common : re-arm reasoning budget after DONE on new <think> (#22323)
DONE state absorbs all tokens including a new start tag, causing any think blocks after the first to run unbudgeted. Observed on unsloth/Qwen3.6-27B-GGUF which interleaves multiple <think> blocks per response.
Fixed by advancing start_matcher in DONE branch and re-arming to COUNTING with a fresh budget on match. Adds regression test (test-reasoning-budget: test 6).
Matt Corallo [Tue, 28 Apr 2026 15:31:04 +0000 (15:31 +0000)]
vulkan: Coalesce Q4_K/Q5_K scale loads (#21751)
Some SPIR-V compilers (notably mesa) don't handle the current
vulkan Q4_K/Q5_K scale load pattern in mul_mat particularly well.
While reading three `u8`s from the 12-byte scale array should (at
least on some hardware) result in loading the full 12 bytes in a
single LOAD followed by whatever extraction is needed, at least
the ANV Intel driver really can't practically perform this
optimization.
`mesa`'s unsigned upper bound logic doesn't handle tracking bounds
through ternary, resulting in the `(is < 4) ? ... : is - 4` having
an infinite upper bound (as it cannot prove `is - 4` doesn't
underflow). While this could still be rectified if mesa looked at
the array bounds, it currently doesn't and `glslc` currently emits
SPIR-V that doesn't allow for this optimization anyway (though
maybe it will at some point, see
https://github.com/KhronosGroup/glslang/issues/4206).
In mul_mat_vecq we took a different approach to loading the same
fields. We read the first two bytes we needed from `scale` then
took a branch before deciding whether we needed to read a third
byte. In mesa this did, indeed, lead to a top-level branch with
conditional loads. As such these loads ended up not being
coalesced either (at least in the ANV driver) resulting in
additional instructions in our hot loop.
Instead, here, we go ahead and force loading the full 12 bytes and
extract the bits we need from the packed-u32s instead. In mul_mat
there's a few less ternaries and only one extra shift, so even on
drivers that did optimize the previous loads properly the only
material change should be pulling a few extra bytes into registers
(which on most hardware won't cost anything anyway, though
ironically on Intel it theoretically could). In mul_mat_vecq this
requires a bit of extra math and may read bytes from the u32 that
weren't needed, but it seems likely avoiding the branch is a win
on most platforms.
On Intel Xe2/mesa 26.0.4 with the optimizations from
https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15162,
for shader matmul_id_subgroup_q4_k_f32_f16acc_aligned_l:
* Instruction Count: 2753 -> 2688
* SEND Count: 269 -> 261
* Cycle Count: 273976 -> 266138
* Max live registers: 248 -> 246
* Non SSA regs after NIR: 381 -> 382
for shader matmul_id_subgroup_q5_k_f32_f16acc_aligned_l:
* Instruction Count: 2767 -> 2702
* SEND Count: 271 -> 263
* Cycle Count: 274140 -> 268144
* Max live registers: 248 -> 246
* Non SSA regs after NIR: 381 -> 382
for shader mul_mat_vec_id_q4_k_q8_1_f32:
* Instruction Count: 1930 -> 1646
* SEND Count: 116 -> 71
* Cycle Count: 1348306 -> 843350
* Max live registers: 78 -> 84
* Non SSA regs after NIR: 300 -> 135
for shader mul_mat_vec_id_q5_k_q8_1_f32:
* Instruction Count: 2207 -> 1922
* SEND Count: 131 -> 86
* Cycle Count: 1392012 -> 1037836
* Max live registers: 90 -> 90
* Non SSA regs after NIR: 300 -> 135
for shader mul_mat_vec_q4_k_q8_1_f32:
* Instruction Count: 2029 -> 1749
* SEND Count: 111 -> 66
* Cycle Count: 1347278 -> 840118
* Max live registers: 74 -> 80
* Non SSA regs after NIR: 299 -> 134
for shader mul_mat_vec_q5_k_q8_1_f32:
* Instruction Count: 2307 -> 2022
* SEND Count: 126 -> 81
* Cycle Count: 1379820 -> 954042
* Max live registers: 86 -> 86
* Non SSA regs after NIR: 299 -> 134
ggml : revert to -lm linking instead of find_library (#22355)
* ggml : revert to -lm linking instead of find_library
`find_library(MATH_LIBRARY m)` was introduced recently, but it breaks
CUDA compilation with GGML_STATIC. I could not find any valid use case
where we would prefer `find_library` over the standard `-lm` approach.
