* cuda : prints wip
* cuda : new cublas gemm branch for multi-batch quantized src0
* cuda : add F32 sgemm branch
* cuda : fine-tune >= VOLTA params + use MMQ only for small batches
* cuda : remove duplicated cuBLAS GEMM code
* cuda : add CUDA_USE_TENSOR_CORES and GGML_CUDA_FORCE_MMQ macros
* build : add compile option to force use of MMQ kernels
* implementing parallel decoding in server example
* crash fixed
* save dev progress
* refactored sampling function
* completion endpoint working
* multiple client support
* grammar + no stream completion
* cached prompt support
* chat.mjs support cached prompt + some fixes
* server ui now support multiple clients
* unused change reverted
* fixed timings per slot
* add context swap
* add changes to README.md
* llava multimodal integration
* fixed tokens probs
* add multimodal input - alfa
* refactor code + remove unused comments + improved README.md
* fix compilation errors with llvm
* notify the user from server ui that multimodality is unavialable
* some ci fixes
* fix ci make build undefined ref errors
* fix long prompt than ctx proposed in #3639
* fixed premature end due stop word
* context shift fixed
* fix llava implementation
* sync README.md changes
* readme change
* update api like OpenAI
* multimodal support enabled by default
* fix make bui;d errors
* fix multiple clients
* fix zig build
* new sampling API
* latest changes of sampling API
* server : coding-style normalization
* server : coding-style normalization (part 2)
* server : remove beam-search functionality
* server : bug fix in ingest_images
n_tokens is incremented internally by llama_batch_add
* server : use refs + use llama_batch_clear()
* server : snake case
* server : minor sync
* added thread safe pipeline
* server : bach has to be allocated for n_parallel sequences
* server : no need for atomic int - already using mutex
* server : logs + minor code style
* server : fix multibyte handle in partial response (#3706)
* fix image load + view image in chat
* make : silence stb warnings
* clip : link to ggml, not to llama
* server : fix switch fallthrough
* server : fix crash in Debug on macOS (I have no idea why this fixes it!?)
* server : refactor ctx_sampling init + n_ctx + names
* server : bug fix for prompt caching
* Do not save/load image_data to localStorage
* editorconfig : new line in index.html
* server : completion requests remember slot_id
* Update readme to document multimodal in server
* server : minor style
* Update readme to document multimodal in server
* server : hide ctx_sampling->prev behind API (#3696)
* server : apply fix from #3722
* server : fix slot reuse
* server : add comment about changing slot_state to bool
---------
Co-authored-by: FSSRepo <go778sgt@gmail.com>
Co-authored-by: Damian Stewart <d@damianstewart.com>
Co-authored-by: Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
Co-authored-by: Jhen-Jie Hong <iainst0409@gmail.com>
Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
* WIP: start implementing LLaVA
* rm scratch buf for now, will revert after cleanup
* LLaVA image encoder is working. will combine with llama
* Add llava inference code, but it's buggy. debugging
* LLaVA is working e2e, needs to optimize memory allocation + cleanup
* Use ggml_allocr + rm unnecessary code
* fix: crlf -> lf
* fix: new line at EoF
* fix: trailing whitespace
* Add readme
* Update readme
* Some cleanup
* Are you happy editorconfig?
* rm unused batch image preprocessing
* rm unused import
* fix: rm designated initializers
* introduce pad-to-square mode for non-square images
* are you happy editorconfig?
* gitignore /llava
* Handle cases where image file does not exist
* add llava target to Makefile
* add support for 13b model variant
* Maybe seed is unlucky?
* Check if apples are compared to apples
* are you happy editorconfig?
* Use temperature = 0.1 by default
* command line: use gpt_params_parse()
* minor
* handle default n_predict
* fix typo
* llava : code formatting, rename files, fix compile warnings
* do not use Wno-cast-qual for MSVC
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix mirostat state when using multiple sequences
* Fix mirostat by completely refactoring sampling!
* Try to fix zig build.
* Export function to fetch/create default sampler states
Code formatting cleanups and add some comments
Silence a warning about id not being used when logging is disabled
* Apply some renaming suggestions.
Fix comments that were out of sync with the pull.
* Use more consistant naming convention for sampling contexts
* Work on the BPE tokenizer
Tokenizer tests work for Falcon-7B
* Try to fix build problem
* Fix debug assertion failure
* Fix MSVC Unicode BOM problem
* Cleanup and an improvement
* Fix compiler warning
* Cleanup
* Test doesn't work over the full range of Unicodes
* Update .gitignore and Makefile
* Another Makefile rule
* Testing Aquila
* Moving byte decoding back to `token_to_piece` ...
... because everyone is using it.
* Guarding some unusable code pathes
* Streamlining code and adding some more assertions
Important change: I'm classifying added tokens as control tokens now for BPE.
