Gaurav Garg [Wed, 10 Jun 2026 17:51:16 +0000 (23:21 +0530)]
Remove padding and multiple D2D copies for MTP (llama/24086)
* Make ggml_gated_delta_net take only the initial recurrent state (D, 1, n_seqs) and passes the snapshot count K as an op parameter instead of inferring it from state->ne[1].
Remove the padding hack and copy all emitted snapshots into the recurrent cache with a single strided ggml_cpy
* Make GDN changes in all backends. Address review comments.
Oliver Simons [Wed, 10 Jun 2026 12:27:08 +0000 (14:27 +0200)]
CUDA: Fix ssm_scan_f32 data-races (llama/24360)
* Add missing syncthreads before resuing cub_temp_storage
__syncthreads() is required before being allowed to resue TempStorage
smem:
https://nvidia.github.io/cccl/unstable/cub/api/classcub_1_1BlockLoad.html#_CPPv4I0EN3cub9BlockLoad4LoadEv20RandomAccessIteratorRA14ItemsPerThread_1Ti
* Add one more missing __syncthreads
Could also double-buffer, but alternative is to simply ensure all
threads have read smem* before writing to it again in the next loop
iteration
Pascal [Tue, 9 Jun 2026 09:01:37 +0000 (11:01 +0200)]
ggml : add GGML_OP_COL2IM_1D (llama/24206)
* cpu: add GGML_OP_COL2IM_1D
Add the overlap-add (scatter-add) step of a 1D transposed convolution.
A ConvTranspose1d factorizes as a GEMM followed by col2im: a weight
pre-permuted to [IC, K*OC] is contracted against the [IC, T_in] input
with mul_mat to produce a column matrix [K*OC, T_in], and col2im_1d
scatters those columns back into the [T_out, OC] signal, with
T_out = (T_in - 1)*s0 + K - 2*p0.
Keeping the contraction as a plain mul_mat leaves the heavy work on the
optimized (and quantizable) matmul kernels, so col2im_1d only does the
cheap overlap-add.
CPU uses a gather formulation parallelized over output channels,
supporting F32, F16 and BF16 with an F32 accumulator.
* tests: add backend coverage for GGML_OP_COL2IM_1D
Add test_col2im_1d next to the conv_transpose_1d cases, covering F32,
F16 and BF16 across eight geometries: the canonical kernel = 2*stride
DAC upsampling shape, overlap, no overlap, cropping (p0 = 1 and
p0 = stride/2), kernel < stride with zeroed gaps, kernel not a
multiple of stride, and a single column unfold.
Perf mode gets three real vocoder stage shapes reporting memory
bandwidth. max_nmse_err relaxes to 5e-4 for F16 and BF16.
* cpu: harden GGML_OP_COL2IM_1D
ggml_col2im_1d validates s0, oc, p0 and input contiguity at graph
build time, before the oc division, protecting every backend at once.
The kernel asserts the contiguity its flat indexing assumes and its
doc states the full output length including the crop term.
The kernel parallelizes over the time axis: the split stays balanced
down to OC = 1, where the previous channel split was single threaded.
Values are bit identical on the three real vocoder chains, two out of
three improve.
* tests: extend the GGML_OP_COL2IM_1D grid
The eval grid grows to eleven geometries: OC = 1 (mono output stage),
K = 1 with stride > 1 (sparse scatter, every gap position zeroed) and
a crop down to T_out = 2 where all the gather bounds act at once.
* tests: add col2im_1d equivalence test
tests/test-col2im-1d.cpp proves mul_mat + col2im_1d matches the
native ggml_conv_transpose_1d on the CPU backend, F32 bit exact, F16
and BF16 through casts of the column matrix. test-backend-ops cannot
cover this for a CPU only op since the CPU backend is its own
reference there.
* rpc: bump protocol patch version for GGML_OP_COL2IM_1D
GGML_OP_COUNT goes from 96 to 97 with the new op, which trips the
static_assert in ggml-rpc.h. Bump RPC_PROTO_PATCH_VERSION since the
op is appended and no existing op code shifts.
