* metal : add f16 and bf16 support for concat operator
Extend the Metal backend concat operator to support f16 and bf16 tensor
types in addition to the existing f32 and i32 support.
- Template kernel_concat on type T with specializations for float, half,
bfloat, and int
- Add type-specific pipeline getter ggml_metal_library_get_pipeline_concat()
- Update device support check to allow f16 unconditionally and bf16 when
device supports bfloat16
- Update dispatch to select the correct kernel specialization by type
Assisted-by: pi:llama.cpp/Qwen3.6-27B
* metal : extend concat operator to support f16, bf16, i8, i16 and i64
Assisted-by: pi:llama.cpp/Qwen3.6-27B
* add dev2dev memcpy by SYCL API
* mv GGML_SYCL_DEV2DEV_MEMCPY to runntime table
* update the detect method for p2p comm
* fix the erro created during fix confilct
---------
Co-authored-by: Neo Zhang <NA>
* Add interface is_model_splitted() to check the c-graph is splited or not
* Infer and propagate dynamic-dimension indices for all tensors in the GGML graph in api compute_model_outputs()
* Only do this for fallback sub graph
* Move dynamic dims compute in graph missmatch
* ggml-openvino: fix tensor data handling for PERMUTE/VIEW ops in split models
* ggml-openvino:add comments
* ggml-openvino: override VIEW op_case to 0 for split model inputs
* openvino backend: Handle unsupported VIEW shape-mismatch in OpenVINO backend
* Enable additional mul_mat tests and add tensor data saving function (llama/81)
* ggml-openvino: fix CONT/TRANSPOSE mapping and improve dynamic-dimension handling
* OpenVINO: add NORM/TANH support and rework SOFT_MAX translation
* ggml-openvino: extend VIEW handling
* Enable -fa off (llama/118)
* Enable --context-shift
* Fix llm param compute error for normal softmax not the softmax in attention
* OpenVINO backend: fix error for attention size compute in llm param
* use tensor->extra in infer_request i/o
* OpenVINO backend: refacter the compute_llm_params() func add get_attention_pattern_case to easy extand
* OpenVINO backend: clean unused code
* 1to1 match op update (llama/146)
* added translate_1to1_match_1_input function and updated gelu and tanh translations
* Remove unused translation function calls
---------
Co-authored-by: Mustafa Cavus <mustafacavus@intel.com>
* initial gemma4 support
* removed hardcoded names for kv cache slicing
* OpenVINO backend: Add new attention pattern for llm parameters compute
* flash attn Q shape static conversion
* Remove slice in permute translation when n_seq is 1
* return optional in extract_layer_from_name
* OpenVINO backend: refactor VIEW related operation (llama/148)
* OpenVINO backend: refactor VIEW related operation
* Enable VIEW handling in following ops
* OpenVINO backend does not support GGML_OP_NORM & GGML_OP_L2_NORM with VIEW input accuracy issue from OpenVINO
* OpenVINO backend: Add ops l2_norm & pad
* OpenVINO backend does not support CPY with non-contiguous data or mismatched types
* add op SSM_CONV GATED_DELTA_NET
* OpenVINO backend: fix error for bf16 in OV gpu plugin
* reverted static Q input shape for attention layer
* OpenVINO backend: remove hardcode name inp_tokens, which ignore some leaf case
* Disable remote tensor due to bug in ov gpu
* Disable n_token > 1 GATED_DELTA_NET on gpu
* OpenVINO backend: fix the view op dynamic handling issue in gemma4 & enable view + get_row
* OpenVINO backend: clean code
* OpenVINO backend: enable view + norm/rms_norm
* OpenVINO backend: concat op
* OpenVINO backend: argsort op
* OpenVINO backend: enable unary + view & GGML_UNARY_OP_SOFTPLUS
