('split.'):
+            continue
+        if key in old:
+            if not inserted:
+                kv += [[A + k, vt, encode_value(vt, v, at)] for k, vt, v, at in new]
+                inserted = True
+            continue
+        if key == 'tokenizer.chat_template':
+            raw = encode_value(T_STRING, chat_template)  # the repository's, whether or not the old one was patched
+        kv.append([key, t, raw])
+    return kv
+
+
+def bf16_to_f32(h):
+    return struct.unpack('<f', struct.pack('<I', h << 16))[0]
+
+
+def engram_gate(q, k):
+    """F32 q * k from two BF16 tensors (the upstream converter's data_torch.float() * k().float())."""
+    n = len(q) // 2
+    qv, kv = struct.unpack('<%dH' % n, q), struct.unpack('<%dH' % n, k)
+    return struct.pack('<%df' % n, *[bf16_to_f32(a) * bf16_to_f32(b) for a, b in zip(qv, kv)])
+
+
+def canonical_key(name):
+    """Converter order: global tensors, then the blocks in order, the Engram table last; names sorted."""
+    if name == 'per_layer_token_embd.weight':
+        return (2, 0, name)
+    m = re.match(r'blk\.(\d+)\.', name)
+    return (1, int(m.group(1)), name) if m else (0, 0, name)
+
+
+def plan_tensors(meta, sources, engram, bias_vl):
+    """sources: GGUFFiles holding the old backbone tensors; engram: the old Engram GGUFFile."""
+    index = {}
+    for src in sources:
+        for name, (off, dims, gt, _) in src.tensors.items():
+            if name in index:
+                raise ValueError('tensor %s found in %s and %s' % (name, index[name][0].path, src.path))
+            index[name] = (src, off, dims, gt)
+    A = ARCH + '.'
+    n_layer, n_expert = meta.get(A + 'block_count'), meta.get(A + 'expert_count')
+    layers = meta.get(A + 'engram.layer_ids')
+    out = []
+    for name, (src, off, dims, gt) in index.items():
+        if name in CONST_TENSORS or re.match(r'blk\.\d+\.engram_[qk]\.weight$', name):
+            continue
+        if name.endswith('.attn_output_a.weight') and len(dims) == 3:
+            dims = [dims[0], dims[1] * dims[2]]  # the checkpoint's 2D wo_a, as the converter writes it; same bytes
+        out.append(Tensor(name.replace('.engram_wkv.', '.engram_kv.'), dims, gt, src.path, off))
+
+    for layer in layers:
+        q, k = ('blk.%d.engram_%s.weight' % (layer, w) for w in 'qk')
+        if index[q][3] != GGML_BF16 or index[k][3] != GGML_BF16 or index[q][2] != index[k][2]:
+            raise ValueError('blk.%d engram_q / engram_k must be BF16 of the same shape' % layer)
+        out.append(Tensor('blk.%d.engram_gate.weight' % layer, index[q][2], GGML_F32,
+                          data=engram_gate(index[q][0].read_tensor(q), index[k][0].read_tensor(k))))
+
+    if len(bias_vl) != n_layer * n_expert * 4:
+        raise ValueError('exp_probs_b_vl.f32 is for a different model')
+    for layer in range(n_layer):
+        if 'blk.%d.exp_probs_b.bias' % layer not in index:
+            raise ValueError('blk.%d.exp_probs_b.bias is missing' % layer)
+        row = bias_vl[layer * n_expert * 4:(layer + 1) * n_expert * 4]
+        out.append(Tensor('blk.%d.exp_probs_b_vl.bias' % layer, [n_expert], GGML_F32, data=row))
+
+    table, table_layers = engram_table(engram)
+    if table.dims[1] != sum(meta.get(A + 'engram.num_embeddings')) or table_layers not in (None, layers) or \
+            meta.get(A + 'engram.source') != engram.get(A + 'engram.source'):
+        raise ValueError('%s does not belong to this model (other tables or checkpoint revision)' % engram.path)
+    out.append(table)
+    out.sort(key=lambda t: canonical_key(t.name))
+    return out
+
+
+def engram_table(engram):
+    """The Engram table as one tensor, and the layers of the old per-layer tables (None if already converted).
+    The old file holds the per-layer tables back to back, so their concatenation is one byte range."""
+    if TABLE in engram.tensors:
+        off, dims, gt, _ = engram.tensors[TABLE]
+        return Tensor(TABLE, dims, gt, engram.path, off), None
+    names = sorted((n for n in engram.tensors if re.match(r'blk\.\d+\.engram_embd\.weight$', n)),
+                   key=lambda n: engram.tensors[n][0])
+    if not names or len(names) != len(engram.tensors):
+        raise ValueError('%s: unexpected tensors in the Engram file' % engram.path)
+    first_off, dims, gt, _ = engram.tensors[names[0]]
+    end, rows = first_off, 0
+    for name in names:
+        off, d, g, n = engram.tensors[name]
+        if off != end or g != gt or d[0] != dims[0]:
+            raise ValueError('%s: the Engram tables are not contiguous' % engram.path)
+        end, rows = off + n, rows + d[1]
+    layers = [int(re.match(r'blk\.(\d+)\.', n).group(1)) for n in names]
+    return Tensor(TABLE, [dims[0], rows], gt, engram.path, first_off), layers
+
+
+def plan_splits(tensors, max_size):
+    """llama-gguf-split --split-max-size with --no-tensor-first-split: the first file holds only metadata."""
+    splits, cur = [[], []], 0
+    for t in tensors:
+        n = align(t.nbytes)
+        if cur + n > max_size:
+            if not splits[-1]:
+                raise ValueError('tensor %s is larger than the split size and comes first' % t.name)
+            splits.append([])
+            cur = n
+        else:
+            cur += n
+        splits[-1].append(t)
+    return splits
+
+
+def split_kv(no, count, n_tensors):
+    return [['split.no', T_U16, encode_value(T_U16, no)], ['split.count', T_U16, encode_value(T_U16, count)],
+            ['split.tensors.count', T_I32, encode_value(T_I32, n_tensors)]]
+
+
+def copy_range(src_path, offset, n, fout, buf):
+    """Plain reads and writes through buf, as on every file system (os.copy_file_range leaves the copy to the
+    file system's own implementation)."""
+    with open(src_path, 'rb') as fin:
+        fin.seek(offset)
+        while n:
+            got = fin.readinto(buf[:min(n, len(buf))])
+            if not got:
+                raise ValueError('%s ends early' % src_path)
+            fout.write(buf[:got])
+            n -= got
+
+
+def write_gguf(path, kv, tensors):
+    infos, off = [], 0
+    for t in tensors:
+        infos.append((t.name, t.dims, t.type, off))
+        off += align(t.nbytes)
+    tmp = path + '.partial'
+    buf = memoryview(bytearray(COPY_CHUNK))
+    with open(tmp, 'wb') as f:
+        f.write(serialize_header(kv, infos))
+        for t in tensors:
+            if t.data is not None:
+                f.write(t.data)
+            else:
+                copy_range(t.path, t.offset, t.nbytes, f, buf)
+            f.write(b'\0' * (align(t.nbytes) - t.nbytes))
+        f.flush()
+        os.fsync(f.fileno())
+    os.replace(tmp, path)
+
+
+def verify_gguf(path, kv, tensors, expected=None, origin=()):
+    """Re-read the written file: same metadata, tensor table and tensor bytes as planned, and the SHA-256 of the
+    published file (expected; None for a file this repository does not publish). origin: the old files it was made
+    from, for the error message."""
+    g = GGUFFile(path)
+    if g.kv != kv or g.order != [t.name for t in tensors]:
+        raise ValueError('%s: header differs from the plan' % path)
+    if g.size != g.data_start + sum(align(t.nbytes) for t in tensors):
+        raise ValueError('%s: unexpected file size' % path)
+    whole = hashlib.sha256()
+    with open(path, 'rb') as f:
+        read_into(f, g.data_start, whole)
+        for t in tensors:
+            off, dims, gt, n = g.tensors[t.name]
+            if dims != t.dims or gt != t.type or off != f.tell():
+                raise ValueError('%s: tensor %s differs from the plan' % (path, t.name))
+            part = hashlib.sha256()
+            read_into(f, n, whole, part)
+            if part.hexdigest() != t.source_hash():
+                raise ValueError('%s: tensor %s differs from its source%s. The copy or a read failed on this machine, '
+                                 'or the source changed during the conversion; a damaged download would have been '
+                                 'copied as it is' % (path, t.name, '' if t.path is None else ' (%s at %d)'
+                                                      % (t.path, t.offset)))
+            read_into(f, align(n) - n, whole)
+    if expected is not None and whole.hexdigest() != expected:
+        raise ValueError('%s: every tensor equals its source, but the file is not the published one (SHA-256), so the '
+                         'old files it was made from are not the published ones: %s. A damaged download, or a '
+                         'patch_chat_template.py run that did not finish, gives this; download the new files instead'
+                         % (path, ', '.join(os.path.basename(p) for p in origin)))
+
+
+def plan_model(meta, sources, engram, chat_template, bias_vl):
+    """Metadata, all tensors (the Engram table last) and the backbone split of one build."""
+    kv = convert_metadata(meta, lambda t: next(s for s in sources if t in s.tensors).read_tensor(t), chat_template)
+    tensors = plan_tensors(meta, sources, engram, bias_vl)
+    splits = plan_splits(tensors[:-1], SPLIT_MAX_SIZE)
+    if tensors[-1].name != TABLE or len(splits) != N_SPLIT - 1:
+        raise ValueError('the backbone needs %d shards; the shared Engram file is shard %d of %d'
+                         % (len(splits), N_SPLIT, N_SPLIT))
+    return kv, tensors, splits
+
+
+def convert_model(meta, sources, engram, chat_template, bias_vl, name, out_dir, verify, log):
+    """Writes shards 1 to N_SPLIT - 1 of a build; shard N_SPLIT is the shared Engram file."""
+    kv, tensors, splits = plan_model(meta, sources, engram, chat_template, bias_vl)
+    names = ['%s-%05d-of-%05d.gguf' % (name, i + 1, N_SPLIT) for i in range(len(splits))]
+    if verify and not all(n in PUBLISHED for n in names):
+        log('  %s is not a build of this repository: each tensor is checked against its source, but the files cannot '
+            'be checked against published ones' % name)
+    paths = []
+    for i, part in enumerate(splits):
+        path = os.path.join(out_dir, names[i])
+        part_kv = (kv if i == 0 else []) + split_kv(i, N_SPLIT, len(tensors))
+        log('  %s: %d tensors, %.2f GB' % (names[i], len(part), sum(t.nbytes for t in part) / 1e9))
+        write_gguf(path, part_kv, part)
+        if verify:
+            origin = [meta.path] if i == 0 else sorted({t.path for t in part if t.path is not None})
+            verify_gguf(path, part_kv, part, PUBLISHED.get(names[i]), origin)
+        paths.append(path)
+    return paths, tensors
+
+
+def engram_header(table, source, n_tensors):
+    kv = [['general.name', T_STRING, encode_value(T_STRING, 'DeepSeek-V4.1-Flash-1017 Engram tables')],
+          ['general.description', T_STRING, encode_value(T_STRING, 'Shard 6 of 6 of every GSQ-RCO build')],
+          [ARCH + '.engram.source', T_STRING, encode_value(T_STRING, source)]]
+    kv += split_kv(N_SPLIT - 1, N_SPLIT, n_tensors)
+    return kv, serialize_header(kv, [(TABLE, table.dims, table.type, 0)])
+
+
+def convert_engram(engram, table, n_tensors, verify, log):
+    """Rewrite the old Engram file's header in place. The new header ends where the old one did, so the 110 GB of
+    table data stay where they are; the old header is kept in a side file until the new one is checked. With verify,
+    the tables are first checked against the published file: its SHA-256 is that of the new header and the tables."""
+    path = engram.path
+    backup = path + '.old-header'
+    kv, header = engram_header(table, engram.get(ARCH + '.engram.source'), n_tensors)
+    if TABLE in engram.tensors:
+        if engram.kv != kv or engram.data_start != len(header):
+            raise ValueError('%s: unexpected header' % path)
+        if os.path.exists(backup):
+            os.remove(backup)
+        return False
+    if len(header) != engram.data_start or table.offset != engram.data_start:
+        raise ValueError('%s: the new header does not fit the old one' % path)
+    if verify:
+        log('%s: checking the tables against the published file (%.0f GB)'
+            % (os.path.basename(path), engram.size / 1e9))
+        h = hashlib.sha256(header)
+        with open(path, 'rb') as f:
+            f.seek(engram.data_start)
+            read_into(f, engram.size - engram.data_start, h)
+        if h.hexdigest() != PUBLISHED[ENGRAM_FILE]:
+            raise ValueError('%s: its tables are not those of the published file (SHA-256), e.g. from a damaged '
+                             'download; the file was left unchanged. Download %s again' % (path, ENGRAM_FILE))
+    with open(path, 'rb') as f:
+        old = f.read(engram.data_start)
+    with open(backup, 'wb') as f:
+        f.write(old)
+        os.fsync(f.fileno())
+    with open(path, 'r+b') as f:
+        f.write(header)
+        f.flush()
+        os.fsync(f.fileno())
+    g = GGUFFile(path)
+    if g.kv != kv or g.order != [TABLE] or g.tensors[TABLE][:3] != (table.offset, table.dims, table.type) or \
+            g.size != engram.size:
+        with open(path, 'r+b') as f:
+            f.write(old)
+        raise ValueError('%s: the rewritten header did not check out; the old one was restored' % path)
+    os.remove(backup)
+    log('%s: header rewritten in place (%d bytes), the table data are unchanged' % (os.path.basename(path), len(old)))
+    return True
+
+
+def link_engram(engram_path, name, out_dir, log):
+    """llama.cpp opens <name>-0000N-of-0000N.gguf as the last shard: a link to the shared Engram file."""
+    path = os.path.join(out_dir, '%s-%05d-of-%05d.gguf' % (name, N_SPLIT, N_SPLIT))
+    if os.path.lexists(path):
+        if os.path.exists(path) and os.path.samefile(path, engram_path):
+            return
+        raise ValueError('%s exists and is not the Engram file' % path)
+    try:
+        os.symlink(os.path.relpath(engram_path, out_dir), path)
+        kind = 'link'
+    except (OSError, NotImplementedError):
+        os.link(engram_path, path)  # e.g. Windows without symlink rights
+        kind = 'hard link'
+    log('  %s: %s to %s' % (os.path.basename(path), kind, os.path.basename(engram_path)))
+
+
+def convert_mmproj(src, out_path, name, log):
+    """name: the file name the result gets (out_path may be a temporary name)."""
+    if src.get('clip.projector_type') != 'deepseek4v':
+        raise ValueError('%s: projector type %r' % (src.path, src.get('clip.projector_type')))
+    kv = []
+    for key, t, raw in src.kv:
+        if key == 'clip.projector_type':
+            raw = encode_value(T_STRING, 'deepseek41v')
+        elif key == 'clip.vision.image_max_pixels':
+            continue  # deepseek41v fixes the image budget at 1024 tokens; the upstream converter writes no key
+        kv.append([key, t, raw])
+    tensors = [Tensor(n, src.tensors[n][1], src.tensors[n][2], src.path, src.tensors[n][0]) for n in src.order]
+    log('  %s: projector type deepseek41v' % name)
+    if name not in PUBLISHED:
+        log('  %s is not a file of this repository: each tensor is checked against its source, but the file cannot '
+            'be checked against a published one' % name)
+    write_gguf(out_path, kv, tensors)
+    verify_gguf(out_path, kv, tensors, PUBLISHED.get(name), [src.path])
+
+
+# ---------------------------------------------------------------------------------------------------------------
+# discovery
+
+
+def find_sets(paths):
+    files = []
+    for p in paths:
+        if os.path.isdir(p):
+            files += [os.path.join(p, f) for f in sorted(os.listdir(p)) if f.endswith('.gguf')]
+        else:
+            files.append(p)
+    models, shards, engrams, mmprojs, converted, firsts = [], {}, {}, [], [], set()
+    for path in files:
+        for p in {path, os.path.realpath(path)}:
+            journal = os.path.join(os.path.dirname(p), '.%s.chat-template-recovery' % os.path.basename(p))
+            if os.path.isdir(journal):
+                raise ValueError('%s: a patch_chat_template.py run on this file did not finish (%s exists), so its '
+                                 'tensor data are partly moved. Finish it by running the same command again (the '
+                                 'script is in revision %s of the repository), or download the new files'
+                                 % (path, journal, OLD_REVISION))
+        g = GGUFFile(path)
+        arch = g.get('general.architecture')
+        m = SPLIT_RE.match(os.path.basename(path))
+        if arch == ARCH and m:
+            firsts.add((os.path.dirname(path), m.group('prefix')))
+        if arch == ARCH:
+            if g.get(ARCH + '.attention.kv_source_layers') is not None:
+                if m:  # already in the upstream layout; it may still lack the link to the Engram file
+                    converted.append((os.path.dirname(os.path.abspath(path)), m.group('prefix'), g))
+                continue
+            models.append(g)
+        elif arch == ARCH + '-engram' or TABLE in g.tensors:
+            engrams.setdefault(os.path.realpath(path), g)  # links to the Engram file count once
+        elif g.get('clip.projector_type') == 'deepseek4v':
+            mmprojs.append(g)
+        elif m and g.get('split.no') is not None:
+            shards[(os.path.dirname(path), m.group('prefix'), int(m.group('no')))] = g
+    sets = []
+    for g in models:
+        m = SPLIT_RE.match(os.path.basename(g.path))
+        if m is None:
+            sets.append((g, [g], os.path.basename(g.path)[:-5]))
+            continue
+        prefix, count = m.group('prefix'), g.get('split.count')
+        parts = [g] + [shards.get((os.path.dirname(g.path), prefix, i)) for i in range(2, count + 1)]
+        if None in parts:
+            raise ValueError('%s: not all %d shards are present' % (g.path, count))
+        sets.append((g, parts, prefix))
+    orphans = sorted(g.path for (d, prefix, _), g in shards.items()
+                     if (d, prefix) not in firsts and prefix.startswith(BUILD_PREFIX))
+    if orphans:
+        raise ValueError('the first shard of these files is missing, so their build cannot be converted: %s'
+                         % ', '.join(orphans))
+    return sets, list(engrams.values()), mmprojs, converted
+
+
+def repo_file(name, path, inputs, sha256, log):
+    """A small file of the repository with the given SHA-256: the given path, else a copy next to this script or in
+    the model folder (other versions, such as an older download, are skipped), else downloaded from the repository."""
+    if path is not None:
+        candidates = [path]
+    else:
+        folders = [os.path.dirname(os.path.abspath(__file__))]
+        folders += [p if os.path.isdir(p) else os.path.dirname(os.path.abspath(p)) for p in inputs]
+        candidates = list(dict.fromkeys(f for f in (os.path.join(d, name) for d in folders) if os.path.isfile(f)))
+    for f in candidates:
+        with open(f, 'rb') as fin:
+            data = fin.read()
+        if hashlib.sha256(data).hexdigest() == sha256:
+            return data
+        if path is not None:
+            raise ValueError('%s is not the current %s of %s (checksum)' % (path, name, REPO))
+        log('%s: not the current version (checksum), skipped' % f)
+    url = 'https://huggingface.co/%s/resolve/main/%s' % (REPO, name)
+    try:
+        with urllib.request.urlopen(url, timeout=120) as r:
+            data = r.read()
+    except OSError as e:
+        raise ValueError('%s could not be downloaded from %s (%s); put it next to this script' % (name, url, e))
+    if hashlib.sha256(data).hexdigest() != sha256:
+        raise ValueError('%s of %s has changed since this script was written; download the current '
+                         'convert_ds41_upstream.py' % (name, REPO))
+    log('%s: downloaded from %s' % (name, REPO))
+    return data
+
+
+def read_inputs(args, log):
+    """chat_template.jinja (embedded in the new first shards) and exp_probs_b_vl.f32 (the vision router bias)."""
+    template = repo_file('chat_template.jinja', args.chat_template, args.inputs, TEMPLATE_SHA256, log).decode('utf-8')
+    bias_vl = repo_file('exp_probs_b_vl.f32', args.vision_bias, args.inputs, BIAS_VL_SHA256, log)
+    return template, bias_vl
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.split('\n')[0])
+    ap.add_argument('inputs', nargs='+', help='model folder(s) or GGUF files')
+    ap.add_argument('--out', help='folder for the new shards (default: next to the old ones)')
+    ap.add_argument('--engram', help='the Engram GGUF (default: the one found among the inputs)')
+    ap.add_argument('--chat-template', help='chat_template.jinja of the repository (default: next to this script or '
+                    'in the model folder)')
+    ap.add_argument('--vision-bias', help='exp_probs_b_vl.f32 of the repository (default: as --chat-template)')
+    ap.add_argument('--metadata', help=argparse.SUPPRESS)  # header of another first shard (release builds)
+    ap.add_argument('--name', help='output name prefix for a single unsplit input')
+    ap.add_argument('--no-verify', action='store_true', help='skip re-reading and hashing the written shards')
+    ap.add_argument('--delete-old', action='store_true', help='delete the old shards after a verified conversion')
+    ap.add_argument('--dry-run', action='store_true', help='only print what would be done')
+    args = ap.parse_args()
+    log = lambda s: print(s, flush=True)
+    if args.no_verify and not args.dry_run:
+        log('warning: --no-verify: the written files are not read back, nor checked against the published ones')
+
+    sets, engrams, mmprojs, converted = find_sets(args.inputs)
+    if args.engram:
+        engrams = [GGUFFile(args.engram)]
+    if (sets or converted) and len(engrams) != 1:
+        raise SystemExit('expected one Engram file (%s), found %d; pass it with --engram' % (ENGRAM_FILE, len(engrams)))
+    if not sets and not mmprojs and not converted:
+        raise SystemExit('nothing to convert (no DeepSeek-V4.1 GSQ-RCO files found)')
+    engram = engrams[0] if engrams else None
+    chat_template = bias_vl = None
+    if sets:
+        chat_template, bias_vl = read_inputs(args, log)
+    written, old = [], []
+    builds = [(d, prefix) for d, prefix, _ in converted]
+    n_tensors = {g.get('split.tensors.count') for _, _, g in converted}
+    for first, parts, prefix in sets:
+        meta = GGUFFile(args.metadata) if args.metadata else first
+        name = args.name or prefix
+        out_dir = os.path.abspath(args.out or os.path.dirname(os.path.abspath(first.path)))
+        log('%s -> %s' % (prefix, os.path.join(out_dir, name + '-*.gguf')))
+        if args.dry_run:
+            _, tensors, splits = plan_model(meta, parts, engram, chat_template, bias_vl)
+            for i, part in enumerate(splits):
+                log('  shard %d: %d tensors, %.2f GB' % (i + 1, len(part), sum(t.nbytes for t in part) / 1e9))
+            log('  shard %d: %s' % (N_SPLIT, os.path.basename(engram.path)))
+            n_tensors.add(len(tensors))
+            continue
+        os.makedirs(out_dir, exist_ok=True)
+        paths, tensors = convert_model(meta, parts, engram, chat_template, bias_vl, name, out_dir, not args.no_verify, log)
+        written += paths
+        builds.append((out_dir, name))
+        n_tensors.add(len(tensors))
+        if args.delete_old:
+            for p in parts:
+                os.remove(p.path)
+            log('  deleted the %d old shards' % len(parts))
+        else:
+            old += [p.path for p in parts]
+
+    if engram is not None and (builds or sets):
+        if len(n_tensors) != 1:
+            raise ValueError('the builds differ in tensor count, so they cannot share one Engram file')
+        table, _ = engram_table(engram)
+        if args.dry_run:
+            if TABLE not in engram.tensors:
+                log('%s: header rewritten in place, the table data stay' % os.path.basename(engram.path))
+        else:
+            convert_engram(engram, table, n_tensors.pop(), not args.no_verify, log)
+            for out_dir, name in builds:
+                link_engram(engram.path, name, out_dir, log)
+    for src in mmprojs:
+        out_dir = args.out or os.path.dirname(os.path.abspath(src.path))
+        out_path = os.path.join(out_dir, os.path.basename(src.path))
+        if args.dry_run:
+            log('%s: projector type deepseek4v -> deepseek41v' % src.path)
+            continue
+        if os.path.abspath(out_path) == os.path.abspath(src.path):
+            convert_mmproj(src, out_path + '.new', os.path.basename(out_path), log)
+            os.replace(out_path + '.new', out_path)
+        else:
+            convert_mmproj(src, out_path, os.path.basename(out_path), log)
+    if args.dry_run:
+        log('dry run: only the headers were read; the tensor data are read and checked during the conversion')
+    elif old:
+        log('done. llama.cpp no longer needs the old shards:\n  ' + '\n  '.join(old))
+    else:
+        log('done.')
