* WIP Gemma3: not working
* gemma3: build_gemma3 seems to be working now
* Revert changes to convert_hf_to_gguf.py
It wasn't working, so I guess, it is better to leave the
conversion up tp upstream.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Repack a model with the quantize tool
* WIP
* Fixed various issues
As we don't have a way to tell if a repacked quant has been modified,
I had to remove the modification at the expense of a slight decrease
in performance. This affects q8_0_r8, q8_KV_r8, q8_k_r8 on Zen4, and
q4_0_r8 on ARM.
* Create wk_b and wv_b as Q8_0_R8 if the wkv_b type is interleaved
* Fix GCC 13.3 compilation error
* Another one
* Add missing include
* FlashMLA-3: the best of both worlds - CPU only
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Repack a model with the quantize tool
* WIP
* Fixed various issues
As we don't have a way to tell if a repacked quant has been modified,
I had to remove the modification at the expense of a slight decrease
in performance. This affects q8_0_r8, q8_KV_r8, q8_k_r8 on Zen4, and
q4_0_r8 on ARM.
* Create wk_b and wv_b as Q8_0_R8 if the wkv_b type is interleaved
* Fix GCC 13.3 compilation error
* Another one
* Add missing include
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
I broke it with PR #265. I was testing with a model where
the wk_b and wk_v tensors were present, so didn't need to be computed,
so didn't notice that the change I made to ggml_compute_forward_dup_q
breaks that computation.
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA-2: eliminate intermediate f32 tensors
This works on the CPU. PP performance is ~13% better for 16k tokens
and compute buffer is quite a bit smaller.
* FlashMLA-2: enable fast path only on the CPU for now
I did implement the necessary ops on CUDA, but something is
still wrong there, so for now we only use it when running
CPU-only.
* FlashMLA-2: slightly smaller computer buffer size
* Prepare wk_b when loading DeepSeek models (if wk_b is missing)
* Add some comments
* Fix case where wkv_b is quantized with k- or i-quants.
* Fix CUDA
There is an issue with quantized GEMV on CUDA when the left operand
(the matrix) is not contiguous. So, for now, we also create wv_b
during model loading and use that instead of the 3D view of wkv_b.
* FlashMLA-2: avoid conversions to f32 also on CUDA
* Be able to compute for more than 65535 tokens
On CUDA just a quick hack that allows us to cancatenate tensors
with more than 65535 rows along zroth dimension as needed by
FlashMLA-2. Also needed some care in the perplexity tool to
avoid int overflows when evaluating the computed logits.
* Reduce memory usage for FlashMLA-2
Oh, also fix int overflow in the CUDA concat implementation.
It is funny how the llama.cpp 64-bit police has gone (almost) everywhere
and replaced 32-bit ints with 64-bit ints, needed or not,
but hasn't done it where it is actually needed.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA-2: eliminate intermediate f32 tensors
This works on the CPU. PP performance is ~13% better for 16k tokens
and compute buffer is quite a bit smaller.
* FlashMLA-2: enable fast path only on the CPU for now
I did implement the necessary ops on CUDA, but something is
still wrong there, so for now we only use it when running
CPU-only.
* FlashMLA-2: slightly smaller computer buffer size
* Prepare wk_b when loading DeepSeek models (if wk_b is missing)
* Add some comments
* Fix case where wkv_b is quantized with k- or i-quants.
* Fix CUDA
There is an issue with quantized GEMV on CUDA when the left operand
(the matrix) is not contiguous. So, for now, we also create wv_b
during model loading and use that instead of the 3D view of wkv_b.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA-2: eliminate intermediate f32 tensors
This works on the CPU. PP performance is ~13% better for 16k tokens
and compute buffer is quite a bit smaller.
* FlashMLA-2: enable fast path only on the CPU for now
I did implement the necessary ops on CUDA, but something is
still wrong there, so for now we only use it when running
CPU-only.
* FlashMLA-2: slightly smaller computer buffer size
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA(CUDA): WIP to allow q8_0 quantized cache
* WIP
* FlashMLA(CUDA) - allow q8_0 for KV cache
This works, and PP is not bad, but TG is still quite a bit slower.
* FlashMLA(CUDA) - allow q8_0 for KV cache
This is better. ~9% slower than f16 cache for short contexts,
nearly on par at 16k tokens.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This gives us ~20% TG speedup for DeepSeek on CUDA
* Slightly better
* Also do it for plain (not fused) mul_mat_id
* Guard against numerical precision issues for MLA on CUDA
* imatrix: wv_b <-> wkv_b
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This gives us ~20% TG speedup for DeepSeek on CUDA
* Slightly better
* Also do it for plain (not fused) mul_mat_id
* Guard against numerical precision issues for MLA on CUDA
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
PP speed is great, almost on par with standard FA.
