Commit Graph

19 Commits

Author SHA1 Message Date
Georgi Gerganov
f93af02488
sync : ggml (conv 1d + 2d updates, UB fixes) (#3468)
* sync : ggml (conv 1d + 2d updates)

ggml-ci

* ggml : fix UB in q5_0 and q5_1 quantize code

ggml.c:1033:39: runtime error: left shift of 1 by 31 places cannot be represented in type 'int'
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior

ggml.c:1081:39: runtime error: left shift of 1 by 31 places cannot be represented in type 'int'
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior

ggml-ci

* tests : fix UB in test-quantize-perf
2023-10-04 15:29:58 +03:00
Tameem
79f34abddb
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453)
* Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v.

The RVV intrinsics is added for the following quantize row functions
   quantize_row_q8_0
   quantize_row_q8_1

The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1
   ggml_vec_dot_q4_0_q8_0
   ggml_vec_dot_q4_1_q8_1
   ggml_vec_dot_q5_0_q8_0
   ggml_vec_dot_q5_1_q8_1

And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics

Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>

* Added RVV intrinsics support for k_quants

This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64
   ggml_vec_dot_q2_K_q8_K
   ggml_vec_dot_q3_K_q8_K
   ggml_vec_dot_q4_K_q8_K
   ggml_vec_dot_q5_K_q8_K
   ggml_vec_dot_q6_K_q8_K

Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>

---------

Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 21:38:19 +03:00
Eric Sommerlade
b52b29ab9d
arm64 support for windows (#3007)
Co-authored-by: Cebtenzzre <cebtenzzre@gmail.com>
2023-09-12 21:54:20 -04:00
Georgi Gerganov
178b1850eb
k-quants : fix zero-weight guard in Q6_K (ref #3040) 2023-09-06 12:40:57 +03:00
Kawrakow
d59bd97065
Guard against all weights in a super-block being zero (#3010)
* Guard against all weights in a super-block being zero

* Also guard against extremely small weights

Closes #2982 

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-09-05 09:55:33 +02:00
Jhen-Jie Hong
21f3d1be86
k-quants : fix build on armv7 (android only) (#2920)
* k-quants : fix build on armv7

* ggml : cleanup unused arm32 specific impl

* k-quants : avoid some unused vzero / mzero define

* ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 15:23:45 +03:00
Cebtenzzre
ef15649972
build : fix most gcc and clang warnings (#2861)
* fix most gcc and clang warnings

* baby-llama : remove commented opt_params_adam

* fix some MinGW warnings

* fix more MinGW warnings
2023-09-01 16:34:50 +03:00
Ronny Brendel
3af6b86301
ggml : tiny ggml_vec_dot_q4_K_q8_K AVX2 improvement (#2819) 2023-08-28 15:51:08 +03:00
Kawrakow
bac66994cf
Quantization imrovements for k_quants (#2707)
* Improve LLaMA-2 2-, 3- and 4-bit quantization

* Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2
* Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2
* Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of
  attention.wv and feed_forward.w2

This leads to a slight model sized increase as follows:
Q2_K  : 2.684G vs 2.670G
Q3_K_S: 2.775G vs 2.745G
Q3_K_M: 3.071G vs 3.057G
Q4_K_S: 3.592G vs 3.563G

LLaMA-2 PPL for context 512 changes as follows:
Q2_K  : 6.6691 vs 6.8201
Q3_K_S: 6.2129 vs 6.2584
Q3_K_M: 6.0387 vs 6.1371
Q4_K_S: 5.9138 vs 6.0041

There are improvements for LLaMA-1 as well, but they are
way smaller than the above.

* Minor 4-bit quantization improvement

For the same model size as previus commit, we get
PPL = 5.9069 vs 5.9138.

* Some more fine tuning

* Adding make_qkx2_quants

With it, we get PPL = 5.8828 for L2-7B Q4_K_S.

* Another minor improvement

* Q2_K improvement

Smaller model, lower perplexity.
 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201
12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178

It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk,
which are Q2_K

* Iterating

* Revert Q5_K back to make_qkx1_quants

* Better Q6_K

* make_qkx2_quants is better for Q5_K after all

* Fix after rebasing on master

* Fix for changed tensor names

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 19:14:09 +03:00
Lee
a9559bf77b
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405) 2023-07-28 21:17:45 +03:00
katsu560
be2301bcda
k_quants : add AVX support to dot functions with QK_K as 64 (#2339)
* add AVX to ggml_vec_dot_q2_K_q8_K()

* add AVX to ggml_vec_dot_q3_K_q8_K()

* add AVX to ggml_vec_dot_q4_K_q8_K()

* add AVX to ggml_vec_dot_q5_K_q8_K()

* add AVX to ggml_vec_dot_q6_K_q8_K()

* refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 15:13:41 +03:00
Kawrakow
42f70cb2f6
Fix scalar version of Q5_K when QK_K = 64 (#2362)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-24 12:55:02 +03:00
Georgi Gerganov
9225baef71
k-quants : fix indentation 2023-06-26 20:10:52 +03:00
katsu560
5743ca8092
k-quants : add AVX support to dot functions (#1916)
* k_quants : add AVX support

* k_quants : apply review comments
2023-06-26 19:46:07 +03:00
Kawrakow
6769e944c7
k-quants : support for super-block size of 64 (#2001)
* k_quants: WIP super-blocks with 64 weights

* k_quants: WIP super-blocks with 64 weights

Q6_K scalar and AVX2 works

* k_quants: WIP super-blocks with 64 weights

Q4_K scalar and AVX2 works

* k_quants: WIP super-blocks with 64 weights

Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower
than the scalar implementation)

* k_quants: WIP super-blocks with 64 weights

Q3_K scalar and AVX2 works.

