IBM Graniteオープンソース多言語埋め込みモデル - 12言語をサポートし、情報検索と多言語アプリケーションに使用

ホーム

Granite Embedding 278m Multilingual GGUF

bartowskiによって開発

IBM Granite多言語埋め込みモデル、12言語のテキスト埋め込みタスクをサポート、情報検索や多言語アプリケーションシナリオに適しています。

テキスト埋め込み複数言語対応オープンソースライセンス:Apache-2.0 #多言語埋め込み #効率的な検索 #低リソース最適化

ダウンロード数 4,815

リリース時間 : 12/18/2024

モデル概要

これは多言語テキスト埋め込みモデルで、異なる言語のテキストを高次元ベクトル表現に変換でき、クロスランゲージ検索、意味的類似度計算などのタスクに適しています。

モデル特徴

多言語サポート

英語、中国語、アラビア語などの主要言語を含む12言語のテキスト埋め込みをサポート

効率的な量子化

Q4_K_Mなどの推奨バージョンを含む複数の量子化バージョンを提供、モデル品質とサイズのバランスを取ります

ハードウェア最適化

ARM/AVXハードウェア最適化をサポート、モバイルデバイスで良好なパフォーマンスを発揮

優れた検索性能

Miraclなどのデータセットで優れた検索性能指標を示しています

モデル能力

テキスト埋め込み

クロスランゲージ検索

意味的類似度計算

多言語テキスト処理

使用事例

情報検索

クロスランゲージドキュメント検索

埋め込みベクトルを使用して多言語ドキュメントの類似性検索を行う

MiraclデータセットでNDCG@10が0.49372(英語)を達成

意味分析

多言語意味的類似度計算

異なる言語のテキスト間の意味的類似度を計算する

quantized_by: bartowski pipeline_tag: text-generation language:

en
ar
cs
de
es
fr
it
ja
ko
nl
pt
zh tags:
language
granite
embeddings
multilingual license: apache-2.0 base_model: ibm-granite/granite-embedding-278m-multilingual model-index:
name: ibm-granite/granite-embedding-278m-multilingual results:
- task: type: Retrieval dataset: name: Miracl (en) type: miracl/mmteb-miracl config: en split: dev metrics:
  - type: ndcg_at_1 value: 0.45557
  - type: ndcg_at_10 value: 0.49372
  - type: ndcg_at_100 value: 0.5728
  - type: ndcg_at_1000 value: 0.59187
  - type: ndcg_at_20 value: 0.52863
  - type: ndcg_at_3 value: 0.43969
  - type: ndcg_at_5 value: 0.45551
  - type: recall_at_1 value: 0.21785
  - type: recall_at_10 value: 0.59513
  - type: recall_at_100 value: 0.85785
  - type: recall_at_1000 value: 0.96041
  - type: recall_at_20 value: 0.69357
  - type: recall_at_3 value: 0.40403
  - type: recall_at_5 value: 0.48499
- task: type: Retrieval dataset: name: Miracl (ar) type: miracl/mmteb-miracl config: ar split: dev metrics:
  - type: ndcg_at_1 value: 0.57459
  - type: ndcg_at_10 value: 0.64238
  - type: ndcg_at_100 value: 0.6867
  - type: ndcg_at_1000 value: 0.6951
  - type: ndcg_at_20 value: 0.66455
  - type: ndcg_at_3 value: 0.58162
  - type: ndcg_at_5 value: 0.60831
  - type: recall_at_1 value: 0.38064
  - type: recall_at_10 value: 0.75098
  - type: recall_at_100 value: 0.91203
  - type: recall_at_1000 value: 0.96706
  - type: recall_at_20 value: 0.81978
  - type: recall_at_3 value: 0.58618
  - type: recall_at_5 value: 0.66353
- task: type: Retrieval dataset: name: Miracl (bn) type: miracl/mmteb-miracl config: bn split: dev metrics:
  - type: ndcg_at_1 value: 0.60341
  - type: ndcg_at_10 value: 0.68055
  - type: ndcg_at_100 value: 0.72008
  - type: ndcg_at_1000 value: 0.72716
  - type: ndcg_at_20 value: 0.69914
  - type: ndcg_at_3 value: 0.60805
  - type: ndcg_at_5 value: 0.64486
  - type: recall_at_1 value: 0.37948
  - type: recall_at_10 value: 0.80609
  - type: recall_at_100 value: 0.94305
  - type: recall_at_1000 value: 0.98625
  - type: recall_at_20 value: 0.86141
  - type: recall_at_3 value: 0.61095
  - type: recall_at_5 value: 0.71316
- task: type: Retrieval dataset: name: Miracl (de) type: miracl/mmteb-miracl config: de split: dev metrics:
  - type: ndcg_at_1 value: 0.45574
  - type: ndcg_at_10 value: 0.48123
  - type: ndcg_at_100 value: 0.56049
  - type: ndcg_at_1000 value: 0.57979
  - type: ndcg_at_20 value: 0.51785
  - type: ndcg_at_3 value: 0.41243
  - type: ndcg_at_5 value: 0.4386
  - type: recall_at_1 value: 0.20401
  - type: recall_at_10 value: 0.58779
  - type: recall_at_100 value: 0.8584
  - type: recall_at_1000 value: 0.97364
  - type: recall_at_20 value: 0.69061
  - type: recall_at_3 value: 0.36573
  - type: recall_at_5 value: 0.47495
- task: type: Retrieval dataset: name: Miracl (es) type: miracl/mmteb-miracl config: es split: dev metrics:
  - type: ndcg_at_1 value: 0.5571
  - type: ndcg_at_10 value: 0.49688
  - type: ndcg_at_100 value: 0.60493
  - type: ndcg_at_1000 value: 0.62922
  - type: ndcg_at_20 value: 0.54438
  - type: ndcg_at_3 value: 0.47981
  - type: ndcg_at_5 value: 0.46584
  - type: recall_at_1 value: 0.1638
  - type: recall_at_10 value: 0.54155
  - type: recall_at_100 value: 0.85136
  - type: recall_at_1000 value: 0.96951
  - type: recall_at_20 value: 0.65329
  - type: recall_at_3 value: 0.31503
  - type: recall_at_5 value: 0.40356
- task: type: Retrieval dataset: name: Miracl (fa) type: miracl/mmteb-miracl config: fa split: dev metrics:
  - type: ndcg_at_1 value: 0.39873
  - type: ndcg_at_10 value: 0.50226
  - type: ndcg_at_100 value: 0.56517
  - type: ndcg_at_1000 value: 0.57967
  - type: ndcg_at_20 value: 0.5292
  - type: ndcg_at_3 value: 0.42738
  - type: ndcg_at_5 value: 0.45843
  - type: recall_at_1 value: 0.25369
  - type: recall_at_10 value: 0.63776
  - type: recall_at_100 value: 0.87686