+
+
+# exp_probs_b_vl.f32: blk.0..39.exp_probs_b_vl = layers.N.ffn.gate.bias_vl of deepseek-ai/DeepSeek-V4.1-Flash-1017
+# (revision 1bae0e1088993b9b82de08f59de5be5a34e42170), 40 x 384 little-endian F32
+BIAS_VL_SHA256 = '710b11267b4db392d1fdb558f36ba8e0e91a91f4e8180c953c912e18eacf843d'
+# chat_template.jinja of the repository since October 25, 2026 (commit 1c100b9), as in the published first shards
+TEMPLATE_SHA256 = '4d61d22c90ea0cdf15c7e298f7215f23eeffd20f8eed8fb517fda5fb64693d99'
+# SHA-256 of the files this script writes, as published in the repository since October 27, 2026 (commit a419d74).
+# Every published version of the old files converts to these.
+PUBLISHED = {
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf':
+        '3e8ddd11441edd4bc19bacaef5dbe38818af2cbdf0820fb70f1d2d851fcc192a',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00002-of-00006.gguf':
+        'be4ceb1caf1bed28e886860cf6a52adc8f5ca9eb0d2098ade39bb4179d48b323',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00003-of-00006.gguf':
+        '758c66ba65236e8ce0820c21a7b6f9b2947681df8cc509f672a77722e8191b9a',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00004-of-00006.gguf':
+        '04a5bea2b38eb576f49efc19c0a677768d6e2a8ea5772d72fb6cb79e1733a212',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00005-of-00006.gguf':
+        '7c8dbee99890f05dbc420eba6f44393f5228d06248fe61e1ae8ed002449c7549',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00006.gguf':
+        '0182a2a1f42d319a0dbc6e34aa82a063880c069fe3aafbbea86018f386e64e5f',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00006.gguf':
+        '5c88039fa640460a9fc3465a6dfcd162a4083dd738d55e5b52e025be14555a08',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00006.gguf':
+        '0fddebfd05f0a36cfd107ba5dfd2547e64df61adbab39560a5a5f5521ca099a5',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00006.gguf':
+        'c577a43c6e6e615838a1655a0b162541edeac6f36542f08c4a9593a70dcaf500',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00005-of-00006.gguf':
+        '851d8605bf504962054d17ea8a0d9e2e0a228d085d396bd865a4de9f16d90fbd',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf':
+        'f356f198561f56df36f7ab92d7238a926646d76de992c34f15bcb939a5f529b9',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00002-of-00006.gguf':
+        '25af3785ccf925f54137e1ead4c481a8952315dc0b1af6519a99407f67bce8f3',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00003-of-00006.gguf':
+        '8168d0998fc494def2e1952d83c131425a108660aae9d29190f5a8670ab62b48',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00006.gguf':
+        '9eb9b4de1968f150dfc207eb31c27be9cd9b5666a9a6246bde8a22b184b4bee0',
+    'DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00005-of-00006.gguf':
+        '7e35f6b4e6fc46f6c4b152b593d84d683c5ece695ae6ad734f556198c041bec1',
+    'DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf':
+        '3dac4c54893ad7aa89103c0028500eadc78e74112cef3a2930a97e67dc9fb4f3',
+    'mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf':
+        '0fdc4b7693f287e9665940cf42bd39c546ffe4f0def89c569d7d0e45d60d2df1',
+}
+
+if __name__ == '__main__':
+    try:
+        main()
+    except (ValueError, OSError) as e:
+        sys.exit('error: %s' % e)

mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fdc4b7693f287e9665940cf42bd39c546ffe4f0def89c569d7d0e45d60d2df1
+size 972208096

tensor-allocation/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M.rco-allocation.txt

@@ -0,0 +1,1061 @@
+# RCO per-tensor GGUF allocation
+# Repository: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
+# GGUF: DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf (first of 6 shards; this dump covers all of them)
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00002-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00003-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00004-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00005-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf
+# Target bpw over the backbone (all tensors except the Engram table): 1.75 (measured: 1.7499)
+# The Engram table (per_layer_token_embd, shard 6 = the shared Engram file) is IQ4_NL in every build and not part of the search.
+# GGUF version: 3
+# GGUF metadata entries: 77
+# Tensor count: 1043
+# Quant-type counts: BF16=67, F32=531, IQ1_M=15, IQ1_S=83, IQ2_S=2, IQ2_XXS=20, IQ4_NL=1, Q5_K=244, Q8_0=80
+# Format: <GGUF tensor name>: <GGML quantization/storage type>
+# Includes fixed non-RCO tensors (e.g. F32/BF16) so the released GGUF is fully reproducible.
+#
+output.weight: Q5_K
+output_norm.weight: F32
+per_layer_token_embd.weight: IQ4_NL
+token_embd.weight: Q5_K
+blk.0.attn_kv.weight: Q8_0
+blk.0.attn_kv_a_norm.weight: F32
+blk.0.attn_norm.weight: F32
+blk.0.attn_output_a.weight: Q5_K
+blk.0.attn_output_b.weight: Q5_K
+blk.0.attn_q_a.weight: Q8_0
+blk.0.attn_q_a_norm.weight: F32
+blk.0.attn_q_b.weight: Q5_K
+blk.0.attn_sinks.weight: F32
+blk.0.exp_probs_b.bias: F32
+blk.0.exp_probs_b_vl.bias: F32
+blk.0.ffn_down_exps.weight: IQ1_S
+blk.0.ffn_down_shexp.weight: Q5_K
+blk.0.ffn_gate_exps.weight: IQ1_S
+blk.0.ffn_gate_inp.weight: BF16
+blk.0.ffn_gate_shexp.weight: Q5_K
+blk.0.ffn_norm.weight: F32
+blk.0.ffn_up_exps.weight: IQ1_S
+blk.0.ffn_up_shexp.weight: Q5_K
+blk.0.hc_attn_base.weight: F32
+blk.0.hc_attn_fn.weight: F32
+blk.0.hc_attn_scale.weight: F32
+blk.0.hc_ffn_base.weight: F32
+blk.0.hc_ffn_fn.weight: F32
+blk.0.hc_ffn_scale.weight: F32
+blk.1.attn_kv.weight: Q8_0
+blk.1.attn_kv_a_norm.weight: F32
+blk.1.attn_norm.weight: F32
+blk.1.attn_output_a.weight: Q5_K
+blk.1.attn_output_b.weight: Q5_K
+blk.1.attn_q_a.weight: Q8_0
+blk.1.attn_q_a_norm.weight: F32
+blk.1.attn_q_b.weight: Q5_K
+blk.1.attn_sinks.weight: F32
+blk.1.engram_gate.weight: F32
+blk.1.engram_kv.weight: Q5_K
+blk.1.exp_probs_b.bias: F32
+blk.1.exp_probs_b_vl.bias: F32
+blk.1.ffn_down_exps.weight: IQ1_S
+blk.1.ffn_down_shexp.weight: Q5_K
+blk.1.ffn_gate_exps.weight: IQ1_S
+blk.1.ffn_gate_inp.weight: BF16
+blk.1.ffn_gate_shexp.weight: Q5_K
+blk.1.ffn_norm.weight: F32
+blk.1.ffn_up_exps.weight: IQ1_S
+blk.1.ffn_up_shexp.weight: Q5_K
+blk.1.hc_attn_base.weight: F32
+blk.1.hc_attn_fn.weight: F32
+blk.1.hc_attn_scale.weight: F32
+blk.1.hc_ffn_base.weight: F32
+blk.1.hc_ffn_fn.weight: F32
+blk.1.hc_ffn_scale.weight: F32
+blk.2.attn_compressor_gate.weight: BF16
+blk.2.attn_compressor_kv.weight: BF16
+blk.2.attn_compressor_norm.weight: F32
+blk.2.attn_kv.weight: Q8_0
+blk.2.attn_kv_a_norm.weight: F32
+blk.2.attn_norm.weight: F32
+blk.2.attn_output_a.weight: Q5_K
+blk.2.attn_output_b.weight: Q5_K
+blk.2.attn_q_a.weight: Q8_0
+blk.2.attn_q_a_norm.weight: F32
+blk.2.attn_q_b.weight: Q5_K
+blk.2.attn_sinks.weight: F32
+blk.2.exp_probs_b.bias: F32
+blk.2.exp_probs_b_vl.bias: F32
+blk.2.ffn_down_exps.weight: IQ1_S
+blk.2.ffn_down_shexp.weight: Q5_K
+blk.2.ffn_gate_exps.weight: IQ1_S
+blk.2.ffn_gate_inp.weight: BF16
+blk.2.ffn_gate_shexp.weight: Q5_K
+blk.2.ffn_norm.weight: F32
+blk.2.ffn_up_exps.weight: IQ1_S
+blk.2.ffn_up_shexp.weight: Q5_K
+blk.2.hc_attn_base.weight: F32
+blk.2.hc_attn_fn.weight: F32
+blk.2.hc_attn_scale.weight: F32
+blk.2.hc_ffn_base.weight: F32
+blk.2.hc_ffn_fn.weight: F32
+blk.2.hc_ffn_scale.weight: F32
+blk.2.indexer.attn_k.weight: BF16
+blk.2.indexer.attn_q_b.weight: BF16
+blk.2.indexer.k_norm.weight: F32
+blk.2.indexer.proj.weight: BF16
+blk.3.attn_kv.weight: Q8_0
+blk.3.attn_kv_a_norm.weight: F32
+blk.3.attn_norm.weight: F32
+blk.3.attn_output_a.weight: Q5_K
+blk.3.attn_output_b.weight: Q5_K
+blk.3.attn_q_a.weight: Q8_0
+blk.3.attn_q_a_norm.weight: F32
+blk.3.attn_q_b.weight: Q5_K
+blk.3.attn_sinks.weight: F32
+blk.3.exp_probs_b.bias: F32
+blk.3.exp_probs_b_vl.bias: F32
+blk.3.ffn_down_exps.weight: IQ1_S
+blk.3.ffn_down_shexp.weight: Q5_K
+blk.3.ffn_gate_exps.weight: IQ1_S
+blk.3.ffn_gate_inp.weight: BF16
+blk.3.ffn_gate_shexp.weight: Q5_K
+blk.3.ffn_norm.weight: F32
+blk.3.ffn_up_exps.weight: IQ1_S
+blk.3.ffn_up_shexp.weight: Q5_K
+blk.3.hc_attn_base.weight: F32
+blk.3.hc_attn_fn.weight: F32
+blk.3.hc_attn_scale.weight: F32
+blk.3.hc_ffn_base.weight: F32
+blk.3.hc_ffn_fn.weight: F32
+blk.3.hc_ffn_scale.weight: F32
+blk.4.attn_kv.weight: Q8_0
+blk.4.attn_kv_a_norm.weight: F32
+blk.4.attn_norm.weight: F32
+blk.4.attn_output_a.weight: Q5_K
+blk.4.attn_output_b.weight: Q5_K
+blk.4.attn_q_a.weight: Q8_0
+blk.4.attn_q_a_norm.weight: F32
+blk.4.attn_q_b.weight: Q5_K
+blk.4.attn_sinks.weight: F32
+blk.4.exp_probs_b.bias: F32
+blk.4.exp_probs_b_vl.bias: F32
+blk.4.ffn_down_exps.weight: IQ1_S
+blk.4.ffn_down_shexp.weight: Q5_K
+blk.4.ffn_gate_exps.weight: IQ1_S
+blk.4.ffn_gate_inp.weight: BF16
+blk.4.ffn_gate_shexp.weight: Q5_K
+blk.4.ffn_norm.weight: F32
+blk.4.ffn_up_exps.weight: IQ1_S
+blk.4.ffn_up_shexp.weight: Q5_K
+blk.4.hc_attn_base.weight: F32
+blk.4.hc_attn_fn.weight: F32
+blk.4.hc_attn_scale.weight: F32
+blk.4.hc_ffn_base.weight: F32
+blk.4.hc_ffn_fn.weight: F32
+blk.4.hc_ffn_scale.weight: F32
+blk.5.attn_kv.weight: Q8_0
+blk.5.attn_kv_a_norm.weight: F32
+blk.5.attn_norm.weight: F32
+blk.5.attn_output_a.weight: Q5_K
+blk.5.attn_output_b.weight: Q5_K
+blk.5.attn_q_a.weight: Q8_0
+blk.5.attn_q_a_norm.weight: F32
+blk.5.attn_q_b.weight: Q5_K
+blk.5.attn_sinks.weight: F32
+blk.5.exp_probs_b.bias: F32
+blk.5.exp_probs_b_vl.bias: F32
+blk.5.ffn_down_exps.weight: IQ1_S
+blk.5.ffn_down_shexp.weight: Q5_K
+blk.5.ffn_gate_exps.weight: IQ1_S
+blk.5.ffn_gate_inp.weight: BF16
+blk.5.ffn_gate_shexp.weight: Q5_K
+blk.5.ffn_norm.weight: F32
+blk.5.ffn_up_exps.weight: IQ1_S
+blk.5.ffn_up_shexp.weight: Q5_K
+blk.5.hc_attn_base.weight: F32
+blk.5.hc_attn_fn.weight: F32
+blk.5.hc_attn_scale.weight: F32
+blk.5.hc_ffn_base.weight: F32
+blk.5.hc_ffn_fn.weight: F32
+blk.5.hc_ffn_scale.weight: F32
+blk.6.attn_kv.weight: Q8_0
+blk.6.attn_kv_a_norm.weight: F32
+blk.6.attn_norm.weight: F32
+blk.6.attn_output_a.weight: Q5_K
+blk.6.attn_output_b.weight: Q5_K
+blk.6.attn_q_a.weight: Q8_0
+blk.6.attn_q_a_norm.weight: F32
+blk.6.attn_q_b.weight: Q5_K
+blk.6.attn_sinks.weight: F32
+blk.6.exp_probs_b.bias: F32
+blk.6.exp_probs_b_vl.bias: F32
+blk.6.ffn_down_exps.weight: IQ1_S
+blk.6.ffn_down_shexp.weight: Q5_K
+blk.6.ffn_gate_exps.weight: IQ1_S
+blk.6.ffn_gate_inp.weight: BF16
+blk.6.ffn_gate_shexp.weight: Q5_K
+blk.6.ffn_norm.weight: F32
+blk.6.ffn_up_exps.weight: IQ1_S
+blk.6.ffn_up_shexp.weight: Q5_K
+blk.6.hc_attn_base.weight: F32
+blk.6.hc_attn_fn.weight: F32
+blk.6.hc_attn_scale.weight: F32
+blk.6.hc_ffn_base.weight: F32
+blk.6.hc_ffn_fn.weight: F32
+blk.6.hc_ffn_scale.weight: F32
+blk.7.attn_kv.weight: Q8_0
+blk.7.attn_kv_a_norm.weight: F32
+blk.7.attn_norm.weight: F32
+blk.7.attn_output_a.weight: Q5_K
+blk.7.attn_output_b.weight: Q5_K
+blk.7.attn_q_a.weight: Q8_0
+blk.7.attn_q_a_norm.weight: F32
+blk.7.attn_q_b.weight: Q5_K
+blk.7.attn_sinks.weight: F32
+blk.7.exp_probs_b.bias: F32
+blk.7.exp_probs_b_vl.bias: F32
+blk.7.ffn_down_exps.weight: IQ2_S
+blk.7.ffn_down_shexp.weight: Q5_K
+blk.7.ffn_gate_exps.weight: IQ1_S
+blk.7.ffn_gate_inp.weight: BF16
+blk.7.ffn_gate_shexp.weight: Q5_K
+blk.7.ffn_norm.weight: F32
+blk.7.ffn_up_exps.weight: IQ1_S
+blk.7.ffn_up_shexp.weight: Q5_K
+blk.7.hc_attn_base.weight: F32
+blk.7.hc_attn_fn.weight: F32
+blk.7.hc_attn_scale.weight: F32
+blk.7.hc_ffn_base.weight: F32
+blk.7.hc_ffn_fn.weight: F32
+blk.7.hc_ffn_scale.weight: F32
+blk.8.attn_compressor_gate.weight: BF16
+blk.8.attn_compressor_kv.weight: BF16
+blk.8.attn_compressor_norm.weight: F32
+blk.8.attn_kv.weight: Q8_0
+blk.8.attn_kv_a_norm.weight: F32
+blk.8.attn_norm.weight: F32
+blk.8.attn_output_a.weight: Q5_K
+blk.8.attn_output_b.weight: Q5_K
+blk.8.attn_q_a.weight: Q8_0
+blk.8.attn_q_a_norm.weight: F32
+blk.8.attn_q_b.weight: Q5_K
+blk.8.attn_sinks.weight: F32
+blk.8.exp_probs_b.bias: F32
+blk.8.exp_probs_b_vl.bias: F32
+blk.8.ffn_down_exps.weight: IQ1_S
+blk.8.ffn_down_shexp.weight: Q5_K
+blk.8.ffn_gate_exps.weight: IQ1_S
+blk.8.ffn_gate_inp.weight: BF16
+blk.8.ffn_gate_shexp.weight: Q5_K
+blk.8.ffn_norm.weight: F32
+blk.8.ffn_up_exps.weight: IQ1_S
+blk.8.ffn_up_shexp.weight: Q5_K
+blk.8.hc_attn_base.weight: F32
+blk.8.hc_attn_fn.weight: F32
+blk.8.hc_attn_scale.weight: F32
+blk.8.hc_ffn_base.weight: F32
+blk.8.hc_ffn_fn.weight: F32
+blk.8.hc_ffn_scale.weight: F32
+blk.8.indexer.attn_k.weight: BF16
+blk.8.indexer.attn_q_b.weight: BF16
+blk.8.indexer.k_norm.weight: F32
+blk.8.indexer.proj.weight: BF16
+blk.9.attn_kv.weight: Q8_0
+blk.9.attn_kv_a_norm.weight: F32
+blk.9.attn_norm.weight: F32
+blk.9.attn_output_a.weight: Q5_K
+blk.9.attn_output_b.weight: Q5_K
+blk.9.attn_q_a.weight: Q8_0
+blk.9.attn_q_a_norm.weight: F32
+blk.9.attn_q_b.weight: Q5_K
+blk.9.attn_sinks.weight: F32
+blk.9.exp_probs_b.bias: F32
+blk.9.exp_probs_b_vl.bias: F32
+blk.9.ffn_down_exps.weight: IQ1_S
+blk.9.ffn_down_shexp.weight: Q5_K
+blk.9.ffn_gate_exps.weight: IQ2_XXS
+blk.9.ffn_gate_inp.weight: BF16
+blk.9.ffn_gate_shexp.weight: Q5_K
+blk.9.ffn_norm.weight: F32
+blk.9.ffn_up_exps.weight: IQ2_XXS
+blk.9.ffn_up_shexp.weight: Q5_K
+blk.9.hc_attn_base.weight: F32
+blk.9.hc_attn_fn.weight: F32
+blk.9.hc_attn_scale.weight: F32
+blk.9.hc_ffn_base.weight: F32
+blk.9.hc_ffn_fn.weight: F32
+blk.9.hc_ffn_scale.weight: F32
+blk.10.attn_kv.weight: Q8_0
+blk.10.attn_kv_a_norm.weight: F32
+blk.10.attn_norm.weight: F32
+blk.10.attn_output_a.weight: Q5_K
+blk.10.attn_output_b.weight: Q5_K
+blk.10.attn_q_a.weight: Q8_0
+blk.10.attn_q_a_norm.weight: F32
+blk.10.attn_q_b.weight: Q5_K
+blk.10.attn_sinks.weight: F32
+blk.10.exp_probs_b.bias: F32
+blk.10.exp_probs_b_vl.bias: F32
+blk.10.ffn_down_exps.weight: IQ1_S
+blk.10.ffn_down_shexp.weight: Q5_K
+blk.10.ffn_gate_exps.weight: IQ2_XXS
+blk.10.ffn_gate_inp.weight: BF16
+blk.10.ffn_gate_shexp.weight: Q5_K
+blk.10.ffn_norm.weight: F32
+blk.10.ffn_up_exps.weight: IQ2_XXS
+blk.10.ffn_up_shexp.weight: Q5_K
+blk.10.hc_attn_base.weight: F32
+blk.10.hc_attn_fn.weight: F32
+blk.10.hc_attn_scale.weight: F32
+blk.10.hc_ffn_base.weight: F32
+blk.10.hc_ffn_fn.weight: F32
+blk.10.hc_ffn_scale.weight: F32
+blk.11.attn_kv.weight: Q8_0
+blk.11.attn_kv_a_norm.weight: F32
+blk.11.attn_norm.weight: F32
+blk.11.attn_output_a.weight: Q5_K
+blk.11.attn_output_b.weight: Q5_K
+blk.11.attn_q_a.weight: Q8_0
+blk.11.attn_q_a_norm.weight: F32
+blk.11.attn_q_b.weight: Q5_K
+blk.11.attn_sinks.weight: F32
+blk.11.exp_probs_b.bias: F32
+blk.11.exp_probs_b_vl.bias: F32
+blk.11.ffn_down_exps.weight: IQ2_XXS
+blk.11.ffn_down_shexp.weight: Q5_K
+blk.11.ffn_gate_exps.weight: IQ1_S
+blk.11.ffn_gate_inp.weight: BF16
+blk.11.ffn_gate_shexp.weight: Q5_K
+blk.11.ffn_norm.weight: F32
+blk.11.ffn_up_exps.weight: IQ1_S
+blk.11.ffn_up_shexp.weight: Q5_K
+blk.11.hc_attn_base.weight: F32
+blk.11.hc_attn_fn.weight: F32
+blk.11.hc_attn_scale.weight: F32
+blk.11.hc_ffn_base.weight: F32
+blk.11.hc_ffn_fn.weight: F32
+blk.11.hc_ffn_scale.weight: F32
+blk.12.attn_kv.weight: Q8_0
+blk.12.attn_kv_a_norm.weight: F32
+blk.12.attn_norm.weight: F32
+blk.12.attn_output_a.weight: Q5_K