* Adding a comment
* Adding another assertion
* Fixed vocabulary guarding assertions
* Fix PR for recent change
* Fix PR for recent change
* Fix for compiler warning
* Fix PR for recent change
* Fix PR for recent change
* Fix PR for recent change
* Fix for compiler warning
* Fixes for more compiler warnings
* Remove unused code
* Fix initialization of static maps
* Add scores and token types back, adapt gptneox
* Update llama.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update unicode.h
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update unicode.h
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Ported Starcoder and added some assertions
* Fix coding style
* Apply @jploski 's fix for missing tokens
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* vvhg-code-infill (#1)
* infill in separate example (#2)
* reverted changes to main and added infill example
* cleanup
* naming improvement
* make : add missing blank line
* fix missing semicolon
* brought infill up to current main code
* cleanup
---------
Co-authored-by: Cebtenzzre <cebtenzzre@gmail.com>
* fix track_max_mem in forward_batch_wo_cache_flash_attn_train
* remove unnecessary Adam(W) optimizer tensors.
reduces optimizer memory overhead from 7*modelsize to 2*modelsize.
additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t.
bumps training checkpoint file version, but old checkpoints can still be read.
new version with less tensors is saved.
* add gradient clipping to AdamW
* Fix reset of unused g->nodes and g->grads to NULL
* implement gradient checkpointing for training
reduces memory overhead from O(n_layer) to O(sqrt(n_layer))
as explained in readme of https://github.com/cybertronai/gradient-checkpointing
* remove unused compute buffer 3
* add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
* change AdamW decay parameter to work like the torch AdamW decay parameter
It is now relative to Adam learning rate `alpha*sched`.
Before that it was relative to `sched` only.
`alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1]
* change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT
* change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW
btw: the default weight decay parameter for torch.optim.AdamW is 0.01
* bug fixes for cross entropy loss
ggml_cross_entropy_loss: sums where not correctly added in workload of each thread
ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues
guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16
cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup.
so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance.
* fix test-grad0 for cross_entropy_loss
the second argument to cross_entropy_loss must sum up to 1 for each row
* fix test-grad0 for soft_max
dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
* improve finite differences of test-grad0 by using double instead of float
* change cross_entropy_loss to output average over all rows
this helps keeping the loss and gradients in a sane range
* improve gradient checkpointing
sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal.
since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different:
```
given: n, u, v
objective: minimize(a*u+b*v) where a*b=n, a>0, b>0
b=n/a
minimize(a*u+v*n/a)
diff(a*u+v*n/a, a) = u - (v*n/a)/a
diff(a*u+v*n/a, a) == 0
u - (v*n/a)/a == 0
u == v*n/(a*a)
u*a*a = v*n
a*a = v*n/u
a = sqrt(n*v/u)
```
this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage.
* disable gradient checkpointing debug output
* llama : fix rope usage in train-text-from-scratch after ChatGLM change
* add more training parameters:
--enable-restart N Only for Adam optimizer. Enable restarts of cos-decay
--disable-restart N Only for Adam optimizer. Disable restarts of cos-decay
--opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero.
--opt-delta N Maximum delta for delta convergence test. Disabled when <= zero.
--opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero.
--adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero.
--adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha
* replace memcpy with reshape operation so that the graph is not cut at the input
this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it
* remove unused function argument from get_example_targets_batch
* measure and print total training time
* add optimization callback to ggml_opt_resume_g
this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)).
can be used for dynamic learning schedule and setting input data for batches before each iteration
* use optimization callback in training
allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters
reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration
* add minimum number of tensor dimensions to apply weight decay (default 2)
this allows to not apply weight decay to bias parameters
* rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup
* fix increase of model.train_samples and model.train_tokens
now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations
* change sampling parameters for prediction after training to defaults of common.h
and clarify what is context for prediction and what are generated tokens
* tighten abs error bounds for cross_entropy_loss in test-grad0
* add conditional compilation of using F16 exp in flash attention
uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention
* tighten abs error bounds for flash_attn in test-grad0
* tighten abs error bounds for sqrt in test-grad0
* remove out-commented vectorized code of opt_adam
the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead
* ggml : update ggml_rms_norm_back with configurable eps
* llama training : fix ggml_rms_norm_back calls to pass configurable eps
* remove trailing whitespace
* add train function using automatic gradient checkpointing backward pass and allocator
* in train function replace add_inplace by regular add
because using add_inplace seems to result in different gradients
* don't use allocate hash_map on context
because the context has no_alloc=True when using memory allocator resulting in NULL data pointers
* correctly clone reshape and permute operations by also cloning tensor->nb values
* fix variable name and add missing type cast
* terminate recursive tensor cloning when reaching tensor without src tensors
* correctly clone view tensors by setting data pointers
without this the checkpointing would only work when being used together with memory allocator
* fix variable names
* swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn`
* add input tensors as checkpoints
so that recursive tensor cloning of gradient checkpointing terminates on input tensors
* fix variable name and add missing boolean negation
* make sure some tensors are not reallocated by inserting new temporary nodes depending on them:
output and parameter gradient tensors need to be available at the end of the graph execution
parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration
checkpoint tensors are allocated all together to reduce memory allocator fragmentation
afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs
* fix ASSERT to work with zero layers
* add training options whether to use allocator and/or unified training function
* integrate unified training function which may use memory allocator
the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing
* format name of cloned tensors with " (clone)" suffix
* set names for tensors in unified train function for easier debugging
* allocate graph on context using ggml_new_graph
* remove handwritten training functions
* remove unused training parameters "use_scratch" and "use_unified"
* remove trailing whitespace
* remove unused train params: mem_compute1_gb & mem_compute2_gb
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
* remove unused forward_batch function
* add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly
* only use ggml_allocr_alloc when tensor has NULL data and is no view
* fix test when to create temporary backward graph
temporary backward graph is only necessary when using checkpointing
* fix memory "leak" in optimizers
each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.
* reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator
with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.
the computation results are the same
* add API functions to access llama model tensors
* add stub example for finetuning, based on train-text-from-scratch
* move and remove code
* add API functions to access remaining model parameters:
mult, head and rot
* first draft for LORA finetune training
* remove const model and layer arguments in API functions for accessing model tensors
* bug fixes to make finetune compile
automatic allocator does not work yet
* add debug prints for training memory improvements
* fix names of lora tensors
* avoid stack overflow resulting from big ggml_cgraph
replace stack allocation and ggml_build_forward by ggml_new_graph in combination with ggml_build_forward_expand
* replace llama API functions to get model tensors by one function to get model tensor by name
LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
* remove unused call to not existing llama_get_layer_from_model
* implement ggml_compute_forward_out_prod_q_f32
* remove trailing whitespace
* add lora finetune support on quantized base model tensors
* add ggml_add_cast API function
this function works like ggml_add, but accepts a data type for the resulting tensor.
only supported for quantized src0 input.
* use ggml_add_cast in finetuning
lora-applied weights will now have data type F32, which improves gradients when finetuning quantized base models
* bug fix: actually use result type passed to ggml_add_cast
* make sure base model tensors data cannot be used in viewable operations
memory allocator would try to make lora application inplace on base model tensors.
since those are memory mapped this will result in memory access violations
* fix bug in ggml_out_prod which resulted in wrong n_dims of result tensors
* avoid keeping in memory ALL of the gradients
The problem here stems from ggml_graph_reset. This function is called in the optimization function, before each graph computation, to reset the gradients to zero. This required a unique memory slot for each gradient: allocating memory from a previosly freed memory location might lead to non-zero input gradients.
During ggml_compute_backward the gradients are build stepwise by adding or substracting new values, starting from a OP_NONE tensor which needs to contain zero-values. This requires the graph reset.
To avoid this I now remember in ggml_build_backward_expand the original OP_NONE gradient tensors in a hash table, which is passed to ggml_compute_backward. There instead of using add (or sub or similar) I test whether the existing gradient to be changed is a zero-valued-tensor by looking up its existence in the hash table. When it is such a zero-tensor it will not be modified, but replaced by the value to be added, otherwise the regular add (not inplace, allocator will take care of this) will be used. This way none of those zero-tensor values will be necessary in the final backward graph and more importantly they won't need a unique memory slot, just to make them zero.
* remove trailing whitespace
* remove debug prints and function to compute tensor data hash
* improve optimization iteration prints
* adjust maximal values to support finetuning 3B models
* change default finetune params lora_r and lora_alpha to match the n_rank parameters of 4
* bug fix: make sure finetune input gradient is allocated at begin and kept until end
* remove unnecessary src tensor from ggml_get_rows_back
we don't need data of src[2] for computation, only to setup the correct output shape.
remove dependency on src[2], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included.
this is similar to how ggml_reshape does it.
* remove unnecessary src tensor from ggml_repeat & ggml_repeat_back
we don't need data of src[1] for computation, only to setup the correct output shape.
remove dependency on src[1], so that allocator can work more freely.
the computational graph is still completely determined, because the output shape is naturally included
* resolve todo
allocator will only make it inplace when they are of the same type
* mixing multiple LORA adapters is now possible
pass more than one '--lora FNAME' argument to apply more than one LORA.
use '--lora-scaled FNAME S' when you want to specify a user-defined scale for an adapter.
* add option to save finetune output every N iterations
* also save latest finetune output with ITERATION="LATEST" and print where files are saved
saving with LATEST makes it easier to resume training from the latest checkpoint
the string "LATEST" can be configured with command line option "--fn-latest STR"
* update checkpoint train stats before saving via "--save-every"
* add command line option `--rank-wo N` for rank of wo tensor
* update finetune README
* fix dump_non_result_info_yaml to output multiple lora adapters
* bug fix: replace GGML_TYPE_SIZE[t] by ggml_type_size(t)
* replace llama_n_mult by llama_n_ff
* finetune bug fixes to compile with merged in code from master
* remove prediction related code to reduce duplicated code with main
use main instead
* reduce large memory overhead in train-text-from-scratch
all gradients had to be pinned so that graph_reset works correctly.
this is no longer necessary with the changes to ggml_compute_backward introduced in this PR.
