ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-24 08:29:29 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
409bfe6648bf819bcbcea4abd6ba2def291b27f2
ik_llama.cpp/ggml/src/iqk/iqk_quantize.h
Andrew Chan a1c931c30c Trellis quants with CPU inference (#441) 
* WIP

* WIP

* WIP

* Testing Trellis quantization

Using 12 bits per 8 weights I get a better rmse than
iq2_xxs. I still need to see how quantizing the group-of-8
scales will affect accuracy. By AVX2 SIMDifying the search
for the best code, LLaMA-3.1-8B gets quantized in 130 seconds
on the Ryzen-7950X CPU - sluggish but still acceptable.

* Testing Trellis quantization: 4-bit quantized block scales

rmse increases by just 3%, so this is beating iq2_xss in terms
of rmse at the same 2.0625 bpw.

* Testing Trellis quantization: playing with scales and generators

* iq2_kt: quantize / dequantize

I now see that I was comparing apples to oranges:
iq2_xxs was using a weight of sigma^2/4 + x^2, while
the Trellis approach wasn't (weight = 1). Once I use the same weight,
iq2_kt is actually slightly worse than iq2_xxs in terms
of rmse, so does not look promising at this point.
Also, once each group of 8 Trellis values no longer has a
constant sum(q^2) that we can precompute, quantization
becomes significantly slower (476 seconds for LLaMA-3.1-8B).

* iq2_kt: CUDA dequantize

so we can run perplexity calcs.
As already indicated by rmse, the 2-bit trellis approach is
quite a bit worse than iq2_xxs.

* WIP

* WIP

* WIP - try larger blocks

With blocks of 32 and 16 bits per groups of 8 the brute force
seach becomes prohibitive in terms of CPU time (30+ minutes
for 8B LLaMA after SIMDifying with AVX2). The trick is to
group the points in clusters, find the nearest cluster,
and only search within the cluster.

* iq2_kt - this is better

Using blocks of 32 and 16 bits per group of 8 weights
it beats iq2_xxs in terms of PPL by a significant margin.
It is 0.0625 bpw larger, but even if we go to 15 bits per
group od 8 (so 0.0625 bpw less than iq2_xxs), PPL is still
lower.

* iq2_kt - even better

Re-quantize after determining block scales
(at the epxense of much longer quantization time).

* iq2_kt: CUDA dot product

Implemented as DMMV.
Very slow - just 81 t/s for LLaMA-3.1-8B.
Then again, Q2_K_S with forced to use DMMV only
gets 112 t/s vs 145 t/s via MMVQ. My memory is that
when the DMMV kernels were properly maintained/used,
DMMV was about on par with MMVQ for k-quants on my GPU.

* iq2_kt: very slightly faster CUDA dot product

* iq2_kt: f16 CUDA dot product

We arrive at 112 t/s.

* iq2_kt: faster f16 CUDA dot product

We arrive at 139 t/s (no FA), and 149 t/s (FA).

My RTX-4080 is ~20% slower than the RTX-6000 quoted in the
QTIP repository, so with FA (which I'm sure they also used)
we are at around ~180 t/s on their GPU, so almost matching
their performance.

* iq2_kt: faster f16 CUDA dot product

We arrive at 146 t/s (no FA), and 158 t/s (FA).
This is measured for LLaMA-3.1-8B with output.weight
left as f16.

* Minor

* Adding iq3_kt

3.125 bpw. So far does not look good on the PPL vs bpw plot.

* Forgotten change

* WIP

* WIP

* iq3_kt WIP: slowly improving

PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.8322, which is
starting to be competitive/slightly better than other quants.

* WIP

* iq3_kt WIP: slowly improving

PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.7892

* iq3_kt WIP: slowly improving

PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.7689 after shrinking
by 0.015 bpw by using iq4_k instead of q5_k for attn_v.

* iq3_kt WIP: speed up quantization

Nearly 60% improvement of quantization speed by having the
points nelonging to a cluster copied to contiguous memory
during initialization, and then accessed sequantially while
searching for the closest point. LLaMA-3.1-8B now gets
quantized in ~150 seconds on the Ryzen-5975WX.

* iq3_kt speed up quantization

Same trick as last commit applied to iq2_kt. Here we get
an even larger speedup: quantization time on the Ryzen-5975WX
for LLaMA-3.1-8B drops to 195 seconds from 375 seconds!

* iq3_kt: CUDA dot product

* iq2_kt: SOTA

We arrive at
PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.2406
PPL(LLaMA-2-7B,            4096) = 6.4179

* iq2_kt: SOTA

We arrive at
PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.1642
PPL(LLaMA-2-7B,            4096) = 6.3920

* Adding iq4_kt - not competitive at this point

* WIP

* WIP

* iq4_kt: CUDA dot product

* iq4_kt: minor tweaks

* iq2_kt: SOTA

We arrive at
PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.1642
PPL(LLaMA-2-7B,            4096) = 6.3920

* iq2_kt: SOTA

We arrive at
PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.0297
PPL(LLaMA-2-7B,            4096) = 6.3913

Ah, quantization is faster too. About 20% faster.

