* server: fix query params lost when proxying requests in multi-model router mode
* server: re-encode query params using httplib::encode_query_component in proxy
* vulkan: allow using fp16 in scalar flash attention shader
* split rows inside of subgroups for faster synchronization
* use row_split when Br >= 4, change reductions to use shared memory if row_split == 1
* use f32 scalar FA if f16 is not supported by device
* fix amd workgroup size issue
* optimize masksh use
* add medium rows FA shader Br size
* fixes
* add padding to mask shmem buffer
* cache q values into registers for KQ
* fuse lf accumulation, pf and v accumulation into a loop
* stage K loads through shmem
* stage V loads through shmem
* only stage through shmem on Nvidia
* default to Bc 32
* also stage V through shmem when this is done for K
* dynamic subgroups for intel
* use vectorized stores
* use float_type for dequantize4 functions
* use smaller scalar rows size for smaller rows count
* relax flash attention split_k condition to allow non-gqa use
* use minimal subgroup size on Intel
* fix shmem support function
* fix rebase issues
* fixes
* Bc 4 for scalar FA is not a valid configuration
* Use wave32 on AMD RDNA for scalar FA
* add Intel shader core count lookup-table
* fix regressions
* device tuning
* tmpsh size fix
* fix editorconfig
* refactor fa tuning logic into a single place
* fix gqa opt logic
* fix block_rows with small n_rows
* amd tuning
* fix hsk=72/80 issue
* tuning
* allow condition skipping for column check
* use float16 for Of if available
* address feedback
* fix bad RDNA performance on head size <= 128 by limiting occupancy
* allow printing pipeline stats
* cleanup and fixes
* limit occupancy for GCN for small batch FA with large HSK
* disable f16 FA for GCN AMD GPUs on the proprietary driver
* hexagon: refactor set/get/sum-rows ops to use local context
* hexagon: refactor ROPE and Softmax Ops to use local context
Improves performance a bit by precomputing things and saving in the context.
* hexagon: refactor activation ops to use local context struct
* hexagon: refactor unary ops to use local context struct and DMA/VTCM
* hexagon: use aligned hvx_scale function
* hexagon: remove unused fields from op_context
* hexagon: rewrite ROPE to use DMA and VTCM scratchpad
* hex-rope: keep N rows in scratchpad (instead of just two)
* hex-rope: introduce rowidx cache
* hex-rope: remove unused fields
* hex-rope: rewrite dma prefetch logic to allow for multi-row fetch/compute
also removes the need for fastdiv.
* hex-rope: minor formatting
* hex-rope: use indices and unroll the loops
* hex-rope: more updates to cleanup rope-block handling
* hexagon: cleanup supported type/dims checks
* hexagon: all reduce funcs replicated across lanes
There is no need to explicitly replicate the first value.
* snapdragon: update adb and windows scripts to use ubatch-size 256
Updated Ops support handles larger ubatches.
This commit replaces/merges the inspect-org-model.py script with the
contents tensor-info.py script. The merged script has also been updated
to also print tensor sizes which was the only thing that was not done
before (by tensor-info.py that is).
The motivation for this is that tensor-info.py does not load the tensor
weights which can be time consuming for larger models. And also now that
both are doing almost the same thing it makes sense to just have one and
not two scripts to maintain.
* llama : remove write/read of output ids/logits/embeddings
This commit removes the write/read of output ids, logits and
embeddings from the llama context state.
Refs: https://github.com/ggml-org/llama.cpp/pull/18862#issuecomment-3756330941
* completion : add replying of session state
This commit updates the session handing in the completion tool to handle
the that logits are no longer stored in the session file. Instead, we
need to replay the last token to get the logits for sampling.
* common : add common_prompt_batch_decode function
This commit adds a new function which is responsible for decoding prompt
and optionally handle the saving for session data.
* update save-state.cpp to use llama_state_load_file
This commit updates the save-load-state example to utilize the new
llama_state_load_file function for loading the model state from a file.
And it also replays the last token after loading since this state is now
stored before the last token is processed.
