* ggml : add GGML_SCHED_NO_REALLOC option to disable reallocations in ggml_backend_sched
Enabled in ggml-ci for testing.
* llama : update worst-case graph for unified cache
* ci : disable op offload in some tests
* fix spelling
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Fix llama-save-load-state which currently fails by handling the case
when batch.logits is nullptr (like when loading state) by allocating
space for all outputs as CPU logits.
* Qwen3 Next - cleaned up version
* Whitespaces and stuff
* Correct minor errors
* Update src/llama-model.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Misc. fixes.
* Clean up code, add missing hybrid qualifier
* Did someone transpose the SOLVE_TRI result matrix? Perhaps...
* Whitespace
* Proper tensors for cb calls
* Use llama-graph.h vertical alignment
* BROKEN: chunking
* Set new tensors as inputs.
* Proper chunk logic
* It's the circle of life...
* More shenanigans for n_seq > 1
* Nail in the coffin?
* Fix Windows build
* Eh, one fails on Windows, the other fails on Mac... just use general capture.
* quant : cleanup
* model : cleanup
* qwen3 : cleanup
* cont : cleanup
* cont : cleanup
* ggml : revert change
* qwen3 : cleanup
* cont : cleanup
* Readd cmath
* qwen3 : fix typo
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Usual suspects
* fix my bad suggestion
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit updates the backend sampling implementation to support
intermixed usage of backend and CPU samplers within the same batch.
The initial implementation was developed as an all-or-nothing solution:
either perform backend sampling for the entire batch, or perform CPU
sampling for the entire batch.
The motivation for this change is to support batches with mixed
sequences. For example, we may have a backend sampler configured for
sequence 0, while sequence 1 in the same batch uses CPU sampling. This
was not supported in the initial implementation.
This issue manifested in llama-server with the webui: decoding with
backend samplers would work initially, but after changing to CPU
sampling, a slot (sequence) could still be using a backend sampler.
This meant that logits in output_reserve would not be allocated,
resulting in an error.
The solution in this commit inspects the batch to determine which
sampling modes are needed and allocates buffers accordingly. However,
there is a known inefficiency: when we have intermixed backend/CPU
samplers in the same batch, we currently copy all logits to the host,
even for sequences using backend samplers.
Added test_backend_cpu_mixed_batch to verify correct behavior with
mixed backend/CPU samplers in a single batch, including dynamic
sampler switching between decode calls.
This commit adds a check to skip the output reordering logic when
n_outputs == 1. With a single output token, the data is trivially
sorted and the reordering code is currently doing unnecessary work
(resetting and rebuilding output_ids to the same values).
The motivation for this change is improved code clarity and avoiding
confusion when debugging. While the performance impact is probably
negligible, this unnecessary work happens on every decode call in
llama-server when processing batches with single-token outputs.
This commit fixes the issue where both sampled tokens and logits/probs
were not being copied correctly from the backend to the host when
multiple backend samplers were used.
A test for this scenario has also been added to ensure that both types
of data are copied correctly when different backend samplers are
employed.
This commit introduces a sampling_info struct to encapsulate all
backend sampling related data within the llama_context class.
It also updates to use more descriptive names for sampled tokens and
candidates in the backend sampler ggml data structure.
This commit precomputes and caches the full-vocab token id list in
llama_context's constructor, so llama_get_backend_sampled_token_ids_ith
always returns a valid pointer.
The motivation for this is that this enables both common/sampling.cpp
and src/llama-sampling.cpp can simplify their logic.
Not all backends samplers that process logits need to set the
sampled_tokens_id as they may not change the order of the logits, for
example the temperature sampler only scales the logits but does not
change their order. Simliar the logit bias sampler only adds bias to
specific token ids but does not change the order of the logits. In
these cases there will not be a device to host copy of the sampled
token ids, and this is the use case where having this precomputed
list is useful.
This commit adds support for performing sampling operations on the
backend (e.g. GPU) as part of the model computation graph.
The motivation for this feature is to enable sampling to be performed
directly on the backend as part of the computation graph being executed,
allowing for some or all of the sampling to be done on the backend.
For example, the backend sampler chain might select/sample a token
directly in which case only the sampled token needs to be transferred
from device memory to host memory.
It is also possible for the backend samplers to perform filtering of
the logits, or compute and filter the probability distribution, in
which case only the filtered logits or probabilites need to be
transferred back to system memory for further processing by CPU
samplers.
Currently the backend sampling works in a similar manner to how
pooling works, it is a function that is called by build_graph and the
sampler operations become part of the models computation graph.
* kv-cache : pad the size of the small SWA cache for performance
* context : pad the total context to 256
* cont : future-proof the swa pad
* server : adjust test params to new logic
* server : support unified context across slots
* cont : fix speculative decoding initialization
* context : fix n_ctx_per_seq computation
* server : purge slots one by one
* tests : add unified cache server tests
* llama : update per-seq context computation
* test-thread-safety : handle tiny training context of the input model
* server : fix server_tokens clear()
* server : use 4 slots + unified KV by default
* llama : add note about context size queries
* cont : update todos [no ci]
* context : do not cap the size of the context
* tests : adjust parameters to be CI friendlier
* context : add warning
The unexpeced pooling_type warning was incorrectly shown when users did not
specify the --pooling-type parameter. In this case, the parameter
defaults to `LLAMA_POOLING_TYPE_UNSPECIFIED (-1)`, and the code
automatically applies the model's default pooling type.
