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

README.md

@@ -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)
 
@@ -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)
 

ggml-cuda.cu

@@ -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
@@ -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