bartowski commited on
Commit
fa4931c
1 Parent(s): 9dd336a

Llamacpp quants

Browse files
.gitattributes CHANGED
@@ -33,3 +33,27 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ DeepSeek-Coder-V2-Lite-Base-IQ2_M.gguf filter=lfs diff=lfs merge=lfs -text
37
+ DeepSeek-Coder-V2-Lite-Base-IQ2_S.gguf filter=lfs diff=lfs merge=lfs -text
38
+ DeepSeek-Coder-V2-Lite-Base-IQ2_XS.gguf filter=lfs diff=lfs merge=lfs -text
39
+ DeepSeek-Coder-V2-Lite-Base-IQ3_M.gguf filter=lfs diff=lfs merge=lfs -text
40
+ DeepSeek-Coder-V2-Lite-Base-IQ3_XS.gguf filter=lfs diff=lfs merge=lfs -text
41
+ DeepSeek-Coder-V2-Lite-Base-IQ3_XXS.gguf filter=lfs diff=lfs merge=lfs -text
42
+ DeepSeek-Coder-V2-Lite-Base-IQ4_XS.gguf filter=lfs diff=lfs merge=lfs -text
43
+ DeepSeek-Coder-V2-Lite-Base-Q2_K.gguf filter=lfs diff=lfs merge=lfs -text
44
+ DeepSeek-Coder-V2-Lite-Base-Q3_K_L.gguf filter=lfs diff=lfs merge=lfs -text
45
+ DeepSeek-Coder-V2-Lite-Base-Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
46
+ DeepSeek-Coder-V2-Lite-Base-Q3_K_S.gguf filter=lfs diff=lfs merge=lfs -text
47
+ DeepSeek-Coder-V2-Lite-Base-Q4_K_L.gguf filter=lfs diff=lfs merge=lfs -text
48
+ DeepSeek-Coder-V2-Lite-Base-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
49
+ DeepSeek-Coder-V2-Lite-Base-Q4_K_S.gguf filter=lfs diff=lfs merge=lfs -text
50
+ DeepSeek-Coder-V2-Lite-Base-Q5_K_L.gguf filter=lfs diff=lfs merge=lfs -text
51
+ DeepSeek-Coder-V2-Lite-Base-Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
52
+ DeepSeek-Coder-V2-Lite-Base-Q5_K_S.gguf filter=lfs diff=lfs merge=lfs -text
53
+ DeepSeek-Coder-V2-Lite-Base-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
54
+ DeepSeek-Coder-V2-Lite-Base-Q6_K_L.gguf filter=lfs diff=lfs merge=lfs -text
55
+ DeepSeek-Coder-V2-Lite-Base-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
56
+ DeepSeek-Coder-V2-Lite-Base-Q8_1.gguf filter=lfs diff=lfs merge=lfs -text
57
+ DeepSeek-Coder-V2-Lite-Base-f32.gguf/DeepSeek-Coder-V2-Lite-Base-f32-00001-of-00002.gguf filter=lfs diff=lfs merge=lfs -text
58
+ DeepSeek-Coder-V2-Lite-Base-f32.gguf/DeepSeek-Coder-V2-Lite-Base-f32-00002-of-00002.gguf filter=lfs diff=lfs merge=lfs -text
59
+ DeepSeek-Coder-V2-Lite-Base.imatrix filter=lfs diff=lfs merge=lfs -text
DeepSeek-Coder-V2-Lite-Base-IQ2_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c1f881b734a1b61ff30f2a7cf10f76df7e49aa785f8ca23b641fa5355bd209c
3
+ size 6328456960
DeepSeek-Coder-V2-Lite-Base-IQ2_S.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:edd3c4570eededd61d46484096ebfcb3426708772599660ebaa9266b3233e80a
3
+ size 6005212928
DeepSeek-Coder-V2-Lite-Base-IQ2_XS.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1e945187331ec242c3845e3afc129b7e73d0ece7ebdf1e2ee5bd4de846fbfb4b
3
+ size 5967402752
DeepSeek-Coder-V2-Lite-Base-IQ3_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:414c1c657bda6aacb566630f6f00b334d12da19244b47e1fc8fb93f5ef9357c3
3
+ size 7553175296
DeepSeek-Coder-V2-Lite-Base-IQ3_XS.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:295c08b367be194a073741ca694da065a5907b9eab69c086f3d3d32d87445b10
3
+ size 7122857728
DeepSeek-Coder-V2-Lite-Base-IQ3_XXS.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:766040981b09277133aeeff2c5e512ceacf070e0c4dacb43ff01716290264429
3
+ size 6964057856
DeepSeek-Coder-V2-Lite-Base-IQ4_XS.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e7ca21601e36da60774935e72da1f73ba9a776f7d0242ed7eaaa641ab0fca2cc