Example of spurious warning:
```
$ llama-embedding -hf ggml-org/bge-m3-Q8_0-GGUF -p "hello"
...
llama_init_from_model: model default pooling_type is [2], but [-1] was specified
...
```
This fix ensures the warning only appears when users explicitly specify
a pooling type that differs from the model's default (e.g., using
--pooling-type mean on a model that expects CLS pooling).
* model: EmbeddingGemma sentence-transformers dense linear projections support
* model: add support for EmbeddingGemma SentenceTransformers dense linear projections
Adding support for the Dense modules used in EmbeddingGemma models.
EmbeddingGemma is a SentenceTransformers model with additional modules beyond the base Transformer backbone.
See: https://developers.googleblog.com/en/gemma-explained-embeddinggemma-architecture-and-recipe/
* model: add support for EmbeddingGemma SentenceTransformers dense linear projections
- converting model with dense-layers is optional
- introduced dense config params
* Update convert_hf_to_gguf.py
Co-authored-by: Daniel Bevenius <daniel.bevenius@gmail.com>
* fixed formatting issues
* Update src/llama-graph.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* - removed pooling_type_opt, always allow overriding pooling_type
- asserts checking dense features dims
* fix python lint
* fix ubuntu gcc build warning
* - fixed thread-safety test
- moved asserts to load_hparams
* - tidying up code
- simplifying graph-context expecting both dense weights
* minor : add TODO
---------
Co-authored-by: Daniel Bevenius <daniel.bevenius@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* add grok-2 support
* type fix
* type fix
* type fix
* "fix" vocab for invalid sequences
* fix expert tensor mapping and spaces in vocab
* add chat template
* fix norm tensor mapping
* rename layer_out_norm to ffn_post_norm
* ensure ffn_post_norm is mapped
* fix experts merging
* remove erroneous FFN_GATE entry
* concatenate split tensors and add more metadata
* process all expert layers and try cat instead of hstack
* add support for community BPE vocab
* fix expert feed forward length and ffn_down concat
* commit this too
* add ffn_up/gate/down, unsure if sequence is right
* add ffn_gate/down/up to tensor names
* correct residual moe (still not working)
* mess--
* fix embedding scale being applied twice
* add built in chat template
* change beta fast for grok if default value
* remove spm vocab in favor of community bpe vocab
* change attention temp length metadata type to integer
* update attention temp length metadata
* remove comment
* replace M_SQRT2 with std::sqrt(2)
* add yarn metadata, move defaults to hparams
This commit adds check for two function pointers returned from
ggml_backend_reg_get_proc_address.
The motivation for this is that the function pointer could be nullptr if
the get proc address function changes in the future. This is also
consistent with all the other calls to ggml_backend_reg_get_proc_address
in the code base.
* llama : set n_outputs to 1 to avoid 0 outputs mean-pooling
This commit modifies the llama_context constructor to set n_outputs to
1.
The motivation for this is that when using pooling, and specifically
mean pooling, for embeddings having n_outputs set to 0 can lead to the
following error:
```console
$ build/bin/llama-embedding -m models/nomic-embed-text-1.5-Q4_K_M.gguf \
--pooling mean -p "Hello, how are you?"
...
llama_context: CPU output buffer size = 0.12 MiB
/home/danbev/work/ai/llama.cpp/ggml/src/ggml.c:3023: GGML_ASSERT(ggml_can_mul_mat(a, b)) failed
0x0000743c96d107e3 in __GI___wait4 (pid=292978, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
warning: 30 ../sysdeps/unix/sysv/linux/wait4.c: No such file or directory
30 in ../sysdeps/unix/sysv/linux/wait4.c
196 waitpid(child_pid, NULL, 0);
230 ggml_print_backtrace();
3023 GGML_ASSERT(ggml_can_mul_mat(a, b));
1823 cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
18983 llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
1399 auto * gf = model.build_graph(gparams);
292 auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
2329 auto * ctx = new llama_context(*model, params);
913 llama_context * lctx = llama_init_from_model(model, cparams);
105 common_init_result llama_init = common_init_from_params(params);
[Inferior 1 (process 292976) detached]
Aborted (core dumped)
```
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* add comment about not reserving graphs with zero outputs
* add assert in graph_reserve to ensure n_outputs >= 1
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Exposes ggml_backend_sched_split_graph() to allow splitting the graph without allocating compute buffers and uses it to split the graph for the automatic Flash Attention check.
* server : add SWA checkpoints
ggml-ci
* cont : server clean-up
* server : handle state restore fails
* llama : add extended llama_state_seq_ API
* server : do not make checkpoints if --swa-full
ggml-ci
* llama : remove flags value for NONE
* server : configure number of SWA checkpoints with CLI arg
ggml-ci
* args : fix scope of new argument
* examples/finetune -opt SGD (stochastic gradient descent) memory opt
add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.
support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)
llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)
(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val: [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00
SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val: [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)
note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')
-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.
note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence
new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)
cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)
since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)
test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values); tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)
* Vulkan: Implement GGML_OP_OPT_STEP_SGD
* tests: Fix OPT_STEP_SGD test-backend-ops
* SGD op param store weight-decay and not 1-alpha*wd
* minor + cosmetic changes
* fix vulkan sgd
* try CI fix
---------
Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Georgi Gerganov
Diego Devesa
Daniel Bevenius
Piotr Wilkin (ilintar)
Sigbjørn Skjæret
Johannes Gäßler
takuya kodama
Saba Fallah
Haiyue Wang
Jonathan Graehl
compilade