Daniel Bevenius [Tue, 9 Sep 2025 04:06:52 +0000 (06:06 +0200)]
requirements : update transformers/torch for Embedding Gemma (#15828)
* requirements : update transformers/torch for Embedding Gemma
This commit updates the requirements to support converting
Embedding Gemma 300m models.
The motivation for this change is that during development I had a local
copy of the transformers package which is what I used for converting
the models. This was a mistake on my part and I should have also updated
my transformers version to the official release.
I had checked the requirements/requirements-convert_legacy_llama.txt
file and noted that the version was >=4.45.1,<5.0.0 and came to the
conculusion that no updated would be needed, this assumed that
Embedding Gemma would be in a transformers release at the time
Commit fb15d649ed14ab447eeab911e0c9d21e35fb243e ("llama : add support
for EmbeddingGemma 300m (#15798)) was merged. So anyone wanting to
convert themselves would be able to do so. However, Embedding Gemma is
a preview release and this commit updates the requirements to use this
preview release.
* resolve additional python dependencies
* fix pyright errors in tokenizer test and remove unused import
model-conversion : add extra debugging support for model conversion (#15877)
* feat: Extra debugging support for model conversion - added BF16 support for llama-callback-eval and support for dumping intermediate steps in run-org-model.py
Jeff Bolz [Mon, 8 Sep 2025 18:10:07 +0000 (13:10 -0500)]
vulkan: sort graph to allow more parallel execution (#15850)
* vulkan: sort graph to allow more parallel execution
Add a backend proc to allow the backend to modify the graph. The
vulkan implementation looks at which nodes depend on each other
and greedily reorders them to group together nodes that don't
depend on each other. It only reorders the nodes, doesn't change
the contents of any of them.
With #15489, this reduces the number of synchronizations needed.
chat : Deepseek V3.1 reasoning and tool calling support (OpenAI Style) (#15533)
* Add DeepSeek V3.1 thinking mode support
- Added COMMON_CHAT_FORMAT_DEEPSEEK_V3_1 enum value
- Created common_chat_params_init_deepseek_v3_1() function (currently uses R1 implementation)
- Created common_chat_parse_deepseek_v3_1() function that handles V3.1 thinking format:
- Extracts reasoning content before '</think>' tag into reasoning_content
- Extracts regular content after '</think>' tag into content
- No opening '<think>' tag in V3.1 format
- Added detection logic for V3.1 templates based on pattern: 'message['prefix'] is defined and message['prefix'] and thinking'
- Added V3.1 case to parsing switch statement
This addresses the issue where V3.1 outputs reasoning content followed by '</think>' and then regular content without the opening '<think>' tag.
* Another attempt by V3.1 non-thinking
* Fix test, but it's not asserting anything.
* Ignore vim swap files in tests dir
* Update the test
* Try using try_find_literal instead of regex
* passing test
* Revert "Try using try_find_literal instead of regex"
* Strip grammar down to strictly what we expect based on model card. Throw
out parts we cargo culted from R1 that don't make sense.
* Update tests/test-chat-parser.cpp
Co-authored-by: Sigbjørn Skjæret <redacted>
* DeepSeek V3.1 - Add edge case where thinking is forced open, there is
tool calling in the reasoning content, but then the model just stops the
output without closing the </think> tag, so it's not a partial. In this
case, use the tool call in the reasoning content.
Daniel Bevenius [Mon, 8 Sep 2025 07:44:34 +0000 (09:44 +0200)]
convert : force setting sliding_window from original config (#15867)
* convert : force setting sliding_window from original config
This commit modifies the set_gguf_parameters method for EmbeddingGemma
so that it reads the sliding_window parameter from the original model
config.json and uses that value.
The motivation for this change is that the Gemma3TextConfig
constructor adjusts the sliding_window value, which can lead to
inconsistencies when converting models as we expects this value to
match the original model's configuration.
