* ggml_graph_compute: deprecate using ggml_context, try resolve issue #287
* rewrite: no longer consider backward compitability; plan and make_plan
* minor: rename ctx as plan; const
* remove ggml_graph_compute from tests/test-grad0.c, but current change breaks backward
* add static ggml_graph_compute_sugar()
* minor: update comments
* reusable buffers
* ggml : more consistent naming + metal fixes
* ggml : fix docs
* tests : disable grad / opt + minor naming changes
* ggml : add ggml_graph_compute_with_ctx()
- backwards compatible API
- deduplicates a lot of copy-paste
* ci : enable test-grad0
* examples : factor out plan allocation into a helper function
* llama : factor out plan stuff into a helper function
* ci : fix env
* llama : fix duplicate symbols + refactor example benchmark
* ggml : remove obsolete assert + refactor n_tasks section
* ggml : fix indentation in switch
* llama : avoid unnecessary bool
* ggml : remove comments from source file and match order in header
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* use javascript generators as much cleaner API
Also add ways to access completion as promise and EventSource
* export llama_timings as struct and expose them in server
* update readme, update baked includes
* llama : uniform variable names + struct init
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Generalize quantize_fns for simpler FP16 handling
* Remove call to ggml_cuda_mul_mat_get_wsize
* ci : disable FMA for mac os actions
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* convert checks in llama_load_session_file to throw and handle them
* make llama_load_session_file_internal static
* address feedbacks to avoid using exceptions
* Remove multiple shards
* Remove multiple file loaders
* Remove llama_load_tensor_shard class
* Simplify load logic
* Remove dead code guess_n_parts function
* Remove vocab_only from constructor of llama_model_loader
* Remove alignment_prevents_mmap which is not more needed.
* Remove useless check
* add interface for float input
* fixed inpL shape and type
* add examples of input floats
* add test example for embd input
* fixed sampling
* add free for context
* fixed add end condition for generating
* add examples for llava.py
* add READMD for llava.py
* add READMD for llava.py
* add example of PandaGPT
* refactor the interface and fixed the styles
* add cmake build for embd-input
* add cmake build for embd-input
* Add MiniGPT-4 example
* change the order of the args of llama_eval_internal
* fix ci error
* detect NUMA systems and pin work threads to nodes (linux)
* disable mmap prefetch/readahead for NUMA systems
* avoid sending finalize op to thread pool if it does nothing
* silence robot
* fix args
* make --numa a param
* recommendation that n_nodes evenly divide n_threads did not warrant such aggressive enforcement
* lower synchronization overhead
* statically allocate
* move numa state to g_state
* add description for --numa
* ggml : minor style changes
* ggml : minor style + try fix sanitizer build
* llama : allow to initialize backend with NUMA support
* llama : avoid ggml include in llama-util.h
* ggml : style / formatting
* ggml : fix handling of ops with n_threads > n_tasks > 1
* server : utilize numa parameter
---------
Co-authored-by: Georgi Gerganov <ggerganov@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>
* Fix top-p sampling to match the standard definition (smallest set that has probability mass at least p, not largest set with probability mass less than p)
* top-p: correct gt to gte
* add test for correct top-p behavior
* llama : make model stateless and context stateful
* llama : minor cleanup
* llama : update internal API declaration
* Apply suggestions from code review
fix style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Missing model memory release
* Fix style
* Add deprecated warning for public API function llama_init_from_file
* Update public API use cases: move away from deprecated llama_init_from_file
* Deprecate public API function llama_apply_lora_from_file
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* fixed issue: memory is not guaranteed to be aligned properly during ggml_init call from loading saved sessions
* - removed commented out old code from fix
- updated another instance of same issue below original
* 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>
* Convert vector to f16 for dmmv
* compile option
* Added compilation option description to README
* Changed cmake CUDA_ARCHITECTURES from "OFF" to "native"
* Fix examples/metal
* k-quants: prevent usage when tensor size is not divisible by 256
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* metal : handle buffers larger than device's maxBufferLength
* metal : print more verbose device info + handle errors
* metal : fix prints for overlapping views
* metal : minimize view overlap to try to utilize device memory better
* add python wrapper
https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce
* fix decoding error. adds errors=ignore parameter
* add python bindings for functions to get and set the whole llama state
(rng, logits, embedding and kv_cache)
* update python bindings
* add text generating baby-llama from scratch example
* fix race condition bug in ggml_compute_forward_diag_mask_f32
* implement ggml_soft_max_back for more performant backward pass of soft_max
avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss
* improve softmax backward pass
go from quadratic runtime to linear runtime by simplifying the formulas
* fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32
memcpy needs to be synchronized across threads to avoid race conditions.
=> do it in INIT phase
* fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build
* improve performance of mul_mat backward pass
avoid transpose by using mul_mat with swapped arguments
* avoid printing too much newlines in baby-llama-text
* activate threading in baby-llama-text
* add ggml_out_prod and use it for mul_mat backward pass for improved performance
performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests
* better weight initialization improves training convergence at start
* better weight initialization improves training convergence at start
* improve ggml_out_prod performance
- change iteration order (>15s -> 10s runtime)
- parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime)
* add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data
* fix get_samples call, add model tensor names, increase model size, start training samples after newline
* save train trained model to checkpoint and load model to be trained from checkpoint
* use inplace functions where possible
* initialize rng with srand
* use different arguments for input and output checkpoint
* ggml fixes to support backward pass on inplace operations
* remove duplicate include
* fix cross entropy loss
- add target probabilities for each sample which is then used in cross entropy loss
* print used memory before and after optimization
* sample with non-greedy sampling parameters at the end of training
* add cmake target for baby-llama-text
* add ggml_add1_inplace to header
* enable gradient propagation for inplace add1 and scale operations
those functions backward passes don't need the original src0, so they also work when forward is inplace
* implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f)
also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule.
setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer.
since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer.