* Fix issue for test-backend-ops in TOPK_MOE, which compare VIEW ops result, VIEW node in OpenVINO no need compare, the whole graph result is correct
* OpenVINO backend: enable sum_rows
* OpenVINO backend: enable clamp
* OpenVINO backend: enable DIV
* OpenVINO backend: enable GGML_OP_MUL_MAT_ID
* OpenVINO backend: disable MUL_MAT_ID_FUSION case with large mem needed
* OpenVINO backend: Disable GGML_OP_ARGSORT, cause test_backend-ops failed
* OpenVINO backend: fix issue in mul_mat_id
* OpenVINO backend: Disable DIV with broadcast on GPU
* OpenVINO backend: update DIV
* use ov internal op GatedDeltaNet
* OpenVINO backend: enable llama erch test qwen3next
* OpenVINO backend: enable RMS_NORM + VIEW & remove op_case 2 for rope
* OpenVINO backend: fix error
* suggested changes, need review
* suggested changes, need review
* OpenVINO backend: clean unused code & fix build warning
* OpenVINO backend: enable minicpm3 for arch test
* Disable GDN op (llama/177)
* disable gated_delta_net
* update stateful_kv_size correctly in mismatch case
* OpenVINO backend: enable arch test for qwen3vl
* OpenVINO backend: enable cohere2 for arch test
* OpenVINO backend: enable t5 for arch test
* OpenVINO backend: enable jamba for arch test
* OpenVINO backend: remove warning for tmp
* OpenVINO backend: enable kimi-linear for arch test
* Remove unused
* Fix gpt-oss accuracy issue
* OpenVINO backend: enable arctic for arch test
* OpenVINO backend: enable grok for arch test
* Gemma4 initial npu support (llama/179)
* Initiall gemma4 npu support
* temp. fix for gemma4 accuracy bug on npu
* Remove hardcoded names for npu-fold handling
* revert static n tokens for cont translation as it is not needed
* removed unused variable
* ggml-openvino: add GGML_OPENVINO_ENABLE_CACHE env var to control decoder cache. Add environment variable GGML_OPENVINO_ENABLE_CACHE (default: YES). When set to NO, the decoder_cache is bypassed and models are rebuilt from the cgraph on every inference call in both dynamic and static compute paths. This is useful for debugging and verifying correctness without caching interference.
* Revert "Gemma4 initial npu support (#179)"
This reverts commit 0d29a9c4a52dc2c8aa52990f1a3854cfb01768ad.
* OpenVINO backend: disable debug log print
* Update TBB discovery. Delegated to OpenVINOs own config.
* OpenVINO backend: GGML_OPENVINO_ENABLE_CACHE YES -> 1
* OpenVINO backend: fallback FLASH_ATTN_EXT in gemma3n to CPU backend
* Add raw ov infer profiling metric
* Add OV raw infer time metric to static compute path
Co-authored-by: virajwad <84867530+virajwad@users.noreply.github.com>
* Modify precision of static profiling
* update to OV 2026.2, add OV windows CI
* fix editorconfig-checks
* Initiall gemma4 npu support
* temp. fix for gemma4 accuracy bug on npu
* Remove hardcoded names for npu-fold handling
* revert static n tokens for cont translation as it is not needed
* removed unused variable
* test-llama-archs fix
* Fix gemma4 flash_attn fallback
* support im2col
* fix code style
* disable add_rope_sin_cos optimization
* stateless boradcast and rope optimizations
* Enable manual gqa attn by default for stateless gpu
* manual gqa: fixed static batch
* gemma4 llama-bench ctx update fix
* Update OV win CI
* stateful rope fusion temp. fix
* OpenVINO backend: Conslolidate supported ops
* Exclude unsupported GGML_OP_SUB cases
* Exclude unsupported TOPK_MOE cases
* OpenVINO Backend: MUL_MAT enhancements
* Update OV CI
* support f16 mask input for npu
* Make GGML_OPENVINO_* env vars usage uniform
Standardize all GGML_OPENVINO_* env flags:
positive integers >0 to enable. Unset, empty, =0, or non-numeric values to disable.