Daniel Bevenius [Sun, 7 Sep 2025 08:19:45 +0000 (10:19 +0200)]
ggml WebGPU: remove userdata from request adapter callback (#15527)
* ggml WebGPU: remove userdata from request adapter callback
This commit removes the `userdata` parameter from the WebGPU request
adapter callback in `ggml-webgpu.cpp`. Instead, the lambda function
captures the `webgpu_context` directly.
The motivation for this change is to simplify the code and improve
readability.
* inline the callback lambda into the RequestAdapter call
This commit removes the callback lambda variable and inlines it directly
into the RequestAdapter call.
Signed-off-by: Gabe Goodhart <redacted>
* fix(cpp): Update to alora_invocation_tokens on c++ side
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* feat: Add C APIs to get alora invocation token array from lora
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* feat: Initial implementation of alora cache logic in server
This does not yet do the part to identify the invocation tokens and only
apply the lora adapter afterwards, but it does seem to produce correct
results if the invocation tokens are the beginning of the uncached input.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* feat: Identify alora invocation sequences
This currently limits to a single enabled alora per slot. Multiple aloras
with different invocation sequences would be possible, but it would require
a more complex integration of the adapter toggling and is not really a well
studied case for alora since it's unclear if one alora can reuse cache from
previous prefill computed with a different alora.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* feat: Only reuse cache for tokens before the alora invocation start
This is a bit of an edge case, but theoretically a user could try the same
query with the alora disabled (just using the base model), then retry with
the alora. The cached tokens from the first pass should be invalid.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* feat: Handle un-cached tokens that come before the alora activation
The solution is to only fill up to the token before the invocation start in
the batch if there are any tokens to be prefilled between those pulled from
cache and the invocation start. When this is detected, the alora is
temporarily disabled with a scale of 0.0, then immediately re-enabled after
it has been initialized for the internal graph. Since the batch does not
complete the prompt tokens, the remaining prompt tokens are handled in the
next task, pulling all of the non-alora tokens from cache and proceeding
with prefill for the alora tokens.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* fix: Use || instead of 'or'
Too much python :facepalm:
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* fix: Fix off-by-one for limiting cached tokens to before alora start
This was the cause of the inconsistent results from the dummy test script
with and without the turn that runs the prompt without the adapter before
running it with the adapter.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* fix: Support backwards-compatibility for "invocation_string" in adapter_config.json
While this has been replaced in the PEFT PR in favor of
alora_invocation_tokens, the existing adapters in the ibm-granite org on HF
use "invocation_string," so this will enable backwards compatibility and
enable testing now (before PEFT PR changes have percolated everywhere).
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
* fix: Remove duplicate logging
Signed-off-by: Gabe Goodhart <redacted> Co-authored-by: Sigbjørn Skjæret <redacted>
* feat: Report alora_invocation_string and alora_invocation_tokens from /lora-adapters
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <redacted>
---------
Signed-off-by: Gabe Goodhart <redacted> Co-authored-by: Sigbjørn Skjæret <redacted>
Signed-off-by: Gabe Goodhart <redacted>
* fix: Always parse the enable_thinking kwarg to overwrite the default value
From what I can tell, this started as a Qwen3-specific keyword, but from
the use in `chat.cpp` translates this inputs.enable_thinking to the right
thinking kwarg for the given model, this is now more of a standardized
kwarg, so it should always override the default value when sent as part of
the chat_template_kwargs field in the API.
Signed-off-by: Gabe Goodhart <redacted>
* feat: Explicitly reject string values for "enable_thinking"
There are too many possible "truthy" / "falsy" strings and too many
ambiguous strings that don't have a clear truthy/falsy value, so the
simplest thing to do here is to reject the request. Ideally, this would be
a 422 (Unprocessable Entity), but right now it's coming back as a 500.
Daniel Bevenius [Fri, 5 Sep 2025 12:49:21 +0000 (14:49 +0200)]
tests : add --list-ops and --show-coverage options (#15745)
This commit adds two new command-line options to the
test-backend-ops.cpp that allow users to list all available GGML
operations and to show test coverage of these operations.