Jeff Bolz [Mon, 8 Jun 2026 08:40:37 +0000 (03:40 -0500)]
vulkan: Use cm2 decode_vector for mul_mat_id B matrix loads (llama/23991)
This allows vec4 loads of the B elements. Also increase BK to 64 when this is
enabled. Neither of these alone is consistently faster, but together these give
a nice speedup.
In ggml-vulkan.cpp, we need to make sure the B matrix alignment and stride are
multiples of 4.
commit 8b92060 switched ct.convert() to mlprogram, but did not update
the --quantize path. quantize_weights() from
neural_network.quantization_utils only works with the legacy
neuralnetwork format. Running with --quantize crashed with:
Exception: MLModel of type mlProgram cannot be loaded just from the
model spec object. It also needs the path to the weights file.
Fix: pass compute_precision=ct.precision.FLOAT16 into ct.convert() when
--quantize is set. This matches the original intent of nbits=16 (F16
storage) without changing the quantization scheme or model accuracy.
Also fix the three boolean CLI flags (--encoder-only, --quantize,
--optimize-ane) to use a _str_to_bool helper so that both
--flag True
and
--flag False
parse correctly. The type=bool form accepted "False" as True because
bool("False") == True.
Remove the "currently broken" label from --optimize-ane: the ANE path
(WhisperANE with Conv2d attention and LayerNormANE) converts and loads
correctly with both PyTorch 2.x and coremltools 9.x.
Mason Milburn [Fri, 5 Jun 2026 05:10:31 +0000 (01:10 -0400)]
sycl : port multi-column MMVQ from CUDA backend (llama/21845)
mmvq:
Port the ncols_dst optimization from ggml-cuda/mmvq.cu to SYCL.
Read weights once per dispatch instead of once per column.
Covers all standard quant types + reorder paths for Q4_0, Q8_0,
Q3_K, Q4_K, Q5_K, Q6_K. IQ types (except IQ4_XS) excluded due to
incompatible vec_dot signatures.
ggml-sycl:
The weight reorder was only bootstrapped on single-token mat-vec
(ne[1] == 1). Speculative / MTP verify issues only multi-column mat-vec,
so it never triggered the reorder and ran on the slower non-reorder
kernel. Bootstrap it on small multi-column batches (ne[1] <= 8) too.
Kartik Sirohi [Thu, 4 Jun 2026 13:12:38 +0000 (18:42 +0530)]
ggml: vectorize ggml_vec_dot_q4_1_q8_1 with WASM SIMD128 (llama/22209)
* ggml: vectorize ggml_vec_dot_q4_1_q8_1 with WASM SIMD128
Optimize the inner loop of ggml_vec_dot_q4_1_q8_1_generic using
WASM SIMD128 intrinsics, gated behind #ifdef __wasm_simd128__ so
non-wasm builds are completely unaffected.
Approach:
- single wasm_v128_load covers all 32 packed 4-bit weights
- nibbles unpacked via AND/SHR into two u8x16 registers
- widened to i16 before multiply (WASM SIMD has no i8*i8 instruction)
- 4x wasm_i32x4_dot_i16x8 calls accumulate all 32 element pairs
- horizontal reduce via 4x wasm_i32x4_extract_lane
Correctness verified against scalar reference across 10 random seeds
with exact output match.
* ggml: move q4_1_q8_1 WASM SIMD implementation to wasm backend
Relocate the SIMD128 implementation of ggml_vec_dot_q4_1_q8_1 to ggml/src/ggml-cpu/arch/wasm/quants.c to follow architecture-specific layout. Restore the generic implementation in ggml/src/ggml-cpu/quants.c.
Move for loop in the else block.
