Commit Graph

33 Commits

Author SHA1 Message Date
Kawrakow 2bafb03aac Fix bug and D < 128 case for Q8_0 k-cache (#52)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-13 07:19:47 +03:00
Kawrakow e25c2e7ec2 Quantized Flash Attention for all supported CPU platforms (#51)
* NEON Flash Attention: add support for Q8_0, Q4_0, Q4_1

* NEON Flash Attention: quantized K*Q for q4_0

I could finally take advantage of the matrix multiplication
templates. We get quite a bit of speedup that way for q4_0:
For Gemma-2b using mul_mat_qX_0_q8_0<DequantizerQ40, q_step>
results in PP-2048 = 287 t/s vs 268 t/s when converting the
q4_0 k-cache and Q to fp16 and using fp16 multiplication.

* NEON Flash Attention: quantized K*Q for q4_1

* NEON Flash Attention: quantized K*Q for q8_0

This makes quite a bit of difference:
For Gemma2-2b PP-8192 is 228 t/s with quantized K*Q vs
178 t/s when converting things to fp16 and using fp16
matrix multiplication.
We have PP-512 = 307 t/s, so PP-8192 is now ~75% of the
performance of PP-512. In contrast, llama.cpp with Q8_0
cache is 38% of PP-512.

* Zen4 Flash Attention: quantized K*Q for q4_0, q4_1, q8_0

* AVX2 Flash Attention: quantized K*Q for q4_0, q4_1, q8_0

* Tidy up FlashMS

* Delete no longer used stuff

With the usage of quantized matrix multiplications for
quantized k- and/or v-cache, we no longer need the
helper methods loading entire rows.

* Disallow mixing bf16 with other types for kv caches

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-12 19:03:20 +03:00
Kawrakow 7874e4425f AVX2 Flash Attention 2 (#50)
* AVX2 Flash Attention: add ability to use Q8_0 for kv-cache

* AVX2 Flash Attention: add ability to use Q4_0 for kv-cache

* AVX2 Flash Attention: add ability to use Q4_1 for kv-cache

* Fix Zen4

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-11 19:55:42 +03:00
Kawrakow dd52475507 ARM_NEON Flash Attention (#49)
* NEON Flash Attention - first working version

Simply reuse the Zen4/AVX2 implementation, but use
f16 for the K*Q multiplication and V*softmax(K*Q) accumulation.
This makes the FlashMS portion somewhat awkward because we
do not have fast f16 implementations for expf (and tanh when
softcap is enabled), so we need to convert back-and-fort
to f32.

FA is slightly faster than no-FA for the 4B TriLM model,
but lightly slower for Gemma-2b.

* NEON Flash Attention - convert Q to f16 before computing Q*K

* NEON Flash Attention - use fp32 for K*Q operations

Else I get wrong results for LLaMA-3.1-8B (but it works for
Gemma-2b).

* Delete commented out stuff

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-11 10:26:49 +03:00
Kawrakow 808f08787a AVX2 Flash Attention (#48)
* First version of AVX2 Flash attention

I simply took the Zen4 implementation and converted
platform specific stuff to methods of a struct providing
data loading/storing, conversions, multiply, add, etc.

Most likely not optimal as the Zen4 strategy has been
designed based on having 32 512-bit registers, so basically
we can have 4X more data stored in vector registers compared
to AVX2 with 16 x 256-bit.

It still gives a small speedup (~4% at 2048 tokens) for Gemma-2b.

* Fix Zenn4 parts broken via the AVX2 change

* Try smaller q_step - no improvement

* Fix ARM_NEON

I had forgotten to guard the AVX2/Zen4 implementation against __aarch64__

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-10 19:17:04 +03:00
Kawrakow 49cbbc9fee iq2_tn: slightly better performance on AVX2 (#47)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-10 16:21:57 +03:00
Kawrakow 4a5d5e207d Adding IQ1_TN - 1.6875 bpw for TriLM ternary models (#44)
* Adding iq1_tn - 1.6875 bpw for TriLM ternary models

* iq1_tn: NEON

* iq1_tn: faster NEON

* iq2_bn: improve performance on NEON

We now get TG-128 = 100 t/s for Bitnet-3B-1.58b!