* add comment explaining why finetune checkpoints are allocated in one block
* make default value of float member a float literal
* handle rms_norm and rope parameters the same as in train-text-from-scratch
* remove unused code
* remove vocab related code as it is unnecessary
* add LLM_KV_TRAINING_TYPE to train-text-from-scratch checkpoints
so that they can be differentiated from lora finetune checkpoints
* add gguf constants and load/save functions from train-text-from-scratch
* add load & save lora finetune checkpoints via gguf
* add python script to convert old finetune checkpoint files to gguf
* remove old checkpoint save & load code
* remove code to print data checksums which was used to verify correctness of new gguf code
* omit tokenization when training is disabled, only save llama lora adapter
training can be disabled by passing '-n 0' to finetune
* remove trailing whitespace
* update README.md
* implement ggml_compute_forward_repeat_f16
* avoid stack overflow of large cgraphs in test-grad0
* add ggml API functions ggml_unravel_index, ggml_get_i32_nd and its analogs for set and for f32
ggml_get_i32_1d, ggml_set_i32_1d, ggml_get_f32_1d, ggml_set_f32_1d now support non-contiguous tensors.
in case of non-contiguous tensor, the 1d index is unraveled into a multi index using ggml_unravel_index to be passed to '_nd' function equivalent.
this fixes a bug in test-grad0 which happens due to ggml_build_backward not building purely contiguous tensors anymore
* increase test-grad0 context mem size to accommodate for bigger cgraph
* add sanity check to ggml_compute_backward, asserting the correct shape of gradients
* fix ggml_acc_or_set to return tensor of correct shape
* remove unused 'inplace' argument from ggml_compute_backward function
inplace operations to add gradients are no longer created by ggml_compute_backward
use allocator to automatically make inplace operations
* add missing argument 'int i0' to ggml_get_i32_nd & ggml_set_i32_nd header declarations
* fix error message in ggml_allocr_alloc to display actual max_avail
* fix check_gradient
ggml_build_backward_expand was previously replaced by ggml_build_backward, but the assignment of forward graph to backward graph missing
* use tensor->view_src instead of ggml_is_view and get_view_source
* move gradient checkpointing code into ggml, new API function:
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
* replace custom data getters and setters by ggml functions
* train-text-from-scratch can train (full finetune) gguf models
just pass the gguf model via `--checkpoint-in FN`.
after this, to continue training, pass the generated checkpoint instead of the original gguf model.
tested with smaller models, bigger models may exceed available memory.
use (LORA) finetune for those.
* remove trailing whitespace
* add option to save train-text-from-scratch output every N iterations
* update README.md
* fix warnings
* fix warnings
* remove finetune option to disable allocator
the allocator should always be used.
by making sure that it is always used it gets easier to implement automatic memory requirements computation
* add tensor checkpoints only when gradient checkpointing is enabled
* initialize opt ggml context if none was provided
* add ggml-alloc API function 'ggml_allocr_max_size' to get max size of alloc
GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
* finetune: automatically allocate all memory and changes to command line options
remove '--n_examples N' parameter, as it no longer makes sense to call optimization process multiple times in a loop.
add '--only_write_lora' command line option: will skip tokenization and training, to only write a llama.cpp comptabile LORA adapter.
remove memory buffer related command line options.
improve iteration console output.
* add finetune to Makefile
* update README.md
* print time per iteration and estimate remaining time
* increase measured alloc size by tensor_alignment
ggml_allocr_reset will reduce the given size by up to tensor_alignment-1
* fix README.md
* add some more allocator debug prints
* bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue
* revert last commit
"bug fix, probably solves the 'ggml_allocr_alloc: not enough space in the buffer' issue"
"alloc was freeing an externally allocated tensor, because it calculated the end of allocator memory as alloc->data + alloc->max_size instead of alloc->data + alloc->size."
This is intentional to reduce the risk of freeing external tensors when measuring. Unless max_size is not properly calculated, I don't see why this is an issue.
* remove unnecessary "0x" before "%p" output
* move measurement memory segment to upper region of the address space
* update README.md
* fix printf format warnings
* add missing gguf_free in load_checkpoint_lora_file
* load default rms_norm and rope parameters from base model
* add gradient accumulation
specify number accumulation steps with '--grad-acc N'.
this will simulate a bigger batch size of grad_acc*batch.
* fix tracking of train_samples and train_tokens
* build : fix compile warnings
* ggml : fix L-BFGS linesearch loop
* improve finetune time measurement
fix printf warnings on system where int64_t is (long int).
change time datatypes to double because values get big with long training times.
exclude file saving from time measurement.
converge faster to actual time per iteration by removing very small first duration before first iteration was performed.
fix bug in output of total training time, the reported value was 1000 times to small.
* specify default lora rank with '--lora-r N'
'--lora-r N' will specify default rank for all tensors
'--rank-wq N', etc. will override this default rank for specific tensor types.