* examples : set n_seq_max = 2 for ctx3
This commit updates the save-load-state example to set the n_seq_max
parameter to 2 when initializing the ctx3 context.
The motivation for this change is that using 1 as n_parallel/n_seq_max
the context only supports one sequence, but the test laster tries to
use a second sequence which results in the following error:
```console
main : loaded state with 4 tokens
main : seq 0 copied, 225760 bytes
main : kv cache cleared
find_slot: seq_id=1 >= n_seq_max=1 Try using a bigger --parallel value
state_read_meta: failed to find available cells in kv cache
```
This seems to only happen for recurrent/hybrid models.
* Improve CUDA graph capture
Currently, CUDA graphs are eagerly enabled on the first call to ggml_backend_cuda_graph_compute. If the graph properties keep changing (4+ consecutive updates), the graph is permanently disabled. This is suboptimal because:
- The first call always incurs CUDA graph capture overhead even if the graph is unstable
- Once permanently disabled, CUDA graphs never re-enable even after the graph stabilizes (e.g., switching from prompt processing to decode)
The new approach delays CUDA graph activation until warmup completes: the same cgraph must be called at least twice with matching properties before CUDA graph capture begins. This avoids wasted capture overhead on volatile graphs and allows graphs to become eligible once they stabilize.
This also fixes issues such as https://github.com/ggml-org/llama.cpp/discussions/19708
* Update ggml/src/ggml-cuda/ggml-cuda.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Remove EM dashes
* Update ggml/src/ggml-cuda/ggml-cuda.cu
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
* common : fix Step-3.5-Flash format detection and thinking support
Step-3.5-Flash uses the same XML-style tool call format as Qwen3-Coder
(<tool_call><function=...><parameter=...>) but its Jinja template lacks
the bare <function> and plural <parameters> markers that the detection
logic previously required. This caused it to fall through to Hermes 2
Pro, which doesn't call func_args_not_string(), so arguments stayed as
JSON strings and templates using arguments|items crashed.
Additionally, the Qwen3-Coder-XML format handler had no thinking support.
Models like Step-3.5-Flash that unconditionally emit <think> in their
generation prompt need the same thinking_forced_open handling that
Nemotron v3 and Hermes 2 Pro already have, otherwise reasoning_content
is never separated from content in API responses.
Changes:
- Relax Qwen3-Coder XML detection to only require the 3 shared markers
- Tighten Nemotron v3 branch to also require bare <function> and plural
<parameters>, preventing Step-3.5-Flash from being misrouted via <think>
- Add thinking_forced_open support to Qwen3-Coder-XML init function
- Add <think>/</think> to preserved tokens
- Fix build_grammar_xml_tool_call to handle thinking_forced_open in the
grammar root rule, allowing </think> before tool calls
- Add Step-3.5-Flash chat template and format detection test
Builds on: https://github.com/ggml-org/llama.cpp/pull/19283
* chat : route Step-3.5-Flash to Nemotron v3 PEG parser, add tests
Step-3.5-Flash uses the same XML tool call format as Qwen3-Coder and
Nemotron 3 Nano (<tool_call>/<function=...>/<parameter=...>) but with
unconditional <think> output. Route it to the Nemotron v3 PEG parser
for streaming and schema-aware parameter parsing.
Detection: templates with <think> + XML tool tags use Nemotron v3 PEG
parser; templates without <think> (Qwen3-Coder) use GBNF grammar.
Tests cover: basic messages, tool calls with/without thinking content,
parallel tool calls, code string parameters, optional </parameter>
closing tags, and JSON schema response format.
* chat : remove dead thinking code from qwen3_coder_xml
Remove thinking handling code that became unreachable after routing
Step-3.5-Flash to the Nemotron v3 PEG parser. Qwen3-Coder has no
<think> in its template, so the thinking_forced_open logic, preserved
tokens, and grammar prefix were dead paths.
* model: add JAIS-2 architecture support
Add support for the JAIS-2 family of Arabic-English bilingual models
from Inception AI (https://huggingface.co/inceptionai/Jais-2-8B-Chat).
