* Initial XTC commit
Adds XTC sampler, not activated by default, but recommended settings by default.
* Cleanup
* Simplified chances calculation
To be more inline with the original implementation, chance is calculated once at the beginning.
* First fixes by comments
Still need to look into sorting
* Fixed trailing backspaces
* Fixed RNG to be reproduceable
Thanks to @slaren for directions
* Fixed forgotten header
* Moved `min_keep`
Moved from conditions to a simple check at the end.
* Fixed broken randomization
Thanks to @slaren for explanation
* Swapped sorting for a custom algorithm
Shifts tokens to remove the penalized ones, then puts the penalized at the back. Should make `min_keep` still viable.
* Algorithm rework
1. Scan token from top till the first non-penalizable
2. Remove the last captured token (the least probable above threshold)
3. Shift all tokens to override the remaining penalizable
4. Penalize and put them at the the bottom.
* Added XTC to `test-sampling`
* Simplified algorithm and more tests
* Updated info in common and args
* Merged back lost commits in common and arg
* Update dump info in common
* Fixed incorrect min_keep check
* Added XTC to README
* Renamed parameters, fixed info and defaults
* probability is at 0 by default, but XTC is included in sampling queue
* threshold higher than 0.5 switches XTC off
* Initial server support
* Added XTC to server UIs
* Fixed labels in old server UI
* Made algorithm safer and more readable
* Removed xtc_threshold_max
* Fixed arg after update
* Quick fixes by comments
* Simplified algorithm since threshold_max is removed
* Renamed random distribution
* Fixed tests and outdated README
* Small fixes
* server : accept extra_context for the infill endpoint
ggml-ci
* server : update readme [no ci]
* server : use repo-level FIM pattern if possible
ggml-ci
* llama : improve infill support
ggml-ci
* llama : add more FIM token strings
ggml-ci
* server : update prompt on slot restore (#9800)
* gguf : deprecate old FIM token KVs
* ggml : do not use BLAS with types without to_float
* ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies
* ggml : rename ggml_internal_get_type_traits -> ggml_get_type_traits
it's not really internal if everybody uses it
* ggml : add metal backend registry / device
ggml-ci
* metal : fix names [no ci]
* metal : global registry and device instances
ggml-ci
* cont : alternative initialization of global objects
ggml-ci
* llama : adapt to backend changes
ggml-ci
* fixes
* metal : fix indent
* metal : fix build when MTLGPUFamilyApple3 is not available
ggml-ci
* fix merge
* metal : avoid unnecessary singleton accesses
ggml-ci
* metal : minor fix [no ci]
* metal : g_state -> g_ggml_ctx_dev_main [no ci]
* metal : avoid reference of device context in the backend context
ggml-ci
* metal : minor [no ci]
* metal : fix maxTransferRate check
* metal : remove transfer rate stuff
---------
Co-authored-by: slaren <slarengh@gmail.com>
* rerank : use [SEP] token instead of [BOS]
ggml-ci
* common : sanity check for non-NULL tokens
ggml-ci
* ci : adjust rank score interval
ggml-ci
* ci : add shebang to run.sh
ggml-ci
* Add scaffolding for ggml logging macros
* Metal backend now uses GGML logging
* Cuda backend now uses GGML logging
* Cann backend now uses GGML logging
* Add enum tag to parameters
* Use C memory allocation funcs
* Fix compile error
* Use GGML_LOG instead of GGML_PRINT
* Rename llama_state to llama_logger_state
* Prevent null format string
* Fix whitespace
* Remove log callbacks from ggml backends
* Remove cuda log statement
* feat(gguf-py): Add granitemoe architecture
This includes the addition of new tensor names for the new moe layers.
These may not be correct at this point due to the need for the hack in
gguf_writer.py to double-check the length of the shape for these layers.
Branch: GraniteMoE
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(convert_hf_to_gguf): Add GraniteMoeModel
GraniteMoe has the same configuration deltas as Granite
Branch: GraniteMoE
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(granitemoe convert): Split the double-sized input layer into gate and up
After a lot of staring and squinting, it's clear that the standard mixtral
expert implementation is equivalent to the vectorized parallel experts in
granite. The difference is that in granite, the w1 and w3 are concatenated
into a single tensor "input_linear." Rather than reimplementing all of the
math on the llama.cpp side, the much simpler route is to just split this
tensor during conversion and follow the standard mixtral route.