* iq3_kt: small improvements and faster quantization

* iq2_kt: SOTA

We arrive at
PPL(LLaMA-3.1-8B-Instruct, 8192) = 8.9627
PPL(LLaMA-2-7B,            4096) = 6.3825

Quantization is faster too: ~200 seconds for LLaMA-3.1-8B
on Ryzen-5975WX.

* iq3_kt: small progress

* WIP

* iq4_kt: go to 4.0 bpw

15 bits per group of 4, plus 8 bit scales ifor blocks of 32.
This gives a slightly better PPL than iq4_kss.

* iq4_kt: very slightly better

at the expense of much longer quantization time.

* iq4_kt: failed attemt to adjust CUDA dot product

It was working for 4.125 bpw. But after changing to 4.0 bpw
there is something wrong and I don't see the bug.

* DRY

* DRY

* iq4_kt: CUDA dot product works

* DRY

* Report actual bpw

* Minor tweaks

* Checkpoint

Go to groups of 8 for iq3_kt. 2 x 8 = 16 bits for the magnitude
plus 1 bpw for the sign. It goves a visible improvement in the
PPL vs bpw plot, but that comes at the expense of much longer
quantization time (7.5 minutes for LLaMA-3.1-8B on the Ryzen-5975WX).

I also notices that the 3INST generator is not actually generating a
Gaussian distribution. But going to a better generator means
readjusting all the hyper-parameters, so leaving it for later.

* WIP for IQ2_KT

* WIP - working basic iq2_kt

* still super slow (0.17t/s eval)

* flatten 3inst iters + avx2 (0.3t/s eval)

* iq3_kt (0.3t/s eval) and renames

* wip buggy iq4_KT

* fix (0.22t/s eval)