+blk.12.attn_output_b.weight: Q5_K
+blk.12.attn_q_a.weight: Q8_0
+blk.12.attn_q_a_norm.weight: F32
+blk.12.attn_q_b.weight: Q5_K
+blk.12.attn_sinks.weight: F32
+blk.12.exp_probs_b.bias: F32
+blk.12.exp_probs_b_vl.bias: F32
+blk.12.ffn_down_exps.weight: IQ1_M
+blk.12.ffn_down_shexp.weight: Q5_K
+blk.12.ffn_gate_exps.weight: IQ1_S
+blk.12.ffn_gate_inp.weight: BF16
+blk.12.ffn_gate_shexp.weight: Q5_K
+blk.12.ffn_norm.weight: F32
+blk.12.ffn_up_exps.weight: IQ1_S
+blk.12.ffn_up_shexp.weight: Q5_K
+blk.12.hc_attn_base.weight: F32
+blk.12.hc_attn_fn.weight: F32
+blk.12.hc_attn_scale.weight: F32
+blk.12.hc_ffn_base.weight: F32
+blk.12.hc_ffn_fn.weight: F32
+blk.12.hc_ffn_scale.weight: F32
+blk.13.attn_kv.weight: Q8_0
+blk.13.attn_kv_a_norm.weight: F32
+blk.13.attn_norm.weight: F32
+blk.13.attn_output_a.weight: Q5_K
+blk.13.attn_output_b.weight: Q5_K
+blk.13.attn_q_a.weight: Q8_0
+blk.13.attn_q_a_norm.weight: F32
+blk.13.attn_q_b.weight: Q5_K
+blk.13.attn_sinks.weight: F32
+blk.13.exp_probs_b.bias: F32
+blk.13.exp_probs_b_vl.bias: F32
+blk.13.ffn_down_exps.weight: IQ1_M
+blk.13.ffn_down_shexp.weight: Q5_K
+blk.13.ffn_gate_exps.weight: IQ1_S
+blk.13.ffn_gate_inp.weight: BF16
+blk.13.ffn_gate_shexp.weight: Q5_K
+blk.13.ffn_norm.weight: F32
+blk.13.ffn_up_exps.weight: IQ1_S
+blk.13.ffn_up_shexp.weight: Q5_K
+blk.13.hc_attn_base.weight: F32
+blk.13.hc_attn_fn.weight: F32
+blk.13.hc_attn_scale.weight: F32
+blk.13.hc_ffn_base.weight: F32
+blk.13.hc_ffn_fn.weight: F32
+blk.13.hc_ffn_scale.weight: F32
+blk.14.attn_compressor_gate.weight: BF16
+blk.14.attn_compressor_kv.weight: BF16
+blk.14.attn_compressor_norm.weight: F32
+blk.14.attn_kv.weight: Q8_0
+blk.14.attn_kv_a_norm.weight: F32
+blk.14.attn_norm.weight: F32
+blk.14.attn_output_a.weight: Q5_K
+blk.14.attn_output_b.weight: Q5_K
+blk.14.attn_q_a.weight: Q8_0
+blk.14.attn_q_a_norm.weight: F32
+blk.14.attn_q_b.weight: Q5_K
+blk.14.attn_sinks.weight: F32
+blk.14.engram_gate.weight: F32
+blk.14.engram_kv.weight: Q5_K
+blk.14.exp_probs_b.bias: F32
+blk.14.exp_probs_b_vl.bias: F32
+blk.14.ffn_down_exps.weight: IQ1_S
+blk.14.ffn_down_shexp.weight: Q5_K
+blk.14.ffn_gate_exps.weight: IQ1_S
+blk.14.ffn_gate_inp.weight: BF16
+blk.14.ffn_gate_shexp.weight: Q5_K
+blk.14.ffn_norm.weight: F32
+blk.14.ffn_up_exps.weight: IQ1_S
+blk.14.ffn_up_shexp.weight: Q5_K
+blk.14.hc_attn_base.weight: F32
+blk.14.hc_attn_fn.weight: F32
+blk.14.hc_attn_scale.weight: F32
+blk.14.hc_ffn_base.weight: F32
+blk.14.hc_ffn_fn.weight: F32
+blk.14.hc_ffn_scale.weight: F32
+blk.14.indexer.attn_k.weight: BF16
+blk.14.indexer.attn_q_b.weight: BF16
+blk.14.indexer.k_norm.weight: F32
+blk.14.indexer.proj.weight: BF16
+blk.15.attn_kv.weight: Q8_0
+blk.15.attn_kv_a_norm.weight: F32
+blk.15.attn_norm.weight: F32
+blk.15.attn_output_a.weight: Q5_K
+blk.15.attn_output_b.weight: Q5_K
+blk.15.attn_q_a.weight: Q8_0
+blk.15.attn_q_a_norm.weight: F32
+blk.15.attn_q_b.weight: Q5_K
+blk.15.attn_sinks.weight: F32
+blk.15.exp_probs_b.bias: F32
+blk.15.exp_probs_b_vl.bias: F32
+blk.15.ffn_down_exps.weight: IQ1_S
+blk.15.ffn_down_shexp.weight: Q5_K
+blk.15.ffn_gate_exps.weight: IQ1_S
+blk.15.ffn_gate_inp.weight: BF16
+blk.15.ffn_gate_shexp.weight: Q5_K
+blk.15.ffn_norm.weight: F32
+blk.15.ffn_up_exps.weight: IQ1_S
+blk.15.ffn_up_shexp.weight: Q5_K
+blk.15.hc_attn_base.weight: F32
+blk.15.hc_attn_fn.weight: F32
+blk.15.hc_attn_scale.weight: F32
+blk.15.hc_ffn_base.weight: F32
+blk.15.hc_ffn_fn.weight: F32
+blk.15.hc_ffn_scale.weight: F32
+blk.16.attn_kv.weight: Q8_0
+blk.16.attn_kv_a_norm.weight: F32
+blk.16.attn_norm.weight: F32
+blk.16.attn_output_a.weight: Q5_K
+blk.16.attn_output_b.weight: Q5_K
+blk.16.attn_q_a.weight: Q8_0
+blk.16.attn_q_a_norm.weight: F32
+blk.16.attn_q_b.weight: Q5_K
+blk.16.attn_sinks.weight: F32
+blk.16.exp_probs_b.bias: F32
+blk.16.exp_probs_b_vl.bias: F32
+blk.16.ffn_down_exps.weight: IQ1_M
+blk.16.ffn_down_shexp.weight: Q5_K
+blk.16.ffn_gate_exps.weight: IQ1_S
+blk.16.ffn_gate_inp.weight: BF16
+blk.16.ffn_gate_shexp.weight: Q5_K
+blk.16.ffn_norm.weight: F32
+blk.16.ffn_up_exps.weight: IQ1_S
+blk.16.ffn_up_shexp.weight: Q5_K
+blk.16.hc_attn_base.weight: F32
+blk.16.hc_attn_fn.weight: F32
+blk.16.hc_attn_scale.weight: F32
+blk.16.hc_ffn_base.weight: F32
+blk.16.hc_ffn_fn.weight: F32
+blk.16.hc_ffn_scale.weight: F32
+blk.17.attn_kv.weight: Q8_0
+blk.17.attn_kv_a_norm.weight: F32
+blk.17.attn_norm.weight: F32
+blk.17.attn_output_a.weight: Q5_K
+blk.17.attn_output_b.weight: Q5_K
+blk.17.attn_q_a.weight: Q8_0
+blk.17.attn_q_a_norm.weight: F32
+blk.17.attn_q_b.weight: Q5_K
+blk.17.attn_sinks.weight: F32
+blk.17.exp_probs_b.bias: F32
+blk.17.exp_probs_b_vl.bias: F32
+blk.17.ffn_down_exps.weight: IQ1_S
+blk.17.ffn_down_shexp.weight: Q5_K
+blk.17.ffn_gate_exps.weight: IQ1_S
+blk.17.ffn_gate_inp.weight: BF16
+blk.17.ffn_gate_shexp.weight: Q5_K
+blk.17.ffn_norm.weight: F32
+blk.17.ffn_up_exps.weight: IQ1_S
+blk.17.ffn_up_shexp.weight: Q5_K
+blk.17.hc_attn_base.weight: F32
+blk.17.hc_attn_fn.weight: F32
+blk.17.hc_attn_scale.weight: F32
+blk.17.hc_ffn_base.weight: F32
+blk.17.hc_ffn_fn.weight: F32
+blk.17.hc_ffn_scale.weight: F32
+blk.18.attn_kv.weight: Q8_0
+blk.18.attn_kv_a_norm.weight: F32
+blk.18.attn_norm.weight: F32
+blk.18.attn_output_a.weight: Q5_K
+blk.18.attn_output_b.weight: Q5_K
+blk.18.attn_q_a.weight: Q8_0
+blk.18.attn_q_a_norm.weight: F32
+blk.18.attn_q_b.weight: Q5_K
+blk.18.attn_sinks.weight: F32
+blk.18.exp_probs_b.bias: F32
+blk.18.exp_probs_b_vl.bias: F32
+blk.18.ffn_down_exps.weight: IQ1_S
+blk.18.ffn_down_shexp.weight: Q5_K
+blk.18.ffn_gate_exps.weight: IQ2_XXS
+blk.18.ffn_gate_inp.weight: BF16
+blk.18.ffn_gate_shexp.weight: Q5_K
+blk.18.ffn_norm.weight: F32
+blk.18.ffn_up_exps.weight: IQ2_XXS
+blk.18.ffn_up_shexp.weight: Q5_K
+blk.18.hc_attn_base.weight: F32
+blk.18.hc_attn_fn.weight: F32
+blk.18.hc_attn_scale.weight: F32
+blk.18.hc_ffn_base.weight: F32
+blk.18.hc_ffn_fn.weight: F32
+blk.18.hc_ffn_scale.weight: F32
+blk.19.attn_kv.weight: Q8_0
+blk.19.attn_kv_a_norm.weight: F32
+blk.19.attn_norm.weight: F32
+blk.19.attn_output_a.weight: Q5_K
+blk.19.attn_output_b.weight: Q5_K
+blk.19.attn_q_a.weight: Q8_0
+blk.19.attn_q_a_norm.weight: F32
+blk.19.attn_q_b.weight: Q5_K
+blk.19.attn_sinks.weight: F32
+blk.19.exp_probs_b.bias: F32
+blk.19.exp_probs_b_vl.bias: F32
+blk.19.ffn_down_exps.weight: IQ1_M
+blk.19.ffn_down_shexp.weight: Q5_K
+blk.19.ffn_gate_exps.weight: IQ1_S
+blk.19.ffn_gate_inp.weight: BF16
+blk.19.ffn_gate_shexp.weight: Q5_K
+blk.19.ffn_norm.weight: F32
+blk.19.ffn_up_exps.weight: IQ1_S
+blk.19.ffn_up_shexp.weight: Q5_K
+blk.19.hc_attn_base.weight: F32
+blk.19.hc_attn_fn.weight: F32
+blk.19.hc_attn_scale.weight: F32
+blk.19.hc_ffn_base.weight: F32
+blk.19.hc_ffn_fn.weight: F32
+blk.19.hc_ffn_scale.weight: F32
+blk.20.attn_compressor_kv.weight: BF16
+blk.20.attn_compressor_norm.weight: F32
+blk.20.attn_kv.weight: Q8_0
+blk.20.attn_kv_a_norm.weight: F32
+blk.20.attn_norm.weight: F32
+blk.20.attn_output_a.weight: Q5_K
+blk.20.attn_output_b.weight: Q5_K
+blk.20.attn_q_a.weight: Q8_0
+blk.20.attn_q_a_norm.weight: F32
+blk.20.attn_q_b.weight: Q5_K
+blk.20.attn_sinks.weight: F32
+blk.20.exp_probs_b.bias: F32
+blk.20.exp_probs_b_vl.bias: F32
+blk.20.ffn_down_exps.weight: IQ1_M
+blk.20.ffn_down_shexp.weight: Q5_K
+blk.20.ffn_gate_exps.weight: IQ1_S
+blk.20.ffn_gate_inp.weight: BF16
+blk.20.ffn_gate_shexp.weight: Q5_K
+blk.20.ffn_norm.weight: F32
+blk.20.ffn_up_exps.weight: IQ1_S
+blk.20.ffn_up_shexp.weight: Q5_K
+blk.20.hc_attn_base.weight: F32
+blk.20.hc_attn_fn.weight: F32
+blk.20.hc_attn_scale.weight: F32
+blk.20.hc_ffn_base.weight: F32
+blk.20.hc_ffn_fn.weight: F32
+blk.20.hc_ffn_scale.weight: F32
+blk.20.indexer.attn_k.weight: BF16
+blk.20.indexer.attn_q_b.weight: BF16
+blk.20.indexer.k_norm.weight: F32
+blk.20.indexer.proj.weight: BF16
+blk.21.attn_kv.weight: Q8_0
+blk.21.attn_kv_a_norm.weight: F32
+blk.21.attn_norm.weight: F32
+blk.21.attn_output_a.weight: Q5_K
+blk.21.attn_output_b.weight: Q5_K
+blk.21.attn_q_a.weight: Q8_0
+blk.21.attn_q_a_norm.weight: F32
+blk.21.attn_q_b.weight: Q5_K
+blk.21.attn_sinks.weight: F32
+blk.21.exp_probs_b.bias: F32
+blk.21.exp_probs_b_vl.bias: F32
+blk.21.ffn_down_exps.weight: IQ1_M
+blk.21.ffn_down_shexp.weight: Q5_K
+blk.21.ffn_gate_exps.weight: IQ1_S
+blk.21.ffn_gate_inp.weight: BF16
+blk.21.ffn_gate_shexp.weight: Q5_K
+blk.21.ffn_norm.weight: F32
+blk.21.ffn_up_exps.weight: IQ1_S
+blk.21.ffn_up_shexp.weight: Q5_K
+blk.21.hc_attn_base.weight: F32
+blk.21.hc_attn_fn.weight: F32
+blk.21.hc_attn_scale.weight: F32
+blk.21.hc_ffn_base.weight: F32
+blk.21.hc_ffn_fn.weight: F32
+blk.21.hc_ffn_scale.weight: F32
+blk.22.attn_kv.weight: Q8_0
+blk.22.attn_kv_a_norm.weight: F32
+blk.22.attn_norm.weight: F32
+blk.22.attn_output_a.weight: Q5_K
+blk.22.attn_output_b.weight: Q5_K
+blk.22.attn_q_a.weight: Q8_0
+blk.22.attn_q_a_norm.weight: F32
+blk.22.attn_q_b.weight: Q5_K
+blk.22.attn_sinks.weight: F32
+blk.22.exp_probs_b.bias: F32
+blk.22.exp_probs_b_vl.bias: F32
+blk.22.ffn_down_exps.weight: IQ1_M
+blk.22.ffn_down_shexp.weight: Q5_K
+blk.22.ffn_gate_exps.weight: IQ1_S
+blk.22.ffn_gate_inp.weight: BF16
+blk.22.ffn_gate_shexp.weight: Q5_K
+blk.22.ffn_norm.weight: F32
+blk.22.ffn_up_exps.weight: IQ1_S
+blk.22.ffn_up_shexp.weight: Q5_K
+blk.22.hc_attn_base.weight: F32
+blk.22.hc_attn_fn.weight: F32
+blk.22.hc_attn_scale.weight: F32
+blk.22.hc_ffn_base.weight: F32
+blk.22.hc_ffn_fn.weight: F32
+blk.22.hc_ffn_scale.weight: F32
+blk.23.attn_kv.weight: Q8_0
+blk.23.attn_kv_a_norm.weight: F32
+blk.23.attn_norm.weight: F32
+blk.23.attn_output_a.weight: Q5_K
+blk.23.attn_output_b.weight: Q5_K
+blk.23.attn_q_a.weight: Q8_0
+blk.23.attn_q_a_norm.weight: F32
+blk.23.attn_q_b.weight: Q5_K
+blk.23.attn_sinks.weight: F32
+blk.23.exp_probs_b.bias: F32
+blk.23.exp_probs_b_vl.bias: F32
+blk.23.ffn_down_exps.weight: IQ1_S
+blk.23.ffn_down_shexp.weight: Q5_K
+blk.23.ffn_gate_exps.weight: IQ2_XXS
+blk.23.ffn_gate_inp.weight: BF16
+blk.23.ffn_gate_shexp.weight: Q5_K
+blk.23.ffn_norm.weight: F32
+blk.23.ffn_up_exps.weight: IQ2_XXS
+blk.23.ffn_up_shexp.weight: Q5_K
+blk.23.hc_attn_base.weight: F32
+blk.23.hc_attn_fn.weight: F32
+blk.23.hc_attn_scale.weight: F32
+blk.23.hc_ffn_base.weight: F32
+blk.23.hc_ffn_fn.weight: F32
+blk.23.hc_ffn_scale.weight: F32
+blk.24.attn_kv.weight: Q8_0
+blk.24.attn_kv_a_norm.weight: F32
+blk.24.attn_norm.weight: F32
+blk.24.attn_output_a.weight: Q5_K
+blk.24.attn_output_b.weight: Q5_K
+blk.24.attn_q_a.weight: Q8_0
+blk.24.attn_q_a_norm.weight: F32
+blk.24.attn_q_b.weight: Q5_K
+blk.24.attn_sinks.weight: F32
+blk.24.exp_probs_b.bias: F32
+blk.24.exp_probs_b_vl.bias: F32
+blk.24.ffn_down_exps.weight: IQ1_S
+blk.24.ffn_down_shexp.weight: Q5_K
+blk.24.ffn_gate_exps.weight: IQ1_S
+blk.24.ffn_gate_inp.weight: BF16
+blk.24.ffn_gate_shexp.weight: Q5_K
+blk.24.ffn_norm.weight: F32
+blk.24.ffn_up_exps.weight: IQ1_S
+blk.24.ffn_up_shexp.weight: Q5_K
+blk.24.hc_attn_base.weight: F32
+blk.24.hc_attn_fn.weight: F32
+blk.24.hc_attn_scale.weight: F32
+blk.24.hc_ffn_base.weight: F32
+blk.24.hc_ffn_fn.weight: F32
+blk.24.hc_ffn_scale.weight: F32
+blk.24.indexer.attn_q_b.weight: BF16
+blk.24.indexer.proj.weight: BF16
+blk.25.attn_kv.weight: Q8_0
+blk.25.attn_kv_a_norm.weight: F32
+blk.25.attn_norm.weight: F32
+blk.25.attn_output_a.weight: Q5_K
+blk.25.attn_output_b.weight: Q5_K
+blk.25.attn_q_a.weight: Q8_0
+blk.25.attn_q_a_norm.weight: F32
+blk.25.attn_q_b.weight: Q5_K
+blk.25.attn_sinks.weight: F32
+blk.25.exp_probs_b.bias: F32
+blk.25.exp_probs_b_vl.bias: F32
+blk.25.ffn_down_exps.weight: IQ1_M
+blk.25.ffn_down_shexp.weight: Q5_K
+blk.25.ffn_gate_exps.weight: IQ2_XXS
+blk.25.ffn_gate_inp.weight: BF16
+blk.25.ffn_gate_shexp.weight: Q5_K
+blk.25.ffn_norm.weight: F32
+blk.25.ffn_up_exps.weight: IQ2_XXS
+blk.25.ffn_up_shexp.weight: Q5_K
+blk.25.hc_attn_base.weight: F32
+blk.25.hc_attn_fn.weight: F32
+blk.25.hc_attn_scale.weight: F32
+blk.25.hc_ffn_base.weight: F32
+blk.25.hc_ffn_fn.weight: F32
+blk.25.hc_ffn_scale.weight: F32
+blk.26.attn_kv.weight: Q8_0
+blk.26.attn_kv_a_norm.weight: F32
+blk.26.attn_norm.weight: F32
+blk.26.attn_output_a.weight: Q5_K
+blk.26.attn_output_b.weight: Q5_K
+blk.26.attn_q_a.weight: Q8_0
+blk.26.attn_q_a_norm.weight: F32
+blk.26.attn_q_b.weight: Q5_K
+blk.26.attn_sinks.weight: F32
+blk.26.exp_probs_b.bias: F32
+blk.26.exp_probs_b_vl.bias: F32
+blk.26.ffn_down_exps.weight: IQ2_S
+blk.26.ffn_down_shexp.weight: Q5_K
+blk.26.ffn_gate_exps.weight: IQ1_S
+blk.26.ffn_gate_inp.weight: BF16
+blk.26.ffn_gate_shexp.weight: Q5_K
+blk.26.ffn_norm.weight: F32
+blk.26.ffn_up_exps.weight: IQ1_S
+blk.26.ffn_up_shexp.weight: Q5_K
+blk.26.hc_attn_base.weight: F32
+blk.26.hc_attn_fn.weight: F32
+blk.26.hc_attn_scale.weight: F32
+blk.26.hc_ffn_base.weight: F32
+blk.26.hc_ffn_fn.weight: F32
+blk.26.hc_ffn_scale.weight: F32
+blk.27.attn_kv.weight: Q8_0
+blk.27.attn_kv_a_norm.weight: F32
+blk.27.attn_norm.weight: F32
+blk.27.attn_output_a.weight: Q5_K
+blk.27.attn_output_b.weight: Q5_K
+blk.27.attn_q_a.weight: Q8_0
+blk.27.attn_q_a_norm.weight: F32
+blk.27.attn_q_b.weight: Q5_K
+blk.27.attn_sinks.weight: F32
+blk.27.exp_probs_b.bias: F32
+blk.27.exp_probs_b_vl.bias: F32
+blk.27.ffn_down_exps.weight: IQ1_S
+blk.27.ffn_down_shexp.weight: Q5_K
+blk.27.ffn_gate_exps.weight: IQ1_S
+blk.27.ffn_gate_inp.weight: BF16
+blk.27.ffn_gate_shexp.weight: Q5_K
+blk.27.ffn_norm.weight: F32
+blk.27.ffn_up_exps.weight: IQ1_S
+blk.27.ffn_up_shexp.weight: Q5_K
+blk.27.hc_attn_base.weight: F32
+blk.27.hc_attn_fn.weight: F32
+blk.27.hc_attn_scale.weight: F32
+blk.27.hc_ffn_base.weight: F32
+blk.27.hc_ffn_fn.weight: F32
+blk.27.hc_ffn_scale.weight: F32
+blk.28.attn_kv.weight: Q8_0
+blk.28.attn_kv_a_norm.weight: F32
+blk.28.attn_norm.weight: F32
+blk.28.attn_output_a.weight: Q5_K
+blk.28.attn_output_b.weight: Q5_K
+blk.28.attn_q_a.weight: Q8_0
+blk.28.attn_q_a_norm.weight: F32
+blk.28.attn_q_b.weight: Q5_K
+blk.28.attn_sinks.weight: F32
+blk.28.exp_probs_b.bias: F32
+blk.28.exp_probs_b_vl.bias: F32
+blk.28.ffn_down_exps.weight: IQ1_S
+blk.28.ffn_down_shexp.weight: Q5_K
+blk.28.ffn_gate_exps.weight: IQ1_M
+blk.28.ffn_gate_inp.weight: BF16
+blk.28.ffn_gate_shexp.weight: Q5_K
+blk.28.ffn_norm.weight: F32
+blk.28.ffn_up_exps.weight: IQ1_M
+blk.28.ffn_up_shexp.weight: Q5_K
+blk.28.hc_attn_base.weight: F32
+blk.28.hc_attn_fn.weight: F32
+blk.28.hc_attn_scale.weight: F32
+blk.28.hc_ffn_base.weight: F32
+blk.28.hc_ffn_fn.weight: F32
+blk.28.hc_ffn_scale.weight: F32
+blk.28.indexer.attn_q_b.weight: BF16
+blk.28.indexer.proj.weight: BF16
+blk.29.attn_kv.weight: Q8_0
+blk.29.attn_kv_a_norm.weight: F32
+blk.29.attn_norm.weight: F32
+blk.29.attn_output_a.weight: Q5_K
+blk.29.attn_output_b.weight: Q5_K
+blk.29.attn_q_a.weight: Q8_0
+blk.29.attn_q_a_norm.weight: F32
+blk.29.attn_q_b.weight: Q5_K
+blk.29.attn_sinks.weight: F32
+blk.29.exp_probs_b.bias: F32
+blk.29.exp_probs_b_vl.bias: F32
+blk.29.ffn_down_exps.weight: IQ1_S
+blk.29.ffn_down_shexp.weight: Q5_K
+blk.29.ffn_gate_exps.weight: IQ2_XXS
+blk.29.ffn_gate_inp.weight: BF16
+blk.29.ffn_gate_shexp.weight: Q5_K