But TG speed is pathetic. The strangest thing is that
the slowdown is not due to FA, but due to the ffn_gate_exps
gemm, which somehow becomes very slow. WTF?
As I'm unable the resolve the slow ffn_gate_exps GEMM mystery,
for now TG goes via mla=2, PP is via FA.
Also discovered the ggml_cast op, so we don't need the aux
tensors that I had added to the KV cache.
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA-2: faster prompt processing
The current MLA implementation computes
wv_b * (k_cache * softmax(k_cache * (wk_b*q)))
This leads to 3.4X more multiply-adds (madds)
compared to standard attention. Due to the resulting
tensor shapes, TG is still faster than standard attention
because the k_cache*(wk_b*q) and k_cache*(softmax(k_cache * (wk_b*q)))
multiplications become GEMMs, so the additional madds are
more than compensated for due to the much higher performance
of GEMMs compared to GEMVs. But for PP, where we are dealing
with GEMMs in both cases, the additional madds needed for MLA
lead to lower performance, with the performance gap increasing
with context length.
So, then, when we are dealing with PP, we can rearrange the
above to (wv_b * k_cache) * softmax( (wk_b^T*k_cache) * q),
thus transforming it into the standard attention mechanism.
We do need two additional matrix multiplications (which in practice
is done as a single wkv_b * k_cache GEMM) with the *entire*
K cache. But this is still cheaper than MLA, as we end up with
1.8X the madds required by standard attention. Oh, these figures
are for the DeepSeek-V3/R1/Lite attention architecture.
This leads to a significant PP performance increase compared
to standard MLA with FA.
There are many upsides to this:
* If we only apply the above trick when we are processing more than
X tokens (with suitable chosen X), TG performance stays the same
as MLA with FA
* We still need to store just the K-cache, so 576 entries per layer
for DeepSeek-V3/R1/Lite
* We get significantly better PP performance
* We can use MLA+FA on CUDA. It works already with this commit
for PP, something is not yet quite right for TG.
The downside is that it only works with fp16 cache (for now).
This is so because we need to convert the cache to fp32,
else we cannot do the wkv_b * k_cache matrix multiplication
(which in ggml requires the second operand to be fp32).
But converting (copying) to fp32 only works for f16, bf16 and
f32 tensors, so no luck with quantized cache. Another reason
that we need to convert to fp32 is that the cache contains the
RoPE'd portion, which we need to concatenate to the result of
the wkv_b * k_cache matrix multiplication. Also this op
works only when the tensors being concatenated are both fp32.
So much about ggml being a general purpose ML library.
* FlashMLA-2: on the CPU it now works for quantized cache
except for q8_KV (q8_KV has row meta data, and there is still
some confusion with row sizes because of that).
* FlashMLA-2: on the CPU it now works also with q8_KV
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This is a better FA for TG
It should benefit MLA and GQA. Tested to work with
DeepSeek-Lite MLA, not yet for GQA.
For tg64@pp8192 it is ~13% faster than MLA without FA,
and 57% faster that the main branch FA.
* WIP
* Cleanup
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Custom quantization rules with regular expressions
* Add the --custom-q option to the help
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* WIP CUDA FA with Dk != Dv
* WIP
* CUDA FA WIP - It actually works!
No TG yet, but for PP I can run FA with fp16 cache and it gets
the same answer.
* CUDA FA WIP - it now works for Q8_0 + Q8_0 for KV cache
* CUDA FA WIP - TG, not working yet.
* CUDA FA with Dk != Dv: it works now for DeepSeek
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* FlashMLA - it finally works (on the CPU)
* FlashMLA: allow for f16 and bf16 cache in addition to q8_0
* It works with ggml FA, not with iqk FA
* WIP
* FlashMLA: it now works with iqk
I had forgotten to divide the Q stride by sizeof(float) and
that's why, very cobfusingly, it was working for TG but not for PP.
* WIP
* FlashMLA: that should be it for now
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* A better way to measure the cost of ggml_barrier
* Smart expert selection
* Add ser option to llama-bench
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This reduces compute buffer size for MLA
* This should accomplish it for standard attention
* Much better
* Better concat for contiguous tensors
If all the op does is to concatenate the second tensor
to the first, why would we want to have a loop?
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
The `-mla` command line option turns into an int from a bool.
mla = 0: use standard attention
mla = 1: use MLA with transposed cache
mla > 1: use MLA without transposed cache
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Slight MLA TG performance improvement on CUDA
The low MLA performance on CUDA is dues to
the wk_b * q_nope operation.
It turns into n_head matrix multiplications with
n_head separate quantization and GEMV steps.
The associated overhead is just too much for TG
where each GEMV is very fast (512 x 128 = 131 KFLOP
for DeepSeek-Lite, 4X that for DeepSeekV3/R1).
