Atomic-Germ commited on
Commit
dcfcf69
·
verified ·
1 Parent(s): c0a923a

Q4NX/FastFlowLM card: flm-add.py install + run/serve flow, kernels source, model info

Browse files
Files changed (1) hide show
  1. README.md +102 -88
README.md CHANGED
@@ -1,95 +1,109 @@
1
  ---
2
- base_model: richardyoung/fable-qwen2.5-3b-agentic-merged-heretic
3
  language:
4
  - en
5
- library_name: transformers
6
- license: apache-2.0
7
- mradermacher:
8
- readme_rev: 1
9
- quantized_by: mradermacher
10
  tags:
11
- - text-generation-inference
12
- - transformers
13
- - unsloth
14
- - qwen2
15
  - heretic
16
  - uncensored
17
- - decensored
18
- - abliterated
19
- - reproducible
 
 
20
  ---
21
- ## About
22
-
23
- <!-- ### quantize_version: 2 -->
24
- <!-- ### output_tensor_quantised: 1 -->
25
- <!-- ### convert_type: hf -->
26
- <!-- ### vocab_type: -->
27
- <!-- ### tags: nicoboss -->
28
- <!-- ### quants: Q2_K IQ3_M Q4_K_S IQ3_XXS Q3_K_M small-IQ4_NL Q4_K_M IQ2_M Q6_K IQ4_XS Q2_K_S IQ1_M Q3_K_S IQ2_XXS Q3_K_L IQ2_XS Q5_K_S IQ2_S IQ1_S Q5_K_M Q4_0 IQ3_XS Q4_1 IQ3_S -->
29
- <!-- ### quants_skip: -->
30
- <!-- ### skip_mmproj: -->
31
- weighted/imatrix quants of https://huggingface.co/richardyoung/fable-qwen2.5-3b-agentic-merged-heretic
32
-
33
- <!-- provided-files -->
34
-
35
- ***For a convenient overview and download list, visit our [model page for this model](https://hf.tst.eu/model#fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF).***
36
-
37
- static quants are available at https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-GGUF
38
- ## Usage
39
-
40
- If you are unsure how to use GGUF files, refer to one of [TheBloke's
41
- READMEs](https://huggingface.co/TheBloke/KafkaLM-70B-German-V0.1-GGUF) for
42
- more details, including on how to concatenate multi-part files.
43
-
44
- ## Provided Quants
45
-
46
- (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)
47
-
48
- | Link | Type | Size/GB | Notes |
49
- |:-----|:-----|--------:|:------|
50
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.imatrix.gguf) | imatrix | 0.1 | imatrix file (for creating your own quants) |
51
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ1_S.gguf) | i1-IQ1_S | 0.9 | for the desperate |
52
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ1_M.gguf) | i1-IQ1_M | 1.0 | mostly desperate |
53
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ2_XXS.gguf) | i1-IQ2_XXS | 1.0 | |
54
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ2_XS.gguf) | i1-IQ2_XS | 1.1 | |
55
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ2_S.gguf) | i1-IQ2_S | 1.2 | |
56
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ2_M.gguf) | i1-IQ2_M | 1.2 | |
57
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q2_K_S.gguf) | i1-Q2_K_S | 1.3 | very low quality |
58
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q2_K.gguf) | i1-Q2_K | 1.4 | IQ3_XXS probably better |
59
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ3_XXS.gguf) | i1-IQ3_XXS | 1.4 | lower quality |
60
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ3_XS.gguf) | i1-IQ3_XS | 1.5 | |
61
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q3_K_S.gguf) | i1-Q3_K_S | 1.6 | IQ3_XS probably better |
62
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ3_S.gguf) | i1-IQ3_S | 1.6 | beats Q3_K* |
63
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ3_M.gguf) | i1-IQ3_M | 1.6 | |
64
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q3_K_M.gguf) | i1-Q3_K_M | 1.7 | IQ3_S probably better |
65
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q3_K_L.gguf) | i1-Q3_K_L | 1.8 | IQ3_M probably better |
66
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ4_XS.gguf) | i1-IQ4_XS | 1.8 | |
67
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-IQ4_NL.gguf) | i1-IQ4_NL | 1.9 | prefer IQ4_XS |
68
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q4_0.gguf) | i1-Q4_0 | 1.9 | fast, low quality |
69
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q4_K_S.gguf) | i1-Q4_K_S | 1.9 | optimal size/speed/quality |
70
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q4_K_M.gguf) | i1-Q4_K_M | 2.0 | fast, recommended |
71
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q4_1.gguf) | i1-Q4_1 | 2.1 | |
72
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q5_K_S.gguf) | i1-Q5_K_S | 2.3 | |
73
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q5_K_M.gguf) | i1-Q5_K_M | 2.3 | |
74
- | [GGUF](https://huggingface.co/mradermacher/fable-qwen2.5-3b-agentic-merged-heretic-i1-GGUF/resolve/main/fable-qwen2.5-3b-agentic-merged-heretic.i1-Q6_K.gguf) | i1-Q6_K | 2.6 | practically like static Q6_K |
75
-
76
- Here is a handy graph by ikawrakow comparing some lower-quality quant
77
- types (lower is better):
78
-
79
- ![image.png](https://www.nethype.de/huggingface_embed/quantpplgraph.png)
80
-
81
- And here are Artefact2's thoughts on the matter:
82
- https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9
83
-
84
- ## FAQ / Model Request
85
-
86
- See https://huggingface.co/mradermacher/model_requests for some answers to
87
- questions you might have and/or if you want some other model quantized.
88
-
89
- ## Thanks
90
-
91
- I thank my company, [nethype GmbH](https://www.nethype.de/), for letting
92
- me use its servers and providing upgrades to my workstation to enable
93
- this work in my free time. Additional thanks to [@nicoboss](https://huggingface.co/nicoboss) for giving me access to his private supercomputer, enabling me to provide many more imatrix quants, at much higher quality, than I would otherwise be able to.
94
-
95
- <!-- end -->
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
  language:
4
  - en
5
+ pipeline_tag: text-generation
 