This commit is also meant to start a discussion if there is a valid
reason to keep `find_library(MATH_LIBRARY m)`, we should clarify what
problem it was solving and find an alternative fix that does not break
CUDA with GGML_STATIC.
Signed-off-by: Adrien Gallouët <redacted>
* ggml : use MATH_LIBRARY only if defined
New operators:
- GGML_OP_SET: implement via aclnnInplaceCopy on target region
- GGML_OP_CUMSUM: implement via aclnnCumsum
- GGML_OP_FILL: implement via aclnnInplaceFillScalar
- GGML_OP_DIAG: implement via aclnnInplaceCopy on diagonal strides
- GGML_OP_TRI (lower/lower_diag/upper_diag/upper): implement via
aclnnTril(-1/0) and aclnnTriu(0/1) with appropriate diagonal offsets
- GGML_OP_SOLVE_TRI: implement via aclnnTriangularSolve
- GGML_UNARY_OP_SOFTPLUS: implement via aclnnSoftplus
Optimizations:
- GLU (SwiGLU/GeGLU/GeGLU_ERF/GeGLU_QUICK): fuse with aclnnSwiGlu /
aclnnGeGluV3 when applicable; fallback conditions now checked inside
each function rather than at the call site
- CROSS_ENTROPY_LOSS: replace 5-kernel sequence (LogSoftmax→Mul→
ReduceSum×2→Muls) with single aclnnSoftmaxCrossEntropyWithLogits call
- L2_NORM: fix in-place ClampMin on norm result (was clamping wrong
tensor); add eps clamping before division to avoid divide-by-zero
- PAD_REFLECT_1D: eliminate per-ne[3] loop; assert contiguity and call
ReflectionPad1d once on the full 4-D view; remove redundant nb copies
- GET_ROWS: replace IndexSelect with GatherV2 per batch slice; refactor
helper into gather_batched lambda with batch loop inlined
- SET_ROWS: replace IndexCopy with InplaceIndexCopy per batch slice;
refactor helper into scatter_batched lambda with batch loop inlined
- OUT_PROD: replace O(ne[3]*ne[2]*ne[1]) Ger+InplaceAdd loop with
per-slice Matmul loop (src0 @ src1^T); handles strided-broadcast
batch dims where ne02/ne03 may differ from ne2/ne3
- backend memset_tensor: implement via aclrtMemset (was NULL)
Bug fixes:
- COUNT_EQUAL: use non-inplace EqTensor into a same-type temporary
buffer instead of InplaceEqTensor, avoiding corruption of src0
- ACL graph cache (USE_ACL_GRAPH): restore node_type and src_type[]
fields in ggml_graph_node_properties; has_matching_properties() was
missing type checks, causing F16 and BF16 tensors (same nb[0]=2) to
incorrectly share cached graphs and produce wrong results (ERR≈679)
- graph cache op_params matching: compare full GGML_MAX_OP_PARAMS
bytes so that ops differing only in parameters are not incorrectly
replayed from cache
server: (router) Forward form-data to model server (Fixes #22044) (#22118)
* This commit enables the router to forward form-data to model server.
Fixes #22044 (enabling to use the /v1/audio/transcriptions in router mode)
* * Applied the suggestion from Copilots first comment: using the non-throwing json::parse overload.
* Addressed Copilots third comment by extending the files representation to also include filename and content-type
* Addressed Copilots fourth comment by making the RNG thread_local
* Changed variable body from std::string to std::ostringstream in build_multipart_body
as suggested by ngxson in https://github.com/ggml-org/llama.cpp/pull/22118#discussion_r3127099053
* Added sanitize_field lambda in build_multipart_body for key, filename and content_type
as suggested by ngxson in https://github.com/ggml-org/llama.cpp/pull/22118#discussion_r3127104647
* explicitly checking if value/item is string before calling value/item.get<std::string>()
as requested by ngxson in https://github.com/ggml-org/llama.cpp/pull/22118#discussion_r3127111279
* Added double quote to the sanitize lambda and throw on json parse failure
Max Krasnyansky [Mon, 27 Apr 2026 05:06:39 +0000 (22:06 -0700)]
common: fix missing exports in llama-common (#22340)
* common: refactor common/debug to move abort_on_nan into base_callback_data
Passing bool abort_on_nan as template parameter for common_debug_cb_eval is unnecessary and creates an issue with LTO.