3
+ size 8571593472
DeepSeek-Coder-V2-Lite-Base-Q2_K.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:368f160c33fda8a7974b4798d2c5fec46dd52743a60eb736ef9a7234efa5f210
3
+ size 6430464768
DeepSeek-Coder-V2-Lite-Base-Q3_K_L.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:535007ada823f4ccb2afea0ba310465876b7e5450b99b9a7f3dadbc266ef1af7
3
+ size 8459398912
DeepSeek-Coder-V2-Lite-Base-Q3_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:392b8ed176eec3b9060e757b54846b46004bd4d9d05040af645f300972b2d62c
3
+ size 8126607104
DeepSeek-Coder-V2-Lite-Base-Q3_K_S.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c62e3317c2fb8f6aca2a9cbfca3d795a34af8ca38bcda82b8d8a7df9c5c5991
3
+ size 7487663872
DeepSeek-Coder-V2-Lite-Base-Q4_K_L.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb21ebc391c4e417f141585bdfa362db3493376f11e08b6d2c05a2e6312db91a
3
+ size 10913280768
DeepSeek-Coder-V2-Lite-Base-Q4_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9c606f987ace7e95bbcd93e381c32fa0438d3e2da7a94601a0ed3d18a4ed15b7
3
+ size 10364416768
DeepSeek-Coder-V2-Lite-Base-Q4_K_S.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b2fe7dcac20764fa8ac4072b1136b732b3acf5657b32700d17378d49fa3cff6
3
+ size 9533608704
DeepSeek-Coder-V2-Lite-Base-Q5_K_L.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c6bb6761e602c02326f7ed6a44963d39bd94fef19dd7257171f13cde617459ae
3
+ size 12373963520
DeepSeek-Coder-V2-Lite-Base-Q5_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:becdf516214e67e1907a89664f7cd8ea60059d8558a59c9a3e4345cc71f1abd5
3
+ size 11851313920
DeepSeek-Coder-V2-Lite-Base-Q5_K_S.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9cffc72550d55bd31efa15b92af278dd550697783b548be0b5ea1a522d1f0a8
3
+ size 11143058176
DeepSeek-Coder-V2-Lite-Base-Q6_K.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a2e25280075d769abdb111de8211d9d3367f2ae0d0e6166a288ee6e8ed0345d
3
+ size 14066972416
DeepSeek-Coder-V2-Lite-Base-Q6_K_L.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dde3ab6bcd3593d3333079a4c39c3a20e93fdfb52ce278bbd267a2fb30cc25e9
3
+ size 14561769216
DeepSeek-Coder-V2-Lite-Base-Q8_0.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22cef614d5b3e1d00ece317ded91c48046f1a6f24618639dea110fbc22e5f86c
3
+ size 16702518016
DeepSeek-Coder-V2-Lite-Base-Q8_1.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0516a0d5f1e89b081cebed7f353fb7230913f3040a924988607890832aee04a7
3
+ size 17095734016
DeepSeek-Coder-V2-Lite-Base-f32.gguf/DeepSeek-Coder-V2-Lite-Base-f32-00001-of-00002.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df693bd686a00818db81192d4198aa86ab9c4bed34fd280dda86795f21f9c8ed
3
+ size 39436017312
DeepSeek-Coder-V2-Lite-Base-f32.gguf/DeepSeek-Coder-V2-Lite-Base-f32-00002-of-00002.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8771f89f923c63def9e33f060fbf60cf0b191708a45bf4d0d16e25e41116bc7c
3
+ size 23393915904
DeepSeek-Coder-V2-Lite-Base.imatrix ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d366d4860d972d0fd5a0ca7c9447ab58d26b99217c5f5c798dfa0d6e3206a61b
3
+ size 38356422
README.md ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: deepseek-license
4
+ license_link: LICENSE
5
+ quantized_by: bartowski
6
+ pipeline_tag: text-generation
7
+ ---
8
+
9
+ ## Llamacpp imatrix Quantizations of DeepSeek-Coder-V2-Lite-Base
10
+
11
+ Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b3166">b3166</a> for quantization.