This fixes cases where text values or empty strings enabled features.
* OpenVINO backend: Enhance envvar handling
* more cleanup
* move ggml_openvino_env_flag to appropriate place
* OpenVINO backend: add REPEAT translator, Q5_1 weights, and GLU view-input fix
* ggml-openvino: fix -Werror=cast-qual in extract_q5_1_data
* Update openvino.Dockerfile
Use BuildKit cache mounts for faster Docker rebuilds.
Use apt instead of dpkg, remove unused .ddeb downloads, add DLLAMA_BUILD_TESTS=OFF.
* ggml-openvino: centralize env var access via *getenv_str/getenv_int helpers
Replace getenv and legacy flags with _str and _int helpers.Minor cleanup, doc updates.
* OpenVINO backend: Enable GGML_OP_ADD_ID
* Uptade openvino backend clamg-format
* clang-format
* Update OPENVINO.md (llama/211)
* OpenVINO backend: fix accuracy issue for op CONCAT with i64 precision
* Remove strict concurrency for gpu-openvino-low-perf
* Update openvino CI keynames; add ccache-clear
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Fix formatting
---------
Co-authored-by: Xuejun Zhai <Xuejun.Zhai@intel.com>
Co-authored-by: Mustafa Cavus <mustafa.cavus@intel.com>
Co-authored-by: Mustafa Cavus <mustafacavus@intel.com>
Co-authored-by: Xuejun <XuejunZhai@intel.com>
Co-authored-by: Wang Yang <yang4.wang@intel.com>
Co-authored-by: Ravi Panchumarthy <ravi.panchumarthy@intel.com>
Co-authored-by: virajwad <84867530+virajwad@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Mostafa Faheem <mostafaaafaheem@gmail.com>
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* SYCL: fix a bug with async memcpy
* make mmid_row_mapping_host persistent
* comment on stream->wait
* Apply suggestion from @sanmai
* Apply suggestion from @sanmai
* Apply suggestion from @sanmai
This introduces an optional feature to allocate large GPU buffers (≥ 1GB)
using USM system allocations if supported by the device. It allows using
buffers from the system allocator then letting the system manage memory
migrations between host and device as necessary.
This feature is disabled by default and requires the GGML_SYCL_USM_SYSTEM
environment variable to enable. If USM system allocations are not supported
by the device or the system, we fallback to regular allocations.
This feature can allow VRAM overcommit. For example, the test below fails
on B580 due to lack of memory for allocation, but it passes when enabling
USM system allocations:
./examples/sycl/test.sh -m Qwen3.5-27B-Q3_K_M.gguf -lv 4
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
* sycl: support reordered Q4_K and Q5_K MoE MUL_MAT_ID
Extend reordered-weight handling to fused MoE MUL_MAT_ID for Q4_K and Q5_K expert tensors and add Q5_K reordered DMMV coverage. Unsupported 3D reorder cases now fall back instead of aborting.
* sycl: extend MoE reorder to Q6_K mul_mat_id
* vulkan: add GGML_OP_COL2IM_1D, follow-up to the CPU op
* vulkan: col2im_1d bounded gather loop instead of full-K scan with modulo
* vulkan: col2im_1d address review from @jeffbolznv
* vulkan: col2im_1d return nullptr for unsupported types, address review from @0cc4m
* [SYCL] Centralize Level Zero detection in ggml_sycl_init
* use the same wording
* get back the warning
* [SYCL] Remove per-allocation getenv() for GGML_SYCL_ENABLE_LEVEL_ZERO
* bring back the comment
* move it up to make sure devices call the shots
* move the env detection early
* replace g_ggml_sycl_enable_level_zero with a direct call to .ext_oneapi_level_zero
* update the comment
* switch back to g_ggml_sycl_enable_level_zero with a sentinel
* remove the check
* Reduce the diff
* reword, move lower
* move things aroudn
* remove forward declaration if favor of a full replace
* pre-cache results of zeDeviceGetProperties
* put ggml_sycl_get_env back
* replace get_sycl_env with ggml_sycl_get_env
* add whitespace back
* Apply suggestion from @sanmai
* vulkan: support non-contig unary/glu ops
Change unary/glu ops to pass in all strides and use fastdiv for the index
calculation. Put all unary ops in one file, similar to glu, to share the
code. codex went ahead and added expm1 without me asking, but I had to
make it do a real precision analysis rather than just making stuff up.
