llama.cpp/ggml-quants-k.h
Kawrakow 99009e72f8
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684)
* Starting to add k-quantization to ggml

I think it is better to have quantization separate from
ggml. For now just adding the k-quants there, but it would be
better to also factor out the existing ggml quantizations.

* Adding Q3_K and Q8_K (de)-quantization

* Q3_K now working on CUDA and AVX2/scalar

CUDA is not ideal - ~50% slower than Q4_0 for
single token prediction, about the same in batch
mode (perplexity). CPU single token is ~55 ms
(on Ryzen 7950X).

* Some improvement for Q3_K on CUDA

It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0.

* Some more CUDA optimizations for Q3_K

Single token is now 20.5 ms/token (~20% slower than Q4_0).
Perplexity is on par with Q4_0.

* Adding Q4_K - scalar, AVX2, CUDA

Performance is the same or perhaps very slightly better than Q4_0 on the CPU.
On the GPU, single token prediction is ~10% better than Q4_0,
batch mode (perplexity is about the same).

* Adding Q6_K - scalar, AVX2, CUDA

Performance is ~40% lower compared to Q4_K on the CPU.
This is to be expected, considering that we are memory bound
on the CPU and the 6-bit model is ~44% larger than the 4-bit.
On the GPU, single token prediction is ~6% lower than Q4_0,
batch mode (perplexity) is even closer (but still slower).

* Adding Q5_K - scalar, AVX2, CUDA

Performance is ~20% lower compared to Q4_K on the CPU.
This is to be expected, considering that we are memory bound
on the CPU and the 5-bit model is ~22% larger than the 4-bit.
On the GPU, single token prediction is about the same as Q4_0
for both, single token and batch prediction.

* Per convention, all QX_K quantizations use Q5_K for output.weight

* Adding quantization mixes

* Quantization mixes: didn't quite get what I wanted in the last commit

* Q4_K dot product for ARM_NEON

* Q6_K dot product for ARM_NEON

* Q5_K dot product for ARM_NEON

* Adding Q3_K dot for ARM_NEON

It is 22% slower than Q4_K, despite the smaller model size.
On x86_64, where we are memory bound, the Q3_K model is
quite a bit faster than Q4_K.

* A very slightly faster ARM_NEON Q3_K dot

* Adding Q2_K - just CUDA for now

Token prediction is pretty good - about 15.5 ms on a RTX 4080.
Perplexity is about the same as Q4_K.

* Adding scalar and AVX2 Q2_K dot

* Adding ARM_NEON Q2_K dot

About the same performance as Q4_K.

* A slightly faster ARM_NEON Q2_K dot

Single token prediction is now ~36 ms on M2 Max.
The code is much simpler too.

* Fixed bug in Q2_K CUDA dot product kernel

Stranegly enough, for the few prompts I tried with the 7B model
the responses looked perfectly reasonable. Only realized something
is not quite right when I tried the larger models and started getting
nonse back.

In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X
box iusing CUDA and model fully loaded on the GPU are
  ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B.
The max number of layers that fit in VRAM for The 65B is 32.
With that, we get ~330 ms per token, which is not that much faster
than just running on the CPU (~470 ms per token).

* Don't print zeros/NaNs when no count histogram has been collected

* A 10% faster CUDA vector dot kernel for Q3_K

Q3_K is now running at ~18.5 ms / token on CUDA,
so the gap to Q4_0 is only 10%.
It seems memory acccess pattern is more important for
performance than the amount of computation the kernel
does.

* A slightly daster Q4_K AVX2 dot product

For perplexity, where we are less memory bound, time per
pass drops by ~5%. Barely measurable difference for single
token prediction.

* A slightly faster ARM_NEON A4_K dot product

* Minor

* Fix quantization error test

We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit
quantization variants.

* Fix docker build

I have been sloppy with vector reinterpret casts on ARM_NEON.
It seems clang is very forgiving in that regard.

* Added forgotten ggml.o dependence on k_quants.h to the Makefile

* Had unintentionally committed the Makefile with -Ofast enabled

* ggml : rename k_quants -> ggml-quants-k, use lowercase in code

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 22:56:18 +03:00

122 lines
5.7 KiB
C

#pragma once
#include "ggml.h"
#include <stdint.h>
#include <assert.h>
#include <stddef.h>
// Super-block size
#define QK_K 256
//
// Super-block quantization structures
//
// 2-bit quantization
// weight is represented as x = a * q + b
// 16 blocks of 16 elemenets each
// Effectively 2.5625 bits per weight
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
} block_q2_k;
static_assert(sizeof(block_q2_k) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_k block size/padding");
// 3-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elemenets each
// Effectively 3.4375 bits per weight
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
ggml_fp16_t d; // super-block scale
} block_q3_k;
static_assert(sizeof(block_q3_k) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_k block size/padding");
// 4-bit quantization
// 16 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 4.5 bits per weight
typedef struct {
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_k;
static_assert(sizeof(block_q4_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_k block size/padding");
// 5-bit quantization
// 16 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 5.5 bits per weight
typedef struct {
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_k;
static_assert(sizeof(block_q5_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_k block size/padding");
// 6-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elemenets each
// Effectively 6.5625 bits per weight
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
ggml_fp16_t d; // super-block scale
} block_q6_k;
static_assert(sizeof(block_q6_k) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_k block size/padding");
// This is only used for intermediate quantization and dot products
typedef struct {
float d; // delta
int8_t qs[QK_K]; // quants
int16_t bsums[QK_K/16]; // sum of quants in groups of 16
} block_q8_k;
static_assert(sizeof(block_q8_k) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_k block size/padding");
// Quantization
void quantize_row_q2_k_reference(const float * restrict x, block_q2_k * restrict y, int k);
void quantize_row_q3_k_reference(const float * restrict x, block_q3_k * restrict y, int k);
void quantize_row_q4_k_reference(const float * restrict x, block_q4_k * restrict y, int k);
void quantize_row_q5_k_reference(const float * restrict x, block_q5_k * restrict y, int k);
void quantize_row_q6_k_reference(const float * restrict x, block_q6_k * restrict y, int k);
void quantize_row_q8_k_reference(const float * restrict x, block_q8_k * restrict y, int k);
void quantize_row_q2_k(const float * restrict x, void * restrict y, int k);
void quantize_row_q3_k(const float * restrict x, void * restrict y, int k);
void quantize_row_q4_k(const float * restrict x, void * restrict y, int k);
void quantize_row_q5_k(const float * restrict x, void * restrict y, int k);
void quantize_row_q6_k(const float * restrict x, void * restrict y, int k);
void quantize_row_q8_k(const float * restrict x, void * restrict y, int k);
// Dequantization
void dequantize_row_q2_k(const block_q2_k * restrict x, float * restrict y, int k);
void dequantize_row_q3_k(const block_q3_k * restrict x, float * restrict y, int k);
void dequantize_row_q4_k(const block_q4_k * restrict x, float * restrict y, int k);
void dequantize_row_q5_k(const block_q5_k * restrict x, float * restrict y, int k);
void dequantize_row_q6_k(const block_q6_k * restrict x, float * restrict y, int k);
void dequantize_row_q8_k(const block_q8_k * restrict x, float * restrict y, int k);
// Dot product
void ggml_vec_dot_q2_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q3_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q4_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q5_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q6_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
// Quantization with histogram collection
size_t ggml_quantize_q2_k(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q3_k(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q4_k(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q5_k(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q6_k(const float * src, void * dst, int n, int k, int64_t * hist);