maddes8cht commited on
Commit
96ed249
1 Parent(s): e97f2aa

"Update README.md"

Browse files
Files changed (1) hide show
  1. README.md +7 -19
README.md CHANGED
@@ -24,21 +24,7 @@ I'm constantly enhancing these model descriptions to provide you with the most r
24
  - Model creator: [mosaicml](https://huggingface.co/mosaicml)
25
  - Original model: [mpt-30b-instruct](https://huggingface.co/mosaicml/mpt-30b-instruct)
26
 
27
- # Important Update for Falcon Models in llama.cpp Versions After October 18, 2023
28
-
29
- As noted on the [Llama.cpp GitHub repository](https://github.com/ggerganov/llama.cpp#hot-topics), all new Llama.cpp releases after October 18, 2023, will require a re-quantization due to the new BPE tokenizer.
30
-
31
- **Good news!** I am glad that my re-quantization process for Falcon Models is nearly complete. Download the latest quantized models to ensure compatibility with recent llama.cpp software.
32
-
33
- **Key Points:**
34
-
35
- - **Stay Informed:** Keep an eye on software application release schedules using llama.cpp libraries.
36
- - **Monitor Upload Times:** Re-quantization is *almost* done. Watch for updates on my Hugging Face Model pages.
37
-
38
- **Important Compatibility Note:** Old software will work with old Falcon models, but expect updated software to exclusively support the new models.
39
-
40
- This change primarily affects **Falcon** and **Starcoder** models, with other models remaining unaffected.
41
-
42
 
43
 
44
 
@@ -50,19 +36,21 @@ The core project making use of the ggml library is the [llama.cpp](https://githu
50
 
51
  # Quantization variants
52
 
53
- There is a bunch of quantized files available. How to choose the best for you:
54
 
55
  # Legacy quants
56
 
57
  Q4_0, Q4_1, Q5_0, Q5_1 and Q8 are `legacy` quantization types.
58
  Nevertheless, they are fully supported, as there are several circumstances that cause certain model not to be compatible with the modern K-quants.
59
- Falcon 7B models cannot be quantized to K-quants.
 
 
60
 
61
  # K-quants
62
 
63
- K-quants are based on the idea that the quantization of certain parts affects the quality in different ways. If you quantize certain parts more and others less, you get a more powerful model with the same file size, or a smaller file size and lower memory load with comparable performance.
64
  So, if possible, use K-quants.
65
- With a Q6_K you should find it really hard to find a quality difference to the original model - ask your model two times the same question and you may encounter bigger quality differences.
66
 
67
 
68
 
 
24
  - Model creator: [mosaicml](https://huggingface.co/mosaicml)
25
  - Original model: [mpt-30b-instruct](https://huggingface.co/mosaicml/mpt-30b-instruct)
26
 
27
+ MPT-7b and MPT-30B are part of the family of Mosaic Pretrained Transformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
 
29
 
30
 
 
36
 
37
  # Quantization variants
38
 
39
+ There is a bunch of quantized files available to cater to your specific needs. Here's how to choose the best option for you:
40
 
41
  # Legacy quants
42
 
43
  Q4_0, Q4_1, Q5_0, Q5_1 and Q8 are `legacy` quantization types.
44
  Nevertheless, they are fully supported, as there are several circumstances that cause certain model not to be compatible with the modern K-quants.
45
+ ## Note:
46
+ Now there's a new option to use K-quants even for previously 'incompatible' models, although this involves some fallback solution that makes them not *real* K-quants. More details can be found in affected model descriptions.
47
+ (This mainly refers to Falcon 7b and Starcoder models)
48
 
49
  # K-quants
50
 
51
+ K-quants are designed with the idea that different levels of quantization in specific parts of the model can optimize performance, file size, and memory load.
52
  So, if possible, use K-quants.
53
+ With a Q6_K, you'll likely find it challenging to discern a quality difference from the original model - ask your model two times the same question and you may encounter bigger quality differences.
54
 
55
 
56