Branch: GraniteMoE
Co-Authored-By: alex.brooks@ibm.com
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(granitemoe): Implement granitemoe
GraniteMoE follows the mixtral architecture (once the input_linear layers
are split into gate_exps/up_exps). The main delta is the addition of the
same four multipliers used in Granite.
Branch: GraniteMoE
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* Typo fix in docstring
Co-Authored-By: ggerganov@gmail.com
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(conversion): Simplify tensor name mapping in conversion
Branch: GraniteMoE
Co-Authored-By: git@compilade.net
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(convert): Remove unused tensor name mappings
Branch: GraniteMoE
Co-Authored-By: git@compilade.net
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(convert): Sanity check on merged FFN tensor sizes
Branch: GraniteMoE
Co-Authored-By: git@compilade.net
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Allow "output" layer in granite moe architecture (convert and cpp)
Branch: GraniteMoE
Co-Authored-By: git@compilade.net
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(granite): Add missing 'output' tensor for Granite
This is a fix for the previous `granite` architecture PR. Recent snapshots
have included this (`lm_head.weights`) as part of the architecture
Branch: GraniteMoE
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit updates the llama_sampler_sample function to use reserve and
emplace_back for the vector of llama_token_data structs.
The motivation for this change is to avoid the creation of n_vocab
default-constructed llama_token_data structs which are then
immediately overwritten.
* llama: fixed n_vocab for `no_vocab` models
* llama: updated error output for `llama_decode_internal` and `llama_encode_internal`
* llama: log warning if there's no vocab_size in metadata
* llama: correct vocab size for logging
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* feat(gguf-py): Add Granite model and params to gguf-py
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(convert_hf_to_gguf): Add registration and param setup for Granite
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama.cpp): Add config parsing for Granite multiplier params
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama.cpp): First pass at full port of granite deviations from llama
Something is still not working right since the results are mostly terrible,
but on occasion it's producing relevant results at this point, so
_something_ is working.
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama.cpp): Determine granite language 3b instruct by vocab size
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(convert_hf_to_gguf): Use LlamaModel as base for GraniteModel
The defaults in LlamaModel are needed for Granite as well
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama.cpp): Switch Granite param names to use _scale for consistency
Other scalar multipliers are called *_scale, so this provides a more
consistent naming convention.
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(convert_hf_to_gguf/gguf-py): _multiplier -> _scale
The transformers names with _multiplier will now be converted to the _scale
equivalent during conversion.
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama.cpp): Use separate switch clause for granite in llm_load_hparams
Branch: GraniteLM
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit renames n_embed to n_embd in llm_build_rwkv6_time_mix.
The motivation for this change is consistency with the other rwkv6
functions like build_rwkv6 (and other parts of the code base).
This commit makes the cell_id variable const in the inp_s_mask block.
The motivation for this change is consistency with the code in the
inp_s_copy block.
* llama : llama_perf + option to disable timings during decode
ggml-ci
* common : add llama_arg
* Update src/llama.cpp
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
* perf : separate functions in the API
ggml-ci
* perf : safer pointer handling + naming update
ggml-ci
* minor : better local var name
* perf : abort on invalid sampler pointer
ggml-ci
---------
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
This commit updates the comment, which seems to contain a typo or be an
outdated comment, in the copy_mask_state function changing the variable
n_rs to n_kv.
I believe this change is correct and what the comment wants to
convey is to copy the states that are not going to be used in the
upcoming processing, which are the tokens states from n_seqs up to
the number of possible token states n_kv.
* common : do not add null tokens during warmup
ggml-ci
* llama : check that the input tokens are valid
ggml-ci
* tests : fix batch size of bert model
ggml-ci
- Add `struct llama_sampler` and `struct llama_sampler_i`
- Add `llama_sampler_` API
- Add `llama_sampler_chain_` API for chaining multiple samplers
- Remove `LLAMA_API_INTERNAL`
- Add `llama_perf_` API and remove old `llama_print_timings` and `llama_reset_timings`
* ggml-quants : 1.625 bpw ternary packing for BitNet 1.58b
* ggml-quants : faster 1.625 bpw AVX2 vec_dot
Not using a lookup table anymore makes it match q4_0 speed.
* gguf-py : fix formatting
* llama : remove spaces on empty line
* ggml-quants : subtract 1 when back in epi8
This makes the 1.625 bpw type go faster than q4_0. Still not the fastest.