The motivation for this is that it can be useful to quickly see which
operations are currently covered by tests and which are not. Also it
migth be useful when using the `support` mode.
Erik Scholz [Fri, 5 Sep 2025 09:34:28 +0000 (11:34 +0200)]
gguf: gguf_writer refactor (#15691)
* gguf: split gguf writer into base and buf impl
* gguf: templated gguf write out
* gguf: file based writer (avoid writing everything to memory first!)
* examples(llama2c): fix log not being the same level and compiler nits
Daniel Bevenius [Fri, 5 Sep 2025 02:36:23 +0000 (04:36 +0200)]
model-conversion : add --embeddings flag to modelcard.template [no ci] (#15801)
This commit updates the modelcard.template file used in the model
conversion scripts for embedding models to include the llama-server
--embeddings flag in the recommended command to run the model.
The motivation for this change was that when using the model-conversion
"tool" to upload the EmbeddingGemma models to Hugging Face this flag was
missing and the embedding endpoint was there for not available when
copy&pasting the command.
Daniel Bevenius [Thu, 4 Sep 2025 16:10:29 +0000 (18:10 +0200)]
llama : add support for EmbeddingGemma 300m (#15798)
This commit add support for the EmbeddingGemma 300m. This model supports
sliding window attention (SWA) and a new swq_type is introduced to
support symmetric SWA masking.
This commit also extracts the code from the function
llama_is_masked_swa in llama-impl.h, so that the logic can be shared
by both llm_graph_input_attn_no_cache::set_input and
llama_kv_cache::set_input_kq_mask.
With this commit the EmbeddingGemma 300m model can be converted to
to GGUF and used with llama.cpp.
Once the model has been uploaded to HuggingFace it can be used like
this:
```console
./build/bin/llama-cli -hf ggml-org/embeddinggemma-300m-GGUF:Q8_0
```
Daniel Bevenius [Thu, 4 Sep 2025 13:40:44 +0000 (15:40 +0200)]
llama : set n_outputs to 1 to avoid 0 outputs mean-pooling (#15791)
* 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 <redacted>
* add comment about not reserving graphs with zero outputs
* add assert in graph_reserve to ensure n_outputs >= 1
Chenguang Li [Thu, 4 Sep 2025 12:20:14 +0000 (20:20 +0800)]
CANN: Refactor ND to NZ workspace to be per-device (#15763)
* CANN:Refactor ND to NZ workspace to be per-device in Ascend backend
- Replaced the previous single global ND→NZ workspace with a per-device
cache using unordered_map keyed by device ID.
- Functions `release_nz_workspace`, `relloc_nz_workspace`, and
`get_nz_workspace` now manage workspace independently for each device,
preventing memory conflicts in multi-device / pipeline parallel scenarios.
- This change fixes potential precision issues caused by workspace
overwrites when multiple devices perform ND→NZ conversions concurrently.
This commit updates ggml_vk_instance_validation_ext_available() to
check for VK_EXT_validation_features instead of
VK_KHR_portability_enumeration.
Based on how the returned boolean is used later in the code (to enable
both the validation layer and the VK_EXT_validation_features extension),
it appears the function may have been intended to check for the
validation layer features extension.
* remove try/catch
This was a left over from a previous iteration where I was explicitly
quering for a specific validation layer first, which would throw.
Daniel Bevenius [Wed, 3 Sep 2025 11:35:49 +0000 (13:35 +0200)]
llama : fix incorrect model type for Gemma 270M (#15764)
This commit fixes the model type for the Gemma 270M model in
llama_model.cpp which should be LLM_TYPE_270M. I incorrectly added this
previously as LLM_TYPE_537M which was wrong.