unary.comp initially couldn't use generic_unary_head because there wasn't
space for xielu's additional constants. Fixing this required packing the
fastdiv 'L' values.
* attempt to workaround compiler bug
* resolve conflict from #23991
* use expm1
* 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.
* Fix CI build errors
* vulkan: add support for valve fp16 dot2 extension
* use macro for dot2 path choice
* properly check for the feature
* add dot_product abstraction to reduce preprocessor branching
* 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.
* Only run webgpu CI on my fork
* Add webgpu only workflow
* handle buffer overlap case for concat operator
* restore build-webgpu.yml
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Run clang-format
* Update ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Reese Levine <reeselevine1@gmail.com>
* Only run webgpu CI on my fork
* Add webgpu only workflow
* Implement 2d workgroups for more operations
* fix
* Fix type
* Move back to global_invocation_id
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.
* cuda: reset device in get_memory function if no backend is active
* also count device and host buffers
* exclude hip and musa from counting and device reset
* use device mutex instead of atomic
* undo backend_free function move
* vulkan: add fwht support for Intel with shmem reduction
* don't use N as workgroup size
* disable subgroup shuffle on MoltenVK AMD
* disable fwht shader on Intel Windows due to driver bug
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.
* 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
Benchmark (node v25, emcc -O3 -msimd128, 64 blocks x QK8_1=32,
200k iterations):
| impl | ns/call | speedup |
|--------|---------|---------|
| scalar | 880.7 | 1.00x |
| simd | 257.8 | 3.42x |
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
* Start work on flash_attn refactor
* Refactor
* Split k/v quantization
* Refactor and abstract quantization logic for flash_attn and mul_mat
* Add quantization support to tile path
* formatting
* Move to functions, add a check
* 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 <osimons@nvidia.com>
---------
Co-authored-by: Oliver Simons <osimons@nvidia.com>
* cuda: reserve space for quantize kv-cache at startup
* address review comments
* remove forward decl
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* remove assert in ggml-cuda.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* 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
* opencl: add general q5_0 support
* opencl: add general q5_1 support
* opencl: support non-uniform workgrp size
---------
Co-authored-by: Li He <lih@qti.qualcomm.com>
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.
* 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).
* remove 'needed'
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.
* 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
* correct the link
* vulkan: add flash attention bf16 kv support
* vulkan: bf16 FA coopmat1 support
* vulkan: bf16 FA coopmat2 support
* fix FA bf16 f32 fallback
* fix FA bf16 coopmat1 shader
* fix FA bf16 coopmat2 shader
* code cleanup
* cleanup comment change
* address feedback
* add O_TYPE for cm2 FA
* use O_TYPE for gqaStore function
* reduce BFLOAT16 ifdefs
* 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:
2698b43803/include/boost/container_hash/detail/hash_mix.hpp (L19-L65)
* Update ggml/src/ggml-cuda/common.cuh
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Address review comments, make seed non-zero
* Apply code-formatting
* Replace std::size_t -> size_t for consistency
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* mmvq Optim: add MMVQ_PARAMETERS_TURING(mmvq_parameter_table_id) for SM75 TURING
* avoid a mismatch for JIT compilation of Turing device code for Ampere or newer
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Copilot <copilot@github.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* hex-fa: clean up qf32/fp32 handling and stride handling
* hex-fa: fix corner case fp NAN issues that were cause bad output from gemma4 on v79
* hex-fa: vectorize leftover handling
* hex-fa: avoid HVX fallback during token gen HMX has more FP16 compute capacity
* hmx-mm: remove dead code
* hmx-mm: use fastdiv in x4x2 dequant
* hmx-mm: sandwich dequant and scatter to improve perf
* hmx-mm: fixed rebase conflicts
* hmx-mm: further improve weight dequant by doing early type dispatch and precomputing fastdiv
* hmx-mm: an even earlier dispatch for per-type dequant
* hmx-mm: dequant linear types like q4_0 and q4_1 without the LUTs
This is a bit faster than LUT.