* ggml-quants : Q2_2 now faster than Q4_K on with AVX2
* ggml-quants : cleanup Q1_3 code formatting
* ggml-quants : ARM NEON vec_dot for q2_2 and q1_3
* ggml-quants : use ceiling division when quantizing q1_3
* convert-hf : simplify BitNet pre-quantization
This still results in the exact same tensor weights and scales,
but it reveals some weirdness in the current algorithm.
* convert-hf : allow converting the weird BitNet 1.3B
Its FFN size is 5460 which is not convenient.
The offending tensors are kept in F16,
which makes the final model 5.01 bpw.
* bitnet : replace 1.58b with b1.58, as in the paper
* ggml-quants : fix build failure on Windows
* ggml-quants : attempt to fix Arm 32-bit support
* ggml : add some informative comments in q1_3 vec_dot
* ggml : add TQ1_0 and TQ2_0 ternary quantization types
* ggml : even faster TQ2_0
* ggml : also faster TQ1_0
Same optimization as for TQ2_0 by offsetting the sum instead of the weights.
This makes TQ1_0 almost as fast as Q8_0 on AVX2.
* ggml : fix build issues in certain environments
* ggml : add NEON vec_dot implementation for TQ1_0 and TQ2_0
* ggml : avoid directly using vmlal_high_s8, for 32-bit ARM compat
The compiler seems smart enough to use the same instruction
even when using vget_high_s8 instead.
* ggml : remove q1_3 and q2_2
No more 1.625 bpw and 2.000 bpw,
now instead using 1.6875 bpw and 2.0625 bpw
with TQ1_0 and TQ2_0, respectively.
* llama : remove the separate scale tensors of BitNet b1.58
They won't be needed, since the remaining ternary quant types have
built-in scales.
* ggml-quants : rename fields of TQ1_0 and TQ2_0 structs for consistency
* ggml-quants : allow using vdotq_s32 in TQ2_0 vec_dot
Not yet tested on hardware which supports it,
might not work or might not even compile. But also it might.
It should make the performance better on recent ARM CPUs.
* ggml-quants : remove comment about possible format change of TQ2_0
Making it slightly more convenient for AVX512
but less convenient for everything else is not worth the trouble.
* gguf-py : Numpy (de)quantization for TQ1_0 and TQ2_0
* ggml-quants : use roundf instead of nearest_int for TQ1_0 and TQ2_0
This does not change anything for ternary models,
since their values should never end up being in halfway cases anyway.
* convert : allow direct conversion to TQ1_0 and TQ2_0
The token embeddings and output tensors are kept in F16
to allow quantizing them to Q4_K and Q6_K with llama-quantize.
* llama : handle fallback for TQ1_0 and TQ2_0 with Q4_0
Q4_0 is not completely symmetric (so not lossless for ternary models),
but it should be good enough.
* ggml-quants : allow using ARM dot product instructions for TQ1_0
* ggml-quants : deduplicate TQ1_0 and TQ2_0 __ARM_FEATURE_DOTPROD support
* ggml : remove unused ggml_mul special case
It would otherwise conflict with the more general
optimization coming with Mamba-2.
* ggml : handle TQ1_0 and TQ2_0 in dequantization-based operators
* test-backend-ops : add TQ1_0 and TQ2_0 comments for later
Not yet adding uncommented, because some backends like SYCL and Metal
do not properly handle unknown types in supports_op for GGML_OP_MUL_MAT.
(and Metal also doesn't handle it with GGML_OP_GET_ROWS)
Support for TQ1_0 and TQ2_0 for other backends than CPU
will be added in follow-up pull requests.
* Introduce ggml_compute_threadpool
- OpenMP functional: check
- Vanilla ggml functional: Check
- ggml w/threadpool functional: Check
- OpenMP no regression: No glaring problems
- Vanilla ggml no regression: No glaring problems
- ggml w/threadpool no regression: No glaring problems
* Minor fixes
* fixed use after release bug
* fixed a harmless race condition
* Fix Android bulid issue
* fix more race conditions
* fix deadlock for cases where cgraph.n_nodes == 1
and fix --poll case
* threadpool: use cpu_get_num_math to set the default number of threadpool threads
This way we avoid using E-Cores and Hyperthreaded siblings.
* bench: create fresh threadpool for each test
For benchmarking it's better to start a fresh pool for each test with the exact number of threads
needed for that test. Having larger pools is suboptimal (causes more load, etc).
* atomics: always use stdatomics with clang and use relaxed memory order when polling in ggml_barrier
This also removes sched_yield() calls from ggml_barrier() to match OpenMP behavior.