 
 
 
6
  tags:
7
+ - qwen2.5
8
+ - agentic
 
 
9
  - heretic
10
  - uncensored
11
+ - fastflowlm
12
+ - q4nx
13
+ - npu
14
+ base_model:
15
+ - richardyoung/fable-qwen2.5-3b-agentic-merged-heretic
16
  ---
17
+
18
+ # Qwen2.5-3B-Fable (Agentic/H eretic) - Q4NX for FastFlowLM (AMD Ryzen AI XDNA2)
19
+
20
+ Fable-Qwen2.5-3B, an agentic merged/heretic fine-tune of Qwen2.5-3B, converted to Q4NX for FastFlowLM. Runs on the Qwen2.5 engine (3B kernels).
21
+
22
+ ## What is Q4NX?
23
+
24
+ Q4NX is FastFlowLM's native packed-quantization format - a rearranged Q4_1
25
+ layout tuned for the NPU matrix engine's tile sizes and memory access
26
+ patterns. It is **not** a GGUF file and it does not run on llama.cpp or
27
+ Ollama; it is meant exclusively for the [FastFlowLM](https://fastflowlm.com)
28
+ engine on AMD Ryzen AI NPUs.
29
+
30
+ ## Requirements
31
+
32
+ - FastFlowLM >= 0.9.45 (`flm` CLI)
33
+ - AMD Ryzen AI processor with **XDNA2 (NPU2)** - Strix Point / Ryzen AI 300
34
+ series or later
35
+ - Linux with the XRT NPU stack installed
36
+ - ~8 GB of unified system memory (Q4NX weights + activations + KV cache)
37
+
38
+ ## Files
39
+
40
+ | File | Purpose |
41
+ |---|---|
42
+ | model.q4nx | Quantized Q4NX weights |
43
+ | config.json | FastFlowLM model configuration |
44
+ | tokenizer.json | Tokenizer |
45
+ | tokenizer_config.json | Special tokens and chat template |
46
+ | chat_template.jinja | Chat template (optional) |
47
+ | flm-add.py | Installer script - registers this model with FastFlowLM |
48
+
49
+
50
+ ## Install and run
51
+
52
+ This repository ships `flm-add.py`, a small installer that copies the model
53
+ into the FastFlowLM user directory and registers the tag `qwen2.5-fable:3b`. It never
54
+ modifies the system FastFlowLM install.
55
+
56
+ ```bash
57
+ # one-time environment (add these to ~/.bashrc)
58
+ export FLM_CONFIG_PATH="$HOME/.config/flm/model_list.json"
59
+ export FLM_XCLBIN_PATH="$HOME/.config/flm"
60
+
61
+ git lfs install
62
+ git clone https://huggingface.co/Atomic-Germ/Qwen2.5-3B-Fable-Agent-Heretic-NPU2
63
+ cd Qwen2.5-3B-Fable-Agent-Heretic-NPU2
64
+ python3 ./flm-add.py .
65
+ flm run qwen2.5-fable:3b
66
+ ```
67
+
68
+ Run `python3 ./flm-add.py --help` for all options. Without a clone, the same
69
+ command works against the repo id directly:
70
+
71
+ ```bash
72
+ python3 ./flm-add.py Atomic-Germ/Qwen2.5-3B-Fable-Agent-Heretic-NPU2
73
+ ```
74
+
75
+
76
+ ## Kernels
77
+
78
+ FastFlowLM's NPU kernels (xclbins) are closed source and are not shipped in
79
+ this repository. `flm-add.py` links the kernels of the official **`qwen2.5-it:3b`**
80
+ model (`Qwen2.5-3B-Instruct-NPU2`), because this model shares the same engine family
81
+ (`qwen2`) and architecture.
82
+
83
+ ## Serve (OpenAI-compatible)
84
+
85
+ ```bash
86
+ flm serve qwen2.5-fable:3b --port 8080
87
+ ```
88
+
89
+ ```bash
90
+ curl http://127.0.0.1:8080/v1/chat/completions \
91
+ -H 'Content-Type: application/json' \
92
+ -d '{"model":"qwen2.5-fable:3b","messages":[{"role":"user","content":"Hello!"}],"max_tokens":256}'
93
+ ```
94
+
95
+ ## Model
96
+
97
+ - Registry tag: `qwen2.5-fable:3b`
98
+ - Engine family: `qwen2`
99
+ - Kernel source: Qwen2.5-3B-Instruct-NPU2
100
+ - Context length: 32,768 tokens (from config)
101
+ - `model.q4nx` size: 2.59 GB
102
+ - Base model: [richardyoung/fable-qwen2.5-3b-agentic-merged-heretic](https://huggingface.co/richardyoung/fable-qwen2.5-3b-agentic-merged-heretic)
103
+ - License: apache-2.0
104
+
105
+ ## Original model card
106
+
107
+ See the upstream model card for training details, benchmarks, and upstream
108
+ usage. This repository only contains the Q4NX conversion for FastFlowLM.
109
+ - Upstream card: [richardyoung/fable-qwen2.5-3b-agentic-merged-heretic](https://huggingface.co/richardyoung/fable-qwen2.5-3b-agentic-merged-heretic)