+blk.29.ffn_norm.weight: F32
+blk.29.ffn_up_exps.weight: IQ2_XXS
+blk.29.ffn_up_shexp.weight: Q5_K
+blk.29.hc_attn_base.weight: F32
+blk.29.hc_attn_fn.weight: F32
+blk.29.hc_attn_scale.weight: F32
+blk.29.hc_ffn_base.weight: F32
+blk.29.hc_ffn_fn.weight: F32
+blk.29.hc_ffn_scale.weight: F32
+blk.30.attn_kv.weight: Q8_0
+blk.30.attn_kv_a_norm.weight: F32
+blk.30.attn_norm.weight: F32
+blk.30.attn_output_a.weight: Q5_K
+blk.30.attn_output_b.weight: Q5_K
+blk.30.attn_q_a.weight: Q8_0
+blk.30.attn_q_a_norm.weight: F32
+blk.30.attn_q_b.weight: Q5_K
+blk.30.attn_sinks.weight: F32
+blk.30.exp_probs_b.bias: F32
+blk.30.exp_probs_b_vl.bias: F32
+blk.30.ffn_down_exps.weight: IQ1_M
+blk.30.ffn_down_shexp.weight: Q5_K
+blk.30.ffn_gate_exps.weight: IQ2_XXS
+blk.30.ffn_gate_inp.weight: BF16
+blk.30.ffn_gate_shexp.weight: Q5_K
+blk.30.ffn_norm.weight: F32
+blk.30.ffn_up_exps.weight: IQ2_XXS
+blk.30.ffn_up_shexp.weight: Q5_K
+blk.30.hc_attn_base.weight: F32
+blk.30.hc_attn_fn.weight: F32
+blk.30.hc_attn_scale.weight: F32
+blk.30.hc_ffn_base.weight: F32
+blk.30.hc_ffn_fn.weight: F32
+blk.30.hc_ffn_scale.weight: F32
+blk.31.attn_kv.weight: Q8_0
+blk.31.attn_kv_a_norm.weight: F32
+blk.31.attn_norm.weight: F32
+blk.31.attn_output_a.weight: Q5_K
+blk.31.attn_output_b.weight: Q5_K
+blk.31.attn_q_a.weight: Q8_0
+blk.31.attn_q_a_norm.weight: F32
+blk.31.attn_q_b.weight: Q5_K
+blk.31.attn_sinks.weight: F32
+blk.31.exp_probs_b.bias: F32
+blk.31.exp_probs_b_vl.bias: F32
+blk.31.ffn_down_exps.weight: IQ1_S
+blk.31.ffn_down_shexp.weight: Q5_K
+blk.31.ffn_gate_exps.weight: IQ1_S
+blk.31.ffn_gate_inp.weight: BF16
+blk.31.ffn_gate_shexp.weight: Q5_K
+blk.31.ffn_norm.weight: F32
+blk.31.ffn_up_exps.weight: IQ1_S
+blk.31.ffn_up_shexp.weight: Q5_K
+blk.31.hc_attn_base.weight: F32
+blk.31.hc_attn_fn.weight: F32
+blk.31.hc_attn_scale.weight: F32
+blk.31.hc_ffn_base.weight: F32
+blk.31.hc_ffn_fn.weight: F32
+blk.31.hc_ffn_scale.weight: F32
+blk.32.attn_kv.weight: Q8_0
+blk.32.attn_kv_a_norm.weight: F32
+blk.32.attn_norm.weight: F32
+blk.32.attn_output_a.weight: Q5_K
+blk.32.attn_output_b.weight: Q5_K
+blk.32.attn_q_a.weight: Q8_0
+blk.32.attn_q_a_norm.weight: F32
+blk.32.attn_q_b.weight: Q5_K
+blk.32.attn_sinks.weight: F32
+blk.32.exp_probs_b.bias: F32
+blk.32.exp_probs_b_vl.bias: F32
+blk.32.ffn_down_exps.weight: IQ1_S
+blk.32.ffn_down_shexp.weight: Q5_K
+blk.32.ffn_gate_exps.weight: IQ2_XXS
+blk.32.ffn_gate_inp.weight: BF16
+blk.32.ffn_gate_shexp.weight: Q5_K
+blk.32.ffn_norm.weight: F32
+blk.32.ffn_up_exps.weight: IQ2_XXS
+blk.32.ffn_up_shexp.weight: Q5_K
+blk.32.hc_attn_base.weight: F32
+blk.32.hc_attn_fn.weight: F32
+blk.32.hc_attn_scale.weight: F32
+blk.32.hc_ffn_base.weight: F32
+blk.32.hc_ffn_fn.weight: F32
+blk.32.hc_ffn_scale.weight: F32
+blk.32.indexer.attn_q_b.weight: BF16
+blk.32.indexer.proj.weight: BF16
+blk.33.attn_kv.weight: Q8_0
+blk.33.attn_kv_a_norm.weight: F32
+blk.33.attn_norm.weight: F32
+blk.33.attn_output_a.weight: Q5_K
+blk.33.attn_output_b.weight: Q5_K
+blk.33.attn_q_a.weight: Q8_0
+blk.33.attn_q_a_norm.weight: F32
+blk.33.attn_q_b.weight: Q5_K
+blk.33.attn_sinks.weight: F32
+blk.33.exp_probs_b.bias: F32
+blk.33.exp_probs_b_vl.bias: F32
+blk.33.ffn_down_exps.weight: IQ1_S
+blk.33.ffn_down_shexp.weight: Q5_K
+blk.33.ffn_gate_exps.weight: IQ1_M
+blk.33.ffn_gate_inp.weight: BF16
+blk.33.ffn_gate_shexp.weight: Q5_K
+blk.33.ffn_norm.weight: F32
+blk.33.ffn_up_exps.weight: IQ1_M
+blk.33.ffn_up_shexp.weight: Q5_K
+blk.33.hc_attn_base.weight: F32
+blk.33.hc_attn_fn.weight: F32
+blk.33.hc_attn_scale.weight: F32
+blk.33.hc_ffn_base.weight: F32
+blk.33.hc_ffn_fn.weight: F32
+blk.33.hc_ffn_scale.weight: F32
+blk.34.attn_kv.weight: Q8_0
+blk.34.attn_kv_a_norm.weight: F32
+blk.34.attn_norm.weight: F32
+blk.34.attn_output_a.weight: Q5_K
+blk.34.attn_output_b.weight: Q5_K
+blk.34.attn_q_a.weight: Q8_0
+blk.34.attn_q_a_norm.weight: F32
+blk.34.attn_q_b.weight: Q5_K
+blk.34.attn_sinks.weight: F32
+blk.34.exp_probs_b.bias: F32
+blk.34.exp_probs_b_vl.bias: F32
+blk.34.ffn_down_exps.weight: IQ1_S
+blk.34.ffn_down_shexp.weight: Q5_K
+blk.34.ffn_gate_exps.weight: IQ1_S
+blk.34.ffn_gate_inp.weight: BF16
+blk.34.ffn_gate_shexp.weight: Q5_K
+blk.34.ffn_norm.weight: F32
+blk.34.ffn_up_exps.weight: IQ1_S
+blk.34.ffn_up_shexp.weight: Q5_K
+blk.34.hc_attn_base.weight: F32
+blk.34.hc_attn_fn.weight: F32
+blk.34.hc_attn_scale.weight: F32
+blk.34.hc_ffn_base.weight: F32
+blk.34.hc_ffn_fn.weight: F32
+blk.34.hc_ffn_scale.weight: F32
+blk.35.attn_kv.weight: Q8_0
+blk.35.attn_kv_a_norm.weight: F32
+blk.35.attn_norm.weight: F32
+blk.35.attn_output_a.weight: Q5_K
+blk.35.attn_output_b.weight: Q5_K
+blk.35.attn_q_a.weight: Q8_0
+blk.35.attn_q_a_norm.weight: F32
+blk.35.attn_q_b.weight: Q5_K
+blk.35.attn_sinks.weight: F32
+blk.35.exp_probs_b.bias: F32
+blk.35.exp_probs_b_vl.bias: F32
+blk.35.ffn_down_exps.weight: IQ1_S
+blk.35.ffn_down_shexp.weight: Q5_K
+blk.35.ffn_gate_exps.weight: IQ1_S
+blk.35.ffn_gate_inp.weight: BF16
+blk.35.ffn_gate_shexp.weight: Q5_K
+blk.35.ffn_norm.weight: F32
+blk.35.ffn_up_exps.weight: IQ1_S
+blk.35.ffn_up_shexp.weight: Q5_K
+blk.35.hc_attn_base.weight: F32
+blk.35.hc_attn_fn.weight: F32
+blk.35.hc_attn_scale.weight: F32
+blk.35.hc_ffn_base.weight: F32
+blk.35.hc_ffn_fn.weight: F32
+blk.35.hc_ffn_scale.weight: F32
+blk.36.attn_kv.weight: Q8_0
+blk.36.attn_kv_a_norm.weight: F32
+blk.36.attn_norm.weight: F32
+blk.36.attn_output_a.weight: Q5_K
+blk.36.attn_output_b.weight: Q5_K
+blk.36.attn_q_a.weight: Q8_0
+blk.36.attn_q_a_norm.weight: F32
+blk.36.attn_q_b.weight: Q5_K
+blk.36.attn_sinks.weight: F32
+blk.36.exp_probs_b.bias: F32
+blk.36.exp_probs_b_vl.bias: F32
+blk.36.ffn_down_exps.weight: IQ1_S
+blk.36.ffn_down_shexp.weight: Q5_K
+blk.36.ffn_gate_exps.weight: IQ1_M
+blk.36.ffn_gate_inp.weight: BF16
+blk.36.ffn_gate_shexp.weight: Q5_K
+blk.36.ffn_norm.weight: F32
+blk.36.ffn_up_exps.weight: IQ1_M
+blk.36.ffn_up_shexp.weight: Q5_K
+blk.36.hc_attn_base.weight: F32
+blk.36.hc_attn_fn.weight: F32
+blk.36.hc_attn_scale.weight: F32
+blk.36.hc_ffn_base.weight: F32
+blk.36.hc_ffn_fn.weight: F32
+blk.36.hc_ffn_scale.weight: F32
+blk.36.indexer.attn_q_b.weight: BF16
+blk.36.indexer.proj.weight: BF16
+blk.37.attn_kv.weight: Q8_0
+blk.37.attn_kv_a_norm.weight: F32
+blk.37.attn_norm.weight: F32
+blk.37.attn_output_a.weight: Q5_K
+blk.37.attn_output_b.weight: Q5_K
+blk.37.attn_q_a.weight: Q8_0
+blk.37.attn_q_a_norm.weight: F32
+blk.37.attn_q_b.weight: Q5_K
+blk.37.attn_sinks.weight: F32
+blk.37.exp_probs_b.bias: F32
+blk.37.exp_probs_b_vl.bias: F32
+blk.37.ffn_down_exps.weight: IQ1_S
+blk.37.ffn_down_shexp.weight: Q5_K
+blk.37.ffn_gate_exps.weight: IQ1_S
+blk.37.ffn_gate_inp.weight: BF16
+blk.37.ffn_gate_shexp.weight: Q5_K
+blk.37.ffn_norm.weight: F32
+blk.37.ffn_up_exps.weight: IQ1_S
+blk.37.ffn_up_shexp.weight: Q5_K
+blk.37.hc_attn_base.weight: F32
+blk.37.hc_attn_fn.weight: F32
+blk.37.hc_attn_scale.weight: F32
+blk.37.hc_ffn_base.weight: F32
+blk.37.hc_ffn_fn.weight: F32
+blk.37.hc_ffn_scale.weight: F32
+blk.38.attn_kv.weight: Q8_0
+blk.38.attn_kv_a_norm.weight: F32
+blk.38.attn_norm.weight: F32
+blk.38.attn_output_a.weight: Q5_K
+blk.38.attn_output_b.weight: Q5_K
+blk.38.attn_q_a.weight: Q8_0
+blk.38.attn_q_a_norm.weight: F32
+blk.38.attn_q_b.weight: Q5_K
+blk.38.attn_sinks.weight: F32
+blk.38.exp_probs_b.bias: F32
+blk.38.exp_probs_b_vl.bias: F32
+blk.38.ffn_down_exps.weight: IQ1_S
+blk.38.ffn_down_shexp.weight: Q5_K
+blk.38.ffn_gate_exps.weight: IQ2_XXS
+blk.38.ffn_gate_inp.weight: BF16
+blk.38.ffn_gate_shexp.weight: Q5_K
+blk.38.ffn_norm.weight: F32
+blk.38.ffn_up_exps.weight: IQ2_XXS
+blk.38.ffn_up_shexp.weight: Q5_K
+blk.38.hc_attn_base.weight: F32
+blk.38.hc_attn_fn.weight: F32
+blk.38.hc_attn_scale.weight: F32
+blk.38.hc_ffn_base.weight: F32
+blk.38.hc_ffn_fn.weight: F32
+blk.38.hc_ffn_scale.weight: F32
+blk.39.attn_kv.weight: Q8_0
+blk.39.attn_kv_a_norm.weight: F32
+blk.39.attn_norm.weight: F32
+blk.39.attn_output_a.weight: Q5_K
+blk.39.attn_output_b.weight: Q5_K
+blk.39.attn_q_a.weight: Q8_0
+blk.39.attn_q_a_norm.weight: F32
+blk.39.attn_q_b.weight: Q5_K
+blk.39.attn_sinks.weight: F32
+blk.39.exp_probs_b.bias: F32
+blk.39.exp_probs_b_vl.bias: F32
+blk.39.ffn_down_exps.weight: IQ2_XXS
+blk.39.ffn_down_shexp.weight: Q5_K
+blk.39.ffn_gate_exps.weight: IQ1_S
+blk.39.ffn_gate_inp.weight: BF16
+blk.39.ffn_gate_shexp.weight: Q5_K
+blk.39.ffn_norm.weight: F32
+blk.39.ffn_up_exps.weight: IQ1_S
+blk.39.ffn_up_shexp.weight: Q5_K
+blk.39.hc_attn_base.weight: F32
+blk.39.hc_attn_fn.weight: F32
+blk.39.hc_attn_scale.weight: F32
+blk.39.hc_ffn_base.weight: F32
+blk.39.hc_ffn_fn.weight: F32
+blk.39.hc_ffn_scale.weight: F32

tensor-allocation/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS.rco-allocation.txt

@@ -0,0 +1,1061 @@
+# RCO per-tensor GGUF allocation
+# Repository: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
+# GGUF: DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00006.gguf (first of 6 shards; this dump covers all of them)
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00005-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf
+# Target bpw over the backbone (all tensors except the Engram table): 2.25 (measured: 2.2501)
+# The Engram table (per_layer_token_embd, shard 6 = the shared Engram file) is IQ4_NL in every build and not part of the search.
+# GGUF version: 3
+# GGUF metadata entries: 77
+# Tensor count: 1043
+# Quant-type counts: BF16=67, F32=531, IQ2_S=16, IQ2_XS=27, IQ2_XXS=76, IQ3_XXS=1, IQ4_NL=1, Q5_K=244, Q8_0=80
+# Format: <GGUF tensor name>: <GGML quantization/storage type>
+# Includes fixed non-RCO tensors (e.g. F32/BF16) so the released GGUF is fully reproducible.
+#
+output.weight: Q5_K
+output_norm.weight: F32
+per_layer_token_embd.weight: IQ4_NL
+token_embd.weight: Q5_K
+blk.0.attn_kv.weight: Q8_0
+blk.0.attn_kv_a_norm.weight: F32
+blk.0.attn_norm.weight: F32
+blk.0.attn_output_a.weight: Q5_K
+blk.0.attn_output_b.weight: Q5_K
+blk.0.attn_q_a.weight: Q8_0
+blk.0.attn_q_a_norm.weight: F32
+blk.0.attn_q_b.weight: Q5_K
+blk.0.attn_sinks.weight: F32
+blk.0.exp_probs_b.bias: F32
+blk.0.exp_probs_b_vl.bias: F32
+blk.0.ffn_down_exps.weight: IQ2_XXS
+blk.0.ffn_down_shexp.weight: Q5_K
+blk.0.ffn_gate_exps.weight: IQ2_XXS
+blk.0.ffn_gate_inp.weight: BF16
+blk.0.ffn_gate_shexp.weight: Q5_K
+blk.0.ffn_norm.weight: F32
+blk.0.ffn_up_exps.weight: IQ2_XXS
+blk.0.ffn_up_shexp.weight: Q5_K
+blk.0.hc_attn_base.weight: F32
+blk.0.hc_attn_fn.weight: F32
+blk.0.hc_attn_scale.weight: F32
+blk.0.hc_ffn_base.weight: F32
+blk.0.hc_ffn_fn.weight: F32
+blk.0.hc_ffn_scale.weight: F32
+blk.1.attn_kv.weight: Q8_0
+blk.1.attn_kv_a_norm.weight: F32
+blk.1.attn_norm.weight: F32
+blk.1.attn_output_a.weight: Q5_K
+blk.1.attn_output_b.weight: Q5_K
+blk.1.attn_q_a.weight: Q8_0
+blk.1.attn_q_a_norm.weight: F32
+blk.1.attn_q_b.weight: Q5_K
+blk.1.attn_sinks.weight: F32
+blk.1.engram_gate.weight: F32
+blk.1.engram_kv.weight: Q5_K
+blk.1.exp_probs_b.bias: F32
+blk.1.exp_probs_b_vl.bias: F32
+blk.1.ffn_down_exps.weight: IQ2_XXS
+blk.1.ffn_down_shexp.weight: Q5_K
+blk.1.ffn_gate_exps.weight: IQ2_XXS
+blk.1.ffn_gate_inp.weight: BF16
+blk.1.ffn_gate_shexp.weight: Q5_K
+blk.1.ffn_norm.weight: F32
+blk.1.ffn_up_exps.weight: IQ2_XXS
+blk.1.ffn_up_shexp.weight: Q5_K
+blk.1.hc_attn_base.weight: F32
+blk.1.hc_attn_fn.weight: F32
+blk.1.hc_attn_scale.weight: F32
+blk.1.hc_ffn_base.weight: F32
+blk.1.hc_ffn_fn.weight: F32
+blk.1.hc_ffn_scale.weight: F32
+blk.2.attn_compressor_gate.weight: BF16
+blk.2.attn_compressor_kv.weight: BF16
+blk.2.attn_compressor_norm.weight: F32
+blk.2.attn_kv.weight: Q8_0
+blk.2.attn_kv_a_norm.weight: F32
+blk.2.attn_norm.weight: F32
+blk.2.attn_output_a.weight: Q5_K
+blk.2.attn_output_b.weight: Q5_K
+blk.2.attn_q_a.weight: Q8_0
+blk.2.attn_q_a_norm.weight: F32
+blk.2.attn_q_b.weight: Q5_K
+blk.2.attn_sinks.weight: F32
+blk.2.exp_probs_b.bias: F32
+blk.2.exp_probs_b_vl.bias: F32
+blk.2.ffn_down_exps.weight: IQ2_XXS
+blk.2.ffn_down_shexp.weight: Q5_K
+blk.2.ffn_gate_exps.weight: IQ2_XXS
+blk.2.ffn_gate_inp.weight: BF16
+blk.2.ffn_gate_shexp.weight: Q5_K
+blk.2.ffn_norm.weight: F32
+blk.2.ffn_up_exps.weight: IQ2_XXS
+blk.2.ffn_up_shexp.weight: Q5_K
+blk.2.hc_attn_base.weight: F32
+blk.2.hc_attn_fn.weight: F32
+blk.2.hc_attn_scale.weight: F32
+blk.2.hc_ffn_base.weight: F32
+blk.2.hc_ffn_fn.weight: F32
+blk.2.hc_ffn_scale.weight: F32
+blk.2.indexer.attn_k.weight: BF16
+blk.2.indexer.attn_q_b.weight: BF16
+blk.2.indexer.k_norm.weight: F32
+blk.2.indexer.proj.weight: BF16
+blk.3.attn_kv.weight: Q8_0
+blk.3.attn_kv_a_norm.weight: F32
+blk.3.attn_norm.weight: F32
+blk.3.attn_output_a.weight: Q5_K
+blk.3.attn_output_b.weight: Q5_K
+blk.3.attn_q_a.weight: Q8_0
+blk.3.attn_q_a_norm.weight: F32
+blk.3.attn_q_b.weight: Q5_K
+blk.3.attn_sinks.weight: F32
+blk.3.exp_probs_b.bias: F32
+blk.3.exp_probs_b_vl.bias: F32
+blk.3.ffn_down_exps.weight: IQ2_XXS
+blk.3.ffn_down_shexp.weight: Q5_K
+blk.3.ffn_gate_exps.weight: IQ2_XXS
+blk.3.ffn_gate_inp.weight: BF16
+blk.3.ffn_gate_shexp.weight: Q5_K
+blk.3.ffn_norm.weight: F32
+blk.3.ffn_up_exps.weight: IQ2_XXS
+blk.3.ffn_up_shexp.weight: Q5_K
+blk.3.hc_attn_base.weight: F32
+blk.3.hc_attn_fn.weight: F32
+blk.3.hc_attn_scale.weight: F32
+blk.3.hc_ffn_base.weight: F32
+blk.3.hc_ffn_fn.weight: F32
+blk.3.hc_ffn_scale.weight: F32
+blk.4.attn_kv.weight: Q8_0
+blk.4.attn_kv_a_norm.weight: F32
+blk.4.attn_norm.weight: F32
+blk.4.attn_output_a.weight: Q5_K
+blk.4.attn_output_b.weight: Q5_K
+blk.4.attn_q_a.weight: Q8_0
+blk.4.attn_q_a_norm.weight: F32
+blk.4.attn_q_b.weight: Q5_K
+blk.4.attn_sinks.weight: F32
+blk.4.exp_probs_b.bias: F32
+blk.4.exp_probs_b_vl.bias: F32
+blk.4.ffn_down_exps.weight: IQ2_XXS
+blk.4.ffn_down_shexp.weight: Q5_K
+blk.4.ffn_gate_exps.weight: IQ2_XXS
+blk.4.ffn_gate_inp.weight: BF16
+blk.4.ffn_gate_shexp.weight: Q5_K
+blk.4.ffn_norm.weight: F32
+blk.4.ffn_up_exps.weight: IQ2_XXS
+blk.4.ffn_up_shexp.weight: Q5_K
+blk.4.hc_attn_base.weight: F32
+blk.4.hc_attn_fn.weight: F32
+blk.4.hc_attn_scale.weight: F32
+blk.4.hc_ffn_base.weight: F32
+blk.4.hc_ffn_fn.weight: F32
+blk.4.hc_ffn_scale.weight: F32
+blk.5.attn_kv.weight: Q8_0
+blk.5.attn_kv_a_norm.weight: F32
+blk.5.attn_norm.weight: F32
+blk.5.attn_output_a.weight: Q5_K
+blk.5.attn_output_b.weight: Q5_K
+blk.5.attn_q_a.weight: Q8_0
+blk.5.attn_q_a_norm.weight: F32
+blk.5.attn_q_b.weight: Q5_K
+blk.5.attn_sinks.weight: F32
+blk.5.exp_probs_b.bias: F32
+blk.5.exp_probs_b_vl.bias: F32
+blk.5.ffn_down_exps.weight: IQ2_XXS
+blk.5.ffn_down_shexp.weight: Q5_K
+blk.5.ffn_gate_exps.weight: IQ2_XXS
+blk.5.ffn_gate_inp.weight: BF16
+blk.5.ffn_gate_shexp.weight: Q5_K
+blk.5.ffn_norm.weight: F32
+blk.5.ffn_up_exps.weight: IQ2_XXS
+blk.5.ffn_up_shexp.weight: Q5_K
+blk.5.hc_attn_base.weight: F32
+blk.5.hc_attn_fn.weight: F32
+blk.5.hc_attn_scale.weight: F32
+blk.5.hc_ffn_base.weight: F32
+blk.5.hc_ffn_fn.weight: F32
+blk.5.hc_ffn_scale.weight: F32
+blk.6.attn_kv.weight: Q8_0
+blk.6.attn_kv_a_norm.weight: F32
+blk.6.attn_norm.weight: F32
+blk.6.attn_output_a.weight: Q5_K
+blk.6.attn_output_b.weight: Q5_K