* hex-cmake: one more tweak for lto
---------
Co-authored-by: Trivikram Reddy <tamarnat@qti.qualcomm.com>
* Updated vec.h/vec.cpp code to accumulate to F32 rather than F16
Change-Id: I0cb789347f2bf60ffaf9047319f727e788c825f8
Signed-off-by: Martin Klacer <martin.klacer@arm.com>
Co-authored-by: Milos Puzovic <Milos.Puzovic@arm.com>
* OP_GATED_DELTA_NET impl
* add back lanes_per_column declaration
* removed has_subgroup_arithmetic and has_subgroup_clustered_reduce
* removed trailing spaces and fixes indentation. Hard coded subgroup size for Adreno and Intel. Return not supported when K>1 state snapshot
* support for K>1 state snapshot
* removed picky indent multiple of 4 fixes
* removed return that won\'t be executed
* 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.
---------
Co-authored-by: Todor Boinovski <todorb@qti.qualcomm.com>
* 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.
* feat: extend repeat op for vulkan
* feat: add repeat_f16 vulkan pipeline
* fix: ensure same dst and src types
* fix: use type_size instead of data types
* fix: use int16 and int32 for repeat shader op
* chore: rename repeat_f* to repeat_i*
* chore: rename repeat vulkan pipelines
* ggml-zendnn: fixed naming of matmul function
* ggml-zendnn: fixed naming of mul_mat_id function
* ggml-zendnn: fixed print in mul_mat_id
---------
Co-authored-by: plotnikov.v10 <plotnikov.v10@wb.ru>
* vulkan: add CONV_SHAPE_64x128 for medium-K conv2d
* vulkan: skip conv2d bounds checks when shapes align with tile sizes
* vulkan: use WG_SIZE=128 for CONV_SHAPE_64x32 conv2d
* vulkan: stage cm2 conv2d accumulator through shmem before global store
* vulkan: add coopmat1 conv2d path
* fallback when using too much shared memory. clean up comments
* Require 16x16x16 and subgroup size 32 or 64
* check whether shared memory is sufficient before overwriting conv2d params with coopmat1 values
* 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.
* ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K
* Fix to editorconfig checking pass
* Remove mul-mat-legacy pipeline
* Fix to use vendor name as is and add dot_product/vendor to shader_lib_ctx
* Only run webgpu CI on my fork
* Add webgpu only workflow
* refactor batch_compute_passes to a per-thread variable, and submit individual passes when it is set to false and no GPU profiling is enabled
* restore build.yml
* TP: fix ggml context size calculation, memory leak
* move split state cache back into the context
* revert to constant ggml context size for cgraphs
* increase headroom for statically allocated tensors
* remove obsolete include
* ggml: implement `gguf_init_from_buffer`
* test: `gguf_init_from_buffer`
* fix: memory breakdown for a model loaded with `no_alloc` from a file is consistent with being loaded from a buffer
* fix: use `GGML_UNUSED`
Co-authored-by: Copilot <copilot@github.com>
* fix: remove `total_size` from `gguf_reader`
* fix: file offset calculation, rename `offset` to `data_offset`
Co-authored-by: Copilot <copilot@github.com>
* refactor: extract model loader bug fixes to another PR
* feat: add `gguf_init_from_callback`
* fix: always require a max expected size
* fix: change `gguf_reader_callback_t`'s `output` type to `void *`, change `max_expected_size` and offsets to `uint64_t`
* fix: harden against offset overflow in buffer read
* fix: remove seek behavior from the callback
* feat: `max_chunk_read == 0` means `SIZE_MAX`
* fix: seeking in a gguf file with no tensors
---------
Co-authored-by: Copilot <copilot@github.com>
- 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.