* naming and remove unused fn

* cleanup

* more cleanup

* delete unused and noncompiling mmvq functions

* Some performance tweaks

* Slighty faster iq2_kt

* port Trellis struct to iq3_kt, iq4_kt

* oops untracked files

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2025-05-23 09:17:52 +03:00

302 lines
27 KiB
C
Raw Blame History

//
// Copyright (C) 2024-2025 Iwan Kawrakow
// MIT license
// SPDX-License-Identifier: MIT
//
#pragma once
#include <stdint.h>
#include <stddef.h>
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#ifdef __cplusplus
#define GGML_RESTRICT
extern "C" {
#else
#define GGML_RESTRICT restrict
#endif
void quantize_row_iq2_k_ref(const float * GGML_RESTRICT x, block_iq2_k * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_k(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_k(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_k(const block_iq2_k * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_k_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq3_k_ref(const float * GGML_RESTRICT x, block_iq3_k * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_k(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq3_k(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq3_k(const block_iq3_k * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq3_k_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_k_ref(const float * GGML_RESTRICT x, block_iq4_k * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_k(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_k(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_k(const block_iq4_k * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_k_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq5_k_ref(const float * GGML_RESTRICT x, block_iq5_k * GGML_RESTRICT y, int64_t k);
void quantize_row_iq5_k(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq5_k(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq5_k(const block_iq5_k * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq5_k_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq6_k_ref(const float * GGML_RESTRICT x, block_iq6_k * GGML_RESTRICT y, int64_t k);
void quantize_row_iq6_k(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq6_k(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq6_k(const block_iq6_k * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq6_k_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_ks_ref(const float * GGML_RESTRICT x, block_iq4_ks * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_ks(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_ks(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_ks(const block_iq4_ks * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_ks_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_kss_ref(const float * GGML_RESTRICT x, block_iq4_kss * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_kss(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_kss(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_kss(const block_iq4_kss * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_kss_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_ks_ref(const float * GGML_RESTRICT x, block_iq2_ks * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_ks(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_ks(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_ks(const block_iq2_ks * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_ks_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_kt_ref(const float * GGML_RESTRICT x, block_iq2_kt * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_kt(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_kt(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_kt(const block_iq2_kt * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_kt_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq3_kt_ref(const float * GGML_RESTRICT x, block_iq3_kt * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_kt(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq3_kt(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq3_kt(const block_iq3_kt * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq3_kt_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_kt_ref(const float * GGML_RESTRICT x, block_iq4_kt * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_kt(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_kt(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_kt(const block_iq4_kt * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_kt_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq5_ks_ref(const float * GGML_RESTRICT x, block_iq5_ks * GGML_RESTRICT y, int64_t k);
void quantize_row_iq5_ks(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq5_ks(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq5_ks(const block_iq5_ks * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq5_ks_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_nl_r4_ref(const float * GGML_RESTRICT x, block_iq4_nl_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_nl_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_nl_r4(const block_iq4_nl_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_nl_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q4_0_r8_ref(const float * GGML_RESTRICT x, block_iq4_nl_r8 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_0_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q4_0_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q4_0_r8(const block_iq4_nl_r8 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q4_0_r8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q8_0_r8_ref(const float * GGML_RESTRICT x, block_q8_0_r8 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q8_0_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q8_0_r8(const block_q8_0_r8 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q8_0_r8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q5_0_r4_ref(const float * GGML_RESTRICT x, block_q5_0_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q5_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q5_0_r4(const block_q5_0_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q5_0_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q6_0_r4_ref(const float * GGML_RESTRICT x, block_q6_0_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_0_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q6_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q6_0_r4(const block_q6_0_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q6_0_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_xs_r8_ref(const float * GGML_RESTRICT x, block_iq4_xs_r8 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_xs_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_xs_r8(const block_iq4_xs_r8 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_xs_r8_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_bn_ref (const float * GGML_RESTRICT x, block_iq2_bn * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_bn (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_bn (const block_iq2_bn * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_bn (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void vec_dot_iq2_bn_q8_K64(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_bn_r4_ref (const float * GGML_RESTRICT x, block_iq2_bn * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_bn_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_bn_r4(const block_iq2_bn * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_bn_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void vec_dot_iq2_bn_r4_q8_K64(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q3_k_r4_ref(const float * GGML_RESTRICT x, block_q3_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q3_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q3_k_r4(const block_q3_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q3_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q2_k_r4_ref(const float * GGML_RESTRICT x, block_q2_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q2_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q2_k_r4(const block_q2_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q2_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q4_k_r4_ref(const float * GGML_RESTRICT x, block_q4_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q4_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q4_k_r4(const block_q4_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q4_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q5_k_r4_ref(const float * GGML_RESTRICT x, block_q5_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q5_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q5_k_r4(const block_q5_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q5_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q6_k_r4_ref(const float * GGML_RESTRICT x, block_q6_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q6_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q6_k_r4(const block_q6_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q6_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq5_k_r4_ref(const float * GGML_RESTRICT x, block_iq5_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq5_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq5_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq5_k_r4(const block_iq5_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq5_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_k_r4_ref(const float * GGML_RESTRICT x, block_iq4_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_k_r4(const block_iq4_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq3_k_r4_ref(const float * GGML_RESTRICT x, block_iq3_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq3_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq3_k_r4(const block_iq3_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq3_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_k_r4_ref(const float * GGML_RESTRICT x, block_iq2_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_k_r4(const block_iq2_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq4_ks_r4_ref(const float * GGML_RESTRICT x, block_iq4_ks_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_ks_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq4_ks_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq4_ks_r4(const block_iq4_ks_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq4_ks_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq5_ks_r4_ref(const float * GGML_RESTRICT x, block_iq5_ks_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq5_ks_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq5_ks_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq5_ks_r4(const block_iq5_ks_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq5_ks_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq2_xxs_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_xxs_r4(const block_iq2_xxs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_xxs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_xs_r4_ref(const float * GGML_RESTRICT x, block_iq2_xs_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_xs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_xs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_xs_r4(const block_iq2_xs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_xs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq2_s_r4_ref(const float * GGML_RESTRICT x, block_iq2_s_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq2_s_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq2_s_r4(const block_iq2_s_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq2_s_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq3_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq3_xxs_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq3_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq3_xxs_r4(const block_iq3_xxs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq3_xxs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq3_s_r4_ref(const float * GGML_RESTRICT x, block_iq3_s_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq3_s_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq3_s_r4(const block_iq3_s_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq3_s_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq1_s_r4_ref(const float * GGML_RESTRICT x, block_iq1_s_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq1_s_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq1_s_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq1_s_r4(const block_iq1_s_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq1_s_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_iq1_m_r4_ref(const float * GGML_RESTRICT x, block_iq1_m_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_iq1_m_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_iq1_m_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_iq1_m_r4(const block_iq1_m_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_iq1_m_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q8_k_r8_ref(const float * GGML_RESTRICT x, block_q8_k_r8 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_k_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q8_k_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q8_k_r8(const block_q8_k_r8 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q8_k_r8_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q8_KV_ref(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_KV(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q8_KV(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q8_KV(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q8_KV_q8_KV(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q8_KV_r8_ref(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_KV_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q8_KV_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q8_KV_r8(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q8_KV_r8_q8_KV(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void iqk_quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void quantize_row_q8_K64_ref(const float * GGML_RESTRICT x, block_q8_K64 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K64(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K128(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K16(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K32(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_KR8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0_x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1_x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_2_x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void repack_f32_bf16_r16 (const void * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row);
void repack_bf16_bf16_r16(const void * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row);
void iqk_repack_tensor(struct ggml_tensor * tensor);
bool iqk_modify_tensor(struct ggml_tensor * tensor);
int iqk_repacked_type(const struct ggml_tensor * tensor); // int instead of ggml_type so we don't need to include ggml.h
bool iqk_should_modify_tensor(const struct ggml_tensor * tensor);
// So we can re-pack Microsoft's BitNet I2_S quants
void dequantize_row_ms_i2s(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void iqk_quantize_any(int from_type, int to_type,
int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3,
uint64_t nb0, uint64_t nb1, uint64_t nb2, uint64_t nb3,
const void * GGML_RESTRICT x, void * GGML_RESTRICT y, void * work_buffer,
to_float_t to_float, from_float_t from_float, int ith, int nth);
#ifdef __cplusplus
}
#endif
Reference in New Issue View Git Blame Copy Permalink