It should just be a member of the base_callback_data instead.
* cont : cleanup
* common : use pimpl in debug.h to reduce header dependencies
Move common_debug_cb_user_data's data members (std::regex,
std::vector<uint8_t>) into a private impl struct in debug.cpp.
This removes the includes of common.h and <regex> from debug.h,
reducing transitive dependencies for any translation unit that
includes the header.
Fix recurrent state serialization for partial reads and writes (#22362)
The previous code worked only for full tensor reads and writes and was hitting `GGML_ASSERT(size == ggml_nbytes(tensor)); ` assert when tested with llama-server.
metal : optimize Metal Tensor API usage for GGML_OP_MUL_MAT (#20962)
* Optimize Metal Tensor API usage for matmul2d
Separates the Metal Tensor API (matmul2d) path in kernel_mul_mm into its own standalone kernel, gated by GGML_METAL_HAS_TENSOR.
The legacy simdgroup_matrix kernel is preserved under #else.
Previously both paths were interleaved via #ifdef blocks within a single kernel, forcing the tensor path to share the legacy kernel's data layout and threadgroup memory scheme. Splitting the kernel enabled memory and dispatch optimizations that weren't possible when the two paths shared code structure.
Change the default `ftype` in `llama_model_quantize_params` from
`LLAMA_FTYPE_MOSTLY_Q5_1` to `LLAMA_FTYPE_MOSTLY_Q8_0`.
In case some external program naively uses the default quantization
params, we should probably default to a known-good type like Q8_0 rather
than Q5_1, which is rather old.
cli: Remove redundant local sampling variables (#20429) (#22264)
This change implements the third requested change in issue 20429.
Because defaults.sampling contains the reasoning budget token count and
the reasoning budget message, it's not necessary to assign them to
struct variables.
Song Li [Thu, 23 Apr 2026 16:39:07 +0000 (12:39 -0400)]
server: fix heap-buffer-overflow from negative n_discard (CVE-2026-21869) (#22267)
* server: clamp n_discard to non-negative at JSON parse boundary (CVE-2026-21869)
A negative n_discard from client JSON causes heap-buffer-overflow in
update_slots() context-shift loop (CWE-787, CVSS 8.8). Clamp to 0 at
ingress; n_discard=0 already triggers auto-discard (n_left/2).
When testing claude code against llama.cpp, I noticed that only
n_past 18577 was used even when context was 60k or more. The log
in llama-server says:
```
slot update_slots: id 3 | task 10342 | old: ... ; cch= | defa0;You are
slot update_slots: id 3 | task 10342 | new: ... ; cch= | 1c8b4;
```
I observed that the cch value changed every time. Reading about that,
the x-anthropic-billing-header system message seems to be specially
handled inside of the anthropic api. I could remove it, but there
is a meaningful string sometimes included at the end. So instead,
I just replace the changing cch checksum with fffff.
I'm treating this as an anthropic message body API detail - I think this
is the right way to do this, but by all means please correct me!
It's always 5 hexadecimal characters, but I've written the replacement
defensively in case they change the protocol.
Daniel Bevenius [Thu, 23 Apr 2026 13:07:38 +0000 (15:07 +0200)]
model-conversion : fix mmproj output file name [no ci] (#22274)
* model-conversion : fix mmproj output file name [no ci]
This commit updates the convert-model.sh script to properly handle
mmproj output files.
The motivation for this that currently the same name as the original
model is used as the mmproj file, which causes the original model to
be overwritten and no mmproj-<model_name>.gguf to be created.
Create an MMVQ kernel so ggml_sycl_mul_mat_id can consolidate
n_experts_used matmuls in a single kernel launch. The kernel
also reads expert IDs directly, removing a per-call host sync.