* ggml: use generic q4_1_q8_1 fallback in wasm backend
Andreas Kieslinger [Wed, 3 Jun 2026 11:56:42 +0000 (13:56 +0200)]
Avoid PDL race conditions by disabling __restrict__ when PDL is used (llama/24030)
* Removes __restrict__ from PDL kernel headers due to incompatibility with
PDL. Adds preprocessor directives based on arch in kernel body to add
__restrict__ to retain performance on older architectures.
* Simplifies new __restrict__ usage via macro
* Add hopper to PDL __restrict__ fix.
Co-authored-by: Oliver Simons <redacted>
---------
Max Krasnyansky [Tue, 2 Jun 2026 06:40:08 +0000 (23:40 -0700)]
hexagon: MUL_MAT, MUL_MAT_ID, FLASH_ATTN and GDN cleanup and optimizations for latest models (llama/23989)
* hex-mm: initial support for F32 * F32 -> F32 matmuls
* hex-rms-norm: fix src1 stride use in fused rms_norm_mul
* hex-ops: clear spad pointers in the ops that clober it
This fixes an odd case where fused rms-norm-mul was failing but only in qwen3.5-2B and only at searth op-bath sizes.
* hmx-mm: add support for F32 * F32 -> F32 matmul_2d on HMX
Decided to use Q4_0 * F32 -> F32 matmul for this.
Q4_0 gets dequantized and tiled into F16, and here we quantize and tile F32 into F16.
Super simple and pretty efficient.
* hmx-mm: route f16 2D matmuls through the same kernel used for all other types
* hmx-mm: re-introduce pipelined vs non-pipelined mode that we used to have but is much more generic way
This update futher improves matmul performance and at the same time removes most of the redudant logic
we had in different paths.
* hmx-fa: slighlty improved pipeline simimar to matmul updates
* hmx-mm: initial version of MAT_MUL_ID support for HMX
* hmx-mm: fixed mxfp4 handling for MUL_MAT_ID
* hex-gdn: optimize GATED_DELTA_NET
DMA prefetch/double-buff, vectorize everything with HVX, in other words -- the usual :)
* hmx-mm: missed one more case where we can use fastmod
* hexagon: update DCVS settings for a slight perf bump
* hmx-fa: use fastdiv in hmx-flash-attn
* hmx-fa: precompute slope values to avoid disrupting the inner loop
* hvx-utils/fa: new HVX helpers for powf and logf and using those to speed up FA alibi
* hex-ops: fixed a bug in fusion logic that was messing up the order of the src tensors when some srcs are empty
* hex-fa: correctly fallback to HVX if we have sinks or the dims are not quite right
Shrivas Shankar [Mon, 1 Jun 2026 12:40:28 +0000 (07:40 -0500)]
metal: template GLU kernels to support f16/f32 (llama/23882)
Drops the hardcoded f32 GLU kernels in favor of a single template. We now load/store in the native tensor type (half or float) to save memory bandwidth, but keep the actual ALU compute in float to avoid exploding math in geglu/swiglu. Also opened up the dispatch gate to allow f16 inputs.
Jeff Bolz [Mon, 1 Jun 2026 12:04:01 +0000 (07:04 -0500)]
vulkan: don't hold the device mutex while compiling pipelines (llama/23641)
* vulkan: don't hold the device mutex while compiling pipelines
We need to hold a lock while we traverse all pipelines and lazily initialize
them, but we don't need to hold it while the pipeline is being compiled. And
it doesn't need to be the same lock as the device mutex. We call load_shaders
each time a pipeline is needed, so we only need to compile that one pipeline
(and, for example, don't want to end up compiling a pipeline that another
thread should be compiling).
Matt Corallo [Mon, 1 Jun 2026 09:46:48 +0000 (09:46 +0000)]
vulkan: Block-load Q3_K/Q6_K block data and subtract on 32b ints (llama/23056)
Q2_K/Q3_K/Q6_K do much better when using MMVQ on Intel BMG even
though they're only 2-byte aligned, and Q3_K still wins on
NVIDIA as well.
mesa isn't all that great at coalescing back-to-back loads from
alternating arrays, so we force it instead. Further, we can do
subtraction directly on a full int32_t rather than an i8vec4
with bit twiddling because the high bit is always free to start.