+blk.6.attn_q_a.weight: Q8_0
+blk.6.attn_q_a_norm.weight: F32
+blk.6.attn_q_b.weight: Q5_K
+blk.6.attn_sinks.weight: F32
+blk.6.exp_probs_b.bias: F32
+blk.6.exp_probs_b_vl.bias: F32
+blk.6.ffn_down_exps.weight: IQ2_XXS
+blk.6.ffn_down_shexp.weight: Q5_K
+blk.6.ffn_gate_exps.weight: IQ2_XXS
+blk.6.ffn_gate_inp.weight: BF16
+blk.6.ffn_gate_shexp.weight: Q5_K
+blk.6.ffn_norm.weight: F32
+blk.6.ffn_up_exps.weight: IQ2_XXS
+blk.6.ffn_up_shexp.weight: Q5_K
+blk.6.hc_attn_base.weight: F32
+blk.6.hc_attn_fn.weight: F32
+blk.6.hc_attn_scale.weight: F32
+blk.6.hc_ffn_base.weight: F32
+blk.6.hc_ffn_fn.weight: F32
+blk.6.hc_ffn_scale.weight: F32
+blk.7.attn_kv.weight: Q8_0
+blk.7.attn_kv_a_norm.weight: F32
+blk.7.attn_norm.weight: F32
+blk.7.attn_output_a.weight: Q5_K
+blk.7.attn_output_b.weight: Q5_K
+blk.7.attn_q_a.weight: Q8_0
+blk.7.attn_q_a_norm.weight: F32
+blk.7.attn_q_b.weight: Q5_K
+blk.7.attn_sinks.weight: F32
+blk.7.exp_probs_b.bias: F32
+blk.7.exp_probs_b_vl.bias: F32
+blk.7.ffn_down_exps.weight: IQ2_XXS
+blk.7.ffn_down_shexp.weight: Q5_K
+blk.7.ffn_gate_exps.weight: IQ2_XXS
+blk.7.ffn_gate_inp.weight: BF16
+blk.7.ffn_gate_shexp.weight: Q5_K
+blk.7.ffn_norm.weight: F32
+blk.7.ffn_up_exps.weight: IQ2_XXS
+blk.7.ffn_up_shexp.weight: Q5_K
+blk.7.hc_attn_base.weight: F32
+blk.7.hc_attn_fn.weight: F32
+blk.7.hc_attn_scale.weight: F32
+blk.7.hc_ffn_base.weight: F32
+blk.7.hc_ffn_fn.weight: F32
+blk.7.hc_ffn_scale.weight: F32
+blk.8.attn_compressor_gate.weight: BF16
+blk.8.attn_compressor_kv.weight: BF16
+blk.8.attn_compressor_norm.weight: F32
+blk.8.attn_kv.weight: Q8_0
+blk.8.attn_kv_a_norm.weight: F32
+blk.8.attn_norm.weight: F32
+blk.8.attn_output_a.weight: Q5_K
+blk.8.attn_output_b.weight: Q5_K
+blk.8.attn_q_a.weight: Q8_0
+blk.8.attn_q_a_norm.weight: F32
+blk.8.attn_q_b.weight: Q5_K
+blk.8.attn_sinks.weight: F32
+blk.8.exp_probs_b.bias: F32
+blk.8.exp_probs_b_vl.bias: F32
+blk.8.ffn_down_exps.weight: IQ2_S
+blk.8.ffn_down_shexp.weight: Q5_K
+blk.8.ffn_gate_exps.weight: IQ2_XXS
+blk.8.ffn_gate_inp.weight: BF16
+blk.8.ffn_gate_shexp.weight: Q5_K
+blk.8.ffn_norm.weight: F32
+blk.8.ffn_up_exps.weight: IQ2_XXS
+blk.8.ffn_up_shexp.weight: Q5_K
+blk.8.hc_attn_base.weight: F32
+blk.8.hc_attn_fn.weight: F32
+blk.8.hc_attn_scale.weight: F32
+blk.8.hc_ffn_base.weight: F32
+blk.8.hc_ffn_fn.weight: F32
+blk.8.hc_ffn_scale.weight: F32
+blk.8.indexer.attn_k.weight: BF16
+blk.8.indexer.attn_q_b.weight: BF16
+blk.8.indexer.k_norm.weight: F32
+blk.8.indexer.proj.weight: BF16
+blk.9.attn_kv.weight: Q8_0
+blk.9.attn_kv_a_norm.weight: F32
+blk.9.attn_norm.weight: F32
+blk.9.attn_output_a.weight: Q5_K
+blk.9.attn_output_b.weight: Q5_K
+blk.9.attn_q_a.weight: Q8_0
+blk.9.attn_q_a_norm.weight: F32
+blk.9.attn_q_b.weight: Q5_K
+blk.9.attn_sinks.weight: F32
+blk.9.exp_probs_b.bias: F32
+blk.9.exp_probs_b_vl.bias: F32
+blk.9.ffn_down_exps.weight: IQ2_XS
+blk.9.ffn_down_shexp.weight: Q5_K
+blk.9.ffn_gate_exps.weight: IQ2_XS
+blk.9.ffn_gate_inp.weight: BF16
+blk.9.ffn_gate_shexp.weight: Q5_K
+blk.9.ffn_norm.weight: F32
+blk.9.ffn_up_exps.weight: IQ2_XS
+blk.9.ffn_up_shexp.weight: Q5_K
+blk.9.hc_attn_base.weight: F32
+blk.9.hc_attn_fn.weight: F32
+blk.9.hc_attn_scale.weight: F32
+blk.9.hc_ffn_base.weight: F32
+blk.9.hc_ffn_fn.weight: F32
+blk.9.hc_ffn_scale.weight: F32
+blk.10.attn_kv.weight: Q8_0
+blk.10.attn_kv_a_norm.weight: F32
+blk.10.attn_norm.weight: F32
+blk.10.attn_output_a.weight: Q5_K
+blk.10.attn_output_b.weight: Q5_K
+blk.10.attn_q_a.weight: Q8_0
+blk.10.attn_q_a_norm.weight: F32
+blk.10.attn_q_b.weight: Q5_K
+blk.10.attn_sinks.weight: F32
+blk.10.exp_probs_b.bias: F32
+blk.10.exp_probs_b_vl.bias: F32
+blk.10.ffn_down_exps.weight: IQ2_S
+blk.10.ffn_down_shexp.weight: Q5_K
+blk.10.ffn_gate_exps.weight: IQ2_XXS
+blk.10.ffn_gate_inp.weight: BF16
+blk.10.ffn_gate_shexp.weight: Q5_K
+blk.10.ffn_norm.weight: F32
+blk.10.ffn_up_exps.weight: IQ2_XXS
+blk.10.ffn_up_shexp.weight: Q5_K
+blk.10.hc_attn_base.weight: F32
+blk.10.hc_attn_fn.weight: F32
+blk.10.hc_attn_scale.weight: F32
+blk.10.hc_ffn_base.weight: F32
+blk.10.hc_ffn_fn.weight: F32
+blk.10.hc_ffn_scale.weight: F32
+blk.11.attn_kv.weight: Q8_0
+blk.11.attn_kv_a_norm.weight: F32
+blk.11.attn_norm.weight: F32
+blk.11.attn_output_a.weight: Q5_K
+blk.11.attn_output_b.weight: Q5_K
+blk.11.attn_q_a.weight: Q8_0
+blk.11.attn_q_a_norm.weight: F32
+blk.11.attn_q_b.weight: Q5_K
+blk.11.attn_sinks.weight: F32
+blk.11.exp_probs_b.bias: F32
+blk.11.exp_probs_b_vl.bias: F32
+blk.11.ffn_down_exps.weight: IQ2_XS
+blk.11.ffn_down_shexp.weight: Q5_K
+blk.11.ffn_gate_exps.weight: IQ2_XXS
+blk.11.ffn_gate_inp.weight: BF16
+blk.11.ffn_gate_shexp.weight: Q5_K
+blk.11.ffn_norm.weight: F32
+blk.11.ffn_up_exps.weight: IQ2_XXS
+blk.11.ffn_up_shexp.weight: Q5_K
+blk.11.hc_attn_base.weight: F32
+blk.11.hc_attn_fn.weight: F32
+blk.11.hc_attn_scale.weight: F32
+blk.11.hc_ffn_base.weight: F32
+blk.11.hc_ffn_fn.weight: F32
+blk.11.hc_ffn_scale.weight: F32
+blk.12.attn_kv.weight: Q8_0
+blk.12.attn_kv_a_norm.weight: F32
+blk.12.attn_norm.weight: F32
+blk.12.attn_output_a.weight: Q5_K
+blk.12.attn_output_b.weight: Q5_K
+blk.12.attn_q_a.weight: Q8_0
+blk.12.attn_q_a_norm.weight: F32
+blk.12.attn_q_b.weight: Q5_K
+blk.12.attn_sinks.weight: F32
+blk.12.exp_probs_b.bias: F32
+blk.12.exp_probs_b_vl.bias: F32
+blk.12.ffn_down_exps.weight: IQ2_XS
+blk.12.ffn_down_shexp.weight: Q5_K
+blk.12.ffn_gate_exps.weight: IQ2_XS
+blk.12.ffn_gate_inp.weight: BF16
+blk.12.ffn_gate_shexp.weight: Q5_K
+blk.12.ffn_norm.weight: F32
+blk.12.ffn_up_exps.weight: IQ2_XS
+blk.12.ffn_up_shexp.weight: Q5_K
+blk.12.hc_attn_base.weight: F32
+blk.12.hc_attn_fn.weight: F32
+blk.12.hc_attn_scale.weight: F32
+blk.12.hc_ffn_base.weight: F32
+blk.12.hc_ffn_fn.weight: F32
+blk.12.hc_ffn_scale.weight: F32
+blk.13.attn_kv.weight: Q8_0
+blk.13.attn_kv_a_norm.weight: F32
+blk.13.attn_norm.weight: F32
+blk.13.attn_output_a.weight: Q5_K
+blk.13.attn_output_b.weight: Q5_K
+blk.13.attn_q_a.weight: Q8_0
+blk.13.attn_q_a_norm.weight: F32
+blk.13.attn_q_b.weight: Q5_K
+blk.13.attn_sinks.weight: F32
+blk.13.exp_probs_b.bias: F32
+blk.13.exp_probs_b_vl.bias: F32
+blk.13.ffn_down_exps.weight: IQ2_XXS
+blk.13.ffn_down_shexp.weight: Q5_K
+blk.13.ffn_gate_exps.weight: IQ2_XXS
+blk.13.ffn_gate_inp.weight: BF16
+blk.13.ffn_gate_shexp.weight: Q5_K
+blk.13.ffn_norm.weight: F32
+blk.13.ffn_up_exps.weight: IQ2_XXS
+blk.13.ffn_up_shexp.weight: Q5_K
+blk.13.hc_attn_base.weight: F32
+blk.13.hc_attn_fn.weight: F32
+blk.13.hc_attn_scale.weight: F32
+blk.13.hc_ffn_base.weight: F32
+blk.13.hc_ffn_fn.weight: F32
+blk.13.hc_ffn_scale.weight: F32
+blk.14.attn_compressor_gate.weight: BF16
+blk.14.attn_compressor_kv.weight: BF16
+blk.14.attn_compressor_norm.weight: F32
+blk.14.attn_kv.weight: Q8_0
+blk.14.attn_kv_a_norm.weight: F32
+blk.14.attn_norm.weight: F32
+blk.14.attn_output_a.weight: Q5_K
+blk.14.attn_output_b.weight: Q5_K
+blk.14.attn_q_a.weight: Q8_0
+blk.14.attn_q_a_norm.weight: F32
+blk.14.attn_q_b.weight: Q5_K
+blk.14.attn_sinks.weight: F32
+blk.14.engram_gate.weight: F32
+blk.14.engram_kv.weight: Q5_K
+blk.14.exp_probs_b.bias: F32
+blk.14.exp_probs_b_vl.bias: F32
+blk.14.ffn_down_exps.weight: IQ2_XS
+blk.14.ffn_down_shexp.weight: Q5_K
+blk.14.ffn_gate_exps.weight: IQ2_XXS
+blk.14.ffn_gate_inp.weight: BF16
+blk.14.ffn_gate_shexp.weight: Q5_K
+blk.14.ffn_norm.weight: F32
+blk.14.ffn_up_exps.weight: IQ2_XXS
+blk.14.ffn_up_shexp.weight: Q5_K
+blk.14.hc_attn_base.weight: F32
+blk.14.hc_attn_fn.weight: F32
+blk.14.hc_attn_scale.weight: F32
+blk.14.hc_ffn_base.weight: F32
+blk.14.hc_ffn_fn.weight: F32
+blk.14.hc_ffn_scale.weight: F32
+blk.14.indexer.attn_k.weight: BF16
+blk.14.indexer.attn_q_b.weight: BF16
+blk.14.indexer.k_norm.weight: F32
+blk.14.indexer.proj.weight: BF16
+blk.15.attn_kv.weight: Q8_0
+blk.15.attn_kv_a_norm.weight: F32
+blk.15.attn_norm.weight: F32
+blk.15.attn_output_a.weight: Q5_K
+blk.15.attn_output_b.weight: Q5_K
+blk.15.attn_q_a.weight: Q8_0
+blk.15.attn_q_a_norm.weight: F32
+blk.15.attn_q_b.weight: Q5_K
+blk.15.attn_sinks.weight: F32
+blk.15.exp_probs_b.bias: F32
+blk.15.exp_probs_b_vl.bias: F32
+blk.15.ffn_down_exps.weight: IQ2_XXS
+blk.15.ffn_down_shexp.weight: Q5_K
+blk.15.ffn_gate_exps.weight: IQ2_XXS
+blk.15.ffn_gate_inp.weight: BF16
+blk.15.ffn_gate_shexp.weight: Q5_K
+blk.15.ffn_norm.weight: F32
+blk.15.ffn_up_exps.weight: IQ2_XXS
+blk.15.ffn_up_shexp.weight: Q5_K
+blk.15.hc_attn_base.weight: F32
+blk.15.hc_attn_fn.weight: F32
+blk.15.hc_attn_scale.weight: F32
+blk.15.hc_ffn_base.weight: F32
+blk.15.hc_ffn_fn.weight: F32
+blk.15.hc_ffn_scale.weight: F32
+blk.16.attn_kv.weight: Q8_0
+blk.16.attn_kv_a_norm.weight: F32
+blk.16.attn_norm.weight: F32
+blk.16.attn_output_a.weight: Q5_K
+blk.16.attn_output_b.weight: Q5_K
+blk.16.attn_q_a.weight: Q8_0
+blk.16.attn_q_a_norm.weight: F32
+blk.16.attn_q_b.weight: Q5_K
+blk.16.attn_sinks.weight: F32
+blk.16.exp_probs_b.bias: F32
+blk.16.exp_probs_b_vl.bias: F32
+blk.16.ffn_down_exps.weight: IQ2_S
+blk.16.ffn_down_shexp.weight: Q5_K
+blk.16.ffn_gate_exps.weight: IQ2_XXS
+blk.16.ffn_gate_inp.weight: BF16
+blk.16.ffn_gate_shexp.weight: Q5_K
+blk.16.ffn_norm.weight: F32
+blk.16.ffn_up_exps.weight: IQ2_XXS
+blk.16.ffn_up_shexp.weight: Q5_K
+blk.16.hc_attn_base.weight: F32
+blk.16.hc_attn_fn.weight: F32
+blk.16.hc_attn_scale.weight: F32
+blk.16.hc_ffn_base.weight: F32
+blk.16.hc_ffn_fn.weight: F32
+blk.16.hc_ffn_scale.weight: F32
+blk.17.attn_kv.weight: Q8_0
+blk.17.attn_kv_a_norm.weight: F32
+blk.17.attn_norm.weight: F32
+blk.17.attn_output_a.weight: Q5_K
+blk.17.attn_output_b.weight: Q5_K
+blk.17.attn_q_a.weight: Q8_0
+blk.17.attn_q_a_norm.weight: F32
+blk.17.attn_q_b.weight: Q5_K
+blk.17.attn_sinks.weight: F32
+blk.17.exp_probs_b.bias: F32
+blk.17.exp_probs_b_vl.bias: F32
+blk.17.ffn_down_exps.weight: IQ2_XS
+blk.17.ffn_down_shexp.weight: Q5_K
+blk.17.ffn_gate_exps.weight: IQ2_S
+blk.17.ffn_gate_inp.weight: BF16
+blk.17.ffn_gate_shexp.weight: Q5_K
+blk.17.ffn_norm.weight: F32
+blk.17.ffn_up_exps.weight: IQ2_S
+blk.17.ffn_up_shexp.weight: Q5_K
+blk.17.hc_attn_base.weight: F32
+blk.17.hc_attn_fn.weight: F32
+blk.17.hc_attn_scale.weight: F32
+blk.17.hc_ffn_base.weight: F32
+blk.17.hc_ffn_fn.weight: F32
+blk.17.hc_ffn_scale.weight: F32
+blk.18.attn_kv.weight: Q8_0
+blk.18.attn_kv_a_norm.weight: F32
+blk.18.attn_norm.weight: F32
+blk.18.attn_output_a.weight: Q5_K
+blk.18.attn_output_b.weight: Q5_K
+blk.18.attn_q_a.weight: Q8_0
+blk.18.attn_q_a_norm.weight: F32
+blk.18.attn_q_b.weight: Q5_K
+blk.18.attn_sinks.weight: F32
+blk.18.exp_probs_b.bias: F32
+blk.18.exp_probs_b_vl.bias: F32
+blk.18.ffn_down_exps.weight: IQ2_XS
+blk.18.ffn_down_shexp.weight: Q5_K
+blk.18.ffn_gate_exps.weight: IQ2_XXS
+blk.18.ffn_gate_inp.weight: BF16
+blk.18.ffn_gate_shexp.weight: Q5_K
+blk.18.ffn_norm.weight: F32
+blk.18.ffn_up_exps.weight: IQ2_XXS
+blk.18.ffn_up_shexp.weight: Q5_K
+blk.18.hc_attn_base.weight: F32
+blk.18.hc_attn_fn.weight: F32
+blk.18.hc_attn_scale.weight: F32
+blk.18.hc_ffn_base.weight: F32
+blk.18.hc_ffn_fn.weight: F32
+blk.18.hc_ffn_scale.weight: F32
+blk.19.attn_kv.weight: Q8_0
+blk.19.attn_kv_a_norm.weight: F32
+blk.19.attn_norm.weight: F32
+blk.19.attn_output_a.weight: Q5_K
+blk.19.attn_output_b.weight: Q5_K
+blk.19.attn_q_a.weight: Q8_0
+blk.19.attn_q_a_norm.weight: F32
+blk.19.attn_q_b.weight: Q5_K
+blk.19.attn_sinks.weight: F32
+blk.19.exp_probs_b.bias: F32
+blk.19.exp_probs_b_vl.bias: F32
+blk.19.ffn_down_exps.weight: IQ2_XXS
+blk.19.ffn_down_shexp.weight: Q5_K
+blk.19.ffn_gate_exps.weight: IQ2_XXS
+blk.19.ffn_gate_inp.weight: BF16
+blk.19.ffn_gate_shexp.weight: Q5_K
+blk.19.ffn_norm.weight: F32
+blk.19.ffn_up_exps.weight: IQ2_XXS
+blk.19.ffn_up_shexp.weight: Q5_K
+blk.19.hc_attn_base.weight: F32
+blk.19.hc_attn_fn.weight: F32
+blk.19.hc_attn_scale.weight: F32
+blk.19.hc_ffn_base.weight: F32
+blk.19.hc_ffn_fn.weight: F32
+blk.19.hc_ffn_scale.weight: F32
+blk.20.attn_compressor_kv.weight: BF16
+blk.20.attn_compressor_norm.weight: F32
+blk.20.attn_kv.weight: Q8_0
+blk.20.attn_kv_a_norm.weight: F32
+blk.20.attn_norm.weight: F32
+blk.20.attn_output_a.weight: Q5_K
+blk.20.attn_output_b.weight: Q5_K
+blk.20.attn_q_a.weight: Q8_0
+blk.20.attn_q_a_norm.weight: F32
+blk.20.attn_q_b.weight: Q5_K
+blk.20.attn_sinks.weight: F32
+blk.20.exp_probs_b.bias: F32
+blk.20.exp_probs_b_vl.bias: F32
+blk.20.ffn_down_exps.weight: IQ2_XXS
+blk.20.ffn_down_shexp.weight: Q5_K
+blk.20.ffn_gate_exps.weight: IQ2_XXS
+blk.20.ffn_gate_inp.weight: BF16
+blk.20.ffn_gate_shexp.weight: Q5_K
+blk.20.ffn_norm.weight: F32
+blk.20.ffn_up_exps.weight: IQ2_XXS
+blk.20.ffn_up_shexp.weight: Q5_K
+blk.20.hc_attn_base.weight: F32
+blk.20.hc_attn_fn.weight: F32
+blk.20.hc_attn_scale.weight: F32
+blk.20.hc_ffn_base.weight: F32
+blk.20.hc_ffn_fn.weight: F32
+blk.20.hc_ffn_scale.weight: F32
+blk.20.indexer.attn_k.weight: BF16
+blk.20.indexer.attn_q_b.weight: BF16
+blk.20.indexer.k_norm.weight: F32
+blk.20.indexer.proj.weight: BF16
+blk.21.attn_kv.weight: Q8_0
+blk.21.attn_kv_a_norm.weight: F32
+blk.21.attn_norm.weight: F32
+blk.21.attn_output_a.weight: Q5_K
+blk.21.attn_output_b.weight: Q5_K
+blk.21.attn_q_a.weight: Q8_0
+blk.21.attn_q_a_norm.weight: F32
+blk.21.attn_q_b.weight: Q5_K
+blk.21.attn_sinks.weight: F32
+blk.21.exp_probs_b.bias: F32
+blk.21.exp_probs_b_vl.bias: F32
+blk.21.ffn_down_exps.weight: IQ2_S
+blk.21.ffn_down_shexp.weight: Q5_K
+blk.21.ffn_gate_exps.weight: IQ2_XXS
+blk.21.ffn_gate_inp.weight: BF16
+blk.21.ffn_gate_shexp.weight: Q5_K
+blk.21.ffn_norm.weight: F32
+blk.21.ffn_up_exps.weight: IQ2_XXS
+blk.21.ffn_up_shexp.weight: Q5_K
+blk.21.hc_attn_base.weight: F32
+blk.21.hc_attn_fn.weight: F32
+blk.21.hc_attn_scale.weight: F32
+blk.21.hc_ffn_base.weight: F32
+blk.21.hc_ffn_fn.weight: F32
+blk.21.hc_ffn_scale.weight: F32
+blk.22.attn_kv.weight: Q8_0
+blk.22.attn_kv_a_norm.weight: F32
+blk.22.attn_norm.weight: F32
+blk.22.attn_output_a.weight: Q5_K
+blk.22.attn_output_b.weight: Q5_K
+blk.22.attn_q_a.weight: Q8_0
+blk.22.attn_q_a_norm.weight: F32
+blk.22.attn_q_b.weight: Q5_K
+blk.22.attn_sinks.weight: F32
+blk.22.exp_probs_b.bias: F32
+blk.22.exp_probs_b_vl.bias: F32
+blk.22.ffn_down_exps.weight: IQ2_S
+blk.22.ffn_down_shexp.weight: Q5_K
+blk.22.ffn_gate_exps.weight: IQ2_XXS
+blk.22.ffn_gate_inp.weight: BF16
+blk.22.ffn_gate_shexp.weight: Q5_K
+blk.22.ffn_norm.weight: F32
+blk.22.ffn_up_exps.weight: IQ2_XXS
+blk.22.ffn_up_shexp.weight: Q5_K
+blk.22.hc_attn_base.weight: F32
+blk.22.hc_attn_fn.weight: F32
+blk.22.hc_attn_scale.weight: F32
+blk.22.hc_ffn_base.weight: F32
+blk.22.hc_ffn_fn.weight: F32
+blk.22.hc_ffn_scale.weight: F32
+blk.23.attn_kv.weight: Q8_0
+blk.23.attn_kv_a_norm.weight: F32
+blk.23.attn_norm.weight: F32
+blk.23.attn_output_a.weight: Q5_K
+blk.23.attn_output_b.weight: Q5_K
+blk.23.attn_q_a.weight: Q8_0
+blk.23.attn_q_a_norm.weight: F32
+blk.23.attn_q_b.weight: Q5_K
+blk.23.attn_sinks.weight: F32
+blk.23.exp_probs_b.bias: F32
+blk.23.exp_probs_b_vl.bias: F32
+blk.23.ffn_down_exps.weight: IQ3_XXS
+blk.23.ffn_down_shexp.weight: Q5_K
+blk.23.ffn_gate_exps.weight: IQ2_XS
