Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 #1508

Merged
merged 5 commits into from
May 19, 2023
Merged

ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 #1508

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
d627025
ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
ggerganov May 17, 2023
3094f64
llama : bump LLAMA_FILE_VERSION to 3
ggerganov May 19, 2023
8b71329
cuda : update Q4 and Q8 dequantize kernels
ggerganov May 19, 2023
a443497
ggml : fix AVX dot products
ggerganov May 19, 2023
c6d8255
readme : update performance table + hot topics
ggerganov May 19, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 11 additions & 10 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++

**Hot topics:**

- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)

Expand Down Expand Up @@ -333,16 +334,16 @@ Several quantization methods are supported. They differ in the resulting model d

| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |
|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
| 7B | perplexity | 5.9066 | 6.1565 | 6.0910 | 5.9862 | 5.9481 | 5.9069 |
| 7B | file size | 13.0G | 4.0G | 4.8G | 4.4G | 4.8G | 7.1G |
| 7B | ms/tok @ 4th | 128 | 50 | 54 | 75 | 83 | 75 |
| 7B | ms/tok @ 8th | 123 | 44 | 52 | 53 | 58 | 72 |
| 7B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 |
| 13B | perplexity | 5.2543 | 5.3860 | 5.3607 | 5.2856 | 5.2706 | 5.2548 |
| 13B | file size | 25.0G | 7.6G | 9.1G | 8.4G | 9.1G | 14G |
| 13B | ms/tok @ 4th | 239 | 93 | 101 | 150 | 164 | 141 |
| 13B | ms/tok @ 8th | 240 | 81 | 96 | 96 | 104 | 136 |
| 13B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 |
| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G |
| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 |
| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 |
| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |
| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G |
| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 |
| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 |
| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |

### Perplexity (measuring model quality)

Expand Down
14 changes: 7 additions & 7 deletions ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,19 +42,19 @@ typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y,
#define QK4_0 32
#define QR4_0 2
typedef struct {
float d; // delta
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");

#define QK4_1 32
#define QR4_1 2
typedef struct {
float d; // delta
float m; // min
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");

#define QK5_0 32
#define QR5_0 2
Expand All @@ -78,10 +78,10 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
#define QK8_0 32
#define QR8_0 1
typedef struct {
float d; // delta
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");

#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
#define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec
Expand Down
Loading