On Intel BMG on mesa, the switch to MMVQ provides an immediate
~57% perf increase in tg128 for unsloth/Qwen3.5-9B-GGUF:Q3_K and
~78% perf increase in tg128 for unsloth/Qwen3.5-9B-GGUF:Q6_K.
The futher switch to block loads leads to a ~24% perf increase in
tg128 for unsloth/Qwen3.5-9B-GGUF:Q3_K and a ~48% perf increase in
tg128 for unsloth/Qwen3.5-9B-GGUF:Q6_K.
Finally, Xe2 wins on MMVQ even for small k, so we take the NVIDIA
override for K quants on Xe2 as well.
Neo Zhang [Mon, 1 Jun 2026 06:53:04 +0000 (14:53 +0800)]
Add more types in GET_ROWS OP (llama/23710)
* add to support Q1_0, NVFP4, IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ1_S, IQ1_M, IQ3_S, IQ4_NL, IQ4_XS, I32, MXFP4, Q2_K, Q3_K, Q5_K, and Q6_K in GET_ROWS OP
Oliver Simons [Fri, 29 May 2026 10:28:18 +0000 (12:28 +0200)]
CUDA: Check PTX version on host side to guard PDL dispatch (llama/23530)
* CUDA: Check PTX version on host side to guard PDL dispatch
Checking on `__CUDA_ARCH_LIST__` alone is insufficient for JIT, as this
variable doesn't differentiate between compiling for say sm_90, sm_90a
or sm_90f (so forward-jittable PTX vs. arch/family-specific PTX).
Thus, one can have a bug when compiling with
`DCMAKE_CUDA_ARCHITECTURES="89;90a"`, where current code would wrongly
dispatch to PDL on sm_90/sm_120 in forward-JIT mode.
This PR fixes this issue by checking `cudaFuncAttributes::ptxVersion` of
the incoming kernel at runtime. A check on ptxVersion alone is
sufficient, as device-codes will always be >= ptxVersion (and any
violation of this would be a severe bug in CUDA/nvcc), see:
https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/#gpu-code-code-code
* Implement MurmurHash3 mixer for better hash distribution
Magic constants were taken from boost:
https://github.com/boostorg/container_hash/blob/2698b43803c012601e6bb1a6116e83767b97986c/include/boost/container_hash/detail/hash_mix.hpp#L19-L65
* Update ggml/src/ggml-cuda/common.cuh
Co-authored-by: Johannes Gäßler <redacted>
* Address review comments, make seed non-zero
Daniel Bevenius [Thu, 4 Jun 2026 12:25:15 +0000 (14:25 +0200)]
ci : use emscripten-core and pin version (#3857)
This commit updates the setup emscripten sdk jobs to use emscripten-core
instead of mymindstorm and also pins the commit sha for the version
instead of using a version tag.
Georgi Gerganov [Thu, 4 Jun 2026 06:35:58 +0000 (09:35 +0300)]
ci : refactor + optimize (#3847)
* ci : add ccache clear action
* ci : split self-hosted GPU jobs into build-self-hosted.yml
Extract self-hosted runner jobs from build.yml into a dedicated
build-self-hosted.yml following the llama.cpp pattern:
- gpu-cuda (NVIDIA Linux)
- gpu-vulkan-nvidia-cm (NVIDIA Linux)
- gpu-vulkan-nvidia-cm2 (NVIDIA Linux + COOPMAT2)
- gpu-metal (macOS ARM64)
- gpu-vulkan (macOS ARM64)
GitHub-hosted CPU jobs remain in build.yml.
