* 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
A major rewrite for the server example.
Note that if you have built something on the previous server API, it will probably be incompatible.
Check out the examples for how a typical chat app could work.
This took a lot of effort, there are 24 PR's closed in the submitter's repo alone, over 160 commits and a lot of comments and testing.
Summary of the changes:
- adds missing generation parameters: tfs_z, typical_p, repeat_last_n, repeat_penalty, presence_penalty, frequency_penalty, mirostat, penalize_nl, seed, ignore_eos
- applies missing top k sampler
- removes interactive mode/terminal-like behavior, removes exclude parameter
- moves threads and batch size to server command-line parameters
- adds LoRA loading and matches command line parameters with main example
- fixes stopping on EOS token and with the specified token amount with n_predict
- adds server timeouts, host, and port settings
- adds expanded generation complete response; adds generation settings, stop reason, prompt truncated, model used, and final text
- sets defaults for unspecified parameters between requests
- removes /next-token endpoint and as_loop parameter, adds stream parameter and server-sent events for streaming
- adds CORS headers to responses
- adds request logging, exception printing and optional verbose logging
- adds better stopping words handling when matching multiple tokens and while streaming, or when it finishes on a partial stop string
- adds printing an error when it can't bind to the host/port specified
- fixes multi-byte character handling and replaces invalid UTF-8 characters on responses
- prints timing and build info on startup
- adds logit bias to request parameters
- removes embedding mode
- updates documentation; adds streaming Node.js and Bash examples
- fixes code formatting
- sets server threads to 1 since the current global state doesn't work well with simultaneous requests
- adds truncation of the input prompt and better context reset
- removes token limit from the input prompt
- significantly simplified the logic and removed a lot of variables
---------
Co-authored-by: anon998 <131767832+anon998@users.noreply.github.com>
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Felix Hellmann <privat@cirk2.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Lesaun Harvey <Lesaun@gmail.com>
Small, non-functional changes were made to non-compliant files.
These include breaking up long lines, whitespace sanitation and
unused import removal.
Maximum line length in python files was set to a generous 125 chars,
in order to minimize number of changes needed in scripts and general
annoyance. The "txt" prompts directory is excluded from the checks
as it may contain oddly formatted files and strings for a good reason.
Signed-off-by: Jiri Podivin <jpodivin@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>
* make finetuning example accessible
* fixed: targed was in wrong line
* fixed: name of executable was wrong
* fixed: naming of binary
* fixed: model path was wrong
* fixed clean target
* Update examples/train-text-from-scratch/README.md
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update baby-llama.cpp
Seems to be an error in the implementation of the operator!= function. It attempts to compare the this pointer (a llama_hparams_lora object) with the other pointer (a llama_hparams object) using memcmp. This can lead to incorrect results because the sizes of the objects being compared (sizeof(llama_hparams) and sizeof(llama_hparams_lora)) are different, should now be able to compare two llama_hparams_lora objects for inequality.
* Update baby-llama.cpp
* Update baby-llama.cpp
* 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>