* fix gradient accumulation bug where the same batch was used for each microstep
* fix gradient accumulation bug where the same batch was used for each microstep
* support grouped-query-attention in ggml_flash_attn and ggml_flash_attn_back
k and v can now be repeated in q along ne[2]
in forward pass just use modulo to compute k and v indices, like ik2 = iq2 % nek2.
in backard pass this won't work as easy, because multiple threads will compete to accumulate to the same k->grad[:,ik1,ik2,ik3] and v->grad[:,iv1,iv2,iv3].
so we change the parallelization over q rows to be over k rows. this ensures non-overlapping (ik2,ik3) across threads.
in each thread we then iterate over the number of repetitions of k/v in q to compute iq2 as iq2 = ik2 + irep*nek2.
since ne2 is not the same for q,k and v we also change how the gradients are concatenated into the result tensor.
additionally the offsets of gradq, gradk and gradv in the result tensor are now memory aligned.
we also simplify the compute_backward part of flash_attn to use ggml_reshape instead of switching over the number of dimensions.
this needs a small change to ggml_reshape, removing the assertion of second argument to be contiguous.
since only the shape (ne) of the second reshape argument is of relevance, its memory layout (nb) is irrelevant -> it can very well be non-contiguous.
change test-grad0 to also test for repeated k/v in q.
this changes the rng and now results in small gradient differences in softmax. these solely come from using f16 exp table lookup in forward softmax: when temporarily changing softmax to use actual exp function, the reported gradient differences go away. gradient differences coming solely from f16 table lookup are acceptable.
added a note to explain this.
* add llama API functions to get grouped-query-attention n_head parameter 'n_head_kv'.
* fix finetune to support grouped-query-attention (using flash-attention)
note: ggml changes to ggml_out_prod are necessary to support grouped-query-attention without flash-attention.
* support broadcastable a in out_prod(a, b) and backward pass of broadcasting mul_mat(a, b)
* test broadcasting mul_mat backward pass
* decouple random number generator of each operation test
when changing one test the rng of others tests is not influenced anymore
* add comment briefly describing what ggml_repeat_back does
* simplify broadcasting mul_mat backward using ggml_repeat_back
* add cgraph evaluation order member and corresponding enum type
this controls in which order ggml_build_forward visits source nodes.
by default the nodes are visited left to right, i.e. src[0] first.
in some cases it is beneficial for ggml-alloc to visit in a different order.
two possible orders are supported: left-to-right (src[0] first) and right-to-left (src[0] last).
* measure max compute size for each cgraph eval order and use best order
this can bring huge memory savings:
e.g. codellama-34b with n_ctx=64, n_batch=1 goes from 92927.8mb down to 4627.6 MB
* remove unused command line options
* add sample start patterns and options to force new or by default resume last shuffling
* update shuffle rng state on reshuffle
* exclude known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* remove probably unnecessary exception type flags from stringstream
* pass correct max number of tokens to llama_tokenize
* account for possible leading whitespace that will be added by tokenizer
e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
* use unrolled vec_mad in out_prod
y is vec_mad result vec.
x is vec_mad input vec.
v is vec_mad input scalar.
ggml_vec_mad_f32_unroll will internally loop over x and v with same y.
GGML_VEC_MAD_UNROLL is by default defined to 32.
This value is empirical optimized using performance test runs of out-prod in openllama-3b finetune with 256 context length and batch size 1. It gives 23% performance boost for out_prod.
Full measurements of out-prod runtime in ms:
unroll_xv unroll_yv
1 67014.643 87826.469
2 77117.552 89077.656
4 72091.311 109121.657
8 61077.543 88678.334
16 56914.67 79514.947
24 59024.595 84350.254
28 55952.446 83368.73
32 51476.658 85177.745
36 55973.792 84659.92
40 55139.616 93844.738
48 60736.392 93330.267
64 99856.878 116994.99
Second column is when unrollying yv instead of xv
* set lora_alpha to value of lora_r if it is not set via command line
otherwise only changing lora_r will change scaling of lora adapter used in prediction
* reshuffle original sample order instead of the previous shuffled order
otherwise resumed reshuffle will not result in same sample order
* block tiling for out-prod inspired by mul-mat
block sizes are empirically optimized
roughly doubles the flops of out-prod
* exclude some more known zero values from computations in flash_attn_f32 & flash_attn_back_f32
* add static keywords
* remove outcommented old code
* update train-text-from-scratch with tokenization, sample selection and shuffling from finetune
* remove lbfgs related train parameters
* move common train functions into common/train.[h|cpp]
* move train state into struct train_state
* move train data saving code into callback to unify code of opt_callback
train_params are still different in finetune and train-text-from-scratch, so it can't yet be moved to train.h|cpp
* move common train params into common/train
* move common opt_callback into common/train
* fix consume_common_train_arg
* save and load head_count_kv in lora checkpoints
* increase train_samples by used_samples instead of number of batches
on batch can contain more than one sample when option "fill_with_next_samples" is used
* fix usage of llama_tokenize
* remove static from process_escape since we need it exposed in header
* fix code formating of long function declarations
* fix condition in load_train_state_gguf
* use die("msg") instead of replace GGML_ASSERT(!"msg") or throw std::runtime_error("msg")
* fix saving and loading of training type
* remove terminating '\0' from tokenization
(llama_tokenize is now passed the string length instead of relying on terminating '\0')
* fix compile warnings
* fix compile warnings
* use new/delete for train_state instead of malloc/free
using malloc may result in seg faults when trying to assign string fields
* assert that sample_count > 0, avoiding division by zero
* fix frand to return value in interval [0,1)
* add train option "--sample-random-offsets"
Use samples beginning at random offsets.
