Max Krasnyansky [Wed, 27 May 2026 17:46:11 +0000 (10:46 -0700)]
hexagon: add support for Q4_1 in MUL_MAT and MUL_MAT_ID (#23647)
* hex-mm: add support for Q4_1 matmul/matvec, hvx-only for now
* hmx-mm: add support for Q4_1
* hex-mm: use Q8_1 dynamic quantization to avoid having to compute sums in the vec_dot
* hexagon: fix repack scratch buffer overflow
* hex-mm: fix Q4_1 repack buffer sizing
* hexagon: flip the build order for mm and fa (seems to help LTO)
* hex-mm: add vec_dot 4x1s and minor HMX cleanup after adding Q4_1
* hex-mm: fix fp16 vec_dot fallback to 2x1 and another issue that could cause incorrect output
* hexagon: resurrect early-wake and add support for polling for op-batch completions
With Q4_1 ggml-hexagon now claims pretty much the entire graphs which gives the CPU more time to chilax.
This is a good thing! But it does add extra latency for the pure benchmark runs.
Early wakeup helps recover the latency a bit in the normals runs and op-batch polling is just for benchmarking.
Matt Corallo [Wed, 27 May 2026 15:19:23 +0000 (15:19 +0000)]
vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (#22887)
* vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32
Against mesa git, this shows a 4.8% performance improvement for
tg128 on Qwen3.5-9B:BF16 on Intel BMG.
Note that this breaks some tests until the last commit which fixes
OOB A reads.
* vulkan: Use aligned loads in mul_mat_vec when available
Against mesa git, this shows a 3.3% performance improvement for
tg128 on Qwen3.5-9B:BF16 on Intel BMG.
* Make explicit that `num_rows` is <= `NUM_ROWS` in mul_mat_vec
Mesa's UUB logic can't see through conditionals, limiting its
ability to understand the bounds on the `num_rows` field in the
cleanup run. Making it explicit that `num_rows` is, indeed, always
<= `NUM_ROWS` helps mesa make slightly better codegen.
Against mesa git, this currently shows a 1% performance improvement
in tg128 on Qwen3.5-9B:BF16 on Intel BMG.
* vulkan: Fix OOB A reads in MUL_MAT_VEC for odd sizes
There was a TODO to fix the OOB reads from the A matrix which we do
here.
It is within performance noise (+<0.1%) in tg128 for
Qwen3.5-9B:BF16 on Intel BMG.
zhangtao2-1 [Wed, 27 May 2026 05:08:33 +0000 (13:08 +0800)]
convert: add MiniCPM5 tokenizer support (#23384)
Add minicpm5 pre-tokenizer hash via convert_hf_to_gguf_update.py and
implement hardcoded regex handling in llama-vocab.cpp, consistent with
other BPE pre-tokenizers.
Georgi Gerganov [Tue, 26 May 2026 16:49:41 +0000 (19:49 +0300)]
ci : move [no release] check to dedicated check_release job (#23734)
* ci : move [no release] check to dedicated check_release job
Move the workflow-level \`if\` condition that skips builds when the commit
message contains \`[no release]\` into a lightweight \`check_release\` job.
All build jobs now depend on it via \`needs\` and check its output.
This ensures the skip logic is evaluated at the job level rather than at
the workflow level, which is the recommended approach for conditional jobs.
Assisted-by: llama.cpp:local pi
* cont : use `fast` runner
Max Krasnyansky [Tue, 26 May 2026 13:20:05 +0000 (06:20 -0700)]
hexagon: add support for CONCAT op (#23648)
* hexagon: add support for CONCAT with optimized concat_2d_transposed
qwen3.5 models are quite heavy on the CONCAT with large and transposed src1.
* hex-concat: use fastdiv in generic version
* hex-concat: make checks for transposed a bit more readable
* hex-concat: reoder dma ops for better pipelining
* hex-cont/cpy: optimize CPY and CONT ops
The primary change is to avoid scalar divs in the inner loops.
We were calling hvx_copy_uu(... type_size) where type_size is non a constexpr.
This causes runtime divs by that value which is normally just 4 or 2 (f32/f16).
* hex-get-rows: optimize GET_ROWS for large rows
We now use DMA for larger rows and also split them into chunks to improve perf for Qwen3.5 and other models
that do lots of GET_ROWS with huge (2MB+ rows).
Also bump the DMA queue depth now that we can take advantage of it.
* hex-concat: unroll the inner loops of concat_2d
* hex-concat: more updates to concat_2d to improve perf a bit further
* hex-cpy: fixed n_rows per thread checks in the copy ops
* hmx-fa: fix alignment issues while computing dma sizes
* hex-set-rows: add early returns for idle threads
* hvx-rope: minor optimization to replace loops with fastdiv logic
* hex-rope: replace scalar tail processing with HVX
* hex-rope: optimize rope cache init with HVX
Add hvx-utils sin/cos helpers that use an aprox method (similar to rsqrt, inverse, etc)
Use the helpers to optimize ROPE.