The motivation for this is that it causes the model to not be identified
properly when using tools like llama-bench. For example:
```console
$ ./build/bin/llama-bench -m models/gemma-3-270m-Q8_0.gguf
| model | size | ...
| ------------------------------ | ---------: | ...
| gemma3 ?B Q8_0 | 271.81 MiB | ...
| gemma3 ?B Q8_0 | 271.81 MiB | ...
```
With the changes in this commit the output will be:
```console
$ ./build/bin/llama-bench -m models/gemma-3-270m-Q8_0.gguf
| model | size | ...
| ------------------------------ | ---------: | ...
| gemma3 270M Q8_0 | 271.81 MiB | ...
| gemma3 270M Q8_0 | 271.81 MiB | ...
```
This commit updates the bash scripts to use env instead of using
hardcoded /bin/bash in the shebang line.
The motivation for this is that some systems may have bash installed
in a different location, and using /usr/bin/env bash ensures that
the script will use the first bash interpreter found in the user's
PATH, making the scripts more portable across different environments.
* model-conversion : rename script to .py [no ci]
This commit renames run-casual-gen-embeddings-org.sh to
run-casual-gen-embeddings-org.py to reflect its Python nature.
This commit adds a curl script to the model-conversion examples
which is currently missing. This script is required for the running the
embedding server targets to test llama-server embeddings functionality.
Oliver Simons [Tue, 2 Sep 2025 17:40:37 +0000 (19:40 +0200)]
chore: Update `.clang-format` to use `BinPackArguments=true` (#15744)
This seems to correspond with what we want to do, see
[here](https://github.com/ggml-org/llama.cpp/pull/15715#discussion_r2315613796)
and [clang-format docs](https://clang.llvm.org/docs/ClangFormatStyleOptions.html#binpackarguments)
Previously, the slope tensor was set to fp16 to improve efficiency.
While this worked correctly in FA, it caused precision issues in soft_max.
This change applies different data types for different operators
to balance both accuracy and performance.
Chenguang Li [Tue, 2 Sep 2025 06:07:48 +0000 (14:07 +0800)]
CANN: Support eager execution mode under ACL graph compilation (#15712)
* [CANN] Support eager execution mode under ACL graph compilation
Add support for running operators in eager mode while ACL graph
compilation is enabled. This allows bypassing graph execution
and directly submitting ops, which is useful for debugging and
reducing graph build overhead in certain scenarios.
Signed-off-by: noemotiovon <redacted>
* fix typo
Signed-off-by: noemotiovon <redacted>
* rename to acl_graph_mode
CUDA: fix build error from ambiguous __half conversions in conv2d (#15690)
* CUDA: fix build error from ambiguous __half conversions in conv2d
Building conv2d with half precision failed because `__half` defines
multiple implicit conversion operators (to float, int, short, etc.),
causing ambiguous overload resolution when multiplying with float.
Introduce a templated `to_float` helper that explicitly converts
`__half` via `__half2float`, while passing through float unchanged.
Use this helper in conv2d accumulation to ensure unambiguous and
correct promotion to float.
Fixes some build errors with half-precision kernels on CUDA.
ggml-ci
* CUDA: Replace custom to_float helper with unified ggml_cuda_cast and add half‑>float conversion
* CUDA: Add missing convert.cuh header
* CUDA: remove unnecessary extension in ggml_cuda_cast
* CUDA: Address review comment, remove second type template argument
CANN: fix RoPE cache issue on multi-device (#15629)
* CANN: fix RoPE cache issue on multi-device
RoPE cache only needs to be computed once per token.
However, in multi-device scenarios, not every device starts
computation from layer 0, which may lead to unallocated memory
issues and precision errors.
This commit records the first layer of each device to avoid
the above issues.
* CANN: Optimize first-layer detection method
* CANN: Remove trailing whitespace
* CANN: Only cache the data that can be determined as unchanged through the parameters.
Diego Devesa [Sun, 31 Aug 2025 13:49:03 +0000 (06:49 -0700)]
llama : separate compute buffer reserve from fattn check (#15696)
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.
overview / pkg / ggml / sources / llama.cpp / log