* iq1_tn: improve AVX2

PP-512 goes to 533 t/s up from 455.
TG-128 @ 2 threads goes to 16.6 t/s up from 14.2.
However, we seem to have a bottleneck somewhere as
TG saturates at 8 threads.

* iq1_tn: improve Zen4

PP-512 goes to 485 t/s up from 352. With FA we get 545 t/s up from 380.
TG-128 @ 1 thread goes to 12.4 t/s up from 10.4.
However, we seem to have a bottleneck somewhere as
TG saturates at 8 threads.

* iq2_bn: improve on Zen4

We now get PP-512 = 614 t/s up from 542 t/s

* iq2_bn: improve AVX2 implementation

We now get PP-512 = 753 t/s up from 680 t/s.

* Remove unnecessary barrier in ggml_compute_forward_mul_mat

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-09 14:56:34 +03:00
Kawrakow f2ef628e4e iq2_tn: slightly faster PP (#43)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-08 12:41:44 +03:00
Kawrakow 18f5bb47d8 Add support for bf16 to iqk_mul_mat (#39)
* WIP: adding BF16 support to iqk_mul_mat

* Minor

* Improve TG speed (when not memory bound)

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-05 07:48:27 +03:00
Kawrakow 02e4cc0c18 Zen4 Flash Attention - bf16 support (#38)
* Zen4 Flash Attnetion: WIP bf16

* Zen4 Flash Attnetion: bf16 seems to be working

* Zen4 Flash Attnetion: improving bf16

* Zen4 Flash Attnetion: improving bf16

It is better (slightly faster) to first convert Q
to bf16 before processing each block of q_step rows.
This requires D*q_step*sizeof(bf16) bytes, so at
most 4 kb for the head sizes we support, so we can
just allocate on the stack instead of reserving and
passing a work buffer in ggml.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-05 07:46:47 +03:00
Kawrakow d47e1c63b3 Performance improvements for legacy quants on ARM_NEON (#37)
* WIP: trying to improve legacy quants

* WIP: trying to improve legacy quants

With this commit PP-512 for LlaMA-3.1-8B goes from
72 t/s to 87.2 t/s for q4_0, and from 61.5 t/s to 73.9 t/s
for q4_1, so 20+% improvement for both.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-04 07:24:04 +03:00
Kawrakow 0449090ae8 Zen4 Flash Attnetion 2 (#36)
* Zen4 Flash Attnetion: WIP generalize to other types

Now loading of data from K and V is done via a template parameter,
so this should make it easy to generalize to typ[es other than
F16 for the K and V cache.

* Zen4 Flash Attnetion: it works for q4_0 and q8_0

* Zen4 Flash Attnetion: small q8_0 performance improvement

* Zen4 Flash Attnetion: add q4_1

* Delete unused stuff

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-04 07:20:55 +03:00
Kawrakow 724854e7db Fix Zen4 Flash Attention (#35)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-02 15:54:24 +03:00
Kawrakow 59c2e77869 Zen4 Flash Attention (#32)
* Zen4 flash attention: moving useful parts from the kq_fused_softmax branch

* Add flash attention with soft-cap and fix D = 256 case

* Flash attention refinements

* Update FlashAttn comment

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-09-01 16:08:21 +03:00
Kawrakow af2bb96de5 Fix Zen4 implementation of iq3_k, iq4_k, iq5_k
See comments in f3a823ce729a7db33e7d4375eae7291bbe6196db
2024-08-09 16:00:31 +02:00
Kawrakow 3f67708b91 iq6_k: AVX2 2024-08-09 16:00:31 +02:00
Kawrakow ed462a512a iq6_k: NEON
Respectable performance, only slightly slower than Q6_K.
2024-08-09 16:00:31 +02:00
Kawrakow ef32a01c2a iq6_k: slightly better Zen4 iqk_mul_mat
We now arrive at pp-512 = 147 t/s for LLaMA-3.1-8B.
TG-128 is 9.5 t/s. This is better than last commit,
but still kind of slow compared to Q6_K.

My last commit message is wrong: also iq3_k needs a fix
for overflow.
2024-08-09 16:00:31 +02:00
Kawrakow 0bee1c0c0a iq6_k: Zen4 iqk_mul_mat
We need to do 4 shuffles to get the non-uniform values, so this
makes it slower than other iqX_k quants.

And then I realized that I was using the standard Zen4 template for
all iqX_k quants. The standard template converts the 32-bit integers
obtained after _mm512_dpbusds_epi32 back to 16 bits, and then multiples
with 16-bit block scales. But this can overfow for iq4_k, iq5_k, and
iq6_k. I guess, I did not notice with iq4_k and iq5_k because the
PPL difference to CUDA was relatively small, and I attributed it to
Q8_K not being accurate enough for the activations. But for iq6_k
the PPL difference was much too big to be attributable to Q8_K
inaccuracies, so that's when I realized that I cannot be packing
the _mm512_dpbusds_epi32 result into 16 bit for 4-,5-,6-bit iqX_k
quants.

For now I fixed it for iq6_k, but the outcome is that it is
significantly slower than Q6_K: I get PP-512 = 125 t/s for
LLaMA-3.1-8B vs 180 t/s for Q6_K, so I need to look for a better
approach.
2024-08-09 16:00:31 +02:00
Kawrakow 58a323f585 Adding IQ2_TN for use with ternary models (#13)
* iq2_tn: TriLM specific 2.0625 bpw quantization

Quantize/dequantize/scale dot product.

I get 46 t/s for the TriLM-3.9B with any SIMD!
Finally a compiler doing a decent job auto-vectorizing the
scalar implementation.

* iq2_tn: AVX512

Just reusing the k-quants template gets us to PP-512 = 376 t/s,
TG-128 = 47.6 t/s for TriLM-3.9B.

* iq2_tn: AVX512

With this tweak we get to PP-512 = 431 t/s.

* iq2_tn: AVX512

With this tweak we get TG-128 = 19.58 / 35.18 t/s for 1 / 2 threads.
At 4 threads we saturate at 48.41 t/s, and then performance slowly
degrades with increasing number of threads.

* iq2_tn: AVX2

PP512 = 440 t/s on the Ryzen-5975WX.
We should be able to do better.

* iq2_tn: initial NEON version

* iq2_tn: NEON

For TriLM-3.9B running on the M2-Max we get PP-512 = 193.5 t/s,
TG-128 = 75.5 t/s. This is in line with what we have for
iq2_bn ant 3.3B Bitnet.

* iq2_tn: Metal

For TriLM-3.9B on a 30-core M2-Max we get PP-512 = 890 t/s,
TG-128 = 98.5 t/s.

* iq2_tn: CUDA

For TriLM-3.9B running on RTX-4080 we get PP-512 = 9936 t/s,
TG-128 = 299.2 t/s.

* iq2_tn: AVX2 PP improvement

We now get PP-512 = 490.73 t/s for TriLM-3.9B on the Ryzen-5975WX.
We have PP-512 = 636.61 t/s for Bintnet-3B quantized with iq2_bn.
Bintnet-3B is actually 3.4B, TriLM-3.9B is 3.99B, so we would
expect 3.43/3.99 * 636 = 546 t/s, so it seems we still have something
that is not quite optimal in iq2_tn.