Assisted-by: llama.cpp:local pi
* ci : split release jobs into release.yml
Extract release-related jobs from build.yml into a dedicated
release.yml following the llama.cpp pattern:
- determine-tag
- windows (Win32/x64, SDL2)
- windows-blas (Win32/x64, OpenBLAS)
- windows-cublas (x64, CUDA 11.8/12.4)
- ios-xcode-build
- bindings-java (depends on windows)
- release (artifact aggregation + GitHub release)
CoreML job stays in build.yml with its own local tag calculation.
Assisted-by: llama.cpp:local pi
* ci : remove bindings-java job from release.yml
Assisted-by: llama.cpp:local pi
* cont : add manual trigger for build.yml
* cont : remove obsolete ifs
* ci : extract sanitizer job to bild-sanitize.yml
* ci : extract linux jobs into build-linux.yml
* ci : extract macos jobs to build-macos.yml
* ci : extract gcc jobs to build-gcc.yml
* ci : extract clang jobs to build-clang.yml
* ci : extract sycl jobs to build-sycl.yml
* ci : extract windows jobs to build-windows.yml
* ci : extract emscripten job to build-wasm.yml
* ci : extract android jobs into build-android.yml
* ci : extract quantize job to quantize.yml
* ci : extract coreml job into coreml.yml
* ci : extract vad job to vad.yml
* ci : extract cpu jobs to build-cpu.yml
* ci : make naming of yml files consistent
* ci : add --fail to curl download and propagate
This commit adds the --fail option to the model download scripts so that
if the model download returns a server error this is picked up. This is
then detected in run.sh and a error message is displayed and the script
stops and returns an error.
The motivation for this is that currently it is possible for the model
download to fail but this script proceeds and instead of a model file
the contents will be an html page probably with the error. This will
then cause the model to not be able to load due to a missing magic
number. I'm not sure we can do much about the downloading failing,
perhaps a retry but at least this will give a clearer error message.
* ci : enable command traces to see download command in use
* ci : add retry functionality to download model script
This commit adds curl retry options to the model download script.
The motivation is that currently when CI jobs run huggingface rate limit
the requests and return:
```console
curl: (22) The requested URL returned error: 429
```
This is an attempt to work around this and if it does not work then we
can an authorization token.
* ci : extract freebsd job to build-freebsd.yml
This job has been commented out as it has been flaky in the past. I'll
monitor this and if it continues to be unreliable we can disable it in
the github actions GUI instead of commenting it out like we did before.
* ci : add ccache to jobs (non-docker builds)
The ccache will only be saved on pushed to master.
* ci : bump ccache-action version to v1.2.21
The motivation for this is that the save parameter does not seem to work
with the current version.
* ci : add ccache to docker jobs in build-linux.yml
* ci : add debug statements to linux docker build
* ci : set CCACHE_DIR for build-linux.yml
* ci : add ccache to the remaining docker jobs
* ci : remove build-linux.yml
This commit remove build-linux.yml as the same jobs are also run by
build-gcc.yml, with the exception that build-gcc.yml also run ctest).
So keeping build-gcc.yml and removing the redundant build-linux.yml.