The way it was done there was also a copy of each q_nope
row before quantization, which I have now eliminated.
This results in a ~2.5% speedup.
What needs to happen instead is to launch a single
computation that quantizes all heads, and then have
a kernel that does the GEMV for all heads instead of
n_head sequential GEMVs.
* Slightly better
* CUDA: Quantize non-contiguous tensors
* Much better MLA
It is a total hack, but it works.
* Cleanup
Remove duplicated gemv's.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Give the user the option to override where model weights are stored
* Fix ggml_nbytes() problem and cleanup
For a tensor with zero elements ggml_nbytes() was returning
uint64_t::max, and this was causing graph allocation failure.
* Add timing info to CUDA graph evaluation
* Add more timing info
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fusing MoE up * unary(gate)
* Fusing MoE up * unary(gate): CUDA
We get ~13% speedup for PP-512 and ~2% for TG-128
for DeepSeek-Lite
* On CUDA also fuse MoE down * (up * unary(gate))
in case the MUL_MAT_ID op for the down experts is the next
op in the graph.
* Command line option to enable fused MoE up*unary(gate)
* Add fmoe option to llama-bench
* Adding forgotten gelu, relu, silu on ARM
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* examples : add new sweep-bench benchmark
* Change documentation to reference ik_llama.cpp
* Made it compile with ik_llama
* Fix JSONL output
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* This seems to be a better way
to do the attention matrix multiplications in the TG case.
* Cleanup
* Fuse up and gate gemms in MoE models
Small (~1-2%) but measurable performan ce gain
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This seems to be a better way
to do the attention matrix multiplications in the TG case.
* Cleanup
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Adding q8_KV - Basics + AVX2 gemm/gemv
* q8_KV: Better AVX2 gemm
* q8_KV: Better Zen4 gemm
We get 225.7 t/s for L3-8B. In comparison q8_0 without
run-tinme-repacking is at 169 t/s.
* q8_KV: AVX2 gemm/gemv
We get 254 t/s for L3-8B vs 194 t/s for q8_0 without rtr.
* q8_KV: be able to use it for K cache
This required quite a few fixes in ggml and llama.cpp:
* ggml: do not calculate row size as n/block_size*type_size. I had
removed most of it when implementing the quants with per row scale,
bit it was stull lurking in ggml_copy. Not sure if these were the last
remnants of ggmil-style row sizes, or if there are still places left
* llama.cpp: get rid of the the 1d K cache assumption. Create and manage
the K-cache as a 2D tensor so we can have per row meta data as needed
by q8_KV.
Using q8_KV for K-cache results in non-negligible performance gains.
More details to follow, but for DeepSeek-Lite with MLA, we get
18% speedup for PP-8192 compared to q8_0 K-cache.
* q8_KV: be able to use it for K cache in FA
* q8_KV: repack it for K*Q in FA
* q8_KV: slightly faster gemv on Zen4
* q8_KV: slightly faster gemv on Zen4
* q8_KV: ARM_NEON
We get PP-512 = 167 t/s for L3-8B without interleaving!
We do the interleaving on the fly, so I wonder if this
could be done for other quants as well.
* q8_KV: use it in FA on NEON
* q8_KV_r8 - repacked q8_KV
On Zen4 it is slower than q8_k_r8 (292 vs 370 t/s)
This makes no sense whatsoever as the q8_KV_r8 GEMM is
basically the q8_k_r8 GEMM with the unnecessary block stuff
removed (so, one would think that it would be faster).
* q8_KV_r8: don't use nrc_y = 16 on Zen4
This is faster - 350 t/s. Why?
Much better than the 290 t/s we had before, but still slower
than the 370 t/s for q8_k_r8.
* q8_KV: nrc_y = 16 also doesn't pay off in FA
* Minor
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
I added a change in the last PR how activations are quantized.
It looked like it is working and slightly improving performance.
But I now hit an edge case where I get gibberish that goes away if
I remove the change. I absolutely don't see what goes wrong, so
leaving the change in commented out for now.
This allows us to optimize TG performance for GQA models.
E.g., for IQ4_XS L3-8B with 8k TG-64 goes from 8.6 to 10.26 t/s.
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Do not allocate / report caches that are not used
It is either the standard KV cache or MLA cache, not both.
* Rename X_pe to X_rope
Much easier to follow, at least for my brain, when we have
X_rope : rotational position encoding
X_nope : no position encoding
instead of X_pe and X_nope, where I was wondering wtf is 'pe'
and 'nope'.
* WIP
* WIP
* WIP
* WIP
* Warn user when disabling MLA
* MLA: compile time option to not use transposed KV cache
Cuts KV cache size in nearly half at the expense of slower
TG performance for long contexts (it becomes similar to
no-MLA).
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