+blk.23.ffn_gate_inp.weight: BF16
+blk.23.ffn_gate_shexp.weight: Q5_K
+blk.23.ffn_norm.weight: F32
+blk.23.ffn_up_exps.weight: IQ2_XS
+blk.23.ffn_up_shexp.weight: Q5_K
+blk.23.hc_attn_base.weight: F32
+blk.23.hc_attn_fn.weight: F32
+blk.23.hc_attn_scale.weight: F32
+blk.23.hc_ffn_base.weight: F32
+blk.23.hc_ffn_fn.weight: F32
+blk.23.hc_ffn_scale.weight: F32
+blk.24.attn_kv.weight: Q8_0
+blk.24.attn_kv_a_norm.weight: F32
+blk.24.attn_norm.weight: F32
+blk.24.attn_output_a.weight: Q5_K
+blk.24.attn_output_b.weight: Q5_K
+blk.24.attn_q_a.weight: Q8_0
+blk.24.attn_q_a_norm.weight: F32
+blk.24.attn_q_b.weight: Q5_K
+blk.24.attn_sinks.weight: F32
+blk.24.exp_probs_b.bias: F32
+blk.24.exp_probs_b_vl.bias: F32
+blk.24.ffn_down_exps.weight: IQ2_S
+blk.24.ffn_down_shexp.weight: Q5_K
+blk.24.ffn_gate_exps.weight: IQ2_XS
+blk.24.ffn_gate_inp.weight: BF16
+blk.24.ffn_gate_shexp.weight: Q5_K
+blk.24.ffn_norm.weight: F32
+blk.24.ffn_up_exps.weight: IQ2_XS
+blk.24.ffn_up_shexp.weight: Q5_K
+blk.24.hc_attn_base.weight: F32
+blk.24.hc_attn_fn.weight: F32
+blk.24.hc_attn_scale.weight: F32
+blk.24.hc_ffn_base.weight: F32
+blk.24.hc_ffn_fn.weight: F32
+blk.24.hc_ffn_scale.weight: F32
+blk.24.indexer.attn_q_b.weight: BF16
+blk.24.indexer.proj.weight: BF16
+blk.25.attn_kv.weight: Q8_0
+blk.25.attn_kv_a_norm.weight: F32
+blk.25.attn_norm.weight: F32
+blk.25.attn_output_a.weight: Q5_K
+blk.25.attn_output_b.weight: Q5_K
+blk.25.attn_q_a.weight: Q8_0
+blk.25.attn_q_a_norm.weight: F32
+blk.25.attn_q_b.weight: Q5_K
+blk.25.attn_sinks.weight: F32
+blk.25.exp_probs_b.bias: F32
+blk.25.exp_probs_b_vl.bias: F32
+blk.25.ffn_down_exps.weight: IQ2_S
+blk.25.ffn_down_shexp.weight: Q5_K
+blk.25.ffn_gate_exps.weight: IQ2_S
+blk.25.ffn_gate_inp.weight: BF16
+blk.25.ffn_gate_shexp.weight: Q5_K
+blk.25.ffn_norm.weight: F32
+blk.25.ffn_up_exps.weight: IQ2_S
+blk.25.ffn_up_shexp.weight: Q5_K
+blk.25.hc_attn_base.weight: F32
+blk.25.hc_attn_fn.weight: F32
+blk.25.hc_attn_scale.weight: F32
+blk.25.hc_ffn_base.weight: F32
+blk.25.hc_ffn_fn.weight: F32
+blk.25.hc_ffn_scale.weight: F32
+blk.26.attn_kv.weight: Q8_0
+blk.26.attn_kv_a_norm.weight: F32
+blk.26.attn_norm.weight: F32
+blk.26.attn_output_a.weight: Q5_K
+blk.26.attn_output_b.weight: Q5_K
+blk.26.attn_q_a.weight: Q8_0
+blk.26.attn_q_a_norm.weight: F32
+blk.26.attn_q_b.weight: Q5_K
+blk.26.attn_sinks.weight: F32
+blk.26.exp_probs_b.bias: F32
+blk.26.exp_probs_b_vl.bias: F32
+blk.26.ffn_down_exps.weight: IQ2_XS
+blk.26.ffn_down_shexp.weight: Q5_K
+blk.26.ffn_gate_exps.weight: IQ2_XS
+blk.26.ffn_gate_inp.weight: BF16
+blk.26.ffn_gate_shexp.weight: Q5_K
+blk.26.ffn_norm.weight: F32
+blk.26.ffn_up_exps.weight: IQ2_XS
+blk.26.ffn_up_shexp.weight: Q5_K
+blk.26.hc_attn_base.weight: F32
+blk.26.hc_attn_fn.weight: F32
+blk.26.hc_attn_scale.weight: F32
+blk.26.hc_ffn_base.weight: F32
+blk.26.hc_ffn_fn.weight: F32
+blk.26.hc_ffn_scale.weight: F32
+blk.27.attn_kv.weight: Q8_0
+blk.27.attn_kv_a_norm.weight: F32
+blk.27.attn_norm.weight: F32
+blk.27.attn_output_a.weight: Q5_K
+blk.27.attn_output_b.weight: Q5_K
+blk.27.attn_q_a.weight: Q8_0
+blk.27.attn_q_a_norm.weight: F32
+blk.27.attn_q_b.weight: Q5_K
+blk.27.attn_sinks.weight: F32
+blk.27.exp_probs_b.bias: F32
+blk.27.exp_probs_b_vl.bias: F32
+blk.27.ffn_down_exps.weight: IQ2_XS
+blk.27.ffn_down_shexp.weight: Q5_K
+blk.27.ffn_gate_exps.weight: IQ2_XS
+blk.27.ffn_gate_inp.weight: BF16
+blk.27.ffn_gate_shexp.weight: Q5_K
+blk.27.ffn_norm.weight: F32
+blk.27.ffn_up_exps.weight: IQ2_XS
+blk.27.ffn_up_shexp.weight: Q5_K
+blk.27.hc_attn_base.weight: F32
+blk.27.hc_attn_fn.weight: F32
+blk.27.hc_attn_scale.weight: F32
+blk.27.hc_ffn_base.weight: F32
+blk.27.hc_ffn_fn.weight: F32
+blk.27.hc_ffn_scale.weight: F32
+blk.28.attn_kv.weight: Q8_0
+blk.28.attn_kv_a_norm.weight: F32
+blk.28.attn_norm.weight: F32
+blk.28.attn_output_a.weight: Q5_K
+blk.28.attn_output_b.weight: Q5_K
+blk.28.attn_q_a.weight: Q8_0
+blk.28.attn_q_a_norm.weight: F32
+blk.28.attn_q_b.weight: Q5_K
+blk.28.attn_sinks.weight: F32
+blk.28.exp_probs_b.bias: F32
+blk.28.exp_probs_b_vl.bias: F32
+blk.28.ffn_down_exps.weight: IQ2_XS
+blk.28.ffn_down_shexp.weight: Q5_K
+blk.28.ffn_gate_exps.weight: IQ2_XXS
+blk.28.ffn_gate_inp.weight: BF16
+blk.28.ffn_gate_shexp.weight: Q5_K
+blk.28.ffn_norm.weight: F32
+blk.28.ffn_up_exps.weight: IQ2_XXS
+blk.28.ffn_up_shexp.weight: Q5_K
+blk.28.hc_attn_base.weight: F32
+blk.28.hc_attn_fn.weight: F32
+blk.28.hc_attn_scale.weight: F32
+blk.28.hc_ffn_base.weight: F32
+blk.28.hc_ffn_fn.weight: F32
+blk.28.hc_ffn_scale.weight: F32
+blk.28.indexer.attn_q_b.weight: BF16
+blk.28.indexer.proj.weight: BF16
+blk.29.attn_kv.weight: Q8_0
+blk.29.attn_kv_a_norm.weight: F32
+blk.29.attn_norm.weight: F32
+blk.29.attn_output_a.weight: Q5_K
+blk.29.attn_output_b.weight: Q5_K
+blk.29.attn_q_a.weight: Q8_0
+blk.29.attn_q_a_norm.weight: F32
+blk.29.attn_q_b.weight: Q5_K
+blk.29.attn_sinks.weight: F32
+blk.29.exp_probs_b.bias: F32
+blk.29.exp_probs_b_vl.bias: F32
+blk.29.ffn_down_exps.weight: IQ2_XS
+blk.29.ffn_down_shexp.weight: Q5_K
+blk.29.ffn_gate_exps.weight: IQ2_XXS
+blk.29.ffn_gate_inp.weight: BF16
+blk.29.ffn_gate_shexp.weight: Q5_K
+blk.29.ffn_norm.weight: F32
+blk.29.ffn_up_exps.weight: IQ2_XXS
+blk.29.ffn_up_shexp.weight: Q5_K
+blk.29.hc_attn_base.weight: F32
+blk.29.hc_attn_fn.weight: F32
+blk.29.hc_attn_scale.weight: F32
+blk.29.hc_ffn_base.weight: F32
+blk.29.hc_ffn_fn.weight: F32
+blk.29.hc_ffn_scale.weight: F32
+blk.30.attn_kv.weight: Q8_0
+blk.30.attn_kv_a_norm.weight: F32
+blk.30.attn_norm.weight: F32
+blk.30.attn_output_a.weight: Q5_K
+blk.30.attn_output_b.weight: Q5_K
+blk.30.attn_q_a.weight: Q8_0
+blk.30.attn_q_a_norm.weight: F32
+blk.30.attn_q_b.weight: Q5_K
+blk.30.attn_sinks.weight: F32
+blk.30.exp_probs_b.bias: F32
+blk.30.exp_probs_b_vl.bias: F32
+blk.30.ffn_down_exps.weight: IQ2_XXS
+blk.30.ffn_down_shexp.weight: Q5_K
+blk.30.ffn_gate_exps.weight: IQ2_S
+blk.30.ffn_gate_inp.weight: BF16
+blk.30.ffn_gate_shexp.weight: Q5_K
+blk.30.ffn_norm.weight: F32
+blk.30.ffn_up_exps.weight: IQ2_S
+blk.30.ffn_up_shexp.weight: Q5_K
+blk.30.hc_attn_base.weight: F32
+blk.30.hc_attn_fn.weight: F32
+blk.30.hc_attn_scale.weight: F32
+blk.30.hc_ffn_base.weight: F32
+blk.30.hc_ffn_fn.weight: F32
+blk.30.hc_ffn_scale.weight: F32
+blk.31.attn_kv.weight: Q8_0
+blk.31.attn_kv_a_norm.weight: F32
+blk.31.attn_norm.weight: F32
+blk.31.attn_output_a.weight: Q5_K
+blk.31.attn_output_b.weight: Q5_K
+blk.31.attn_q_a.weight: Q8_0
+blk.31.attn_q_a_norm.weight: F32
+blk.31.attn_q_b.weight: Q5_K
+blk.31.attn_sinks.weight: F32
+blk.31.exp_probs_b.bias: F32
+blk.31.exp_probs_b_vl.bias: F32
+blk.31.ffn_down_exps.weight: IQ2_XXS
+blk.31.ffn_down_shexp.weight: Q5_K
+blk.31.ffn_gate_exps.weight: IQ2_XS
+blk.31.ffn_gate_inp.weight: BF16
+blk.31.ffn_gate_shexp.weight: Q5_K
+blk.31.ffn_norm.weight: F32
+blk.31.ffn_up_exps.weight: IQ2_XS
+blk.31.ffn_up_shexp.weight: Q5_K
+blk.31.hc_attn_base.weight: F32
+blk.31.hc_attn_fn.weight: F32
+blk.31.hc_attn_scale.weight: F32
+blk.31.hc_ffn_base.weight: F32
+blk.31.hc_ffn_fn.weight: F32
+blk.31.hc_ffn_scale.weight: F32
+blk.32.attn_kv.weight: Q8_0
+blk.32.attn_kv_a_norm.weight: F32
+blk.32.attn_norm.weight: F32
+blk.32.attn_output_a.weight: Q5_K
+blk.32.attn_output_b.weight: Q5_K
+blk.32.attn_q_a.weight: Q8_0
+blk.32.attn_q_a_norm.weight: F32
+blk.32.attn_q_b.weight: Q5_K
+blk.32.attn_sinks.weight: F32
+blk.32.exp_probs_b.bias: F32
+blk.32.exp_probs_b_vl.bias: F32
+blk.32.ffn_down_exps.weight: IQ2_S
+blk.32.ffn_down_shexp.weight: Q5_K
+blk.32.ffn_gate_exps.weight: IQ2_XXS
+blk.32.ffn_gate_inp.weight: BF16
+blk.32.ffn_gate_shexp.weight: Q5_K
+blk.32.ffn_norm.weight: F32
+blk.32.ffn_up_exps.weight: IQ2_XXS
+blk.32.ffn_up_shexp.weight: Q5_K
+blk.32.hc_attn_base.weight: F32
+blk.32.hc_attn_fn.weight: F32
+blk.32.hc_attn_scale.weight: F32
+blk.32.hc_ffn_base.weight: F32
+blk.32.hc_ffn_fn.weight: F32
+blk.32.hc_ffn_scale.weight: F32
+blk.32.indexer.attn_q_b.weight: BF16
+blk.32.indexer.proj.weight: BF16
+blk.33.attn_kv.weight: Q8_0
+blk.33.attn_kv_a_norm.weight: F32
+blk.33.attn_norm.weight: F32
+blk.33.attn_output_a.weight: Q5_K
+blk.33.attn_output_b.weight: Q5_K
+blk.33.attn_q_a.weight: Q8_0
+blk.33.attn_q_a_norm.weight: F32
+blk.33.attn_q_b.weight: Q5_K
+blk.33.attn_sinks.weight: F32
+blk.33.exp_probs_b.bias: F32
+blk.33.exp_probs_b_vl.bias: F32
+blk.33.ffn_down_exps.weight: IQ2_XXS
+blk.33.ffn_down_shexp.weight: Q5_K
+blk.33.ffn_gate_exps.weight: IQ2_XXS
+blk.33.ffn_gate_inp.weight: BF16
+blk.33.ffn_gate_shexp.weight: Q5_K
+blk.33.ffn_norm.weight: F32
+blk.33.ffn_up_exps.weight: IQ2_XXS
+blk.33.ffn_up_shexp.weight: Q5_K
+blk.33.hc_attn_base.weight: F32
+blk.33.hc_attn_fn.weight: F32
+blk.33.hc_attn_scale.weight: F32
+blk.33.hc_ffn_base.weight: F32
+blk.33.hc_ffn_fn.weight: F32
+blk.33.hc_ffn_scale.weight: F32
+blk.34.attn_kv.weight: Q8_0
+blk.34.attn_kv_a_norm.weight: F32
+blk.34.attn_norm.weight: F32
+blk.34.attn_output_a.weight: Q5_K
+blk.34.attn_output_b.weight: Q5_K
+blk.34.attn_q_a.weight: Q8_0
+blk.34.attn_q_a_norm.weight: F32
+blk.34.attn_q_b.weight: Q5_K
+blk.34.attn_sinks.weight: F32
+blk.34.exp_probs_b.bias: F32
+blk.34.exp_probs_b_vl.bias: F32
+blk.34.ffn_down_exps.weight: IQ2_XS
+blk.34.ffn_down_shexp.weight: Q5_K
+blk.34.ffn_gate_exps.weight: IQ2_XXS
+blk.34.ffn_gate_inp.weight: BF16
+blk.34.ffn_gate_shexp.weight: Q5_K
+blk.34.ffn_norm.weight: F32
+blk.34.ffn_up_exps.weight: IQ2_XXS
+blk.34.ffn_up_shexp.weight: Q5_K
+blk.34.hc_attn_base.weight: F32
+blk.34.hc_attn_fn.weight: F32
+blk.34.hc_attn_scale.weight: F32
+blk.34.hc_ffn_base.weight: F32
+blk.34.hc_ffn_fn.weight: F32
+blk.34.hc_ffn_scale.weight: F32
+blk.35.attn_kv.weight: Q8_0
+blk.35.attn_kv_a_norm.weight: F32
+blk.35.attn_norm.weight: F32
+blk.35.attn_output_a.weight: Q5_K
+blk.35.attn_output_b.weight: Q5_K
+blk.35.attn_q_a.weight: Q8_0
+blk.35.attn_q_a_norm.weight: F32
+blk.35.attn_q_b.weight: Q5_K
+blk.35.attn_sinks.weight: F32
+blk.35.exp_probs_b.bias: F32
+blk.35.exp_probs_b_vl.bias: F32
+blk.35.ffn_down_exps.weight: IQ2_XXS
+blk.35.ffn_down_shexp.weight: Q5_K
+blk.35.ffn_gate_exps.weight: IQ2_XXS
+blk.35.ffn_gate_inp.weight: BF16
+blk.35.ffn_gate_shexp.weight: Q5_K
+blk.35.ffn_norm.weight: F32
+blk.35.ffn_up_exps.weight: IQ2_XXS
+blk.35.ffn_up_shexp.weight: Q5_K
+blk.35.hc_attn_base.weight: F32
+blk.35.hc_attn_fn.weight: F32
+blk.35.hc_attn_scale.weight: F32
+blk.35.hc_ffn_base.weight: F32
+blk.35.hc_ffn_fn.weight: F32
+blk.35.hc_ffn_scale.weight: F32
+blk.36.attn_kv.weight: Q8_0
+blk.36.attn_kv_a_norm.weight: F32
+blk.36.attn_norm.weight: F32
+blk.36.attn_output_a.weight: Q5_K
+blk.36.attn_output_b.weight: Q5_K
+blk.36.attn_q_a.weight: Q8_0
+blk.36.attn_q_a_norm.weight: F32
+blk.36.attn_q_b.weight: Q5_K
+blk.36.attn_sinks.weight: F32
+blk.36.exp_probs_b.bias: F32
+blk.36.exp_probs_b_vl.bias: F32
+blk.36.ffn_down_exps.weight: IQ2_XXS
+blk.36.ffn_down_shexp.weight: Q5_K
+blk.36.ffn_gate_exps.weight: IQ2_XXS
+blk.36.ffn_gate_inp.weight: BF16
+blk.36.ffn_gate_shexp.weight: Q5_K
+blk.36.ffn_norm.weight: F32
+blk.36.ffn_up_exps.weight: IQ2_XXS
+blk.36.ffn_up_shexp.weight: Q5_K
+blk.36.hc_attn_base.weight: F32
+blk.36.hc_attn_fn.weight: F32
+blk.36.hc_attn_scale.weight: F32
+blk.36.hc_ffn_base.weight: F32
+blk.36.hc_ffn_fn.weight: F32
+blk.36.hc_ffn_scale.weight: F32
+blk.36.indexer.attn_q_b.weight: BF16
+blk.36.indexer.proj.weight: BF16
+blk.37.attn_kv.weight: Q8_0
+blk.37.attn_kv_a_norm.weight: F32
+blk.37.attn_norm.weight: F32
+blk.37.attn_output_a.weight: Q5_K
+blk.37.attn_output_b.weight: Q5_K
+blk.37.attn_q_a.weight: Q8_0
+blk.37.attn_q_a_norm.weight: F32
+blk.37.attn_q_b.weight: Q5_K
+blk.37.attn_sinks.weight: F32
+blk.37.exp_probs_b.bias: F32
+blk.37.exp_probs_b_vl.bias: F32
+blk.37.ffn_down_exps.weight: IQ2_XXS
+blk.37.ffn_down_shexp.weight: Q5_K
+blk.37.ffn_gate_exps.weight: IQ2_XXS
+blk.37.ffn_gate_inp.weight: BF16
+blk.37.ffn_gate_shexp.weight: Q5_K
+blk.37.ffn_norm.weight: F32
+blk.37.ffn_up_exps.weight: IQ2_XXS
+blk.37.ffn_up_shexp.weight: Q5_K
+blk.37.hc_attn_base.weight: F32
+blk.37.hc_attn_fn.weight: F32
+blk.37.hc_attn_scale.weight: F32
+blk.37.hc_ffn_base.weight: F32
+blk.37.hc_ffn_fn.weight: F32
+blk.37.hc_ffn_scale.weight: F32
+blk.38.attn_kv.weight: Q8_0
+blk.38.attn_kv_a_norm.weight: F32
+blk.38.attn_norm.weight: F32
+blk.38.attn_output_a.weight: Q5_K
+blk.38.attn_output_b.weight: Q5_K
+blk.38.attn_q_a.weight: Q8_0
+blk.38.attn_q_a_norm.weight: F32
+blk.38.attn_q_b.weight: Q5_K
+blk.38.attn_sinks.weight: F32
+blk.38.exp_probs_b.bias: F32
+blk.38.exp_probs_b_vl.bias: F32
+blk.38.ffn_down_exps.weight: IQ2_S
+blk.38.ffn_down_shexp.weight: Q5_K
+blk.38.ffn_gate_exps.weight: IQ2_XS
+blk.38.ffn_gate_inp.weight: BF16
+blk.38.ffn_gate_shexp.weight: Q5_K
+blk.38.ffn_norm.weight: F32
+blk.38.ffn_up_exps.weight: IQ2_XS
+blk.38.ffn_up_shexp.weight: Q5_K
+blk.38.hc_attn_base.weight: F32
+blk.38.hc_attn_fn.weight: F32
+blk.38.hc_attn_scale.weight: F32
+blk.38.hc_ffn_base.weight: F32
+blk.38.hc_ffn_fn.weight: F32
+blk.38.hc_ffn_scale.weight: F32
+blk.39.attn_kv.weight: Q8_0
+blk.39.attn_kv_a_norm.weight: F32
+blk.39.attn_norm.weight: F32
+blk.39.attn_output_a.weight: Q5_K
+blk.39.attn_output_b.weight: Q5_K
+blk.39.attn_q_a.weight: Q8_0
+blk.39.attn_q_a_norm.weight: F32
+blk.39.attn_q_b.weight: Q5_K
+blk.39.attn_sinks.weight: F32
+blk.39.exp_probs_b.bias: F32
+blk.39.exp_probs_b_vl.bias: F32
+blk.39.ffn_down_exps.weight: IQ2_S
+blk.39.ffn_down_shexp.weight: Q5_K
+blk.39.ffn_gate_exps.weight: IQ2_XXS
+blk.39.ffn_gate_inp.weight: BF16
+blk.39.ffn_gate_shexp.weight: Q5_K
+blk.39.ffn_norm.weight: F32
+blk.39.ffn_up_exps.weight: IQ2_XXS
+blk.39.ffn_up_shexp.weight: Q5_K
+blk.39.hc_attn_base.weight: F32
+blk.39.hc_attn_fn.weight: F32
+blk.39.hc_attn_scale.weight: F32
+blk.39.hc_ffn_base.weight: F32
+blk.39.hc_ffn_fn.weight: F32
+blk.39.hc_ffn_scale.weight: F32

tensor-allocation/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS.rco-allocation.txt

@@ -0,0 +1,1061 @@
+# RCO per-tensor GGUF allocation
+# Repository: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
+# GGUF: DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf (first of 6 shards; this dump covers all of them)
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00002-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00003-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00005-of-00006.gguf
+# Source: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF/resolve/main/DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf
+# Target bpw over the backbone (all tensors except the Engram table): 2.00 (measured: 2.0038)
+# The Engram table (per_layer_token_embd, shard 6 = the shared Engram file) is IQ4_NL in every build and not part of the search.
+# GGUF version: 3
+# GGUF metadata entries: 77
+# Tensor count: 1043
+# Quant-type counts: BF16=67, F32=531, IQ1_M=60, IQ2_XS=18, IQ2_XXS=42, IQ4_NL=1, Q5_K=244, Q8_0=80
+# Format: <GGUF tensor name>: <GGML quantization/storage type>
+# Includes fixed non-RCO tensors (e.g. F32/BF16) so the released GGUF is fully reproducible.