This is similar to the CUDA backend's ggml_cuda_mul_mat_vec_q*
paths.
All types supported in the current MMVQ are supported here as well:
Q2_K, Q3_K, Q4_K, Q5_K, Q6_K, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0
It will fall back to the existing per-expert path when src0 has been rewritten
by opt_for_reorder(), and for any shape the fused path doesn't handle.
test-backend-ops passes for supported type/shape combos.
In #11362 hip graph was disabled by default as, at the time, its performance impact was negative. Due to improvements in rocm and our usage and construction of graphs this is no longer true, so lets change the default.
sycl: Improve mul_mat_id memory efficiency and add BF16 fast path (#22119)
* sycl: size mul_mat_id staging buffers by routed rows
Previously src1_contiguous/dst_contiguous in ggml_sycl_mul_mat_id were
sized to ggml_nelements(src1/dst), which over-allocates when ne12 > 1
and can fail with UR_RESULT_ERROR_OUT_OF_HOST_MEMORY on Level Zero for
MoE models (notably with --cpu-moe). Size them by the actual number of
routed rows (ids->ne[1] * n_ids) instead.
* sycl: add bf16 mul_mat fast path via DNNL
When src0 is BF16 (commonly the case for lm_head / output.weight), the
existing f16 path is skipped because bf16 isn't covered, and the f32
fallback dequantizes the entire src0 slab to f32 in a single pool alloc
(row_diff*ne00 floats). For large-vocab models this can reach several
GB and fail with UR_RESULT_ERROR_OUT_OF_HOST_MEMORY on Level Zero.
Add a bf16xbf16 -> f32 DNNL matmul fast path that uses the bf16 storage
in place and only materializes a small src1 bf16 conversion buffer. bf16
matmul accumulates in f32, so it's correct even when the op requests
GGML_PREC_F32 (as lm_head does).
- gemm.hpp: map bfloat16 to dnnl::memory::data_type::bf16.
- convert.{hpp,cpp}: expose ggml_get_to_bf16_sycl for f32/f16/bf16 -> bf16.
- ggml-sycl.cpp: take the bf16 path early in ggml_sycl_op_mul_mat_sycl
when DNNL and GGML_SYCL_HAS_BF16 are both available.
mtmd, llama : Update HunyuanVL vision-language model support (#22037)
* mtmd, llama : add HunyuanVL vision-language model support
- add LLM_ARCH_HUNYUAN_VL with M-RoPE (XD-RoPE) support
- add PROJECTOR_TYPE_HUNYUANVL with PatchMerger vision encoder
- add HunyuanVL-specific M-RoPE position encoding for image tokens
- add GGUF conversion for HunyuanVL vision and text models
- add smoke test in tools/mtmd/tests.sh
* fix: fix HunyuanVL XD-RoPE h/w section order
* fix: Remove redundant code
* convert : fix HunyuanOCR / HunyuanVL conversion
- Tested locally: both HunyuanOCR and HunyuanVL-4B convert to GGUF
- successfully and produce correct inference output on Metal (F16 / Q8_0).
* clip : fix -Werror=misleading-indentation in bilinear resize
* fix CI: convert_hf_to_gguf type check error
- convert_hf_to_gguf.py: give HunyuanVLTextModel.__init__ an explicit `dir_model: Path` parameter so ty can infer the type for load_hparams instead of reporting `Unknown | None`.
This change refactors the reasoning_budget_message parameter from the
common params into the sampling parameters specifically. It also removes
the reasoning_budget common parameter and standardizes on the existing
reasoning_budget_tokens parameter in the sampling configuration.
Issue: https://github.com/ggml-org/llama.cpp/issues/20429
Original PR: https://github.com/ggml-org/llama.cpp/pull/20297
Kwa Jie Hao [Tue, 21 Apr 2026 18:02:49 +0000 (02:02 +0800)]
mtmd: Add support for Reka Edge 2603 (#21616)
* feat: (vocab) fix stray text appended in llama_decode_text
Remove accidental concatenation of the full `text` string when
formatting UNK_BYTE hex escapes. Only the closing "]" should be appended.