12
+
13
+ Original model: https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Base
14
+
15
+ All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
16
+
17
+ ## Prompt format
18
+
19
+ ```
20
+ <|begin▁of▁sentence|>{system_prompt}
21
+
22
+ User: {prompt}
23
+
24
+ Assistant: <|end▁of▁sentence|>Assistant:
25
+ ```
26
+
27
+ ## Download a file (not the whole branch) from below:
28
+
29
+ | Filename | Quant type | File Size | Description |
30
+ | -------- | ---------- | --------- | ----------- |
31
+ | [DeepSeek-Coder-V2-Lite-Base-Q8_1.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q8_1.gguf) | Q8_1 | 17.09GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Extremely high quality, generally unneeded but max available quant. |
32
+ | [DeepSeek-Coder-V2-Lite-Base-Q8_0.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q8_0.gguf) | Q8_0 | 16.70GB | Extremely high quality, generally unneeded but max available quant. |
33
+ | [DeepSeek-Coder-V2-Lite-Base-Q6_K_L.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF//main/DeepSeek-Coder-V2-Lite-Base-Q6_K_L.gguf) | Q6_K_L | 14.56GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Very high quality, near perfect, *recommended*. |
34
+ | [DeepSeek-Coder-V2-Lite-Base-Q6_K.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q6_K.gguf) | Q6_K | 14.06GB | Very high quality, near perfect, *recommended*. |
35
+ | [DeepSeek-Coder-V2-Lite-Base-Q5_K_L.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q5_K_L.gguf) | Q5_K_L | 12.37GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. High quality, *recommended*. |
36
+ | [DeepSeek-Coder-V2-Lite-Base-Q5_K_M.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q5_K_M.gguf) | Q5_K_M | 11.85GB | High quality, *recommended*. |
37
+ | [DeepSeek-Coder-V2-Lite-Base-Q5_K_S.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q5_K_S.gguf) | Q5_K_S | 11.14GB | High quality, *recommended*. |
38
+ | [DeepSeek-Coder-V2-Lite-Base-Q4_K_L.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q4_K_L.gguf) | Q4_K_L | 10.91GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Good quality, uses about 4.83 bits per weight, *recommended*. |
39
+ | [DeepSeek-Coder-V2-Lite-Base-Q4_K_M.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q4_K_M.gguf) | Q4_K_M | 10.36GB | Good quality, uses about 4.83 bits per weight, *recommended*. |
40
+ | [DeepSeek-Coder-V2-Lite-Base-Q4_K_S.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q4_K_S.gguf) | Q4_K_S | 9.53GB | Slightly lower quality with more space savings, *recommended*. |
41
+ | [DeepSeek-Coder-V2-Lite-Base-IQ4_XS.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ4_XS.gguf) | IQ4_XS | 8.57GB | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
42
+ | [DeepSeek-Coder-V2-Lite-Base-Q3_K_L.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q3_K_L.gguf) | Q3_K_L | 8.45GB | Lower quality but usable, good for low RAM availability. |
43
+ | [DeepSeek-Coder-V2-Lite-Base-Q3_K_M.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q3_K_M.gguf) | Q3_K_M | 8.12GB | Even lower quality. |
44
+ | [DeepSeek-Coder-V2-Lite-Base-IQ3_M.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ3_M.gguf) | IQ3_M | 7.55GB | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
45
+ | [DeepSeek-Coder-V2-Lite-Base-Q3_K_S.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q3_K_S.gguf) | Q3_K_S | 7.48GB | Low quality, not recommended. |
46
+ | [DeepSeek-Coder-V2-Lite-Base-IQ3_XS.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ3_XS.gguf) | IQ3_XS | 7.12GB | Lower quality, new method with decent performance, slightly better than Q3_K_S. |
47
+ | [DeepSeek-Coder-V2-Lite-Base-IQ3_XXS.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ3_XXS.gguf) | IQ3_XXS | 6.96GB | Lower quality, new method with decent performance, comparable to Q3 quants. |
48
+ | [DeepSeek-Coder-V2-Lite-Base-Q2_K.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-Q2_K.gguf) | Q2_K | 6.43GB | Very low quality but surprisingly usable. |
49
+ | [DeepSeek-Coder-V2-Lite-Base-IQ2_M.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ2_M.gguf) | IQ2_M | 6.32GB | Very low quality, uses SOTA techniques to also be surprisingly usable. |
50
+ | [DeepSeek-Coder-V2-Lite-Base-IQ2_S.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ2_S.gguf) | IQ2_S | 6.00GB | Very low quality, uses SOTA techniques to be usable. |
51
+ | [DeepSeek-Coder-V2-Lite-Base-IQ2_XS.gguf](https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF/blob/main/DeepSeek-Coder-V2-Lite-Base-IQ2_XS.gguf) | IQ2_XS | 5.96GB | Very low quality, uses SOTA techniques to be usable. |
52
+
53
+ ## Downloading using huggingface-cli
54
+
55
+ First, make sure you have hugginface-cli installed:
56
+
57
+ ```
58
+ pip install -U "huggingface_hub[cli]"
59
+ ```
60
+
61
+ Then, you can target the specific file you want:
62
+
63
+ ```
64
+ huggingface-cli download bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF --include "DeepSeek-Coder-V2-Lite-Base-Q4_K_M.gguf" --local-dir ./
65
+ ```
66
+
67
+ If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:
68
+
69
+ ```
70
+ huggingface-cli download bartowski/DeepSeek-Coder-V2-Lite-Base-GGUF --include "DeepSeek-Coder-V2-Lite-Base-Q8_0.gguf/*" --local-dir DeepSeek-Coder-V2-Lite-Base-Q8_0
71
+ ```
72
+
73
+ You can either specify a new local-dir (DeepSeek-Coder-V2-Lite-Base-Q8_0) or download them all in place (./)
74
+
75
+ ## Which file should I choose?
76
+
77
+ A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
78
+
79
+ The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.
80
+
81
+ If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.
82
+
83
+ If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.
84
+
85
+ Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
86
+
87
+ If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.
88
+
89
+ If you want to get more into the weeds, you can check out this extremely useful feature chart:
90
+
91
+ [llama.cpp feature matrix](https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix)
92
+
93
+ But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.
94
+
95
+ These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
96
+
97
+ The I-quants are *not* compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.
98
+
99
+ Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski