The prompt cache constitutes a nice speed up when using the same prompt
prefix across multiple evaluations, but when using it, it will also be
updated, which is not always desirable. One use case is to have a large
prompt containing some context and usage rules, and a second part
containing variable data of the problem being studied. In this case it's
desirable to be able to save the first part once, and to always reuse it
as-is without updating it with the second part.
The new argument --prompt-cache-ro enables this read-only mode on the
prompt cache. The prompt's contents that match the cache are loaded
from the cache but the rest is not modified. This allowed to reduce a
total analysis time from 112s to 49.7s here, without having to backup
and restore a copy of the prompt, which takes significant time at 500
MB.
Signed-off-by: Willy Tarreau <w@1wt.eu>
* 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 98c267fc77fe811082f672538fc91bcfc9072d63.
* 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
Set `LLAMA_BUILD_SERVER` in workflow so the `server` example gets build. This currently only applies to Windows builds because it seems like only Windows binary artifacts are included in releases.
Add `server` example target to `Makefile` (still uses `LLAMA_BUILD_SERVER` define and does not build by default)
Fix issue where `vdot` binary wasn't removed when running `make clean`.
Fix compile warnings in `server` example.
Add `.hpp` files to trigger workflow (the server example has one).
Improvements to loading the session with `--prompt-cache` in the `main` example.
1. Fix an issue where the `--seed` parameter was ignored when loading a cached prompt.
2. When loading a cached prompt, you previously had to specify the saved prompt (or a prefix of it) again. This pull changes that behavior to default to the prompt that was cached if a prompt wasn't specified by the user.
* Added httplib support
* Added readme for server example
* fixed some bugs
* Fix the build error on Macbook
* changed json11 to nlohmann-json
* removed some whitespaces
* remove trailing whitespace
* added support custom prompts and more functions
* some corrections and added as cmake option
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* fix get_num_physical_cores()
had been broken on complex topologies because "cpu cores" in /proc/cpuinfo is per-"physical id"
* Add spaces to maintain consistent formatting
---------
Co-authored-by: slaren <ddevesa@gmail.com>
* implement 8 of 14 missing backward pass operations used by llama
- GGML_OP_ADD_AT
- GGML_OP_CPY
- GGML_OP_MUL_MAT (src0.grad)
- GGML_OP_PERMUTE
- GGML_OP_RESHAPE
- GGML_OP_SCALE
- GGML_OP_TRANSPOSE
- GGML_OP_VIEW
implement additional ggml operation GGML_OP_ADD_AT, which is necessary for backward pass of GGML_OP_VIEW.
this operation adds src1 to src0 with data offset, i.e. to view(src0, ..., offset).
the values are return in a tensor size of src0. values outside of [data+offset:data+offset+nbytes(src1)] are just the original values from src0.
still missing backward passes for llama:
- GGML_OP_DIAG_MASK_INF
- GGML_OP_GET_ROWS
- GGML_OP_RMS_NORM
- GGML_OP_ROPE
- GGML_OP_SILU
- GGML_OP_SOFT_MAX
* implement 5 of 6 missing backward pass operations used by llama
- GGML_OP_DIAG_MASK_INF
- GGML_OP_GET_ROWS
- GGML_OP_RMS_NORM
- GGML_OP_SILU
- GGML_OP_SOFT_MAX
add necessary ggml operations GGML_OP_ADD1, GGML_OP_SILU_BACK, GGML_OP_RMS_NORM_BACK, GGML_OP_DIAG_MASK_ZERO, and GGML_OP_ROPE_BACK
GGML_OP_ADD1 is necessary to add a scalar value in the backward pass of GGML_OP_SOFT_MAX
GGML_OP_ADD1 could also be replaced by using GGML_OP_ADD and GGML_OP_REPEAT, but the performance would be worse. additionally GGML_OP_REPEAT will return unexpected value when the the input to GGML_OP_SOFT_MAX contains only a single scalar. in this case GGML_OP_REPEAT will not return the value that should be repeated (src1) but the value which shape the result should take (src0). So in this case it can not replace GGML_OP_ADD1.
GGML_OP_SILU_BACK, GGML_OP_RMS_NORM_BACK and GGML_OP_ROPE_BACK are necessary for backward pass of GGML_OP_SILU, GGML_OP_RMS_NORM and GGML_OP_ROPE. The backward pass for these functions cannot be easily composed of existing operations. Since the backward pass builds a computation graph we need operations forward pass implementations of the the required backward passes. Sounds a bit confusing at first, I know...
GGML_OP_DIAG_MASK_ZERO is necessary for backward pass of GGML_OP_DIAG_MASK_INF.
Some operations where previously inplace-only. for backward pass there needs to be non-inplace variants.
staying consistent with other operations that have non-inplace and inplace variants, the operations are changed to non-inplace and
functions with "_inplace" are added which are inplace.
in llama we need to call the inplace variants so that it is implemented as before.
for llama backward pass we need to use the non-inplace variants.
still not completely implemented backward passes for llama:
- GGML_OP_ROPE: needs forward pass for GGML_OP_ROPE_BACK
- GGML_OP_GET_ROWS: only necessary for tokenizer
* norm & rms_norm can not be threaded:
after investigation rms norm for quite some time I come to the conclusion that neither norm, nor rms_norm can be threaded, because we need mean over all items, not just of the slices each thread sees.
* remove already resolved TODO
* implement backward pass of ggml_rope and ggml_rope_back
* implement backward pass for ggml_get_rows and for new operation ggml_get_rows_back
* add test-grad0.c
* use GGML_PRINT_DEBUG for debug messages which will otherwise flood the console
* test both gradients of mul_mat
* disable graph dot export as it floods console
* bug fixes for silu_back
* successfully test silu backward
* bug fix for scale backward pass
use sum instead of mean for gradient of scalar scale parameter
* successfully test scale backward
* improve performance of sum backward pass
use add1(x,y) instead of add(x,repeat(y,x))
* improve performance of sqr backward pass
use scale(x,y) instead of mul(x,repeat(y,x))
* successfully test rope backward
* bug fix for cpy backward pass
* successfully test cpy backward
* bug fix for reshape backward pass
* successfully test reshape backward
* add test-opt.c
this uses ggml_opt to train a,b for minimal e=sum(sqr(c - a*b)) for random initial a,b,c
* correctly implement softmax backward pass using new operation ggml_diag
ggml_diag constructs diagonal matrices with entries.
ggml_diag(shape[a,1,c,d]) -> shape[a,a,c,d]
* successfully test soft_max backward
* align shape annotations
* add shape annotations for llama
* de-duplicate ggml_forward_dup code taking care of contiguous tensors of same type.
with this we can duplicate tensor of any typ as long as they are contiguous.
* fix ggml_compute_forward_dup_same_cont for when nelements < nthreads
when more threads are used than elements exist ie1 was less than ie0, resulting in invalid negative byte count argument in memcpy
* bug fix for add_at forward
required for view backward pass
src0 values must be copied to dst, because during addition we don't touch all dst elements in contrast to the normal add function.
* successfully test view backward
* minor code format improvement
* fix ggml_forward_add functions to work correctly with transposed tensors
uses the same logic as in ggml_compute_forward_add_q_f32, but make it consistent across all ggml_compute_forward_add_... functions.
this also slightly changes the mem access pattern of the different threads to works as in ggml_compute_forward_add_q_f32.
* fix ggml_forward_add1 functions to work correctly with transposed tensors
uses the same logic as in ggml_compute_forward_add1_q_f32, but make it consistent across all ggml_compute_forward_add1_... functions.
this also slightly changes the mem access pattern of the different threads to works as in ggml_compute_forward_add1_q_f32.
* test-grad0.c : add print_elements to help with debugging
* successfully test permute backward
* some minor test-grad0 fixes
* fix sub, mul and div functions to work correctly with transposed tensors
uses the same logic as in add
* implement ggml_cont backward pass
* successfully test transpose backward and permute for all permutations
also test sub, mul and div up to max n_dims
* test-grad0.c add TODO for view_2d and view_3d
add_at (required for view backward pass) is a bit tricky for n_dims > 1.
* fix comments
* successfully test diag_mask_inf and diag_mask_zero backward
* test-grad0 : fix test for div
nargs and ndims was swapped, corrupting the stack
* fix diag_mask to work with non-inplace input
* move dup call into the actual add_at functions
* fix get rows backward pass
* successfully test get_rows backward
* fix view backward pass
add nb parameters to add_at like in view.
together with offset they define how to view dst and src0 during the add_at operation.
* successfully test backward pass of view_1d, view_2d and view_3d
* fix backward pass for rms_norm
I would have used formulas from other frameworks, but they differed so I could not decide which is correct.
Instead it was derived here in comment using manual forward-backward automatic differention of rms_norm and simplification.
* successfully test backward pass of rms_norm
some tests may fail when gradients are large.
could not find a satisfying configuration to check for abs error and relative error that passes all tests while still actually testing the results with tight enough error bounds.
when looking at the values the "failed" tests look actually ok. for example:
rms_norm: ndims=2, i=0, k=2, x0=0.000153, xm=0.000053, xp=0.000253, f0=0.278594, f1=0.086213, g0=961.905457, g1=966.064941, eps=0.000100, error_abs=4.159485, error_rel=0.004324
it is due to the test logic in check_gradients that they fail.
* add todos for llama backward pass
- implementation for ADD1 backward pass should probably use sum instead of mean (but this backward pass is not required)
- repeat is not yet tested and looks like it only works for single element src0 inputs.
* add operation ggml_sum_rows
ggml_sum_rows(shape[a,b,c,d]) -> shape[1,b,c,d]
* add missing GGML_OP_SUM_ROWS
* fix backward pass for repeat
requires ggml_sum_rows
* successfully test backward pass of repeat
* update quantization types in switch-case of add_at and add1
* add baby-llama example training a very small llama model from scratch to output a sinusoidal wave.
had to increase maximum number of optimization parameters to train from scratch.
* fix softmax in baby-llama example
* switching from training with adam to lbfgs produces much better results in the baby-llama example
* train with two examples, creating new tensors each time..
* fix bug when using ggml_opt to optimize params in one context and use a renewable context for eval and opt
when not keeping gradients of model parameters they are overwritten by tensors created by opt, which may be invalid after opt context is renewed.
so we need to keep the original gradients and make dups for opt
* train on multiple examples, generate & print tokens with trained model afterwards
ctx0 for evaluation and optimization is renewed for each sample
* add ggml_reshape_1d, ggml_reshape_4d and ggml_view_4d
* fix soft_max backward pass for input->ne[1] != 1
* add ggml_log operation necessary for cross entropy loss
* add test for ggml_log gradients
* implement backward pass for ggml_sum_rows, necessary for cross entropy loss
* implement ggml_repeat support for rank > 2 tensors
* add test for ggml_sum_rows gradients
* fix training get_example_targets
predict the next token, not the current token!
* add square_error_loss and cross_entropy_loss functions
* optimize loss over multiple samples
this increases computation graph, need parallel batched forward for more efficiency.
* fix backward pass for add_at and change arguments to have same order as in view
* add ggml_set(ctx, a, b) to set b in view of a and return modified a
necessary to set values into kv_self cache and properly propagate the gradients
* fix kv_self gradients for training
use ggml_set instead of ggml_cpy to set kv_self cache with properly propagating gradients
* replace inplace operations for training with copying operations to allow gradient propagation
* add GGML_ASSERT to catch ggml_rope and back value errors
* add trainable lora-only model with all big matrices C split into A,B with A*B=C
this is not a lora-finetune, but the whole model changed to have only low-rank "lora" matrices.
training this instead of the normal model resulted in much worse results though...
* vastly improve training results
instead of logit targets 0 and 1 use -1 and +1.
* shorten code using a variable
* change name of GGML_OP_ADD_AT to GGML_OP_ACC
* smaller default values for baby llama model parameters
* update static assert of GGML_OP_COUNT
* remove shape annotations in llama_eval_internal
* revert disabling of threading for rms_norm and norm
* rename print functions in baby-llama example
* fix call to ggml_set_name
* add missing include for strcmp, etc
* remove trailing whitespace
* reduce number of test-grad0 iterations
avoid exceeding timeout of automated tests
* remove busy loop that was used as sleep for slower sinus wave generation
* disable slow tests grad0 and opt to avoid exceeding timeouts
* c++ in baby-llama example
use c++ includes instead of c includes
use std::min, std::max instead of MIN, MAX macros
* c++ in baby-llama example
use c++ includes instead of c includes
use std::min, std::max instead of MIN, MAX macros
* ggml : fix compiler warnings + cosmetic changes
* ggml : fix nullptr derefs in GGML_OP_CONT and GGML_OP_RESHAPE back
* swap arguments to vDSP_vdiv call
documentation for vDSP_vdiv states: "Note that B comes before A!"
* swap arguments to vDSP_vdiv call
documentation for vDSP_vdiv states: "Note that B comes before A!"
* ggml : swap vDSP_vsub args as per documentation
* add parallel batched forward function for baby-llama training
* cleanup code for batched training
* remove trailing whitespace
* minor : fix compiler warnings + indentation style
* ggml : fix null ptr deref in backward pass
* ggml : remove Q4_2 remnants
* ggml : fix clang-tidy warnings
* baby-llama : couple of clang-tidy warnings
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* main : add option to save full output to session
* split behavior into --session and --prompt-cache
* restore original implementation with new names
* PR comments
* move the check for incompatible parameters to gpt_params_parse
* Fix whitespace
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
---------
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
* fix reverse prompt and multi line
* Code Formatting
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