* iq2_tn: small NEON improvement

For TriLM-3.9B we now get PP-512 = 206.6 t/s and TG-128 = 76.4 t/s.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-08-07 07:56:09 +02:00
Kawrakow 2927d4f841 iq3_k: NEON 2024-08-01 09:38:06 +02:00
Kawrakow 9c1eea6048 iq3_k: AVX2 iqk_mul_mat
We get PP-512 = 196 t/s for LLaMA-3.1-8B on the Ryzen-5975WX.
2024-08-01 09:38:06 +02:00
Kawrakow a9fa3b1563 iq3_k: AVX512 iqk_mul_mat
We get PP-512 = 180 t/s, TG-128(4 threads) = 16.35 on the Ryzen-7950X
for LLaMA-3.1-8B.
In comparison, iq3_s has PP-512 = 96 t/s, TG-128 = 7.6 t/s with
iqk_mul_mat, and PP-512 = 28 t/s, TG-128 = 6.8 t/s in mainline llama.cpp
2024-08-01 09:38:06 +02:00
Kawrakow 0ab3f0ff86 iq5_k: NEON 2024-08-01 09:38:06 +02:00
Kawrakow daf608e227 iq5_k: AVX512 2024-08-01 09:38:06 +02:00
Kawrakow e9c3ebcbe9 iq5_k: AVX2 2024-08-01 09:38:06 +02:00
Kawrakow f7ab9a13df iq2_k: NEON 2024-08-01 09:38:06 +02:00
Kawrakow cc8e351b68 iq2_k: slightly faster AVX512 2024-08-01 09:38:06 +02:00
Kawrakow 764d4675b8 iq2_k: simplify AVX512 2024-08-01 09:38:06 +02:00
Kawrakow 21319d6fca iq2_k: AVX2 2024-08-01 09:38:06 +02:00
Kawrakow 3f7dad3000 iq2_k: Basics
Quantize/dequantize, CUDA deqantize, AVX512 iqk_mul_mat.
2024-08-01 09:38:06 +02:00
Kawrakow 007d2a56b3 IQ4_K: SOTA 4-bit quantization (#6)
* iq4_k: basics

* quantize/dequantize works
* CUDA dequantize works and one can run PPL calcs. I get
  PPL = 6.5258 for LlaMA-3.1-8B, which is 1.77% above fp16.
  In comparison, q4_K_S (same size) is 2.88% above fp16.
* TG on CUDA does not work. Johannes has changed the way i-quant dot
  products are done, so need to sort out what he had in mind
* iqk_mul_mat is not implemented.

* iq4_k: TG now works on CUDA

* iq4_k: AVX512 implementation

For LLaMA-3.1-8B we get PP-512 = 182.6 t/s, TG-128 = 13.6 t/s,
so almost the same as q4_K_S.

* iq4_k: AVX2 implementation

For LLaMA-3.1-8B we get PP-512 = 203.1 t/s, TG-128 = 12.9 t/s
on the Ryzen-5975X.

* iq4_k: NEON implementation

For LLaMA-3.1-8B we get PP-512 = 60.7 t/s, TG-128 = 25.0 t/s
on the M2-Max. TG is on par with q4_K_S, PP is ~10% slower.

* iq4_k: Metal implementation

For LLaMA-3.1-8B we get PP-512 = 445 t/s, TG-128 = 46.3 t/s
on a 30-core M2-Max GPU. This is to be compared with (currently)
PP-512 = 460 t/s, TG-128 = 51 t/s for q4_K_S.

* iq4_k: scalar dot product

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-07-28 12:11:59 +02:00
Kawrakow 0ceeb11721 Merge mainline llama.cpp (#3)
* Merging mainline - WIP

* Merging mainline - WIP

AVX2 and CUDA appear to work.
CUDA performance seems slightly (~1-2%) lower as it is so often
the case with llama.cpp/ggml after some "improvements" have been made.

* Merging mainline - fix Metal

* Remove check

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-07-27 07:55:01 +02:00