* Implementing strided computation of perplexity
* Alternative way to output PPL results
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* 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>
* Another speed gain for Q4_0 and Q4_1 on Metal
* Have N_DST, etc., be template parameters
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Custom RoPE + bettter memory management for CUDA
* Adjusted look ahead in ggml_cuda_pool_malloc to 5%
This is sufficient it seems.
We end up using about 200 MB less VRAM that way when running
the 13B model with context 8192.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Faster Q2_K on Metal
* Deleting unnoticed and dangereous trailing white space
* Fixed bug in new metal Q2_K implementation
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* 3-5% faster Q4_0 on Metal
* 7-25% faster Q4_1 on Metal
* Oops, forgot to delete the original Q4_1 kernel
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* 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>
* Only use Q6_K for output weights if tensor size is multiple of 256
* Fixed copy/paste mistake
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* k_quants: hopefully much faster Q4_K on older GPUs
On the GTX-1660 that I have available to represent
"old GPUs", token prediction drops from 65.5 ms/tok
to 41.5 ms/tok!
* k_quants: hopefully much faster Q3_K on older GPUs
On the GTX-1660 that I have available to represent
"old GPUs", token prediction drops from 60.3 ms/tok
to 41.0 ms/tok!
* k_quants: faster Q2_K on older GPUs
It looks like I didn't need to change anything
compared to what we already had, so this is just
adding clarifying comments. But I now measure
36.3 ms/tok on the GTX-1660, instead fo the
47.2 ms/tok that I have written in the faster
k-quants PR.
* k_quants: faster Q5_K on older GPUs
68.5 ms/tok -> 62.0 ms/tok on GTX-1660.
For some reason the same access pattern that leads
to such resounding success for Q2_K to Q4_K did not
work at all for Q5_K.
It is also more difficult to measure because for Q5_K_S
we only have 32 layers on the GTX-1660, so output, tok embeddings
and kv cache are done on the CPU.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fix examples/metal
* k-quants: prevent usage when tensor size is not divisible by 256
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* cuda : faster k-quant dot kernels
* Imrove Q2_K dot kernel on older GPUs
We now have a K_QUANTS_PER_ITERATION macro, which should be
set to 1 on older and to 2 on newer GPUs.
With this, we preserve the performance of the original
PR on RTX-4080, and are faster compared to master on
GTX-1660.
* Imrove Q6_K dot kernel on older GPUs
Using the same K_QUANTS_PER_ITERATION macro as last commit,
we preserve performance on RTX-4080 and speed up
Q6_K on a GTX-1660.
* Add LLAMA_CUDA_KQUANTS_ITER to CMakeLists.txt and Makefile
Allowed values are 1 or 2. 2 gives the best performance on
modern GPUs and is set as default. On older GPUs 1 may work
better.
* PR comments
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* metal : improve q4_K
28.3 -> 26.0 ms/token by avoiding a branch in the
calculation of the scales.
* metal : small improvement for Q4_K
* metal : still optimizing Q4_K
This commit pushes it down to 25.3 ms / token.
The crazy idea of using 6 bits for the scales is really costly on
Metal: if I remove the bit fiddling necessary to make the block
scales, time goes almost to the Q4_0 23 ms/token.
Before pushing the k-quants upstream I had a Q4_K variant that
had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight,
was running slightly slower on the CPU (due to the larger model size
and being memory bound there), and the difference was entirely
negligible under CUDA. So, I decided to publish the version with 6-bit
scales. Perhaps I should re-consider and change to 8-bit scales?
* metal : some more optimizations
Q2_K: 25.4 ms/token
Q6_K: 27.3 ms/token
Q4_0: 22.8 ms/token
Q4_1: 23.1 ms/token
* metal : Q3_K support
Something is not quite right yet.
* metal : Q5_K support
Initial version achieves 31.2 ms/token, 210 GB/s
* metal : still not able to figure out why q3_K does not work
* Minor
* metal : yet another failed attempt to make q3_K work
* metal : optimize Q5_K
31.2 ms -> 27.8 ms.
250 GB/s.
* metal : q3_K still not working
Adding a heavily commented q3_K metal kernel to explain
my obviously faulty logic. Perhaps someone could spot the issue?
* metal : q3_K finally working
Not optimized at all.
What was the issue? The scales are not 4-bytes aligned,
and I was accessing them with a uint32_t pointer.
When I tried that on CUDA, I got an error (illegal memory access)
and added a memcpy to a local array of 3 uint32_t's.
But on Metal it told me there is no memcpy, so I tried
accessing directly. There is no error, just garbage results.
At some point I did try accessing the scales with an uint16_t
pointer (the scales are for sure 2-byte aligned), but was
still getting garbage. I guess, there must have been another bug.
No access to scales is via a uint16_t pointer and, after starting
from scratch from the C dequantize function, it finally works.
* metal : Q3_K 1st optimization pass
* metal : Q3_K second optimization pass - 29.6 ms/token
* metal : Q3_K cleanup
* metal : fixed accidentally broken Q2_K
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
The number of buffers in the ggml context was left unitialized.