Architecture characteristics:
- LayerNorm (not RMSNorm) with biases
- ReLU² (ReLU squared) activation function
- Separate Q/K/V projections with biases
- Simple MLP without gate projection (up -> act -> down)
- RoPE positional embeddings
- GPT-2 BPE tokenizer
Supported model sizes:
- Jais-2-8B (32 layers, 26 heads, 3328 hidden)
- Jais-2-70B (68 layers, 56 heads, 7168 hidden)
Tested with quantizations: BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, Q4_K_M, Q4_0, Q3_K_M, Q2_K
Note: JAIS-2 requires F32 precision accumulators for numerical stability
and uses standard attention (not flash attention) on CUDA backends.
* fix: run convert_hf_to_gguf_update.py for jais-2 tokenizer hash
* fix: use NEOX RoPE type for JAIS2
* fix: remove Q/K permutation (NEOX RoPE doesn't need it)
* fix: enable flash attention for JAIS2 (fixed by #19115)
* fix: add dedicated JAIS2 pre-tokenizer type and control vector support
- Add LLAMA_VOCAB_PRE_TYPE_JAIS2 with cascading whitespace regex
- Include original regex from tokenizer.json as comment
- Add build_cvec call for control vector support
* no longer necessary to override set_vocab
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* mtmd : chat : Fix extra \n between text and media marker
Thanks to @tugot17 for detecting and reporting the issue.
For vision models (e.g. LFM2.5-VL-1.6B and Qwen/Qwen3-VL-4B-Instruct) `llama-mtmd-cli` produces identical output to HF implementation.
However `llama-server` doesn't. I traced it down to extra newline
inserted after `<__media__>`.
This happens in `to_json_oaicompat`, that treats media markers as text
and joins all parts with `\n` separator.
PR introduces new type `media_marker` and uses it for media markers.
Extra logic is added to prevent insertion of newlines before and after
media markers.
With this change number of input tokens is identical to HF
implementation and as a result the output is also identical.
I explored other ways to address the issue
* remove completely `\n` between text parts in `to_json_oaicompat`
* merge text messages in server-common.cpp before sending them to `to_json_oaicompat`
Please propose alternative ways of fixing this issue.
* Refactor to use explicite per type ifs
* Update common/chat.cpp
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
* Update common_chat_templates_apply_legacy
---------
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
* model : Add tokenizer from LFM2.5-Audio-1.5B
[LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) introduced lightweight audio tokenizer.
Tokenizer based on LFM2 architecture and acts as "embedding" model with
different input `n_embd` and output `n_embd_out`.
To be used in https://github.com/ggml-org/llama.cpp/pull/18641.
To convert use
```shell
python3 convert_hf_to_gguf.py /path/to/LFM2.5-Audio-1.5B/audio_detokenizer
```
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Formatting
* Rework check for attention layers
* Add LFM2 SWA model support
* Address PR feedback
* Set vocab to none
* Move helper function definitions to cpp file
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit updates get_logits_ith(), and get_embeddings_ith() to use
output_resolve_row() to resolve the batch index to output row index.
The motivation for this is to remove some code duplication between these
functions.
* full modern bert support
* added gelu op in rank pooling for modern bert
* still working on stuff, added mean calculation before classifier head
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* first layer is dense, as per modern bert research paper
* Update src/llama-graph.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* fixed set input for mean pooling to check if pooling type is ranking since modern bert does mean & rank
* Update src/llama-graph.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Pascal
Georgi Gerganov
Tarek Dakhran
Radoslav Gerganov
Ruben Ortlam
Jeff Bolz
Max Krasnyansky
Aleksander Grygier
Adrien Gallouët
Daniel Bevenius
Alberto Cabrera Pérez
Sigbjørn Skjæret
Xuan-Son Nguyen
Aldehir Rojas
Kilian Krampf
HelloKS
Mario Limonciello
Gaurav Garg
crsawyer
Mengsheng Wu
Taimur Ahmad
ddh0
Jesse Posner
abhijitb11
Masashi Yoshimura
megemini
Saba Fallah
3 a l i
Johannes Gäßler
Ryan Mangeno
shalinib-ibm