+#
+output.weight: Q5_K
+output_norm.weight: F32
+per_layer_token_embd.weight: IQ4_NL
+token_embd.weight: Q5_K
+blk.0.attn_kv.weight: Q8_0
+blk.0.attn_kv_a_norm.weight: F32
+blk.0.attn_norm.weight: F32
+blk.0.attn_output_a.weight: Q5_K
+blk.0.attn_output_b.weight: Q5_K
+blk.0.attn_q_a.weight: Q8_0
+blk.0.attn_q_a_norm.weight: F32
+blk.0.attn_q_b.weight: Q5_K
+blk.0.attn_sinks.weight: F32
+blk.0.exp_probs_b.bias: F32
+blk.0.exp_probs_b_vl.bias: F32
+blk.0.ffn_down_exps.weight: IQ1_M
+blk.0.ffn_down_shexp.weight: Q5_K
+blk.0.ffn_gate_exps.weight: IQ1_M
+blk.0.ffn_gate_inp.weight: BF16
+blk.0.ffn_gate_shexp.weight: Q5_K
+blk.0.ffn_norm.weight: F32
+blk.0.ffn_up_exps.weight: IQ1_M
+blk.0.ffn_up_shexp.weight: Q5_K
+blk.0.hc_attn_base.weight: F32
+blk.0.hc_attn_fn.weight: F32
+blk.0.hc_attn_scale.weight: F32
+blk.0.hc_ffn_base.weight: F32
+blk.0.hc_ffn_fn.weight: F32
+blk.0.hc_ffn_scale.weight: F32
+blk.1.attn_kv.weight: Q8_0
+blk.1.attn_kv_a_norm.weight: F32
+blk.1.attn_norm.weight: F32
+blk.1.attn_output_a.weight: Q5_K
+blk.1.attn_output_b.weight: Q5_K
+blk.1.attn_q_a.weight: Q8_0
+blk.1.attn_q_a_norm.weight: F32
+blk.1.attn_q_b.weight: Q5_K
+blk.1.attn_sinks.weight: F32
+blk.1.engram_gate.weight: F32
+blk.1.engram_kv.weight: Q5_K
+blk.1.exp_probs_b.bias: F32
+blk.1.exp_probs_b_vl.bias: F32
+blk.1.ffn_down_exps.weight: IQ1_M
+blk.1.ffn_down_shexp.weight: Q5_K
+blk.1.ffn_gate_exps.weight: IQ1_M
+blk.1.ffn_gate_inp.weight: BF16
+blk.1.ffn_gate_shexp.weight: Q5_K
+blk.1.ffn_norm.weight: F32
+blk.1.ffn_up_exps.weight: IQ1_M
+blk.1.ffn_up_shexp.weight: Q5_K
+blk.1.hc_attn_base.weight: F32
+blk.1.hc_attn_fn.weight: F32
+blk.1.hc_attn_scale.weight: F32
+blk.1.hc_ffn_base.weight: F32
+blk.1.hc_ffn_fn.weight: F32
+blk.1.hc_ffn_scale.weight: F32
+blk.2.attn_compressor_gate.weight: BF16
+blk.2.attn_compressor_kv.weight: BF16
+blk.2.attn_compressor_norm.weight: F32
+blk.2.attn_kv.weight: Q8_0
+blk.2.attn_kv_a_norm.weight: F32
+blk.2.attn_norm.weight: F32
+blk.2.attn_output_a.weight: Q5_K
+blk.2.attn_output_b.weight: Q5_K
+blk.2.attn_q_a.weight: Q8_0
+blk.2.attn_q_a_norm.weight: F32
+blk.2.attn_q_b.weight: Q5_K
+blk.2.attn_sinks.weight: F32
+blk.2.exp_probs_b.bias: F32
+blk.2.exp_probs_b_vl.bias: F32
+blk.2.ffn_down_exps.weight: IQ1_M
+blk.2.ffn_down_shexp.weight: Q5_K
+blk.2.ffn_gate_exps.weight: IQ1_M
+blk.2.ffn_gate_inp.weight: BF16
+blk.2.ffn_gate_shexp.weight: Q5_K
+blk.2.ffn_norm.weight: F32
+blk.2.ffn_up_exps.weight: IQ1_M
+blk.2.ffn_up_shexp.weight: Q5_K
+blk.2.hc_attn_base.weight: F32
+blk.2.hc_attn_fn.weight: F32
+blk.2.hc_attn_scale.weight: F32
+blk.2.hc_ffn_base.weight: F32
+blk.2.hc_ffn_fn.weight: F32
+blk.2.hc_ffn_scale.weight: F32
+blk.2.indexer.attn_k.weight: BF16
+blk.2.indexer.attn_q_b.weight: BF16
+blk.2.indexer.k_norm.weight: F32
+blk.2.indexer.proj.weight: BF16
+blk.3.attn_kv.weight: Q8_0
+blk.3.attn_kv_a_norm.weight: F32
+blk.3.attn_norm.weight: F32
+blk.3.attn_output_a.weight: Q5_K
+blk.3.attn_output_b.weight: Q5_K
+blk.3.attn_q_a.weight: Q8_0
+blk.3.attn_q_a_norm.weight: F32
+blk.3.attn_q_b.weight: Q5_K
+blk.3.attn_sinks.weight: F32
+blk.3.exp_probs_b.bias: F32
+blk.3.exp_probs_b_vl.bias: F32
+blk.3.ffn_down_exps.weight: IQ1_M
+blk.3.ffn_down_shexp.weight: Q5_K
+blk.3.ffn_gate_exps.weight: IQ1_M
+blk.3.ffn_gate_inp.weight: BF16
+blk.3.ffn_gate_shexp.weight: Q5_K
+blk.3.ffn_norm.weight: F32
+blk.3.ffn_up_exps.weight: IQ1_M
+blk.3.ffn_up_shexp.weight: Q5_K
+blk.3.hc_attn_base.weight: F32
+blk.3.hc_attn_fn.weight: F32
+blk.3.hc_attn_scale.weight: F32
+blk.3.hc_ffn_base.weight: F32
+blk.3.hc_ffn_fn.weight: F32
+blk.3.hc_ffn_scale.weight: F32
+blk.4.attn_kv.weight: Q8_0
+blk.4.attn_kv_a_norm.weight: F32
+blk.4.attn_norm.weight: F32
+blk.4.attn_output_a.weight: Q5_K
+blk.4.attn_output_b.weight: Q5_K
+blk.4.attn_q_a.weight: Q8_0
+blk.4.attn_q_a_norm.weight: F32
+blk.4.attn_q_b.weight: Q5_K
+blk.4.attn_sinks.weight: F32
+blk.4.exp_probs_b.bias: F32
+blk.4.exp_probs_b_vl.bias: F32
+blk.4.ffn_down_exps.weight: IQ1_M
+blk.4.ffn_down_shexp.weight: Q5_K
+blk.4.ffn_gate_exps.weight: IQ1_M
+blk.4.ffn_gate_inp.weight: BF16
+blk.4.ffn_gate_shexp.weight: Q5_K
+blk.4.ffn_norm.weight: F32
+blk.4.ffn_up_exps.weight: IQ1_M
+blk.4.ffn_up_shexp.weight: Q5_K
+blk.4.hc_attn_base.weight: F32
+blk.4.hc_attn_fn.weight: F32
+blk.4.hc_attn_scale.weight: F32
+blk.4.hc_ffn_base.weight: F32
+blk.4.hc_ffn_fn.weight: F32
+blk.4.hc_ffn_scale.weight: F32
+blk.5.attn_kv.weight: Q8_0
+blk.5.attn_kv_a_norm.weight: F32
+blk.5.attn_norm.weight: F32
+blk.5.attn_output_a.weight: Q5_K
+blk.5.attn_output_b.weight: Q5_K
+blk.5.attn_q_a.weight: Q8_0
+blk.5.attn_q_a_norm.weight: F32
+blk.5.attn_q_b.weight: Q5_K
+blk.5.attn_sinks.weight: F32
+blk.5.exp_probs_b.bias: F32
+blk.5.exp_probs_b_vl.bias: F32
+blk.5.ffn_down_exps.weight: IQ1_M
+blk.5.ffn_down_shexp.weight: Q5_K
+blk.5.ffn_gate_exps.weight: IQ1_M
+blk.5.ffn_gate_inp.weight: BF16
+blk.5.ffn_gate_shexp.weight: Q5_K
+blk.5.ffn_norm.weight: F32
+blk.5.ffn_up_exps.weight: IQ1_M
+blk.5.ffn_up_shexp.weight: Q5_K
+blk.5.hc_attn_base.weight: F32
+blk.5.hc_attn_fn.weight: F32
+blk.5.hc_attn_scale.weight: F32
+blk.5.hc_ffn_base.weight: F32
+blk.5.hc_ffn_fn.weight: F32
+blk.5.hc_ffn_scale.weight: F32
+blk.6.attn_kv.weight: Q8_0
+blk.6.attn_kv_a_norm.weight: F32
+blk.6.attn_norm.weight: F32
+blk.6.attn_output_a.weight: Q5_K
+blk.6.attn_output_b.weight: Q5_K
+blk.6.attn_q_a.weight: Q8_0
+blk.6.attn_q_a_norm.weight: F32
+blk.6.attn_q_b.weight: Q5_K
+blk.6.attn_sinks.weight: F32
+blk.6.exp_probs_b.bias: F32
+blk.6.exp_probs_b_vl.bias: F32
+blk.6.ffn_down_exps.weight: IQ1_M
+blk.6.ffn_down_shexp.weight: Q5_K
+blk.6.ffn_gate_exps.weight: IQ1_M
+blk.6.ffn_gate_inp.weight: BF16
+blk.6.ffn_gate_shexp.weight: Q5_K
+blk.6.ffn_norm.weight: F32
+blk.6.ffn_up_exps.weight: IQ1_M
+blk.6.ffn_up_shexp.weight: Q5_K
+blk.6.hc_attn_base.weight: F32
+blk.6.hc_attn_fn.weight: F32
+blk.6.hc_attn_scale.weight: F32
+blk.6.hc_ffn_base.weight: F32
+blk.6.hc_ffn_fn.weight: F32
+blk.6.hc_ffn_scale.weight: F32
+blk.7.attn_kv.weight: Q8_0
+blk.7.attn_kv_a_norm.weight: F32
+blk.7.attn_norm.weight: F32
+blk.7.attn_output_a.weight: Q5_K
+blk.7.attn_output_b.weight: Q5_K
+blk.7.attn_q_a.weight: Q8_0
+blk.7.attn_q_a_norm.weight: F32
+blk.7.attn_q_b.weight: Q5_K
+blk.7.attn_sinks.weight: F32
+blk.7.exp_probs_b.bias: F32
+blk.7.exp_probs_b_vl.bias: F32
+blk.7.ffn_down_exps.weight: IQ2_XXS
+blk.7.ffn_down_shexp.weight: Q5_K
+blk.7.ffn_gate_exps.weight: IQ1_M
+blk.7.ffn_gate_inp.weight: BF16
+blk.7.ffn_gate_shexp.weight: Q5_K
+blk.7.ffn_norm.weight: F32
+blk.7.ffn_up_exps.weight: IQ1_M
+blk.7.ffn_up_shexp.weight: Q5_K
+blk.7.hc_attn_base.weight: F32
+blk.7.hc_attn_fn.weight: F32
+blk.7.hc_attn_scale.weight: F32
+blk.7.hc_ffn_base.weight: F32
+blk.7.hc_ffn_fn.weight: F32
+blk.7.hc_ffn_scale.weight: F32
+blk.8.attn_compressor_gate.weight: BF16
+blk.8.attn_compressor_kv.weight: BF16
+blk.8.attn_compressor_norm.weight: F32
+blk.8.attn_kv.weight: Q8_0
+blk.8.attn_kv_a_norm.weight: F32
+blk.8.attn_norm.weight: F32
+blk.8.attn_output_a.weight: Q5_K
+blk.8.attn_output_b.weight: Q5_K
+blk.8.attn_q_a.weight: Q8_0
+blk.8.attn_q_a_norm.weight: F32
+blk.8.attn_q_b.weight: Q5_K
+blk.8.attn_sinks.weight: F32
+blk.8.exp_probs_b.bias: F32
+blk.8.exp_probs_b_vl.bias: F32
+blk.8.ffn_down_exps.weight: IQ2_XXS
+blk.8.ffn_down_shexp.weight: Q5_K
+blk.8.ffn_gate_exps.weight: IQ1_M
+blk.8.ffn_gate_inp.weight: BF16
+blk.8.ffn_gate_shexp.weight: Q5_K
+blk.8.ffn_norm.weight: F32
+blk.8.ffn_up_exps.weight: IQ1_M
+blk.8.ffn_up_shexp.weight: Q5_K
+blk.8.hc_attn_base.weight: F32
+blk.8.hc_attn_fn.weight: F32
+blk.8.hc_attn_scale.weight: F32
+blk.8.hc_ffn_base.weight: F32
+blk.8.hc_ffn_fn.weight: F32
+blk.8.hc_ffn_scale.weight: F32
+blk.8.indexer.attn_k.weight: BF16
+blk.8.indexer.attn_q_b.weight: BF16
+blk.8.indexer.k_norm.weight: F32
+blk.8.indexer.proj.weight: BF16
+blk.9.attn_kv.weight: Q8_0
+blk.9.attn_kv_a_norm.weight: F32
+blk.9.attn_norm.weight: F32
+blk.9.attn_output_a.weight: Q5_K
+blk.9.attn_output_b.weight: Q5_K
+blk.9.attn_q_a.weight: Q8_0
+blk.9.attn_q_a_norm.weight: F32
+blk.9.attn_q_b.weight: Q5_K
+blk.9.attn_sinks.weight: F32
+blk.9.exp_probs_b.bias: F32
+blk.9.exp_probs_b_vl.bias: F32
+blk.9.ffn_down_exps.weight: IQ1_M
+blk.9.ffn_down_shexp.weight: Q5_K
+blk.9.ffn_gate_exps.weight: IQ1_M
+blk.9.ffn_gate_inp.weight: BF16
+blk.9.ffn_gate_shexp.weight: Q5_K
+blk.9.ffn_norm.weight: F32
+blk.9.ffn_up_exps.weight: IQ1_M
+blk.9.ffn_up_shexp.weight: Q5_K
+blk.9.hc_attn_base.weight: F32
+blk.9.hc_attn_fn.weight: F32
+blk.9.hc_attn_scale.weight: F32
+blk.9.hc_ffn_base.weight: F32
+blk.9.hc_ffn_fn.weight: F32
+blk.9.hc_ffn_scale.weight: F32
+blk.10.attn_kv.weight: Q8_0
+blk.10.attn_kv_a_norm.weight: F32
+blk.10.attn_norm.weight: F32
+blk.10.attn_output_a.weight: Q5_K
+blk.10.attn_output_b.weight: Q5_K
+blk.10.attn_q_a.weight: Q8_0
+blk.10.attn_q_a_norm.weight: F32
+blk.10.attn_q_b.weight: Q5_K
+blk.10.attn_sinks.weight: F32
+blk.10.exp_probs_b.bias: F32
+blk.10.exp_probs_b_vl.bias: F32
+blk.10.ffn_down_exps.weight: IQ2_XXS
+blk.10.ffn_down_shexp.weight: Q5_K
+blk.10.ffn_gate_exps.weight: IQ2_XXS
+blk.10.ffn_gate_inp.weight: BF16
+blk.10.ffn_gate_shexp.weight: Q5_K
+blk.10.ffn_norm.weight: F32
+blk.10.ffn_up_exps.weight: IQ2_XXS
+blk.10.ffn_up_shexp.weight: Q5_K
+blk.10.hc_attn_base.weight: F32
+blk.10.hc_attn_fn.weight: F32
+blk.10.hc_attn_scale.weight: F32
+blk.10.hc_ffn_base.weight: F32
+blk.10.hc_ffn_fn.weight: F32
+blk.10.hc_ffn_scale.weight: F32
+blk.11.attn_kv.weight: Q8_0
+blk.11.attn_kv_a_norm.weight: F32
+blk.11.attn_norm.weight: F32
+blk.11.attn_output_a.weight: Q5_K
+blk.11.attn_output_b.weight: Q5_K
+blk.11.attn_q_a.weight: Q8_0
+blk.11.attn_q_a_norm.weight: F32
+blk.11.attn_q_b.weight: Q5_K
+blk.11.attn_sinks.weight: F32
+blk.11.exp_probs_b.bias: F32
+blk.11.exp_probs_b_vl.bias: F32
+blk.11.ffn_down_exps.weight: IQ1_M
+blk.11.ffn_down_shexp.weight: Q5_K
+blk.11.ffn_gate_exps.weight: IQ2_XXS
+blk.11.ffn_gate_inp.weight: BF16
+blk.11.ffn_gate_shexp.weight: Q5_K
+blk.11.ffn_norm.weight: F32
+blk.11.ffn_up_exps.weight: IQ2_XXS
+blk.11.ffn_up_shexp.weight: Q5_K
+blk.11.hc_attn_base.weight: F32
+blk.11.hc_attn_fn.weight: F32
+blk.11.hc_attn_scale.weight: F32
+blk.11.hc_ffn_base.weight: F32
+blk.11.hc_ffn_fn.weight: F32
+blk.11.hc_ffn_scale.weight: F32
+blk.12.attn_kv.weight: Q8_0
+blk.12.attn_kv_a_norm.weight: F32
+blk.12.attn_norm.weight: F32
+blk.12.attn_output_a.weight: Q5_K
+blk.12.attn_output_b.weight: Q5_K
+blk.12.attn_q_a.weight: Q8_0
+blk.12.attn_q_a_norm.weight: F32
+blk.12.attn_q_b.weight: Q5_K
+blk.12.attn_sinks.weight: F32
+blk.12.exp_probs_b.bias: F32
+blk.12.exp_probs_b_vl.bias: F32
+blk.12.ffn_down_exps.weight: IQ1_M
+blk.12.ffn_down_shexp.weight: Q5_K
+blk.12.ffn_gate_exps.weight: IQ2_XXS
+blk.12.ffn_gate_inp.weight: BF16
+blk.12.ffn_gate_shexp.weight: Q5_K
+blk.12.ffn_norm.weight: F32
+blk.12.ffn_up_exps.weight: IQ2_XXS
+blk.12.ffn_up_shexp.weight: Q5_K
+blk.12.hc_attn_base.weight: F32
+blk.12.hc_attn_fn.weight: F32
+blk.12.hc_attn_scale.weight: F32
+blk.12.hc_ffn_base.weight: F32
+blk.12.hc_ffn_fn.weight: F32
+blk.12.hc_ffn_scale.weight: F32
+blk.13.attn_kv.weight: Q8_0
+blk.13.attn_kv_a_norm.weight: F32
+blk.13.attn_norm.weight: F32
+blk.13.attn_output_a.weight: Q5_K
+blk.13.attn_output_b.weight: Q5_K
+blk.13.attn_q_a.weight: Q8_0
+blk.13.attn_q_a_norm.weight: F32
+blk.13.attn_q_b.weight: Q5_K
+blk.13.attn_sinks.weight: F32
+blk.13.exp_probs_b.bias: F32
+blk.13.exp_probs_b_vl.bias: F32
+blk.13.ffn_down_exps.weight: IQ1_M
+blk.13.ffn_down_shexp.weight: Q5_K
+blk.13.ffn_gate_exps.weight: IQ1_M
+blk.13.ffn_gate_inp.weight: BF16
+blk.13.ffn_gate_shexp.weight: Q5_K
+blk.13.ffn_norm.weight: F32
+blk.13.ffn_up_exps.weight: IQ1_M
+blk.13.ffn_up_shexp.weight: Q5_K
+blk.13.hc_attn_base.weight: F32
+blk.13.hc_attn_fn.weight: F32
+blk.13.hc_attn_scale.weight: F32
+blk.13.hc_ffn_base.weight: F32
+blk.13.hc_ffn_fn.weight: F32
+blk.13.hc_ffn_scale.weight: F32
+blk.14.attn_compressor_gate.weight: BF16
+blk.14.attn_compressor_kv.weight: BF16
+blk.14.attn_compressor_norm.weight: F32
+blk.14.attn_kv.weight: Q8_0
+blk.14.attn_kv_a_norm.weight: F32
+blk.14.attn_norm.weight: F32
+blk.14.attn_output_a.weight: Q5_K
+blk.14.attn_output_b.weight: Q5_K
+blk.14.attn_q_a.weight: Q8_0
+blk.14.attn_q_a_norm.weight: F32
+blk.14.attn_q_b.weight: Q5_K
+blk.14.attn_sinks.weight: F32
+blk.14.engram_gate.weight: F32
+blk.14.engram_kv.weight: Q5_K
+blk.14.exp_probs_b.bias: F32
+blk.14.exp_probs_b_vl.bias: F32
+blk.14.ffn_down_exps.weight: IQ1_M
+blk.14.ffn_down_shexp.weight: Q5_K
+blk.14.ffn_gate_exps.weight: IQ1_M
+blk.14.ffn_gate_inp.weight: BF16
+blk.14.ffn_gate_shexp.weight: Q5_K
+blk.14.ffn_norm.weight: F32
+blk.14.ffn_up_exps.weight: IQ1_M
+blk.14.ffn_up_shexp.weight: Q5_K
+blk.14.hc_attn_base.weight: F32
+blk.14.hc_attn_fn.weight: F32
+blk.14.hc_attn_scale.weight: F32
+blk.14.hc_ffn_base.weight: F32
+blk.14.hc_ffn_fn.weight: F32
+blk.14.hc_ffn_scale.weight: F32
+blk.14.indexer.attn_k.weight: BF16
+blk.14.indexer.attn_q_b.weight: BF16
+blk.14.indexer.k_norm.weight: F32
+blk.14.indexer.proj.weight: BF16
+blk.15.attn_kv.weight: Q8_0
+blk.15.attn_kv_a_norm.weight: F32
+blk.15.attn_norm.weight: F32
+blk.15.attn_output_a.weight: Q5_K
+blk.15.attn_output_b.weight: Q5_K
+blk.15.attn_q_a.weight: Q8_0
+blk.15.attn_q_a_norm.weight: F32
+blk.15.attn_q_b.weight: Q5_K
+blk.15.attn_sinks.weight: F32
+blk.15.exp_probs_b.bias: F32
+blk.15.exp_probs_b_vl.bias: F32
+blk.15.ffn_down_exps.weight: IQ1_M
+blk.15.ffn_down_shexp.weight: Q5_K
+blk.15.ffn_gate_exps.weight: IQ2_XXS
+blk.15.ffn_gate_inp.weight: BF16
+blk.15.ffn_gate_shexp.weight: Q5_K
+blk.15.ffn_norm.weight: F32
+blk.15.ffn_up_exps.weight: IQ2_XXS
+blk.15.ffn_up_shexp.weight: Q5_K
+blk.15.hc_attn_base.weight: F32
+blk.15.hc_attn_fn.weight: F32
+blk.15.hc_attn_scale.weight: F32
+blk.15.hc_ffn_base.weight: F32
+blk.15.hc_ffn_fn.weight: F32
+blk.15.hc_ffn_scale.weight: F32
+blk.16.attn_kv.weight: Q8_0
+blk.16.attn_kv_a_norm.weight: F32
+blk.16.attn_norm.weight: F32
+blk.16.attn_output_a.weight: Q5_K
+blk.16.attn_output_b.weight: Q5_K
+blk.16.attn_q_a.weight: Q8_0
+blk.16.attn_q_a_norm.weight: F32
+blk.16.attn_q_b.weight: Q5_K
+blk.16.attn_sinks.weight: F32
+blk.16.exp_probs_b.bias: F32
+blk.16.exp_probs_b_vl.bias: F32
+blk.16.ffn_down_exps.weight: IQ1_M
+blk.16.ffn_down_shexp.weight: Q5_K
+blk.16.ffn_gate_exps.weight: IQ1_M
+blk.16.ffn_gate_inp.weight: BF16
+blk.16.ffn_gate_shexp.weight: Q5_K
+blk.16.ffn_norm.weight: F32
+blk.16.ffn_up_exps.weight: IQ1_M
+blk.16.ffn_up_shexp.weight: Q5_K
+blk.16.hc_attn_base.weight: F32
+blk.16.hc_attn_fn.weight: F32
+blk.16.hc_attn_scale.weight: F32
+blk.16.hc_ffn_base.weight: F32
+blk.16.hc_ffn_fn.weight: F32
+blk.16.hc_ffn_scale.weight: F32
+blk.17.attn_kv.weight: Q8_0
+blk.17.attn_kv_a_norm.weight: F32
+blk.17.attn_norm.weight: F32
+blk.17.attn_output_a.weight: Q5_K
+blk.17.attn_output_b.weight: Q5_K
+blk.17.attn_q_a.weight: Q8_0
+blk.17.attn_q_a_norm.weight: F32
+blk.17.attn_q_b.weight: Q5_K
+blk.17.attn_sinks.weight: F32
+blk.17.exp_probs_b.bias: F32
+blk.17.exp_probs_b_vl.bias: F32
+blk.17.ffn_down_exps.weight: IQ1_M
+blk.17.ffn_down_shexp.weight: Q5_K
+blk.17.ffn_gate_exps.weight: IQ2_XS
+blk.17.ffn_gate_inp.weight: BF16
+blk.17.ffn_gate_shexp.weight: Q5_K
+blk.17.ffn_norm.weight: F32
+blk.17.ffn_up_exps.weight: IQ2_XS
+blk.17.ffn_up_shexp.weight: Q5_K
+blk.17.hc_attn_base.weight: F32
+blk.17.hc_attn_fn.weight: F32
+blk.17.hc_attn_scale.weight: F32
+blk.17.hc_ffn_base.weight: F32
+blk.17.hc_ffn_fn.weight: F32
+blk.17.hc_ffn_scale.weight: F32
+blk.18.attn_kv.weight: Q8_0
+blk.18.attn_kv_a_norm.weight: F32
+blk.18.attn_norm.weight: F32
+blk.18.attn_output_a.weight: Q5_K
+blk.18.attn_output_b.weight: Q5_K
+blk.18.attn_q_a.weight: Q8_0
+blk.18.attn_q_a_norm.weight: F32
+blk.18.attn_q_b.weight: Q5_K
+blk.18.attn_sinks.weight: F32
+blk.18.exp_probs_b.bias: F32
+blk.18.exp_probs_b_vl.bias: F32
+blk.18.ffn_down_exps.weight: IQ2_XXS
+blk.18.ffn_down_shexp.weight: Q5_K
+blk.18.ffn_gate_exps.weight: IQ2_XXS
+blk.18.ffn_gate_inp.weight: BF16
+blk.18.ffn_gate_shexp.weight: Q5_K
+blk.18.ffn_norm.weight: F32
+blk.18.ffn_up_exps.weight: IQ2_XXS
+blk.18.ffn_up_shexp.weight: Q5_K
+blk.18.hc_attn_base.weight: F32
+blk.18.hc_attn_fn.weight: F32
+blk.18.hc_attn_scale.weight: F32
+blk.18.hc_ffn_base.weight: F32
+blk.18.hc_ffn_fn.weight: F32
+blk.18.hc_ffn_scale.weight: F32
+blk.19.attn_kv.weight: Q8_0
+blk.19.attn_kv_a_norm.weight: F32
+blk.19.attn_norm.weight: F32
+blk.19.attn_output_a.weight: Q5_K
+blk.19.attn_output_b.weight: Q5_K
+blk.19.attn_q_a.weight: Q8_0
+blk.19.attn_q_a_norm.weight: F32
+blk.19.attn_q_b.weight: Q5_K
+blk.19.attn_sinks.weight: F32
+blk.19.exp_probs_b.bias: F32
+blk.19.exp_probs_b_vl.bias: F32
+blk.19.ffn_down_exps.weight: IQ1_M
+blk.19.ffn_down_shexp.weight: Q5_K
+blk.19.ffn_gate_exps.weight: IQ2_XS
+blk.19.ffn_gate_inp.weight: BF16
+blk.19.ffn_gate_shexp.weight: Q5_K
+blk.19.ffn_norm.weight: F32
+blk.19.ffn_up_exps.weight: IQ2_XS
+blk.19.ffn_up_shexp.weight: Q5_K
+blk.19.hc_attn_base.weight: F32
+blk.19.hc_attn_fn.weight: F32
+blk.19.hc_attn_scale.weight: F32