  - type: recall_at_1000 value: 0.9671
  - type: recall_at_20 value: 0.72099
  - type: recall_at_3 value: 0.43808
  - type: recall_at_5 value: 0.52378
- task: type: Retrieval dataset: name: Miracl (fi) type: miracl/mmteb-miracl config: fi split: dev metrics:
  - type: ndcg_at_1 value: 0.60818
  - type: ndcg_at_10 value: 0.6746
  - type: ndcg_at_100 value: 0.71516
  - type: ndcg_at_1000 value: 0.7218
  - type: ndcg_at_20 value: 0.69692
  - type: ndcg_at_3 value: 0.6006
  - type: ndcg_at_5 value: 0.63842
  - type: recall_at_1 value: 0.39264
  - type: recall_at_10 value: 0.78577
  - type: recall_at_100 value: 0.93291
  - type: recall_at_1000 value: 0.97493
  - type: recall_at_20 value: 0.85435
  - type: recall_at_3 value: 0.61055
  - type: recall_at_5 value: 0.69774
- task: type: Retrieval dataset: name: Miracl (fr) type: miracl/mmteb-miracl config: fr split: dev metrics:
  - type: ndcg_at_1 value: 0.3965
  - type: ndcg_at_10 value: 0.49891
  - type: ndcg_at_100 value: 0.56492
  - type: ndcg_at_1000 value: 0.57837
  - type: ndcg_at_20 value: 0.53163
  - type: ndcg_at_3 value: 0.39843
  - type: ndcg_at_5 value: 0.44416
  - type: recall_at_1 value: 0.22644
  - type: recall_at_10 value: 0.65169
  - type: recall_at_100 value: 0.89786
  - type: recall_at_1000 value: 0.98081
  - type: recall_at_20 value: 0.75338
  - type: recall_at_3 value: 0.39798
  - type: recall_at_5 value: 0.51001
- task: type: Retrieval dataset: name: Miracl (hi) type: miracl/mmteb-miracl config: hi split: dev metrics:
  - type: ndcg_at_1 value: 0.36857
  - type: ndcg_at_10 value: 0.46141
  - type: ndcg_at_100 value: 0.52565
  - type: ndcg_at_1000 value: 0.54319
  - type: ndcg_at_20 value: 0.49384
  - type: ndcg_at_3 value: 0.39469
  - type: ndcg_at_5 value: 0.4184
  - type: recall_at_1 value: 0.20185
  - type: recall_at_10 value: 0.59474
  - type: recall_at_100 value: 0.83385
  - type: recall_at_1000 value: 0.94813
  - type: recall_at_20 value: 0.69437
  - type: recall_at_3 value: 0.38993
  - type: recall_at_5 value: 0.47881
- task: type: Retrieval dataset: name: Miracl (id) type: miracl/mmteb-miracl config: id split: dev metrics:
  - type: ndcg_at_1 value: 0.46354
  - type: ndcg_at_10 value: 0.47229
  - type: ndcg_at_100 value: 0.5525
  - type: ndcg_at_1000 value: 0.57648
  - type: ndcg_at_20 value: 0.50606
  - type: ndcg_at_3 value: 0.42538
  - type: ndcg_at_5 value: 0.43717
  - type: recall_at_1 value: 0.20787
  - type: recall_at_10 value: 0.54771
  - type: recall_at_100 value: 0.80689
  - type: recall_at_1000 value: 0.94032
  - type: recall_at_20 value: 0.63842
  - type: recall_at_3 value: 0.36229
  - type: recall_at_5 value: 0.44437
- task: type: Retrieval dataset: name: Miracl (ja) type: miracl/mmteb-miracl config: ja split: dev metrics:
  - type: ndcg_at_1 value: 0.56279
  - type: ndcg_at_10 value: 0.6281
  - type: ndcg_at_100 value: 0.67757
  - type: ndcg_at_1000 value: 0.68667
  - type: ndcg_at_20 value: 0.6521
  - type: ndcg_at_3 value: 0.56226
  - type: ndcg_at_5 value: 0.5866
  - type: recall_at_1 value: 0.36648
  - type: recall_at_10 value: 0.7496
  - type: recall_at_100 value: 0.92461
  - type: recall_at_1000 value: 0.97827
  - type: recall_at_20 value: 0.82326
  - type: recall_at_3 value: 0.55845
  - type: recall_at_5 value: 0.63854
- task: type: Retrieval dataset: name: Miracl (ko) type: miracl/mmteb-miracl config: ko split: dev metrics:
  - type: ndcg_at_1 value: 0.52582
  - type: ndcg_at_10 value: 0.59216
  - type: ndcg_at_100 value: 0.65093
  - type: ndcg_at_1000 value: 0.66204
  - type: ndcg_at_20 value: 0.62427
  - type: ndcg_at_3 value: 0.5373
  - type: ndcg_at_5 value: 0.55886
  - type: recall_at_1 value: 0.30521
  - type: recall_at_10 value: 0.71159
  - type: recall_at_100 value: 0.90203
  - type: recall_at_1000 value: 0.96714
  - type: recall_at_20 value: 0.80209
