* 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>
* 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>
* 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>
* 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.
* 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>
* mtl : export the LLaMA computation graph
* ci : disable temporary
* mtl : adapt the MNIST example as starter
* mtl : no need for mtl-export tool, add cli arg for main instead
* mtl : export just a small part of the graph for now to make it easier
* mtl : move MSL code into separate file for easy editing
* mtl : initial get_rows_q4_0 kernel
* mtl : confirmed get_rows_q4_0 is working correctly
* mtl : add rms_norm kernel + confirm working
* mtl : add mul kernel + confirm working
* mtl : initial mul_mat Q4 kernel (wrong results)
* mtl : mul_mat fixes (still wrong)
* mtl : another mul_mat Q4 (still does not work)
* mtl : working mul_mat q4
* ggml : fix handling of "view" ops in ggml_graph_import()
* mtl : add rope kernel
* mtl : add reshape and transpose handling
* ggml : store offset as opt arg for ggml_view_xd() operators
* mtl : add cpy kernel + handle view ops
* mtl : confirm f16 x f32 attention mul mat
* mtl : add scale kernel
* mtl : add diag_mask_inf kernel
* mtl : fix soft_max kernel
* ggml : update ggml_nbytes() to handle non-contiguous tensors
* mtl : verify V tensor contents
* mtl : add f32 -> f32 cpy kernel
* mtl : add silu kernel
* mtl : add non-broadcast mul kernel
* mtl : full GPU inference of the computation graph
* mtl : optimize rms_norm and soft_max kernels
* mtl : add f16 mat x f32 vec multiplication kernel
* mtl : fix bug in f16 x f32 mul mat + speed-up computation
* mtl : faster mul_mat_q4_0_f32 kernel
* mtl : fix kernel signature + roll inner loop
* mtl : more threads for rms_norm + better timing
* mtl : remove printfs from inner loop
* mtl : simplify implementation
* mtl : add save/load vocab to ggml file
* mtl : plug Metal inference into llama.cpp (very quick-n-dirty)
* mtl : make it work with main example
Lots of hacks but at least now it generates text
* mtl : preparing for merge
* mtl : clean-up ggml mtl interface + suport scratch / inplace
* mtl : remove temp / debug code
* metal : final refactoring and simplification
* Revert "ci : disable temporary"
This reverts commit 98c267fc77.
* metal : add comments
* metal : clean-up stuff, fix typos
* readme : add Metal instructions
* readme : add example for main
1. Add a `LLAMA_SUPPORTS_GPU_OFFLOAD` define to `llama.h` (defined when compiled with CLBlast or cuBLAS)
2. Update the argument handling in the common example code to only show the `-ngl`, `--n-gpu-layers` option when GPU offload is possible.
3. Add an entry for the `-ngl`, `--n-gpu-layers` option to the `main` and `server` examples documentation
4. Update `main` and `server` examples documentation to use the new style dash separator argument format
5. Update the `server` example to use dash separators for its arguments and adds `-ngl` to `--help` (only shown when compiled with appropriate support). It will still support `--memory_f32` and `--ctx_size` for compatibility.
6. Add a warning discouraging use of `--memory-f32` for the `main` and `server` examples `--help` text as well as documentation. Rationale: https://github.com/ggerganov/llama.cpp/discussions/1593#discussioncomment-6004356
The underlying representation of multibyte character literals is
implementation-defined. This could, at least in principle, cause
cross-build data export/import issues independent of endianness.
Define magic numbers as integer literals to be on the safe side.
Signed-off-by: Juuso Alasuutari <juuso.alasuutari@gmail.com>
* Sample interface, new samplers.
New samplers:
- locally typical sampling
- tail free sampling
- frequency and presence penalty
- mirostat
Ignore EOS fix: -inf should be used.
* mirostat
* Added --logit-bias and --no-penalize-nl, removed std::span
* Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k)
* Save and load example adjust
* Tests
* Windows build fix
* Windows test fix
The llama_set_state_data function restores the rng state to what it
was at the time llama_copy_state_data was called. But users may want
to restore the state and proceed with a different seed.
* reserve correct size for logits
* add functions to get and set the whole llama state:
including rng, logits, embedding and kv_cache
* remove unused variables
* remove trailing whitespace
* fix comment
* 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>
- Support all three formats (ggml, ggmf, ggjt). (However, I didn't
include the hack needed to support GPT4All files without conversion.
Those can still be used after converting them with convert.py from my
other PR.)
- Support both mmap and read (mmap is used by default, but can be
disabled with `--no-mmap`, and is automatically disabled for pre-ggjt
files or on platforms where mmap is not supported).
- Support multi-file models like before, but automatically determine the
number of parts rather than requiring `--n_parts`.
- Improve validation and error checking.
- Stop using the per-file type field (f16) entirely in favor of just
relying on the per-tensor type/size fields. This has no immediate
benefit, but makes it easier to experiment with different formats, and
should make it easier to support the new GPTQ-for-LLaMa models in the
future (I have some work in progress on that front).
- Support VirtualLock on Windows (using the same `--mlock` option as on
Unix).
- Indicate loading progress when using mmap + mlock. (Which led me
to the interesting observation that on my Linux machine, with a
warm file cache, mlock actually takes some time, whereas mmap
without mlock starts almost instantly...)
- To help implement this, move mlock support from ggml to the
loading code.
- madvise/PrefetchVirtualMemory support (based on #740)
- Switch from ifstream to the `fopen` family of functions to avoid
unnecessary copying and, when mmap is enabled, allow reusing the same
file descriptor for both metadata reads and mmap (whereas the existing
implementation opens the file a second time to mmap).
- Quantization now produces a single-file output even with multi-file
inputs (not really a feature as much as 'it was easier this way').
Implementation notes:
I tried to factor the code into more discrete pieces than before.
Regarding code style: I tried to follow the code style, but I'm naughty
and used a few advanced C++ features repeatedly:
- Destructors to make it easier to ensure everything gets cleaned up.
- Exceptions. I don't even usually use exceptions when writing C++, and
I can remove them if desired... but here they make the loading code
much more succinct while still properly handling a variety of errors,
ranging from API calls failing to integer overflow and allocation
failure. The exceptions are converted to error codes at the
API boundary.)
Co-authored-by: Pavol Rusnak <pavol@rusnak.io> (for the bit I copied from #740)
Command that calculates some statistics over the errors introduced by
quantization, like mean square error, max error and some percentile errors for layer
weights. Should be useful for testing quantization improvements.
Exposes some internal state from ggml and llama for testing
The api provides access methods for retrieving the current memory buffer for the kv_cache and its token number.
It also contains a method for setting the kv_cache from a memory buffer.
This makes it possible to load/save history - maybe support --cache-prompt paramater as well?
Co-authored-by: Pavol Rusnak <pavol@rusnak.io>
This is a breaking change that's going to give you three benefits:
1. Your inference commands should load 100x faster
2. You may be able to safely load models 2x larger
3. You can run many concurrent inference processes
This was accomplished by changing the file format so we can mmap()
weights directly into memory without having to read() or copy them
thereby ensuring the kernel can make its file cache pages directly
accessible to our inference processes; and secondly, that the file
cache pages are much less likely to get evicted (which would force
loads to hit disk) because they're no longer competing with memory
pages that were needlessly created by gigabytes of standard i/o.
The new file format supports single-file models like LLaMA 7b, and
it also supports multi-file models like LLaMA 13B. Our Python tool
now merges the foo.1, foo.2, etc. files back into a single file so
that the C++ code which maps it doesn't need to reshape data every
time. That's made llama.cpp so much simpler. Much of its load code
has now been deleted.
Furthermore, this change ensures that tensors are aligned properly
on a 32-byte boundary. That opens the door to seeing if we can get
additional performance gains on some microprocessors, by using ops
that require memory alignment.
Lastly note that both POSIX and the Windows platform are supported
Fixes#91
* Be more strict about converting float to double
* Test equivalence of round, SILU implementations
Test module is commented out in CMakeLists.txt because the tests may
take a long time, depending on how much the compiler optimizes.
* Fix softmax in perplexity.cpp
* all : prefer float over double where appropriate
* perplexity : add <cmath>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
`llama_sample_top_p_top_k` was missing the struct annotation on line 126.
This causes a compiler issue when being parsed by the Kotlin C interop generator.
This commit fixes the above issue by adding the struct annotation.
* Support calling mlock() on loaded model data on Linux and macOS
This is enabled by a new --mlock command line option.
Using mlock() disables swapping and memory compression for the model
data. Doing so can be useful on systems where the model takes up a
large fraction of system RAM. In my experience, macOS is quite eager to
start compressing llama.cpp's memory, which then makes it halt for a few
seconds while it decompresses, even with a model that uses "only" 25GB
out of 32GB.
Of course, this comes at the cost of forcing the system to swap or
compress other processes' memory instead, so it needs to be used with
care and shouldn't be enabled by default.
In theory it should be possible to support this on Windows as well using
VirtualLock(), but I'm not much of a Windows user.
* Update llama.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* working but ugly
* add arg flag, not working on embedding mode
* typo
* Working! Thanks to @nullhook
* make params argument instead of hardcoded boolean. remove useless time check
* start doing the instructions but not finished. This probably doesnt compile
* Embeddings extraction support
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Major refactoring - introduce C-style API
* Clean up
* Add <cassert>
* Add <iterator>
* Add <algorithm> ....
* Fix timing reporting and accumulation
* Measure eval time only for single-token calls
* Change llama_tokenize return meaning