The offset is only applied to the first sample in each batch context window.
Together with "--fill-with-next-samples" this may help for training endless text generation.
For example given a dataset containing samples "abcd", "ABCD", "0123".
With context size of 8 and options "--fill-with-next-samples", "--no-separate-with-eos", "--no-separate-with-bos",
the context windows of batches could only be filled with "abcdABCD", "ABCDabcd", "0123abcd", etc.
With "--sample-random-offsets" it can also be filled with "23abcdAB", "bcd0123A", etc.
* deduplicate code into function
* remove n_rot hparam, as it must always be hparam.n_embd_head()
* align code
* assert correct base model tensor shapes
* move some params from lora hparams into model hparams and load model params from gguf
this equalizes the model definition in finetune and text-from-scratch and removes the need for additional llama api functions to get model parameters
* remove now unnecessary llama API functions to get model params that where added by this PR
* train-text-from-scratch: automatically allocate model tensors, remove option '--mem-model N'
* train-text-from-scratch: automatically allocate opt context
* train-text-from-scratch: automatically allocate input tensors
* train-text-from-scratch: automatically allocate compute memory
* remove unused options and equalize train-text-from-scratch with finetune
* initialize opt->loss_after with zero
* add export-lora program
* remove trailing whitespace
* add export-lora build in Makefile
* remove unused struct tensor_info from export-lora
* add export-lora build dependency to llama
because it depends on common, which depends on llama
* update finetune README.md
* cancel optimization when specified number of epochs is completed
* improve handling of export-lora arguments
print errors and warnings when files could not be read or created
* Fix export-lora.cpp "not enough space in the context's memory pool" (#1)
* Fix export-lora.cpp "not enough space in the context's memory pool"
Without this patch, export-lora would sometimes error with "not enough space in the context's memory pool (needed 656784, available 656800)".
* increase required context size by 5*GGML_MEM_ALIGN instead of plain 16
---------
Co-authored-by: xaedes <xaedes@gmail.com>
* improve handling of not yet supported tensor types
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: meatbag-18a <145869052+meatbag-18a@users.noreply.github.com>
* Fix für #2721
* Reenable tokenizer test for LLaMa
* Add `console.cpp` dependency
* Fix dependency to `common`
* Fixing wrong fix.
* Make console usage platform specific
Work on compiler warnings.
* Adapting makefile
* Remove trailing whitespace
* Adapting the other parts of the makefile
* Fix typo.
* Do not use _GNU_SOURCE gratuitously.
What is needed to build llama.cpp and examples is availability of
stuff defined in The Open Group Base Specifications Issue 6
(https://pubs.opengroup.org/onlinepubs/009695399/) known also as
Single Unix Specification v3 (SUSv3) or POSIX.1-2001 + XSI extensions,
plus some stuff from BSD that is not specified in POSIX.1.
Well, that was true until NUMA support was added recently,
so enable GNU libc extensions for Linux builds to cover that.
Not having feature test macros in source code gives greater flexibility
to those wanting to reuse it in 3rd party app, as they can build it with
FTMs set by Makefile here or other FTMs depending on their needs.
It builds without issues in Alpine (musl libc), Ubuntu (glibc), MSYS2.
* make : enable Darwin extensions for macOS to expose RLIMIT_MEMLOCK
* make : enable BSD extensions for DragonFlyBSD to expose RLIMIT_MEMLOCK
* make : use BSD-specific FTMs to enable alloca on BSDs
* make : fix OpenBSD build by exposing newer POSIX definitions
* cmake : follow recent FTM improvements from Makefile
* build : on Mac OS enable Metal by default
* make : try to fix build on Linux
* make : move targets back to the top
* make : fix target clean
* llama : enable GPU inference by default with Metal
* llama : fix vocab_only logic when GPU is enabled
* common : better `n_gpu_layers` assignment
* readme : update Metal instructions
* make : fix merge conflict remnants
* gitignore : metal
* make : remove unused -DGGML_BIG_ENDIAN
* make : put preprocessor stuff in CPPFLAGS
* make : pass Raspberry Pi arch flags to g++ as well
* make : support overriding CFLAGS/CXXFLAGS/CPPFLAGS/LDFLAGS
* make : fix inverted conditional
* build ci: run make test
* makefile:
- add all
- add test
* enable tests/test-tokenizer-0-llama
* fix path to model
* remove gcc-8 from macos build test
* Update Makefile
* Update Makefile
* tests : add a C compliance test
* make : build C compliance test by default
* make : fix clean and make sure C test fails on clang
* make : move -Werror=implicit-int to CFLAGS
* make : do not pass headers to the compiler
This fixes building tests with clang.