Jeff Bolz [Tue, 26 May 2026 04:57:56 +0000 (23:57 -0500)]
tests: test-backend-ops -j <N> to run tests in parallel (#23637)
Create a pool of N threads that grab a chunk of up to 100 tests at a time to
iterate through. The number of tests at a time decreases as fewer remain.
Each thread uses its own dev and cpu backend, and set_n_threads_fn is not
called on the cpu backend.
Fix some TSAN issues that arose:
- In init_tensor_uniform, don't use static vector of generators.
- Replace gmtime with versions that don't use a global variable.
- Mutex calls to print_test_result.
Pascal [Mon, 25 May 2026 14:05:04 +0000 (16:05 +0200)]
model: tag ffn_latent as MUL_MAT to fix buft probe (#23664)
ffn_latent_down/up are declared GGML_OP_MUL in LLM_TENSOR_INFOS but
nemotron-h feeds them through ggml_mul_mat. The loader buft probe asks
the backend about the declared op, so it tested an elementwise MUL on a
q8_0 weight. That used to return true unconditionally and the weight
stayed on GPU by luck. Once supports_op told the truth, the probe got a
no and the loader pushed the weight and its matmul to CPU, splitting the
graph. Tagging it MUL_MAT asks the real question, the math is unchanged.
Verified on Nemotron 3 Super 120B Q5_K_M: from 64.9 back to 103.22 t/s.
Jeff Bolz [Mon, 25 May 2026 07:15:46 +0000 (02:15 -0500)]
ggml : Parallelize quant LUT init (#23595)
- Use OpenMP to parallelize iq2xs_init_impl and iq3xs_init_impl.
- Move the OpenMP detection from ggml-cpu to ggml-base.
- Update OpenMP dependencies in ggml-config.cmake.in.
jacekpoplawski [Mon, 25 May 2026 05:56:18 +0000 (07:56 +0200)]
server: fix checkpoints creation (#22929)
* common : add common_chat_split_by_role
* cont : fix spans to reach end of message
* server: fix checkpoints creation
- extract message_spans from chat templates
- find the prompt token position before the latest user message
- split prompt batching at that position
- create a context checkpoint before the latest user input
- avoid periodic mid-prompt checkpoints when that position is known
- handle multimodal prompts when mapping text/template positions to server prompt tokens
- add --checkpoint-min-step to control minimum spacing between checkpoints
* cont : clean-up
* Support autoparser detection for message barriers
* server: fix message span delimiter and update docs
---------
Co-authored-by: Alde Rojas <redacted> Co-authored-by: Georgi Gerganov <redacted> Co-authored-by: Piotr Wilkin <redacted>
Aditya Singh [Sat, 23 May 2026 16:24:39 +0000 (09:24 -0700)]
requirements : bump torch to 2.11.0 (#23503)
* requirements: relax torch~=2.6.0 to torch>=2.6.0 for convert_hf_to_gguf
The ~=2.6.0 operator resolves to >=2.6.0, <2.7.0, which fails on
PyPI for platform/CPython combinations where 2.6.x is not present.
The accompanying comment already says 'PyTorch 2.6.0 or later', so
the looser >=2.6.0 matches the documented intent and unblocks
pip install -r requirements/requirements-convert_hf_to_gguf.txt.
Fixes #23408
* requirements: bump torch floor to 2.11.0 per maintainer
* requirements: pin torch to ==2.11.0 per project policy
* requirements: pin mtmd torch and torchvision to 2.11.0/0.26.0 per project policy
* requirements: suppress check_requirements pin warning on mtmd
The check_requirements script flags '==' on lines in files matched by
*/**/requirements*.txt. Append the documented suppression comment to the
pinned torch and torchvision lines (and to the s390x platform marker lines)
so the check passes while keeping the pins required by project policy.
* ty: silence Tensor/Module union check on model[0].auto_model
With torch 2.11.0 stubs, nn.Sequential.__getitem__ now returns
Tensor | Module rather than Module, so model[0].auto_model fails ty
on the SentenceTransformer code path. The runtime behavior is
unchanged because SentenceTransformer always wraps a Module at
index 0. Adding a targeted unresolved-attribute ignore keeps the
type-check green without altering behavior. A follow-up issue
tracks typing the variable explicitly.
Alexey Kopytko [Fri, 22 May 2026 12:50:17 +0000 (21:50 +0900)]
SYCL: improve MoE prefill throughput (#23142)
- change `k_copy_src1_to_contiguous` so that uses a precomputed contiguous mapping where all rows "owned" by an expert are in one slice with a know starts and ends
- switch the `O(n_as * n_routed_rows)` contraption to a counting sort-based procedure with `O(n_as + n_routed_rows)` complexity
Katostrofik [Fri, 22 May 2026 12:48:24 +0000 (08:48 -0400)]
SYCL: add BF16 to DMMV kernel path (~4x tg speedup on Intel Arc) (#21580)
* SYCL: add BF16 to DMMV kernel path for ~4x token generation speedup
BF16 models had no dedicated token generation kernel — they fell through
to the generic full-GEMM path, resulting in ~14% memory bandwidth
utilization on Intel Arc GPUs. This adds BF16 support to the DMMV
(dequantize mul-mat-vec) path, matching the existing F16 implementation.