* threadpool: make polling the default to match openmp behavior
All command line args now allow for setting poll to 0 (false).
* threadpool: do not wakeup threads in already paused threadpool
* fix potential race condition in check_for_work
* threadpool: do not create two threadpools if their params are identical
* threadpool: reduce pause/resume/wakeup overhead in common cases
We now start threadpool in paused state only if we have two.
The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead.
* threadpool: add support for hybrid polling
poll params (--poll, ...) now specify "polling level", i.e. how aggresively we poll before waiting on cond.var.
poll=0 means no polling, 1 means poll for 128K rounds then wait, 2 for 256K rounds, ...
The default value of 50 (ie 50x128K rounds) seems like a decent default across modern platforms.
We can tune this further as things evolve.
* threadpool: reduce the number of barrier required
New work is now indicated with an atomic counter that is incremented for
each new graph that needs to be computed.
This removes the need for extra barrier for clearing the "new_work" and
removes the special case for trivial graphs.
* threadpool: remove special-casing for disposable threadpools
With the efficient hybrid polling there is no need to make disposable pools any different.
This simplifies the overall logic and reduces branching.
Include n_threads in debug print for disposable threadpool.
Declare pause and stop flags as atomic_bool
This doesn't actually generate any memory barriers and simply informs
the thread sanitizer that these flags can be written & read by different
threads without locking.
* threadpool: do not clear barrier counters between graphs computes (fixes race with small graphs)
This fixes the race condition with very small graphs where the main thread happens to
start a new graph while the workers are just about to exit from barriers.
* threadpool: use relaxed order for chunk sync
Full memory barrier is an overkill for this since each thread works on different chunk
* threadpool: remove abort_callback from threadpool state
* threadpool: better naming for thread/cpumask releated functions
* threadpool: consistent use of int type for n_threads params
* threadpool: add support for ggml_threadpool_params_default/init
Also removes the need for explicit mask_specified param.
all-zero cpumask means use default (usually inherited) cpu affinity mask.
* threadpool: move typedef into ggml.h
* threadpool: fix apply_priority() function name
* threadpool: fix swift wrapper errors due to n_threads int type cleanup
* threadpool: enable --cpu-mask and other threadpool related options only if threadpool is enabled
* threadpool: replace checks for compute_thread ret code with proper status check
* threadpool: simplify threadpool init logic and fix main thread affinity application
Most of the init code is now exactly the same between threadpool and openmp.
* threadpool: update threadpool resume/pause function names
* threadpool: enable openmp by default for now
* threadpool: don't forget to free workers state when omp is enabled
* threadpool: avoid updating process priority on the platforms that do not require it
On Windows we need to change overall process priority class in order to set thread priorities,
but on Linux, Mac, etc we do not need to touch the overall process settings.
* threadpool: update calling thread prio and affinity only at start/resume
This avoids extra syscalls for each graph_compute()
* llama-bench: turn threadpool params into vectors, add output headers, etc
* llama-bench: add support for cool off between tests --delay
This helps for long running tests on platforms that are thermally limited (phones, laptops, etc).
--delay (disabled by default) introduces the sleep for N seconds before starting each test.
* threadpool: move process priority setting into the apps (bench and cli)
This avoids changing the overall process priority on Windows for the apps
that use ggml/llama.cpp directy.
* threadpool: move all pause/resume logic into ggml
* threadpool: futher api cleanup and prep for future refactoring
All threadpool related functions and structs use ggml_threadpool prefix.
* threadpool: minor indent fixes
* threadpool: improve setprioty error message
* Update examples/llama-bench/llama-bench.cpp
Co-authored-by: slaren <slarengh@gmail.com>
* threadpool: fix indent in set_threadpool call
* use int32_t for n_thread type in public llama.cpp API
* threadpool: use _new and _free instead of _create and _release
* fix two more public APIs to use int32_t for n_threads
* build: set _GNU_SOURCE for Adroid
---------
Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com>
Co-authored-by: fmz <quic_fzaghlou@quic.com>
Co-authored-by: Max Krasnyansky <max.krasnyansky@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
This change fixes a bug where replacing text in a very long string could
cause llama.cpp to hang indefinitely. This is because the algorithm used
was quadratic, due to memmove() when s.replace() is called in a loop. It
seems most search results and LLM responses actually provide the O(n**2)
algorithm, which is a great tragedy. Using a builder string fixes things
* llama : advanced batch splits
This includes equal-sequence-length batch splits which are useful
to simplify recurrent model operators.