* k_quants: WIP super-blocks with 64 weights

Q5_K scalar and AVX2 works, and with that all
k_quants are done on AVX2 and scalar

* k_quants: WIP super-blocks with 64 weights

Q6_K working on CUDA. Cannot make it run quite as gast as
with super-blocks with 256 weigths: 8% slower on 4080,
20% slower on the 1660 (but there we fit 1 less layer on the
GPU because pf the larger model size), so some fraction of
these 20% is due to that,

* k_quants: WIP super-blocks with 64 weights

Q4_K working on CUDA. ~10% slower on GTX-1660,
16% slower on 4080.

* k_quants: WIP super-blocks with 64 weights

Q2_K working on CUDA. ~3% slower on GTX-1660,
10% slower on 4080.

* k_quants: WIP super-blocks with 64 weights

Q3_K working on CUDA.

* k_quants: WIP super-blocks with 64 weights

Q5_K working on CUDA, and with this CUDA is done.

* k_quants: WIP super-blocks with 64 weights

Q6_K working on ARM_NEON

* k_quants: WIP super-blocks with 64 weights

Q4_K working on ARM_NEON, but quite a bit slower than 256 weights

* k_quants: WIP super-blocks with 64 weights

Q2_K working on ARM_NEON, but quite a bit slower than 256 weights

* k_quants: WIP super-blocks with 64 weights

Q3_K working on ARM_NEON, but quite a bit slower than 256 weights.

* k_quants: WIP super-blocks with 64 weights

Q5_K working on ARM_NEON, but quite a bit slower than 256 weights.

With that, we have full support for ARM_NEON, although
performance is not quite there.

* k_quants: WIP super-blocks with 64 weights

Slightly more efficient Q3_K and Q5_K

* k_quants: WIP super-blocks with 64 weights

Another small improvement for Q3_K and Q5_K on ARM_NEON

* k_quants: WIP super-blocks with 64 weights

Yet another speedup for Q5_K on ARM_NEON.
We are now within 10% of the QK_K = 256 version.

* k_quants: WIP super-blocks with 64 weights

* We are able to pass preprocessor macros to the Metal
  compiler
* Q6_K works and is actually slightly more efficient than
  the QK_K = 256 version (25.2 ms vs 25.8 ms)

* k_quants: WIP super-blocks with 64 weights

Q4_K works on Metal and is actually slightly faster
than QK_K = 256 (21.95 ms vs 24.0 ms).

* k_quants: WIP super-blocks with 64 weights

Q2_K works on Metal and is very slightly faster
than QK_K = 256 (23.8 ms vs 24.2 ms).

* k_quants: WIP super-blocks with 64 weights

Q3_K works on Metal and is slightly faster
than QK_K = 256 (26.6 ms vs 28.3 ms).

* k_quants: WIP super-blocks with 64 weights

Q5_K works on Metal and is slightly faster
than QK_K = 256 (23.7 ms vs 26.3 ms).

* k_quants: call them _K, not _k, also on Metal

* k_quants: correctly define QK_K in llama.cpp

* Fixed bug in q4_K quantization added with the 64-block addition

* Simplify via lambda

* k_quants: swicth Q3_K to 4-bit scales when QK_K = 64

Otherwise there isn't much benefit from this
quantization type. There is some very slight loss
in accuracy, but we reduce size by ~7%.
E.g., for OpenLLaMA-3B, Q3_K_S perplexity is
8.6131 with 8-bit scales and 8.6352 with 4-bit,
while file size decreases from 1.53G to 1.44G.

* k_quants: switch Q4_K to 4-bit scales when QK_K = 64

 Here the loss in accuracy is greater than for Q3_K,
 but the Q4_K points still move further to the left on
 the perplexity vs size curve.

* k_quants: forgot to add the Metal changes in last commit

* k_quants: change Q5_K to be type 0 when QK_K = 64

Still needs AVX2 implementation

* k_quants: AVX2 implementation for new 64-weight Q5_K

* k_quants: 10% faster ARM_NEON Q5_K dot product

* k_quants: fixed issue caused by merging with master

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 19:43:07 +03:00
Artyom Lebedev
3f1223155a
k-quants : GCC12 compilation fix (#1792) 2023-06-10 22:51:36 +03:00
Georgi Gerganov
0bf7cf1b29
Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)"
This reverts commit 8432d4d9f7.
2023-06-08 20:48:14 +03:00
le.chang
8432d4d9f7
ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738) 2023-06-08 19:47:56 +03:00
Georgi Gerganov
5c64a0952e
k-quants : allow to optionally disable at compile time (#1734)
* k-quants : put behind optional compile flag LLAMA_K_QUANTS

* build : enable k-quants by default
2023-06-07 10:59:52 +03:00