The qk=1 kernel (used for F16 and BF16) iterates with stride
2*GGML_SYCL_DMMV_X (= 64 on Intel targets where WARP_SIZE=16). When
ncols is a multiple of DMMV_X (32) but not of 2*DMMV_X (64), the last
warp iteration accesses elements at col >= ncols, producing NaN for the
final row and wrong values for interior rows.
Fix: tighten can_use_dequantize_mul_mat_vec to require ne[0] %
(2*DMMV_X) == 0 for F16/BF16 types, and update the ASSERT in the BF16
launcher to match. Quantized types use block-structured kernels with
different access patterns and keep the existing DMMV_X check.
Verified: test-backend-ops MUL_MAT passes 913/913 on Intel Arc Pro B70.
Previously failing: m=128/129 n=1 k=1056 cases (NaN and ERR > 0.0005).
Co-Authored-By: Claude Sonnet 4.6 <redacted>
---------
Georgi Gerganov [Fri, 22 May 2026 08:46:26 +0000 (11:46 +0300)]
cmake : add install() for impl libraries + fix apple builds (#23511)
* pi : update
* ci : fix ios build
* ci : fix andoroid
* ci : fix apple builds
* cmake : add install() for impl libraries
Add install(TARGETS <target> LIBRARY) for all -impl libraries that were
changed from STATIC to shared (controlled by BUILD_SHARED_LIBS) in
commit bb28c1fe2. Without this, cmake --install fails to copy the shared
libraries, causing runtime errors like:
llama-server: error while loading shared libraries: libllama-server-impl.so
Pascal [Thu, 21 May 2026 17:39:42 +0000 (19:39 +0200)]
vulkan: fuse snake activation (mul, sin, sqr, mul, add) (#22855)
* vulkan: fuse snake activation (mul, sin, sqr, mul, add)
Add snake.comp shader with F32 / F16 / BF16 pipelines and
ggml_vk_snake_dispatch_fused. The matcher recognizes the naive 5 op
decomposition emitted by audio decoders (BigVGAN, Vocos) for snake
activation y = x + sin(a*x)^2 * inv_b and rewrites it to a single
elementwise kernel.
test_snake_fuse from the CUDA PR now also compares CPU naive vs
Vulkan fused across F32 / F16 / BF16.
* vulkan: address jeffbolznv review for fused snake activation
Rename T / C to ne0 / ne1 in the shader and push constants to match
the standard naming convention used across the Vulkan backend.
Tighten ggml_vk_can_fuse_snake: require x and dst to be contiguous
(the shader uses idx = i0 + i1 * ne0) and require a / inv_b to be
tightly packed on the broadcast dim (the shader reads data_a[i1]).
* vulkan: tighten snake fusion type checks for all operands (address jeffbolznv review)
* vulkan: reject snake fusion when ne[2] or ne[3] > 1 (address jeffbolznv review)
* vulkan: address 0cc4m review for fused snake activation
snake.comp is renamed to follow the ggml DATA_A_* / A_TYPE convention.
A_TYPE now applies to the activation tensor data_a instead of the
broadcast multiplier, and the bindings become data_a (A_TYPE), data_b
(float), data_c (float) and data_d (D_TYPE). A header at the top of
the shader maps each buffer to its role in y = x + sin(b * x)^2 * c.
On the C++ side, ggml_vk_can_fuse_snake reuses the existing snake_pattern
constant instead of duplicating the op list, sin_node is extracted as a
named local alongside the other chain nodes, and the broadcast operands
a and inv_b are now required to be GGML_TYPE_F32 to match the hardcoded
float bindings on data_b and data_c (the previous a->type == x->type
would silently reject any future BF16 or F16 chain once the supports_op
gate for SIN / SQR is lifted). ggml_vk_snake_dispatch_fused gets an
explicit GGML_TYPE_F32 case and GGML_ABORT on default in place of the
silent f32 fallback, and a stale comment about data_a[i1] / data_inv_b[i1]
is refreshed to match the new binding names.
Georgi Gerganov [Thu, 21 May 2026 11:41:50 +0000 (14:41 +0300)]
tests : move save-load-state from examples to tests (#23336)
* tests : move save-load-state from examples to tests
- Move examples/save-load-state/ to tests/test-save-load-state.cpp
- Remove subdirectory reference from examples/CMakeLists.txt
- Add test to tests/CMakeLists.txt as a model test
- Remove CODEOWNERS entry for removed example directory
Assisted-by: llama.cpp:local pi
* cont : update ci
overview / pkg / ggml / sources / llama.cpp / log