* llama : always make recurrent state slots contiguous
* ggml : simplify mamba operators
* llama : fix integer signedness mixing
* llama : logits_all has priority over batch->logits
Otherwise, the server embeddings tests failed.
This was likely an existing problem but was only detected here
because of an additional assertion.
* llama : apply suggestions
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix t5 segfault
* llama : fix Mamba session save and restore
* llama : minor cosmetic changes
* llama : rename llama_reorder_outputs to llama_output_reorder
Also move it closer to llama_output_reserve.
* llama : fix pooled embeddings when using batches with equal_seqs
* minor : add struct members for clarity
ggml-ci
* llama : fix T5 segfault again
* llama : fix Mamba pooled embeddings with multiple sequences
Until the pooled embeddings are refactored to allow splitting
across ubatches for causal embeddings,
recurrent models can only process a single sequence per ubatch
when calculating pooled embeddings.
* llama : add llama_model_is_recurrent to simplify figuring that out
This will make it easier to more cleanly support RWKV-v6 and Mamba-2.
* llama : fix simple splits when the batch contains embeddings
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : std::move llm_bigram_bpe from work_queue
This commit updates the retrieval of llm_bigram_bpe objects from
work_queue.top() by using std::move.
The motivation for this is to avoid the copying of the std::string
`text` member of the llm_bigram_bpe struct.
* squash! llama : std::move llm_bigram_bpe from work_queue
Introduced a MovablePriorityQueue class to allow moving elements
out of the priority queue for llm_bigram_bpe.
* squash! llama : std::move llm_bigram_bpe from work_queue
Rename MovablePriorityQueue to lama_priority_queue.
* squash! llama : std::move llm_bigram_bpe from work_queue
Rename lama_priority_queue -> llama_priority_queue.
* gguf-py : add T5ENCODER model architecture
* common : call llama_decode() during warmup only if the model has decoder
* convert-hf : add T5EncoderModel
* llama : add llama_model_has_decoder() API function
* llama : split build_t5() into build_t5_encoder() and build_t5_decoder()
* llama : add support for LLM_ARCH_T5ENCODER
* llama-embedding : add support for LLAMA_POOLING_TYPE_NONE
* llama-embedding : add support for encoder-only models
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* gguf-py, llama : add constants and methods related to Llama-3.1 <|eom_id|> token
* llama : find Llama-3.1 <|eom_id|> token id during vocab loading
* llama-vocab : add Llama-3.1 <|eom_id|> token to the set of tokens stopping the generation
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* llama : refactor session file management
* llama : saving and restoring state checks for overflow
The size of the buffers should now be given to the functions working
with them, otherwise a truncated file could cause out of bound reads.
* llama : stream from session file instead of copying into a big buffer
Loading session files should no longer cause a memory usage spike.
* llama : llama_state_get_size returns the actual size instead of max
This is a breaking change, but makes that function *much* easier
to keep up to date, and it also makes it reflect the behavior
of llama_state_seq_get_size.
* llama : share code between whole and seq_id-specific state saving
Both session file types now use a more similar format.
* llama : no longer store all hparams in session files
Instead, the model arch name is stored.
The layer count and the embedding dimensions of the KV cache
are still verified when loading.
Storing all the hparams is not necessary.
* llama : fix uint64_t format type
* llama : various integer type cast and format string fixes
Some platforms use "%lu" and others "%llu" for uint64_t.
Not sure how to handle that, so casting to size_t when displaying errors.
* llama : remove _context suffix for llama_data_context
* llama : fix session file loading
llama_state_get_size cannot be used to get the max size anymore.
* llama : more graceful error handling of invalid session files
* llama : remove LLAMA_MAX_RNG_STATE
It's no longer necessary to limit the size of the RNG state,
because the max size of session files is not estimated anymore.
* llama : cast seq_id in comparison with unsigned n_seq_max
* Add llama 3.1 rope scaling factors to llama conversion and inference
This commit generates the rope factors on conversion and adds them to the resulting model as a tensor. At inference time, these factors are passed to the `ggml_rope_ext` rope oepration, improving results for context windows above 8192
* Update convert_hf_to_gguf.py
Co-authored-by: compilade <git@compilade.net>
* address comments
* address comments
* Update src/llama.cpp
Co-authored-by: compilade <git@compilade.net>
* Update convert_hf_to_gguf.py
Co-authored-by: compilade <git@compilade.net>
---------
Co-authored-by: compilade <git@compilade.net>
