Skip to content

Commit a751a45

Browse files
mishig25mishig25
authored
[gguf] document new types (#1305)
1 parent d49f678 commit a751a45

File tree

1 file changed

+8
-1
lines changed

1 file changed

+8
-1
lines changed

‎docs/hub/gguf.md

Lines changed: 8 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -63,11 +63,17 @@ Find more information [here](https://github.com/huggingface/huggingface.js/tree/
6363

6464
| type | source | description |
6565
|---------------------------|--------|-------------|
66+
| F64 | [Wikipedia](https://en.wikipedia.org/wiki/Double-precision_floating-point_format) | 64-bit standard IEEE 754 double-precision floating-point number. |
67+
| I64 | [GH](https://github.com/ggerganov/llama.cpp/pull/6062) | 64-bit fixed-width integer number. |
6668
| F32 | [Wikipedia](https://en.wikipedia.org/wiki/Single-precision_floating-point_format) | 32-bit standard IEEE 754 single-precision floating-point number. |
69+
| I32 | [GH](https://github.com/ggerganov/llama.cpp/pull/6045) | 32-bit fixed-width integer number. |
6770
| F16 | [Wikipedia](https://en.wikipedia.org/wiki/Half-precision_floating-point_format) | 16-bit standard IEEE 754 half-precision floating-point number. |
71+
| BF16 | [Wikipedia](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) | 16-bit shortened version of the 32-bit IEEE 754 single-precision floating-point number. |
72+
| I16 | [GH](https://github.com/ggerganov/llama.cpp/pull/6045) | 16-bit fixed-width integer number. |
6873
| Q8_0 | [GH](https://github.com/huggingface/huggingface.js/pull/615#discussion_r1557654249) | 8-bit round-to-nearest quantization (`q`). Each block has 32 weights. Weight formula: `w = q * block_scale`. Legacy quantization method (not used widely as of today). |
6974
| Q8_1 | [GH](https://github.com/huggingface/huggingface.js/pull/615#discussion_r1557682290) | 8-bit round-to-nearest quantization (`q`). Each block has 32 weights. Weight formula: `w = q * block_scale + block_minimum`. Legacy quantization method (not used widely as of today) |
7075
| Q8_K | [GH](https://github.com/ggerganov/llama.cpp/pull/1684#issue-1739619305) | 8-bit quantization (`q`). Each block has 256 weights. Only used for quantizing intermediate results. All 2-6 bit dot products are implemented for this quantization type. Weight formula: `w = q * block_scale`. |
76+
| I8 | [GH](https://github.com/ggerganov/llama.cpp/pull/6045) | 8-bit fixed-width integer number. |
7177
| Q6_K | [GH](https://github.com/ggerganov/llama.cpp/pull/1684#issue-1739619305) | 6-bit quantization (`q`). Super-blocks with 16 blocks, each block has 16 weights. Weight formula: `w = q * block_scale(8-bit)`, resulting in 6.5625 bits-per-weight. |
7278
| Q5_0 | [GH](https://github.com/huggingface/huggingface.js/pull/615#discussion_r1557654249) | 5-bit round-to-nearest quantization (`q`). Each block has 32 weights. Weight formula: `w = q * block_scale`. Legacy quantization method (not used widely as of today). |
7379
| Q5_1 | [GH](https://github.com/huggingface/huggingface.js/pull/615#discussion_r1557682290) | 5-bit round-to-nearest quantization (`q`). Each block has 32 weights. Weight formula: `w = q * block_scale + block_minimum`. Legacy quantization method (not used widely as of today). |
@@ -77,13 +83,14 @@ Find more information [here](https://github.com/huggingface/huggingface.js/tree/
7783
| Q4_K | [GH](https://github.com/ggerganov/llama.cpp/pull/1684#issue-1739619305) | 4-bit quantization (`q`). Super-blocks with 8 blocks, each block has 32 weights. Weight formula: `w = q * block_scale(6-bit) + block_min(6-bit)`, resulting in 4.5 bits-per-weight. |
7884
| Q3_K | [GH](https://github.com/ggerganov/llama.cpp/pull/1684#issue-1739619305) | 3-bit quantization (`q`). Super-blocks with 16 blocks, each block has 16 weights. Weight formula: `w = q * block_scale(6-bit)`, resulting. 3.4375 bits-per-weight. |
7985
| Q2_K | [GH](https://github.com/ggerganov/llama.cpp/pull/1684#issue-1739619305) | 2-bit quantization (`q`). Super-blocks with 16 blocks, each block has 16 weight. Weight formula: `w = q * block_scale(4-bit) + block_min(4-bit)`, resulting in 2.5625 bits-per-weight. |
86+
| IQ4_NL | [GH](https://github.com/ggerganov/llama.cpp/pull/5590) | 4-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`. |
8087
| IQ4_XS | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 4-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 4.25 bits-per-weight. |
8188
| IQ3_S | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 3-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 3.44 bits-per-weight. |
8289
| IQ3_XXS | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 3-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 3.06 bits-per-weight. |
8390
| IQ2_XXS | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 2-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 2.06 bits-per-weight. |
8491
| IQ2_S | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 2-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 2.5 bits-per-weight. |
8592
| IQ2_XS | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 2-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 2.31 bits-per-weight. |
8693
| IQ1_S | [HF](https://huggingface.co/CISCai/OpenCodeInterpreter-DS-6.7B-SOTA-GGUF/blob/main/README.md?code=true#L59-L70) | 1-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 1.56 bits-per-weight. |
87-
| IQ4_NL | [GH](https://github.com/ggerganov/llama.cpp/pull/5590) | 4-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`. |
94+
| IQ1_M | [GH](https://github.com/ggerganov/llama.cpp/pull/6302) | 1-bit quantization (`q`). Super-blocks with 256 weights. Weight `w` is obtained using `super_block_scale` & `importance matrix`, resulting in 1.75 bits-per-weight. |
8895

8996
*if there's any inaccuracy on the table above, please open a PR on [this file](https://github.com/huggingface/huggingface.js/blob/main/packages/gguf/src/quant-descriptions.ts).*

0 commit comments

Comments
 (0)