+blk.19.hc_ffn_base.weight: F32
+blk.19.hc_ffn_fn.weight: F32
+blk.19.hc_ffn_scale.weight: F32
+blk.20.attn_compressor_kv.weight: BF16
+blk.20.attn_compressor_norm.weight: F32
+blk.20.attn_kv.weight: Q8_0
+blk.20.attn_kv_a_norm.weight: F32
+blk.20.attn_norm.weight: F32
+blk.20.attn_output_a.weight: Q5_K
+blk.20.attn_output_b.weight: Q5_K
+blk.20.attn_q_a.weight: Q8_0
+blk.20.attn_q_a_norm.weight: F32
+blk.20.attn_q_b.weight: Q5_K
+blk.20.attn_sinks.weight: F32
+blk.20.exp_probs_b.bias: F32
+blk.20.exp_probs_b_vl.bias: F32
+blk.20.ffn_down_exps.weight: IQ1_M
+blk.20.ffn_down_shexp.weight: Q5_K
+blk.20.ffn_gate_exps.weight: IQ1_M
+blk.20.ffn_gate_inp.weight: BF16
+blk.20.ffn_gate_shexp.weight: Q5_K
+blk.20.ffn_norm.weight: F32
+blk.20.ffn_up_exps.weight: IQ1_M
+blk.20.ffn_up_shexp.weight: Q5_K
+blk.20.hc_attn_base.weight: F32
+blk.20.hc_attn_fn.weight: F32
+blk.20.hc_attn_scale.weight: F32
+blk.20.hc_ffn_base.weight: F32
+blk.20.hc_ffn_fn.weight: F32
+blk.20.hc_ffn_scale.weight: F32
+blk.20.indexer.attn_k.weight: BF16
+blk.20.indexer.attn_q_b.weight: BF16
+blk.20.indexer.k_norm.weight: F32
+blk.20.indexer.proj.weight: BF16
+blk.21.attn_kv.weight: Q8_0
+blk.21.attn_kv_a_norm.weight: F32
+blk.21.attn_norm.weight: F32
+blk.21.attn_output_a.weight: Q5_K
+blk.21.attn_output_b.weight: Q5_K
+blk.21.attn_q_a.weight: Q8_0
+blk.21.attn_q_a_norm.weight: F32
+blk.21.attn_q_b.weight: Q5_K
+blk.21.attn_sinks.weight: F32
+blk.21.exp_probs_b.bias: F32
+blk.21.exp_probs_b_vl.bias: F32
+blk.21.ffn_down_exps.weight: IQ1_M
+blk.21.ffn_down_shexp.weight: Q5_K
+blk.21.ffn_gate_exps.weight: IQ1_M
+blk.21.ffn_gate_inp.weight: BF16
+blk.21.ffn_gate_shexp.weight: Q5_K
+blk.21.ffn_norm.weight: F32
+blk.21.ffn_up_exps.weight: IQ1_M
+blk.21.ffn_up_shexp.weight: Q5_K
+blk.21.hc_attn_base.weight: F32
+blk.21.hc_attn_fn.weight: F32
+blk.21.hc_attn_scale.weight: F32
+blk.21.hc_ffn_base.weight: F32
+blk.21.hc_ffn_fn.weight: F32
+blk.21.hc_ffn_scale.weight: F32
+blk.22.attn_kv.weight: Q8_0
+blk.22.attn_kv_a_norm.weight: F32
+blk.22.attn_norm.weight: F32
+blk.22.attn_output_a.weight: Q5_K
+blk.22.attn_output_b.weight: Q5_K
+blk.22.attn_q_a.weight: Q8_0
+blk.22.attn_q_a_norm.weight: F32
+blk.22.attn_q_b.weight: Q5_K
+blk.22.attn_sinks.weight: F32
+blk.22.exp_probs_b.bias: F32
+blk.22.exp_probs_b_vl.bias: F32
+blk.22.ffn_down_exps.weight: IQ2_XXS
+blk.22.ffn_down_shexp.weight: Q5_K
+blk.22.ffn_gate_exps.weight: IQ2_XXS
+blk.22.ffn_gate_inp.weight: BF16
+blk.22.ffn_gate_shexp.weight: Q5_K
+blk.22.ffn_norm.weight: F32
+blk.22.ffn_up_exps.weight: IQ2_XXS
+blk.22.ffn_up_shexp.weight: Q5_K
+blk.22.hc_attn_base.weight: F32
+blk.22.hc_attn_fn.weight: F32
+blk.22.hc_attn_scale.weight: F32
+blk.22.hc_ffn_base.weight: F32
+blk.22.hc_ffn_fn.weight: F32
+blk.22.hc_ffn_scale.weight: F32
+blk.23.attn_kv.weight: Q8_0
+blk.23.attn_kv_a_norm.weight: F32
+blk.23.attn_norm.weight: F32
+blk.23.attn_output_a.weight: Q5_K
+blk.23.attn_output_b.weight: Q5_K
+blk.23.attn_q_a.weight: Q8_0
+blk.23.attn_q_a_norm.weight: F32
+blk.23.attn_q_b.weight: Q5_K
+blk.23.attn_sinks.weight: F32
+blk.23.exp_probs_b.bias: F32
+blk.23.exp_probs_b_vl.bias: F32
+blk.23.ffn_down_exps.weight: IQ2_XS
+blk.23.ffn_down_shexp.weight: Q5_K
+blk.23.ffn_gate_exps.weight: IQ2_XXS
+blk.23.ffn_gate_inp.weight: BF16
+blk.23.ffn_gate_shexp.weight: Q5_K
+blk.23.ffn_norm.weight: F32
+blk.23.ffn_up_exps.weight: IQ2_XXS
+blk.23.ffn_up_shexp.weight: Q5_K
+blk.23.hc_attn_base.weight: F32
+blk.23.hc_attn_fn.weight: F32
+blk.23.hc_attn_scale.weight: F32
+blk.23.hc_ffn_base.weight: F32
+blk.23.hc_ffn_fn.weight: F32
+blk.23.hc_ffn_scale.weight: F32
+blk.24.attn_kv.weight: Q8_0
+blk.24.attn_kv_a_norm.weight: F32
+blk.24.attn_norm.weight: F32
+blk.24.attn_output_a.weight: Q5_K
+blk.24.attn_output_b.weight: Q5_K
+blk.24.attn_q_a.weight: Q8_0
+blk.24.attn_q_a_norm.weight: F32
+blk.24.attn_q_b.weight: Q5_K
+blk.24.attn_sinks.weight: F32
+blk.24.exp_probs_b.bias: F32
+blk.24.exp_probs_b_vl.bias: F32
+blk.24.ffn_down_exps.weight: IQ1_M
+blk.24.ffn_down_shexp.weight: Q5_K
+blk.24.ffn_gate_exps.weight: IQ2_XXS
+blk.24.ffn_gate_inp.weight: BF16
+blk.24.ffn_gate_shexp.weight: Q5_K
+blk.24.ffn_norm.weight: F32
+blk.24.ffn_up_exps.weight: IQ2_XXS
+blk.24.ffn_up_shexp.weight: Q5_K
+blk.24.hc_attn_base.weight: F32
+blk.24.hc_attn_fn.weight: F32
+blk.24.hc_attn_scale.weight: F32
+blk.24.hc_ffn_base.weight: F32
+blk.24.hc_ffn_fn.weight: F32
+blk.24.hc_ffn_scale.weight: F32
+blk.24.indexer.attn_q_b.weight: BF16
+blk.24.indexer.proj.weight: BF16
+blk.25.attn_kv.weight: Q8_0
+blk.25.attn_kv_a_norm.weight: F32
+blk.25.attn_norm.weight: F32
+blk.25.attn_output_a.weight: Q5_K
+blk.25.attn_output_b.weight: Q5_K
+blk.25.attn_q_a.weight: Q8_0
+blk.25.attn_q_a_norm.weight: F32
+blk.25.attn_q_b.weight: Q5_K
+blk.25.attn_sinks.weight: F32
+blk.25.exp_probs_b.bias: F32
+blk.25.exp_probs_b_vl.bias: F32
+blk.25.ffn_down_exps.weight: IQ2_XXS
+blk.25.ffn_down_shexp.weight: Q5_K
+blk.25.ffn_gate_exps.weight: IQ2_XXS
+blk.25.ffn_gate_inp.weight: BF16
+blk.25.ffn_gate_shexp.weight: Q5_K
+blk.25.ffn_norm.weight: F32
+blk.25.ffn_up_exps.weight: IQ2_XXS
+blk.25.ffn_up_shexp.weight: Q5_K
+blk.25.hc_attn_base.weight: F32
+blk.25.hc_attn_fn.weight: F32
+blk.25.hc_attn_scale.weight: F32
+blk.25.hc_ffn_base.weight: F32
+blk.25.hc_ffn_fn.weight: F32
+blk.25.hc_ffn_scale.weight: F32
+blk.26.attn_kv.weight: Q8_0
+blk.26.attn_kv_a_norm.weight: F32
+blk.26.attn_norm.weight: F32
+blk.26.attn_output_a.weight: Q5_K
+blk.26.attn_output_b.weight: Q5_K
+blk.26.attn_q_a.weight: Q8_0
+blk.26.attn_q_a_norm.weight: F32
+blk.26.attn_q_b.weight: Q5_K
+blk.26.attn_sinks.weight: F32
+blk.26.exp_probs_b.bias: F32
+blk.26.exp_probs_b_vl.bias: F32
+blk.26.ffn_down_exps.weight: IQ2_XS
+blk.26.ffn_down_shexp.weight: Q5_K
+blk.26.ffn_gate_exps.weight: IQ2_XXS
+blk.26.ffn_gate_inp.weight: BF16
+blk.26.ffn_gate_shexp.weight: Q5_K
+blk.26.ffn_norm.weight: F32
+blk.26.ffn_up_exps.weight: IQ2_XXS
+blk.26.ffn_up_shexp.weight: Q5_K
+blk.26.hc_attn_base.weight: F32
+blk.26.hc_attn_fn.weight: F32
+blk.26.hc_attn_scale.weight: F32
+blk.26.hc_ffn_base.weight: F32
+blk.26.hc_ffn_fn.weight: F32
+blk.26.hc_ffn_scale.weight: F32
+blk.27.attn_kv.weight: Q8_0
+blk.27.attn_kv_a_norm.weight: F32
+blk.27.attn_norm.weight: F32
+blk.27.attn_output_a.weight: Q5_K
+blk.27.attn_output_b.weight: Q5_K
+blk.27.attn_q_a.weight: Q8_0
+blk.27.attn_q_a_norm.weight: F32
+blk.27.attn_q_b.weight: Q5_K
+blk.27.attn_sinks.weight: F32
+blk.27.exp_probs_b.bias: F32
+blk.27.exp_probs_b_vl.bias: F32
+blk.27.ffn_down_exps.weight: IQ1_M
+blk.27.ffn_down_shexp.weight: Q5_K
+blk.27.ffn_gate_exps.weight: IQ2_XS
+blk.27.ffn_gate_inp.weight: BF16
+blk.27.ffn_gate_shexp.weight: Q5_K
+blk.27.ffn_norm.weight: F32
+blk.27.ffn_up_exps.weight: IQ2_XS
+blk.27.ffn_up_shexp.weight: Q5_K
+blk.27.hc_attn_base.weight: F32
+blk.27.hc_attn_fn.weight: F32
+blk.27.hc_attn_scale.weight: F32
+blk.27.hc_ffn_base.weight: F32
+blk.27.hc_ffn_fn.weight: F32
+blk.27.hc_ffn_scale.weight: F32
+blk.28.attn_kv.weight: Q8_0
+blk.28.attn_kv_a_norm.weight: F32
+blk.28.attn_norm.weight: F32
+blk.28.attn_output_a.weight: Q5_K
+blk.28.attn_output_b.weight: Q5_K
+blk.28.attn_q_a.weight: Q8_0
+blk.28.attn_q_a_norm.weight: F32
+blk.28.attn_q_b.weight: Q5_K
+blk.28.attn_sinks.weight: F32
+blk.28.exp_probs_b.bias: F32
+blk.28.exp_probs_b_vl.bias: F32
+blk.28.ffn_down_exps.weight: IQ2_XXS
+blk.28.ffn_down_shexp.weight: Q5_K
+blk.28.ffn_gate_exps.weight: IQ2_XXS
+blk.28.ffn_gate_inp.weight: BF16
+blk.28.ffn_gate_shexp.weight: Q5_K
+blk.28.ffn_norm.weight: F32
+blk.28.ffn_up_exps.weight: IQ2_XXS
+blk.28.ffn_up_shexp.weight: Q5_K
+blk.28.hc_attn_base.weight: F32
+blk.28.hc_attn_fn.weight: F32
+blk.28.hc_attn_scale.weight: F32
+blk.28.hc_ffn_base.weight: F32
+blk.28.hc_ffn_fn.weight: F32
+blk.28.hc_ffn_scale.weight: F32
+blk.28.indexer.attn_q_b.weight: BF16
+blk.28.indexer.proj.weight: BF16
+blk.29.attn_kv.weight: Q8_0
+blk.29.attn_kv_a_norm.weight: F32
+blk.29.attn_norm.weight: F32
+blk.29.attn_output_a.weight: Q5_K
+blk.29.attn_output_b.weight: Q5_K
+blk.29.attn_q_a.weight: Q8_0
+blk.29.attn_q_a_norm.weight: F32
+blk.29.attn_q_b.weight: Q5_K
+blk.29.attn_sinks.weight: F32
+blk.29.exp_probs_b.bias: F32
+blk.29.exp_probs_b_vl.bias: F32
+blk.29.ffn_down_exps.weight: IQ1_M
+blk.29.ffn_down_shexp.weight: Q5_K
+blk.29.ffn_gate_exps.weight: IQ2_XXS
+blk.29.ffn_gate_inp.weight: BF16
+blk.29.ffn_gate_shexp.weight: Q5_K
+blk.29.ffn_norm.weight: F32
+blk.29.ffn_up_exps.weight: IQ2_XXS
+blk.29.ffn_up_shexp.weight: Q5_K
+blk.29.hc_attn_base.weight: F32
+blk.29.hc_attn_fn.weight: F32
+blk.29.hc_attn_scale.weight: F32
+blk.29.hc_ffn_base.weight: F32
+blk.29.hc_ffn_fn.weight: F32
+blk.29.hc_ffn_scale.weight: F32
+blk.30.attn_kv.weight: Q8_0
+blk.30.attn_kv_a_norm.weight: F32
+blk.30.attn_norm.weight: F32
+blk.30.attn_output_a.weight: Q5_K
+blk.30.attn_output_b.weight: Q5_K
+blk.30.attn_q_a.weight: Q8_0
+blk.30.attn_q_a_norm.weight: F32
+blk.30.attn_q_b.weight: Q5_K
+blk.30.attn_sinks.weight: F32
+blk.30.exp_probs_b.bias: F32
+blk.30.exp_probs_b_vl.bias: F32
+blk.30.ffn_down_exps.weight: IQ1_M
+blk.30.ffn_down_shexp.weight: Q5_K
+blk.30.ffn_gate_exps.weight: IQ1_M
+blk.30.ffn_gate_inp.weight: BF16
+blk.30.ffn_gate_shexp.weight: Q5_K
+blk.30.ffn_norm.weight: F32
+blk.30.ffn_up_exps.weight: IQ1_M
+blk.30.ffn_up_shexp.weight: Q5_K
+blk.30.hc_attn_base.weight: F32
+blk.30.hc_attn_fn.weight: F32
+blk.30.hc_attn_scale.weight: F32
+blk.30.hc_ffn_base.weight: F32
+blk.30.hc_ffn_fn.weight: F32
+blk.30.hc_ffn_scale.weight: F32
+blk.31.attn_kv.weight: Q8_0
+blk.31.attn_kv_a_norm.weight: F32
+blk.31.attn_norm.weight: F32
+blk.31.attn_output_a.weight: Q5_K
+blk.31.attn_output_b.weight: Q5_K
+blk.31.attn_q_a.weight: Q8_0
+blk.31.attn_q_a_norm.weight: F32
+blk.31.attn_q_b.weight: Q5_K
+blk.31.attn_sinks.weight: F32
+blk.31.exp_probs_b.bias: F32
+blk.31.exp_probs_b_vl.bias: F32
+blk.31.ffn_down_exps.weight: IQ1_M
+blk.31.ffn_down_shexp.weight: Q5_K
+blk.31.ffn_gate_exps.weight: IQ2_XXS
+blk.31.ffn_gate_inp.weight: BF16
+blk.31.ffn_gate_shexp.weight: Q5_K
+blk.31.ffn_norm.weight: F32
+blk.31.ffn_up_exps.weight: IQ2_XXS
+blk.31.ffn_up_shexp.weight: Q5_K
+blk.31.hc_attn_base.weight: F32
+blk.31.hc_attn_fn.weight: F32
+blk.31.hc_attn_scale.weight: F32
+blk.31.hc_ffn_base.weight: F32
+blk.31.hc_ffn_fn.weight: F32
+blk.31.hc_ffn_scale.weight: F32
+blk.32.attn_kv.weight: Q8_0
+blk.32.attn_kv_a_norm.weight: F32
+blk.32.attn_norm.weight: F32
+blk.32.attn_output_a.weight: Q5_K
+blk.32.attn_output_b.weight: Q5_K
+blk.32.attn_q_a.weight: Q8_0
+blk.32.attn_q_a_norm.weight: F32
+blk.32.attn_q_b.weight: Q5_K
+blk.32.attn_sinks.weight: F32
+blk.32.exp_probs_b.bias: F32
+blk.32.exp_probs_b_vl.bias: F32
+blk.32.ffn_down_exps.weight: IQ1_M
+blk.32.ffn_down_shexp.weight: Q5_K
+blk.32.ffn_gate_exps.weight: IQ2_XXS
+blk.32.ffn_gate_inp.weight: BF16
+blk.32.ffn_gate_shexp.weight: Q5_K
+blk.32.ffn_norm.weight: F32
+blk.32.ffn_up_exps.weight: IQ2_XXS
+blk.32.ffn_up_shexp.weight: Q5_K
+blk.32.hc_attn_base.weight: F32
+blk.32.hc_attn_fn.weight: F32
+blk.32.hc_attn_scale.weight: F32
+blk.32.hc_ffn_base.weight: F32
+blk.32.hc_ffn_fn.weight: F32
+blk.32.hc_ffn_scale.weight: F32
+blk.32.indexer.attn_q_b.weight: BF16
+blk.32.indexer.proj.weight: BF16
+blk.33.attn_kv.weight: Q8_0
+blk.33.attn_kv_a_norm.weight: F32
+blk.33.attn_norm.weight: F32
+blk.33.attn_output_a.weight: Q5_K
+blk.33.attn_output_b.weight: Q5_K
+blk.33.attn_q_a.weight: Q8_0
+blk.33.attn_q_a_norm.weight: F32
+blk.33.attn_q_b.weight: Q5_K
+blk.33.attn_sinks.weight: F32
+blk.33.exp_probs_b.bias: F32
+blk.33.exp_probs_b_vl.bias: F32
+blk.33.ffn_down_exps.weight: IQ2_XS
+blk.33.ffn_down_shexp.weight: Q5_K
+blk.33.ffn_gate_exps.weight: IQ2_XXS
+blk.33.ffn_gate_inp.weight: BF16
+blk.33.ffn_gate_shexp.weight: Q5_K
+blk.33.ffn_norm.weight: F32
+blk.33.ffn_up_exps.weight: IQ2_XXS
+blk.33.ffn_up_shexp.weight: Q5_K
+blk.33.hc_attn_base.weight: F32
+blk.33.hc_attn_fn.weight: F32
+blk.33.hc_attn_scale.weight: F32
+blk.33.hc_ffn_base.weight: F32
+blk.33.hc_ffn_fn.weight: F32
+blk.33.hc_ffn_scale.weight: F32
+blk.34.attn_kv.weight: Q8_0
+blk.34.attn_kv_a_norm.weight: F32
+blk.34.attn_norm.weight: F32
+blk.34.attn_output_a.weight: Q5_K
+blk.34.attn_output_b.weight: Q5_K
+blk.34.attn_q_a.weight: Q8_0
+blk.34.attn_q_a_norm.weight: F32
+blk.34.attn_q_b.weight: Q5_K
+blk.34.attn_sinks.weight: F32
+blk.34.exp_probs_b.bias: F32
+blk.34.exp_probs_b_vl.bias: F32
+blk.34.ffn_down_exps.weight: IQ1_M
+blk.34.ffn_down_shexp.weight: Q5_K
+blk.34.ffn_gate_exps.weight: IQ2_XS
+blk.34.ffn_gate_inp.weight: BF16
+blk.34.ffn_gate_shexp.weight: Q5_K
+blk.34.ffn_norm.weight: F32
+blk.34.ffn_up_exps.weight: IQ2_XS
+blk.34.ffn_up_shexp.weight: Q5_K
+blk.34.hc_attn_base.weight: F32
+blk.34.hc_attn_fn.weight: F32
+blk.34.hc_attn_scale.weight: F32
+blk.34.hc_ffn_base.weight: F32
+blk.34.hc_ffn_fn.weight: F32
+blk.34.hc_ffn_scale.weight: F32
+blk.35.attn_kv.weight: Q8_0
+blk.35.attn_kv_a_norm.weight: F32
+blk.35.attn_norm.weight: F32
+blk.35.attn_output_a.weight: Q5_K
+blk.35.attn_output_b.weight: Q5_K
+blk.35.attn_q_a.weight: Q8_0
+blk.35.attn_q_a_norm.weight: F32
+blk.35.attn_q_b.weight: Q5_K
+blk.35.attn_sinks.weight: F32
+blk.35.exp_probs_b.bias: F32
+blk.35.exp_probs_b_vl.bias: F32
+blk.35.ffn_down_exps.weight: IQ1_M
+blk.35.ffn_down_shexp.weight: Q5_K
+blk.35.ffn_gate_exps.weight: IQ2_XS
+blk.35.ffn_gate_inp.weight: BF16
+blk.35.ffn_gate_shexp.weight: Q5_K
+blk.35.ffn_norm.weight: F32
+blk.35.ffn_up_exps.weight: IQ2_XS
+blk.35.ffn_up_shexp.weight: Q5_K
+blk.35.hc_attn_base.weight: F32
+blk.35.hc_attn_fn.weight: F32
+blk.35.hc_attn_scale.weight: F32
+blk.35.hc_ffn_base.weight: F32
+blk.35.hc_ffn_fn.weight: F32
+blk.35.hc_ffn_scale.weight: F32
+blk.36.attn_kv.weight: Q8_0
+blk.36.attn_kv_a_norm.weight: F32
+blk.36.attn_norm.weight: F32
+blk.36.attn_output_a.weight: Q5_K
+blk.36.attn_output_b.weight: Q5_K
+blk.36.attn_q_a.weight: Q8_0
+blk.36.attn_q_a_norm.weight: F32
+blk.36.attn_q_b.weight: Q5_K
+blk.36.attn_sinks.weight: F32
+blk.36.exp_probs_b.bias: F32
+blk.36.exp_probs_b_vl.bias: F32
+blk.36.ffn_down_exps.weight: IQ1_M
+blk.36.ffn_down_shexp.weight: Q5_K
+blk.36.ffn_gate_exps.weight: IQ2_XXS
+blk.36.ffn_gate_inp.weight: BF16
+blk.36.ffn_gate_shexp.weight: Q5_K
+blk.36.ffn_norm.weight: F32
+blk.36.ffn_up_exps.weight: IQ2_XXS
+blk.36.ffn_up_shexp.weight: Q5_K
+blk.36.hc_attn_base.weight: F32
+blk.36.hc_attn_fn.weight: F32
+blk.36.hc_attn_scale.weight: F32
+blk.36.hc_ffn_base.weight: F32
+blk.36.hc_ffn_fn.weight: F32
+blk.36.hc_ffn_scale.weight: F32
+blk.36.indexer.attn_q_b.weight: BF16
+blk.36.indexer.proj.weight: BF16
+blk.37.attn_kv.weight: Q8_0
+blk.37.attn_kv_a_norm.weight: F32
+blk.37.attn_norm.weight: F32
+blk.37.attn_output_a.weight: Q5_K
+blk.37.attn_output_b.weight: Q5_K
+blk.37.attn_q_a.weight: Q8_0
+blk.37.attn_q_a_norm.weight: F32
+blk.37.attn_q_b.weight: Q5_K
+blk.37.attn_sinks.weight: F32
+blk.37.exp_probs_b.bias: F32
+blk.37.exp_probs_b_vl.bias: F32
+blk.37.ffn_down_exps.weight: IQ1_M
+blk.37.ffn_down_shexp.weight: Q5_K
+blk.37.ffn_gate_exps.weight: IQ2_XXS
+blk.37.ffn_gate_inp.weight: BF16
+blk.37.ffn_gate_shexp.weight: Q5_K
+blk.37.ffn_norm.weight: F32
+blk.37.ffn_up_exps.weight: IQ2_XXS
+blk.37.ffn_up_shexp.weight: Q5_K
+blk.37.hc_attn_base.weight: F32
+blk.37.hc_attn_fn.weight: F32
+blk.37.hc_attn_scale.weight: F32
+blk.37.hc_ffn_base.weight: F32
+blk.37.hc_ffn_fn.weight: F32
+blk.37.hc_ffn_scale.weight: F32
+blk.38.attn_kv.weight: Q8_0
+blk.38.attn_kv_a_norm.weight: F32
+blk.38.attn_norm.weight: F32
+blk.38.attn_output_a.weight: Q5_K
+blk.38.attn_output_b.weight: Q5_K
+blk.38.attn_q_a.weight: Q8_0
+blk.38.attn_q_a_norm.weight: F32
+blk.38.attn_q_b.weight: Q5_K
+blk.38.attn_sinks.weight: F32
+blk.38.exp_probs_b.bias: F32
+blk.38.exp_probs_b_vl.bias: F32
+blk.38.ffn_down_exps.weight: IQ2_XXS
+blk.38.ffn_down_shexp.weight: Q5_K
+blk.38.ffn_gate_exps.weight: IQ2_XS
+blk.38.ffn_gate_inp.weight: BF16
+blk.38.ffn_gate_shexp.weight: Q5_K
+blk.38.ffn_norm.weight: F32
+blk.38.ffn_up_exps.weight: IQ2_XS
+blk.38.ffn_up_shexp.weight: Q5_K
+blk.38.hc_attn_base.weight: F32
+blk.38.hc_attn_fn.weight: F32
+blk.38.hc_attn_scale.weight: F32
+blk.38.hc_ffn_base.weight: F32
+blk.38.hc_ffn_fn.weight: F32
+blk.38.hc_ffn_scale.weight: F32
+blk.39.attn_kv.weight: Q8_0
+blk.39.attn_kv_a_norm.weight: F32
+blk.39.attn_norm.weight: F32
+blk.39.attn_output_a.weight: Q5_K
+blk.39.attn_output_b.weight: Q5_K
+blk.39.attn_q_a.weight: Q8_0
+blk.39.attn_q_a_norm.weight: F32
+blk.39.attn_q_b.weight: Q5_K
+blk.39.attn_sinks.weight: F32
+blk.39.exp_probs_b.bias: F32
+blk.39.exp_probs_b_vl.bias: F32
+blk.39.ffn_down_exps.weight: IQ2_XS
+blk.39.ffn_down_shexp.weight: Q5_K
+blk.39.ffn_gate_exps.weight: IQ2_XS
+blk.39.ffn_gate_inp.weight: BF16
+blk.39.ffn_gate_shexp.weight: Q5_K
+blk.39.ffn_norm.weight: F32
+blk.39.ffn_up_exps.weight: IQ2_XS
+blk.39.ffn_up_shexp.weight: Q5_K
+blk.39.hc_attn_base.weight: F32
+blk.39.hc_attn_fn.weight: F32
+blk.39.hc_attn_scale.weight: F32
+blk.39.hc_ffn_base.weight: F32
+blk.39.hc_ffn_fn.weight: F32
+blk.39.hc_ffn_scale.weight: F32