  - type: recall_at_3 value: 0.515
  - type: recall_at_5 value: 0.6071
- task: type: Retrieval dataset: name: Miracl (ru) type: miracl/mmteb-miracl config: ru split: dev metrics:
  - type: ndcg_at_1 value: 0.47524
  - type: ndcg_at_10 value: 0.52349
  - type: ndcg_at_100 value: 0.59725
  - type: ndcg_at_1000 value: 0.61313
  - type: ndcg_at_20 value: 0.55669
  - type: ndcg_at_3 value: 0.46812
  - type: ndcg_at_5 value: 0.48442
  - type: recall_at_1 value: 0.24337
  - type: recall_at_10 value: 0.62437
  - type: recall_at_100 value: 0.86489
  - type: recall_at_1000 value: 0.95266
  - type: recall_at_20 value: 0.71411
  - type: recall_at_3 value: 0.42927
  - type: recall_at_5 value: 0.51258
- task: type: Retrieval dataset: name: Miracl (sw) type: miracl/mmteb-miracl config: sw split: dev metrics:
  - type: ndcg_at_1 value: 0.5166
  - type: ndcg_at_10 value: 0.61271
  - type: ndcg_at_100 value: 0.66099
  - type: ndcg_at_1000 value: 0.66867
  - type: ndcg_at_20 value: 0.63643
  - type: ndcg_at_3 value: 0.54828
  - type: ndcg_at_5 value: 0.57382
  - type: recall_at_1 value: 0.35277
  - type: recall_at_10 value: 0.74368
  - type: recall_at_100 value: 0.92261
  - type: recall_at_1000 value: 0.97109
  - type: recall_at_20 value: 0.81888
  - type: recall_at_3 value: 0.56739
  - type: recall_at_5 value: 0.6421
- task: type: Retrieval dataset: name: Miracl (te) type: miracl/mmteb-miracl config: te split: dev metrics:
  - type: ndcg_at_1 value: 0.63768
  - type: ndcg_at_10 value: 0.79193
  - type: ndcg_at_100 value: 0.80243
  - type: ndcg_at_1000 value: 0.80438
  - type: ndcg_at_20 value: 0.79549
  - type: ndcg_at_3 value: 0.76031
  - type: ndcg_at_5 value: 0.77915
  - type: recall_at_1 value: 0.63084
  - type: recall_at_10 value: 0.92411
  - type: recall_at_100 value: 0.97363
  - type: recall_at_1000 value: 0.98833
  - type: recall_at_20 value: 0.9374
  - type: recall_at_3 value: 0.84159
  - type: recall_at_5 value: 0.88627
- task: type: Retrieval dataset: name: Miracl (th) type: miracl/mmteb-miracl config: th split: dev metrics:
  - type: ndcg_at_1 value: 0.66712
  - type: ndcg_at_10 value: 0.73324
  - type: ndcg_at_100 value: 0.76633
  - type: ndcg_at_1000 value: 0.77119
  - type: ndcg_at_20 value: 0.75243
  - type: ndcg_at_3 value: 0.67393
  - type: ndcg_at_5 value: 0.70201
  - type: recall_at_1 value: 0.47106
  - type: recall_at_10 value: 0.84294
  - type: recall_at_100 value: 0.95949
  - type: recall_at_1000 value: 0.98874
  - type: recall_at_20 value: 0.90085
  - type: recall_at_3 value: 0.68456
  - type: recall_at_5 value: 0.75915
- task: type: Retrieval dataset: name: Miracl (yo) type: miracl/mmteb-miracl config: yo split: dev metrics:
  - type: ndcg_at_1 value: 0.4958
  - type: ndcg_at_10 value: 0.68705
  - type: ndcg_at_100 value: 0.70664
  - type: ndcg_at_1000 value: 0.71197
  - type: ndcg_at_20 value: 0.698
  - type: ndcg_at_3 value: 0.64793
  - type: ndcg_at_5 value: 0.66709
  - type: recall_at_1 value: 0.46289
  - type: recall_at_10 value: 0.85154
  - type: recall_at_100 value: 0.93557
  - type: recall_at_1000 value: 0.97479
  - type: recall_at_20 value: 0.89076
  - type: recall_at_3 value: 0.7507
  - type: recall_at_5 value: 0.79202
- task: type: Retrieval dataset: name: Miracl (zh) type: miracl/mmteb-miracl config: zh split: dev metrics:
  - type: ndcg_at_1 value: 0.47583
  - type: ndcg_at_10 value: 0.52553
  - type: ndcg_at_100 value: 0.6
  - type: ndcg_at_1000 value: 0.61415
  - type: ndcg_at_20 value: 0.55668
  - type: ndcg_at_3 value: 0.45839
  - type: ndcg_at_5 value: 0.48127
  - type: recall_at_1 value: 0.24488
  - type: recall_at_10 value: 0.63659
  - type: recall_at_100 value: 0.89702
  - type: recall_at_1000 value: 0.97996
  - type: recall_at_20 value: 0.72652
  - type: recall_at_3 value: 0.42827
  - type: recall_at_5 value: 0.52081