This leads to sporadic failures to load the model on
startup. It is actually strange that the failure occurred so
infrequantly.
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* metal : 8% faster q4_0
Avoid copying into local uchar4 anf float4.
* metal : 17% faster Q4_0
Use 64 threads in a thread group.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* metal : add Q2_K implementation
27.1 ms / token on M2 Max 30-core GPU, so about the
same speed as Q4_0. Memory throughput is ~156 GB/s.
The access pattern used in the Q2_K
CUDA implementation resulted in significantly lower
performance (~31 ms/token).
* Fixing merge conflicts
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Metal implementation for Q4_K
Very slow for now:
42 ms / token, Q4_0 runs in 28 ms/token on my
30-core M2 Max GPU.
* Optimizing Q4_K on metal
The first token always takes longer, I guess because
the metal kernel is being jit-compiled.
So, using n = 128 to measure time.
At this point Q4_K takes 29.5 ms / token
compared to 27.2 ms / token for Q4_0.
Quite a bit better than the initial attempt,
but still not good enough.
* Optimizing q4_K metal dot some more
For n = 256 it is now 28.1 ms/token compared to
27 ms/token for q4_0.
* Fix after merge with master
* Metal implementation for Q6_K
Similar to the CUDA implementation.
No idea if this is the optimum for Metal, but the few
alternative variants I tried all had a lower performance.
We get 36.5 ms / token on M2 Max with 30 GPU cores.
This corresponds to ~200 GB/second throughput.
* clang-tidy : add config back
* Much better Q6_K implementation for metal
28.3 ms / token for 7B. Subtracting ~9 ms that is spent in
other compute graph operations, we are left with ~19 ms
for the matrix multiplications. The model is ~5.5 GB,
so we are getting 1000 / 19 * 5.5 = 290 GB/s!
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Metal implementation for Q4_K
Very slow for now:
42 ms / token, Q4_0 runs in 28 ms/token on my
30-core M2 Max GPU.
* Optimizing Q4_K on metal
The first token always takes longer, I guess because
the metal kernel is being jit-compiled.
So, using n = 128 to measure time.
At this point Q4_K takes 29.5 ms / token
compared to 27.2 ms / token for Q4_0.
Quite a bit better than the initial attempt,
but still not good enough.
* Optimizing q4_K metal dot some more
For n = 256 it is now 28.1 ms/token compared to
27 ms/token for q4_0.
* Fix after merge with master
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* 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>
* A faster version for Q4_1 x Q8_0 dot products
The idea nehind being that Q8_0 quantized
values get used many times in the matrix multiplications
where they are involved. In the current implementations,
when we are evaluating the dot products, we need to compute
the sum of the quants in the Q8_0 vector, so the same
operation is repeated many times. Here we pre-compute
the sum during Q8_0 quantization, store it in the
now modified block_q8_0 struct, and then reuse this
result in the subsequent dot products.
In a synthetic benchmark (just compute a bunch of dot
products), this change speeds up the Q4_1 * Q8_0 dot
product by 80%, making the performance identical to
Q4_0 * Q8_0.
In practical application, I see a ~15% gain in speed for
token prediction on M2, and ~5% gain on Ryzen 7950X.
The speed gain in the prompt evaluation is much bigger
(around 50%).
I have only done the change for the scalar version,
ARM_NEON, and AVX2, so we still need an AVX implementation.
* Cleaning up
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Multi-threading quantization.
Not much gain for simple quantizations, bit it will be important
for quantizations that require more CPU cycles.
* Multi-threading for quantize-stats
It now does the job in ~14 seconds on my Mac for
Q4_0, Q4_1 and Q4_2. Single-threaded it was taking
more than 2 minutes after adding the more elaborate
version of Q4_2.
* Reviewer comments
* Avoiding compiler confusion
After changing chunk_size to const int as suggested by
@ggerganov, clang and GCC starting to warn me that I don't
need to capture it in the lambda. So, I removed it from the
capture list. But that makes the MSVC build fail. So,
making it a constexpr to make every compiler happy.
* Still fighting with lambda captures in MSVC
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Q4_2 quantization with rmse-optimized scale and quants
For quantize-stats we get
q4_2: rmse 0.00159301, maxerr 0.17480469, 95pct<0.0030, median<0.0012
For 7B perplexity with BLAS enabled we get 6.2038 after 655 chunks.
Quantization is slow (~90 seconds on my Mac for 7B) as not
multi-threaded as in PR #896.
* ggml : satisfy the sanitizer builds
Not sure why this makes them fail
* Better follow ggml conventions for function names
* Fixed type as per reviewer comment
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
On my Mac, the direct Q4_1 product is marginally slower
(~69 vs ~55 us for Q4_0). The SIMD-ified ggml version
is now almost 2X slower (~121 us).
On a Ryzen 7950X CPU, the direct product for Q4_1 quantization
is faster than the AVX2 implementation (~60 vs ~62 us).
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