* feat(mtmd): add Yasa2 vision encoder support
Add a Yasa2 (ConvNeXtV2-based) vision encoder for reka-edge:
- Register PROJECTOR_TYPE_YASA2 and tensor name definitions
- Add yasa2_block/yasa2_stage model structs
- Implement graph builder with ConvNeXt stages, GRN, adaptive pooling
- Wire into clip.cpp switch statements and mtmd.cpp init_vision
- Use mtmd_image_preprocessor_fixed_size for image preprocessing
- Add chat-reka.cpp/h implementing PEG-based parser for reka-edge format
- Add Reka-Edge.jinja chat template
- Detect reka-edge template in try_specialized_template()
- Add LLAMA_EXAMPLE_MTMD to chat-template-file arg
* feat: add reka vlm to gguf conversion script
Converts Reka Yasa2 hf checkpoints to GGUF format:
- Text decoder: Llama-arch with tiktoken/BPE vocab
- Mmproj (--mmproj): ConvNeXt vision backbone + language_projection
- Generates 2D sincos positional embeddings for vision encoder
* test: add Reka Edge chat template and parser tests
- test-chat-template: oracle tests comparing Jinja engine output vs
common_chat_templates_apply for text, tools, thinking, images, video
- test-chat: PEG parser tests for Reka Edge format, round-trip tests
for image/video content parts, common path integration tests
* scripts: add Reka Edge mixed quantization helper
Q4_0 base quantization with Q8_0 override for the last 8 transformer
blocks (layers 24-31) via --tensor-type regex.
* fix: adapt chat-reka and tests to upstream API
- Use autoparser::generation_params (not templates_params)
- Add p.prefix(generation_prompt) to PEG parser
- Simplify reasoning parser to match LFM2 pattern
- Remove image/video oracle tests (unsupported by oaicompat parser;
no other multimodal models test this path)
* fix: avoid duplicate tensor loading in yasa2 vision encoder
TN_YASA_PATCH_W and TN_PATCH_EMBD both resolve to "v.patch_embd.weight",
causing the same tensor to be loaded twice into ctx_data and overflowing
the memory pool. Reuse the tensors already loaded by the common section.
* chore: update image pre-processing settings
The reka-edge model depends on the following settings in an older
fork of llama.cpp:
1. Fixed square resize
2. BICUBIC
3. add_padding=false
In current llama.cpp, this means setting:
- image_resize_algo = RESIZE_ALGO_BICUBIC
- image_resize_pad = false
* chore: remove reka gguf conversion script
* chore: remove reka quantization script
* chore: remove unnecessary changes from PR scope
This commit removes a couple of unnecessary changes for the PR scope:
1. BPE decoder bug fix - this affects reka edge because there's a bug
in our tokenization that doesn't represent <think> tokens as special
tokens. However this isn't meant to be a thinking model so when run
with --reasoning off the edge case does not affect us
2. --chat-template-file support from llama-mtmd-cli - the focus is on
llama-server and the reka edge gguf contains the necessary metadata
to detect the chat template
3. reka edge oracle test cases - no other model has similar test cases,
so I removed it for standardization
* chore: remove unnecessary ggml_cast
This commit removes unnecessary ggml_cast after updating the
reka vlm -> gguf conversion script on hugging face.
* chore: remove redundant code
* chore: remove unnecessary ggml_cont calls
This commit removes all ggml_cont calls except the four that
precede ggml_reshape_3d/ggml_reshape_4d. Those are necessary
because ggml_reshape recomputes strides assuming contiguous
layout and asserts ggml_is_contiguous.
Other operations (ggml_mean, ggml_add, ggml_mul etc.) use
stride-based indexing and handle non-contiguous inputs
correctly and so we are ok to remove ggml_cont for those.
* chore: remove unnecessary ggml_repeat calls
This commit removes unnecessary ggml_repeat calls because the underlying
ops already broadcast automatically.
Every ggml_repeat in yasa2.cpp was expanding a smaller tensor to match
a larger one's shape before passing both to an elementwise op (ggml_add,
ggml_sub, ggml_mul, or ggml_div). This is unnecessary because all four
of these ops already support broadcasting internally.
* chore: restore ggml_cont needed for cpu operations
* refactor: locate reka chat template handler in chat.cpp
overview / pkg / ggml / sources / llama.cpp / log