Sync master with upstream release b6178 #205
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
* 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>
* fix USE_CUDA_GRAPH=OFF ggml-ci * check capture status * completely disable capturing check instead
…rror. (ggml-org#15304) When attempting to do llama-perplexity on certain tasks which have coupled sequences there is a cryptic error that does not tell you what to do, which is to set the -kvu flag. This adds a hint about that fact.
* ggml: fix ggml_conv_1d_dw bug * Fixed conv1d_dw weight tensor dimension.
ggml-ci
* 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
* vulkan: perf_logger improvements - Account for batch dimension in flops calculation. - Fix how "_VEC" is detected for mat_mul_id. - Fix "n" dimension for mat_mul_id (in case of broadcasting). - Include a->type in name. * use <=mul_mat_vec_max_cols rather than ==1
* model : add harmony parser for gpt-oss * gpt-oss : fix grammar trigger from causing empty stack * gpt-oss: tweak the grammar trigger again * gpt-oss : add support for recipient in role header * gpt-oss : fix ungrouped tool calls in grammar * gpt-oss : loosen function name matching during parse * gpt-oss : clean up workarounds * gpt-oss : add template tests * gpt-oss : simulate thinking and tool call tags * gpt-oss : undo think tags when reasoning_format is none * gpt-oss : set special tokens back to user defined * gpt-oss : update openai-gpt-oss template * server : filter out harmony thought messages * gpt-oss : simplify parsing
add expicit conversion operator to support older versions of rocm Switch over to hip_bf16 from legacy hip_bfloat16 Simplify RDNA3 define Reduce swap over of new hipblas api to rocm 6.5 as this version is used for rocm 7.0 previews --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
… Ubuntu 24.04 (ggml-org#15005) fixes ggml-org#15004 Co-authored-by: Paolo Greppi <paolo.greppi@libpf.com>
This commit adds support for the 18-layer model type in the Gemma3 series, which is the size of the Gemma3-270m model. The motivation for this commit is was the only change required for Gemma3-270m to be converted to GGUF format and used with llama.cpp. Once the model has been converted and uploaded to Huggingface it can be used like this: ```console $ ./build/bin/llama-cli -hf ggml-org/gemma-3-270m-GGUF:Q8_0 ```
* eval-callback : stop on first NaN * cont : log error
* test-opt: fix backend support check * Update tests/test-opt.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* fix ios-xcode-build * use xcode-select with fixed version * switch to macos-15 to get xcode 16.4
* ggml-zdnn: inital backend impl Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: temp change z17 to arch15 Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: fix build bugs Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: tensor->extra logging check Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: add layout name mapping, ztensor information Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: separate logging into its own line Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: add shape comparison Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: add ggml_tensor shape log Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> ggml-zdnn: fix incorrect shape logging Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add output buffer check Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: run compute and store into tensor->extra Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add set_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add more loggers Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: update set_tensor logging to check only for matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: last working matmul version Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add comments to prevent accidentally deleting lines Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: support op out_prod Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: update op out_prod to use tensor->extra Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: rewrite the backend implementation Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: bugfix new impl Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix compiler warnings and bugfixes Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: test ztensor finding in init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: implement at least 1 op to test Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: assign tensor->extra to buffer Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add check for view tensors to prevent init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: rework init_tensor to create new buffers Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: switch to std vector instead of array Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: switch buffers back and set to arbitrary number Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: impl init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: update supports_op matmul matrix Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix incorrect ztensor shape, reduce memory padding Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: code clean up Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: impl matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix compiler error missing type Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix missing data transform call Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add bias init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: tighten memory usage, change string allocation Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add bias ztensor and data free Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add bias data transform Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add more debug info for extra buffer transform Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add logger to check if mat mul ops go through set_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: activate bias transform in matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: move weights transform into mulmat Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add more safeguards in matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix sequencing of transforms Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: bugfix transform ztensor vs origtensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: figure out why sigtrap is happening Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix sigsegv Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: move everything back to local declaration Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: move bias data to local also Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: bring back working matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: rewrite into mre Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix missing vector import Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix missing vector import in header Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt to fix sigsegv Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix missing load tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix invalid ztensor buffer release Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add logging to debug free buffer Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: remove free_buffer debug info Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add parmblkformat detections Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add nnpa installed detection Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add zdnn_init call for static libs Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt at fixing invalid buffer Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: switch to using deque to fix pointer deref problem Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add weights logging to check Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt to use unique ptr Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add tensor to pre_tfm_desc logging Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add inputs logging Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: disable op_none initialisation for testing Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix missing return from init_tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: load ztensors in cgraph exec Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: work on moving output ztensor as well Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: disable logging and breakpoints for full test Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt at manually changing the layout Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt at using default nwhc format instead Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: disable global load ztensor for now Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix errorenous output load tensor Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: add guards to prevent loading ztensor if transformed Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: code cleanup Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: bring load ztensor back to init routine Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: code clean up Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix ztensor deallocation abort stabilise ggml <-> zdnn api Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: clean up matmul selection Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: clean up project structure Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: update documentation, prepare for upstream Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * chore: add codeowners Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: disable batched matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: attempt at fixing tensor views during matmul Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: deny all view tensors directly Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix pr comments Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * docs: update ops docs for zdnn Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: redo test-backend-ops for ops.md Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-zdnn: fix typo in build-s390x.md Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * codeowners: remove taronaeo for now Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Revert "codeowners: remove taronaeo for now" This reverts commit 411ea4e. * ggml-zdnn: remove unused ggml_zdnn macro Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> --------- Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
* opencl: add reference `mul_mv_mxfp4_f32` * opencl: add reference `mul_mv_id` for mxfp4 * Q4_0 tranpose fix for Adreno --------- Co-authored-by: shawngu-quic <shawngu@qti.qualcomm.com>
…d_eos (ggml-org#15326) This commit updates common_chat_templates_apply_jinja to use the the add_bos and add_eos parameters from the chat template instead of the inputs. The motivation for this is that currently if the `add_bos` and `add_eos` from the input parameters are used it is possible to there will be a missmatch between the model and the chat template which can lead to the the removal of duplicate BOS/EOS tokens in chat.cpp `apply` to not happen leading to two BOS tokens being added to the template.
jan-service-account
deleted the
update-dev-from-master-2025-08-16-00-11
branch
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Updates dev branch with latest release (b6178) from ggml-org/llama.cpp