Llamacpp Static Quantizations of granite-embedding-278m-multilingual

Using llama.cpp release b4381 for quantization.

Original model: https://huggingface.co/ibm-granite/granite-embedding-278m-multilingual

Run them in LM Studio

Prompt format

No prompt format found, check original model page

What's new:

Fix tokenizer

Download a file (not the whole branch) from below:

Filename	Quant type	File Size	Split	Description
granite-embedding-278m-multilingual-f16.gguf	f16	0.56GB	false	Full F16 weights.
granite-embedding-278m-multilingual-Q8_0.gguf	Q8_0	0.30GB	false	Extremely high quality, generally unneeded but max available quant.
granite-embedding-278m-multilingual-Q6_K_L.gguf	Q6_K_L	0.28GB	false	Uses Q8_0 for embed and output weights. Very high quality, near perfect, recommended.
granite-embedding-278m-multilingual-Q5_K_L.gguf	Q5_K_L	0.27GB	false	Uses Q8_0 for embed and output weights. High quality, recommended.
granite-embedding-278m-multilingual-Q4_K_L.gguf	Q4_K_L	0.27GB	false	Uses Q8_0 for embed and output weights. Good quality, recommended.
granite-embedding-278m-multilingual-Q3_K_XL.gguf	Q3_K_XL	0.26GB	false	Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability.
granite-embedding-278m-multilingual-Q6_K.gguf	Q6_K	0.24GB	false	Very high quality, near perfect, recommended.
granite-embedding-278m-multilingual-Q5_K_M.gguf	Q5_K_M	0.23GB	false	High quality, recommended.
granite-embedding-278m-multilingual-Q5_K_S.gguf	Q5_K_S	0.22GB	false	High quality, recommended.
granite-embedding-278m-multilingual-Q4_K_M.gguf	Q4_K_M	0.22GB	false	Good quality, default size for most use cases, recommended.
granite-embedding-278m-multilingual-Q4_K_S.gguf	Q4_K_S	0.21GB	false	Slightly lower quality with more space savings, recommended.
granite-embedding-278m-multilingual-Q4_0.gguf	Q4_0	0.21GB	false	Legacy format, offers online repacking for ARM and AVX CPU inference.
granite-embedding-278m-multilingual-IQ4_NL.gguf	IQ4_NL	0.21GB	false	Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference.
granite-embedding-278m-multilingual-IQ4_XS.gguf	IQ4_XS	0.21GB	false	Decent quality, smaller than Q4_K_S with similar performance, recommended.
granite-embedding-278m-multilingual-Q3_K_L.gguf	Q3_K_L	0.21GB	false	Lower quality but usable, good for low RAM availability.
granite-embedding-278m-multilingual-Q3_K_M.gguf	Q3_K_M	0.21GB	false	Low quality.
granite-embedding-278m-multilingual-IQ3_M.gguf	IQ3_M	0.21GB	false	Medium-low quality, new method with decent performance comparable to Q3_K_M.

Embed/output weights

Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.

Downloading using huggingface-cli

Click to view download instructions

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download bartowski/granite-embedding-278m-multilingual-GGUF --include "granite-embedding-278m-multilingual-Q4_K_M.gguf" --local-dir ./

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:

huggingface-cli download bartowski/granite-embedding-278m-multilingual-GGUF --include "granite-embedding-278m-multilingual-Q8_0/*" --local-dir ./

You can either specify a new local-dir (granite-embedding-278m-multilingual-Q8_0) or download them all in place (./)

ARM/AVX information

Previously, you would download Q4_0_4_4/4_8/8_8, and these would have their weights interleaved in memory in order to improve performance on ARM and AVX machines by loading up more data in one pass.

Now, however, there is something called "online repacking" for weights. details in this PR. If you use Q4_0 and your hardware would benefit from repacking weights, it will do it automatically on the fly.

As of llama.cpp build b4282 you will not be able to run the Q4_0_X_X files and will instead need to use Q4_0.

Additionally, if you want to get slightly better quality for , you can use IQ4_NL thanks to this PR which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed incrase.

Click to view Q4_0_X_X information (deprecated

I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.

Click to view benchmarks on an AVX2 system (EPYC7702)

model	size	params	backend	threads	test	t/s	% (vs Q4_0)
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp512	204.03 ± 1.03	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp1024	282.92 ± 0.19	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp2048	259.49 ± 0.44	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg128	39.12 ± 0.27	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg256	39.31 ± 0.69	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg512	40.52 ± 0.03	100%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp512	301.02 ± 1.74	147%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp1024	287.23 ± 0.20	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp2048	262.77 ± 1.81	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg128	18.80 ± 0.99	48%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg256	24.46 ± 3.04	83%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg512	36.32 ± 3.59	90%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp512	271.71 ± 3.53	133%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp1024	279.86 ± 45.63	100%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp2048	320.77 ± 5.00	124%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg128	43.51 ± 0.05	111%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg256	43.35 ± 0.09	110%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg512	42.60 ± 0.31	105%

Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation

Which file should I choose?

Click here for details

A great write up with charts showing various performances is provided by Artefact2 here

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.

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.

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.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

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.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

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.

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.

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.