* make : add missing examples
* make : fix build-info.h dependencies
* use hipblas based on cublas
* Update Makefile for the Cuda kernels
* Expand arch list and make it overrideable
* Fix multi GPU on multiple amd architectures with rocblas_initialize() (#5)
* add hipBLAS to README
* new build arg LLAMA_CUDA_MMQ_Y
* fix half2 decomposition
* Add intrinsics polyfills for AMD
* AMD assembly optimized __dp4a
* Allow overriding CC_TURING
* use "ROCm" instead of "CUDA"
* ignore all build dirs
* Add Dockerfiles
* fix llama-bench
* fix -nommq help for non CUDA/HIP
---------
Co-authored-by: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com>
Co-authored-by: ardfork <134447697+ardfork@users.noreply.github.com>
Co-authored-by: funnbot <22226942+funnbot@users.noreply.github.com>
Co-authored-by: Engininja2 <139037756+Engininja2@users.noreply.github.com>
Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
Co-authored-by: jammm <2500920+jammm@users.noreply.github.com>
Co-authored-by: jdecourval <7315817+jdecourval@users.noreply.github.com>
* llama : add benchmark example
* add to examples CMakeLists.txt
* fix msvc build
* add missing include
* add Bessel's correction to stdev calculation
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* improve markdown formatting
* add missing include
* print warning is NDEBUG is not defined
* remove n_prompt and n_gen from the matrix, use each value separately instead
* better checks for non-optimized builds
* llama.cpp : fix MEM_REQ_SCRATCH0 reusing the value of n_ctx of the first call
* fix json formatting
* add sql output
* add basic cpu and gpu info (linx/cuda only)
* markdown: also show values that differ from the default
* markdown: add build id
* cleanup
* improve formatting
* formatting
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* adds simple llama grammar tests
* fix lint and add Makefile
* 0 terminate code_points
* avoid dangling pointers in candidate cleanup
* cleanup grammar at end of test
* metal: matrix-matrix multiplication kernel
This commit removes MPS and uses custom matrix-matrix multiplication
kernels for all quantization types. This commit also adds grouped-query
attention to support llama2 70B.
* metal: fix performance degradation from gqa
Integers are slow on the GPU, and 64-bit divides are extremely slow.
In the context of GQA, we introduce a 64-bit divide that cannot be
optimized out by the compiler, which results in a decrease of ~8% in
inference performance. This commit fixes that issue by calculating a
part of the offset with a 32-bit divide. Naturally, this limits the
size of a single matrix to ~4GB. However, this limitation should
suffice for the near future.
* metal: fix bugs for GQA and perplexity test.
I mixed up ne02 and nb02 in previous commit.
* fix hellaswag print format, cast away warning in test-double-float
* c++11 cannot use designated initializers
* add static to test-grad0.c internal functions
* use memcpy in test-double-float.c
* port c tests to c++
* use initializer list for ggml_init_params
* ggml : add graph tensor allocator
* ggml : don't calculate data pointer of unallocated tensors when creating a view with an offset
* ggml : refactor ggml_view_Nd into ggml_view_tensor_offset
* makefile: correct deps for server
* server: tighten settings layout a little
* server: expose all currently configured generation params in UI
* server: expose remaining generation params, for the adventurous
* server: embetter mirostat fields
* llama, main : constrain sampling to grammar
* allow loading grammar from file
* fix whitespace errors
* handle & print parser errors
* add comments to grammar syntax and allow newlines where unambiguous
* add missing include
* support alternates in root rule
* fix bugs with empty token and EOS
* adjust JSON grammar
* remove swp file
* rewrite ternary expressions
Co-authored-by: Henri Vasserman <henv@hot.ee>
* use struct for grammar elements and add Unicode support
* add unicode escapes
* add inverse char ranges
* only sample full tokens (no peeking or truncation)
* llama : minor style changes
blindly applied in online editor - hopefully I didn't break something
* update help text
* add warning message if EOS is disabled
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix Makefile for CLBLAST compile support and instructions for compile llama.cpp FreeBSD
* More general use-case for CLBLAST support (Linux and FreeBSD)
A fix in Makefile for FreeBSD users. In the platfrom x86_64 is amd64. This fix resolve compilation using CFLAGS and CXXFLAGS with -march=native and -mtune=native
Add two examples for interactive mode using Llama2 models (thx TheBloke for models)
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Programs in the tests directory are now build with target tests
and placed in the same location.
* clean target was expanded to remove new binaries
* test target binaries are listed in a variable
* Locations of binaries were added to the .gitignore
Signed-off-by: Jiri Podivin <jpodivin@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* MPI support, first cut
* fix warnings, update README
* fixes
* wrap includes
* PR comments
* Update CMakeLists.txt
* Add GH workflow, fix test
* Add info to README
* mpi : trying to move more MPI stuff into ggml-mpi (WIP) (#2099)
* mpi : add names for layer inputs + prep ggml_mpi_graph_compute()
* mpi : move all MPI logic into ggml-mpi
Not tested yet
* mpi : various fixes - communication now works but results are wrong
* mpi : fix output tensor after MPI compute (still not working)
* mpi : fix inference
* mpi : minor
* Add OpenMPI to GH action
* [mpi] continue-on-error: true
* mpi : fix after master merge
* [mpi] Link MPI C++ libraries to fix OpenMPI
* tests : fix new llama_backend API
* [mpi] use MPI_INT32_T
* mpi : factor out recv / send in functions and reuse
* mpi : extend API to allow usage with outer backends (e.g. Metal)
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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
* 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>
* Convert vector to f16 for dmmv
* compile option
* Added compilation option description to README
* Changed cmake CUDA_ARCHITECTURES from "OFF" to "native"
* 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>
* 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>
* 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
* 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
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).