* use inplace operations in cross_entropy_loss
* fix random weight initialization scale
* add missing default parameters for adam optimizer
* add ggml_opt_context, so that we can properly resume training
otherwise the optimizer states, tracking statistics about the error function and its derivates,
will reset to zero each time ggml_opt is called, hindering convergence on resumed training.
now the optimizer context and all its memory is stored in a separate struct.
* fix bug in llama_sample_token_mirostat_v2
when all candidates are filtered out through mu threshold, the following soft_max operation will fail.
so keep at least one.
* add forward function without using cache, for more performant training
during training on whole samples no cache is required.
removing the cache and simplifying the remaining code results in performance and memory usage improvement.
* print suppressed newline tokens as string "\n"
printing too much actual newlines is suppressed to avoid flooding the console.
* store optimizer state in training checkpoint and add learning schedule
persistent optimizer state allows to resume training without resetting the optimizer
learning schedule consists of linear warmup ramp followed by cosine decay with restarts
* remove unused functions
* fix bug in get_samples which corrupted training targets
* save checkpoint only when it was trained
* simplify code
* remove trailing whitespace
* simplify backward pass for SQRT
* replace inefficient repeat backward pass with dedicated repeat_back operation
* add ggml_cross_entropy_loss with backward pass for faster training
cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead.
* add tests for cross_entropy_loss backward pass
finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient.
_probably_ the finite differences fails due to numerical issues
* use ggml_cross_entropy_loss in text training example
* remove trailing whitespace
* slightly improve how cross entropy loss is compute
btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log.
probably the input to log gets closer to zero due to float numerics.
maybe the multiplication by (1.0-eps)/sum is more accurate..
* add llama_get_vocab to get the vocabulary as output parameters
* set default model.type for unknown models with few layers
* add export of training checkpoint to llama compatible model file
* get vocabulary for exporting training checkpoint to llama compatible model file
* implement backward pass of flash attention
* bugfixes for backward pass of flash attention
* test flash attention backward pass
need to set loose error bounds to pass.
the finitie differences are close to numeric limits and often return quite different values than the backward pass.
reducing eps further lets the gradients vanish completely.
likewise setting eps to big results in wronger values.
the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences.
* add option to train with flash attention and move options to the top of the main function
training from scratch also works with flash attention
training convergence and generation results after fix number of iterations are worse than when not using flash attention.
maybe there still lingers a bug in the flash attention backward pass?
but training works, just with slower convergence.
flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx
* add train_params and command line option parser
* remove unnecessary comments
* add train params to specify memory size
* remove python bindings
* rename baby-llama-text to train-text-from-scratch
* replace auto parameters in lambda function
* add #include <climits>
* add explicit cast to fix compile error
"error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]"
* remove trailing whitespace
* add ggml_opt_resume_g which accepts forward and backward cgraphs
* fix formulas in comments
* bug fix for ggml_compute_forward_get_rows_back_f32
the result should be set to zero, not to whatever data is in opt0
* improve training memory usage with scratch buffers
instead of relying on the automatic backward pass, we manually create the graph for the backward pass.
it turns out that all backward pass operations need only temporary memory which can be reused after each layer.
will compute backward pass for ALL model parameters
* add option to use scratch buffers in training or not
make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters.
* ci : disable temporary
* store view offset and permute axes in opt[0] instead of storing it in padding
use memcpy to store offset, because offset is of type size_t.
when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true.
* minor : fix compile warnings + minor style changes
* fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32
* store view offset like in master branch
* bug fix in forward_batch_wo_cache_flash_attn_train
* scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train
data of permute and reshape is the same as their input.
if we want to preserve the output of permute/reshape, we also need to preserve their inputs.
replace reshape(src0, src1) with reshape_nd calls so that we don't need src1.
replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02).
in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls.
for this we need backward pass of broadcasting ggml_mul.
* remove unnecessary scratch buffer 0
buf 0 is persistent memory, so we can just disable scratch for this by using buf -1
* avoid creating unnecessary grad tensors
previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads
this wasted memory, because unnecessary grad for each op were automatically created:
the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ).
this discarded the automatically generated grad resulting in wasted memory.
improved this by changing expand(..) to not use ggml_build_forward_expand.
expand set cgraph->nodes but not the leafs.
cgraph->leafs & cgraph->grads are set in another pass after the last expand call.
* print used training seed
* zero initialize gfbuf and gbbuf
* ci : re-enable workflows + add README for training
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Allow "quantizing" to f16 and f32
Fix an issue where quantizing didn't respect LLAMA_NO_K_QUANTS
Add brief help to the list of quantization types in the quantize tool
Ignore case for quantization type arguments in the quantize tool
* 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>
* Rebase to latest
* Show progress
* Add assert to make sure we only allocate temp buffer for non-CPU backend tensor
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Add support for quantizing already quantized models
* Threaded dequantizing and f16 to f32 conversion
* Clean up thread blocks with spares calculation a bit
* Use std::runtime_error exceptions.
* Use MTLDevice.newBufferWithBytesNoCopy to share buffers between CPU and GPU
* Page-align buffers used by Metal
* Remove trailing whitespace
* Only import unistd.h for Metal builds
* metal : remove unnecessary copies
---------
Co-authored-by: Georgi Gerganov <ggerganov@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>