* ci : add linux build artifacts to release
* ci : revert to hendrikmuhs/ccache-action for win job
This is currently causing the following failure:
```console
sccache C:\PROGRA~1\NVIDIA~1\CUDA\v\bin\nvcc.exe -forward-unknown-to-host-compiler -DGGML_BACKEND_BUILD -DGGML_BACKEND_SHARED -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_SCHED_MAX_COPIES=4 -DGGML_SHARED -D_CRT_SECURE_NO_WARNINGS -D_XOPEN_SOURCE=600 -Dggml_cuda_EXPORTS -DCMAKE_INTDIR=\"Release\" -ID:\a\whisper.cpp\whisper.cpp\ggml\src\ggml-cuda\.. -ID:\a\whisper.cpp\whisper.cpp\ggml\src\..\include -isystem "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v\include" -Xcompiler="-MD -O2 -Ob2" -DNDEBUG -std=c++17 -arch=native -use_fast_math -extended-lambda -Xcompiler /Zc:preprocessor -MD -MT ggml\src\ggml-cuda\CMakeFiles\ggml-cuda.dir\Release\allreduce.cu.obj -MF ggml\src\ggml-cuda\CMakeFiles\ggml-cuda.dir\Release\allreduce.cu.obj.d -x cu -c D:\a\whisper.cpp\whisper.cpp\ggml\src\ggml-cuda\allreduce.cu -o ggml\src\ggml-cuda\CMakeFiles\ggml-cuda.dir\Release\allreduce.cu.obj -Xcompiler=-Fdggml\src\ggml-cuda\CMakeFiles\ggml-cuda.dir\Release\,-FS
sccache: encountered fatal error
sccache: error: Could not parse shell line
sccache: caused by: Could not parse shell line
```
danscMax [Tue, 2 Jun 2026 11:25:29 +0000 (14:25 +0300)]
whisper : catch C++ exceptions in whisper_init_with_params_no_state (#3831)
whisper_model_load() can throw instead of returning false: std::runtime_error
from this file (failed ggml context / no compatible buffer type), or
vk::SystemError / vk::OutOfDeviceMemoryError from the ggml-vulkan backend during
device/buffer allocation.
whisper_init_* are extern "C", so a C++ exception unwinding across that boundary
aborts non-C++ callers (Rust via whisper-rs, Go via cgo) -- on Windows
STATUS_STACK_BUFFER_OVERRUN (0xC0000409) -- even though the function already
returns NULL on failure. Wrap whisper_model_load() in try/catch and route any
throw into the existing NULL-return path.
Patrice Levesque [Tue, 2 Jun 2026 07:22:16 +0000 (03:22 -0400)]
cmake : do not assume /usr/lib library installation. (#3693)
Current `pkgconfig` configuration file installation path and its
contents assume libraries are installed under `/usr/lib` and this is not
always the case, for instance `/usr/lib64` is quite possible under
Gentoo Linux.
Thus use the `CMAKE_INSTALL_LIBDIR` variable instead of a hardcoded
`lib`.
Jaden_Mach [Thu, 28 May 2026 12:50:25 +0000 (08:50 -0400)]
CUDA: route batch>=4 quantized matmul to MMQ on AMD MFMA hardware (llama/23227)
* CUDA: per-quant MMVQ/MMQ batch threshold on AMD MFMA hardware
The dispatcher uses a single global threshold (MMVQ_MAX_BATCH_SIZE = 8)
to choose between mul_mat_vec_q (per-row GEMV) and mul_mat_q (MFMA-tiled
GEMM) for quantized matmul. On AMD CDNA, the optimal crossover differs
substantially by quant family because the per-row GEMV cost is dominated
by dequantisation, not the dot-product itself: K-quants pay a heavier
super-block decode and so MMQ wins sooner; legacy and IQ quants have
lean decode and stay ahead until the batch fully populates an MFMA tile.
This patch introduces ggml_cuda_should_use_mmvq(type, cc, ne11) -> bool,
mirroring the existing ggml_cuda_should_use_mmq, and gates per-quant
thresholds on amd_mfma_available(cc):
Q3_K, Q4_K, Q5_K : MMVQ <= 3 (MMQ wins from batch=4: +5% .. +76%)
Q2_K, Q6_K : MMVQ <= 5 (MMQ wins from batch=6: +8% .. +35%)
others : MMVQ <= 8 (legacy & IQ regress under MMQ; unchanged)
Non-AMD-MFMA paths (NVIDIA, RDNA, CDNA1 without MFMA) are byte-identical
to master. GGML_CUDA_FORCE_MMVQ=1 restores the original global threshold
for A/B testing.
Measured on MI250X (gfx90a, ROCm 7.2.1) with Llama-3.2-3B-Instruct,
llama-bench pp512 across all 20 supported quants, ubatch 1..8, 10 reps.
Full table in PR description.