* xor hack
* block y dim
* loop unrolling
* Fixed cmake LLAMA_CUDA_BY option
* Removed hipblas compatibility code
* Define GGML_CUDA_DMMV_BLOCK_Y if not defined
* Fewer iters, more ops per iter
* Renamed DMMV X/Y compilation options
* Move back to C++ for OpenCL
* Refactor OpenCL code to work more like the CUDA code, add missing functions
* Deduplicate dequant kernels
* Add OpenCL compile options
* Use compile args for preprocessing constants
* Restore default platform + device selection by id behavior
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Henri Vasserman <henv@hot.ee>
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix: blas changes on ci
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add git-based build information for better issue tracking
* macOS fix
* "build (hash)" and "CMAKE_SOURCE_DIR" changes
* Redo "CMAKE_CURRENT_SOURCE_DIR" and clearer build messages
* Fix conditional dependency on missing target
* Broke out build-info.cmake, added find_package fallback, and added build into to all examples, added dependencies to Makefile
* 4 space indenting for cmake, attempt to clean up my mess in Makefile
* Short hash, less fancy Makefile, and don't modify build-info.h if it wouldn't change it
* llama : minor - remove explicity int64_t cast
* ggml : reduce memory buffer for F16 mul_mat when not using cuBLAS
* ggml : add asserts to guard for incorrect wsize
* Allow use of OpenCL GPU-based BLAS using ClBlast instead of OpenBLAS for context processing
* Improve ClBlast implementation, avoid recreating buffers, remove redundant transfers
* Finish merge of ClBlast support
* Move CLBlast implementation to separate file
Add buffer reuse code (adapted from slaren's cuda implementation)
* Add q4_2 and q4_3 CLBlast support, improve code
* Double CLBlast speed by disabling OpenBLAS thread workaround
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <2141330+slaren@users.noreply.github.com>
* Fix device selection env variable names
* Fix cast in opencl kernels
* Add CLBlast to CMakeLists.txt
* Replace buffer pool with static buffers a, b, qb, c
Fix compile warnings
* Fix typos, use GGML_TYPE defines, improve code
* Improve btype dequant kernel selection code, add error if type is unsupported
* Improve code quality
* Move internal stuff out of header
* Use internal enums instead of CLBlast enums
* Remove leftover C++ includes and defines
* Make event use easier to read
Co-authored-by: Henri Vasserman <henv@hot.ee>
* Use c compiler for opencl files
* Simplify code, fix include
* First check error, then release event
* Make globals static, fix indentation
* Rename dequant kernels file to conform with other file names
* Fix import cl file name
---------
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <2141330+slaren@users.noreply.github.com>
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Improve cuBLAS performance by using a memory pool
* Move cuda specific definitions to ggml-cuda.h/cu
* Add CXX flags to nvcc
* Change memory pool synchronization mechanism to a spin lock
General code cleanup
On my Mac, the direct Q4_1 product is marginally slower
(~69 vs ~55 us for Q4_0). The SIMD-ified ggml version
is now almost 2X slower (~121 us).
On a Ryzen 7950X CPU, the direct product for Q4_1 quantization
is faster than the AVX2 implementation (~60 vs ~62 us).
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Initial version of q4_0 matrix multiplication benchmark
* Bugfix: Added dependency to ggml.o to benchmark
* Reviewer requests: added parameter for threads, switched to ggml_time_us()
* Reviewer input: removed rtsc, use epsilon for check
* Review comment: Removed set_locale
* Feature: Param for numer of iterations, Bugfix for use of parameter threads
* Reviewer suggestion: Moved to examples
* Reviewer feedback: Updated clean: and benchmark: sections
---------
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
* 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>
- main -> examples
- utils -> examples (renamed to "common")
- quantize -> examples
- separate tools for "perplexity" and "embedding"
Hope I didn't break something !
* 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
* Add test-tokenizer-0 to do a few tokenizations - feel free to expand
* Added option to convert-pth-to-ggml.py script to dump just the vocabulary
* Added ./models/ggml-vocab.bin containing just LLaMA vocab data (used for tests)
* Added utility to load vocabulary file from previous point (temporary implementation)
* Avoid using std::string_view and drop back to C++11 (hope I didn't break something)
* Rename gpt_vocab -> llama_vocab
* All CMake binaries go into ./bin/ now
* Update Makefile to detect AVX512 support and add compiler flags if it's available
* Based on existing AVX2 implementation, dot product on one 32-value block of 4-bit quantized ints at a time
* Perform 8 bit -> 16 bit sign extension and multiply+add on 32 values at time instead of 16
* Use built-in AVX512 horizontal reduce add to get sum at the end
* Manual unrolling on inner dot product loop to reduce loop counter overhead
* potential out of bounds read
* fix quantize
* style
* Update convert-pth-to-ggml.py
* mild cleanup
* don't need the space-prefixing here rn since main.cpp already does it
* new file magic + version header field
* readme notice
* missing newlines
Co-authored-by: slaren <2141330+slaren@users.noreply.github.com>