- 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
* 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.
Fixes#20478
* SYCL: fix BF16 DMMV out-of-bounds when ncols % 64 != 0
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 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* 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.
* metal : fix GGML_OP_SET kernel threads
* tests : extend test_cpy to support different src/dst shapes
Extend test_cpy to support different source and destination tensor shapes
for CPY operations (reshaping), where the total number of elements must match.
- Renamed ne -> ne_src, added ne_dst parameter (default: use src shape)
- Added 50 new reshaping test cases covering 1D<->2D<->3D<->4D conversions
- Tests exercise 1024 boundary, small shapes, and large dimensionality changes
- Fixed dangling reference bug (storing & to temporary std::array)
- Updated all existing test calls with permute/transpose args for compatibility
Assisted-by: llama.cpp:local pi
* metal : optimize concat kernel with row batching for small widths
When ne0 < 256, batch multiple rows into a single threadgroup to improve
occupancy. This avoids underutilizing the GPU when processing narrow tensors.
- Dispatch nth = min(256, ne0) threads per group
- Calculate nrptg (rows per threadgroup) to fill up to 256 threads
- Update kernel index calculation to handle the row batching
- Add boundary check for i1 >= ne1
Assisted-by: llama.cpp:local pi
* tests : clean-up
* tests : refactor CPY shape tests to use dimension permutations
Replace 75 hardcoded test cases with a loop over permutations of
{3, 5, 7, 32} (total elements: 3360). Each src permutation is tested
against canonical sorted and reverse dst, skipping identical shapes.
Covers F32, F16, and Q4_0 (when both src and dst ne0 == 32).
Assisted-by: llama.cpp:local pi
* hexagon: remove gathers and better handling of vtcm in ssm-conv
* hexagon: relax ssm-conv gating requirements
* hexagon: add new prefill ssm-conv backend test
* hexagon: remove trailing white space
* hex-rope: uninline rope_cache_init, otherwise it breaks after rebaseing with SSM_CONV changes
---------
Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
* opencl: refactor initialization
* opencl: refactor GPU identification
* opencl: rename for consistency
* opencl: cache global mem size in dev_ctx
* opencl: adjust log level
* opencl: load argsort and flash_attn kernels in supports_op
* argsort kernel must be built for supports_op for querying the max
workgroups
* flash_attn kernel has many variants, only load them when needed
* hmx-mm: update debug logging in hmx-mm
* hmx-mm: update dequant logic to use HVX_vector_x2/4
* hmx-mm: remove non-pipelined version of the quantize matmul
It seems that we don't reall need non-pipelined version
* hmx-mm: use activation depth mode and update naming
Co-authored-by: Kim-Chyan Gan <kgan@qti.qualcomm.com>
* hex-mm: minor hmx matmul naming updates
* hmx-mm: remove unused vars
* snapdragon: scripts bump default ubatch-size to 1K
* hexagon: combine HMX and power and clock settings into a single set_power call
* hmx-mm: remove leftover of the scale repl helper
* hexagon: fix editconf error
---------
Co-authored-by: Kim-Chyan Gan <kgan@qti.qualcomm.com>
* Adds initial PDL setup.
* Adds PDL barriers based on simple heuristic: place "sync" before first input pointer access, and "launch" after last write, e.g. to tensors like dst.
* Further optimization pass of the first half of kernels
* Optimized PDL barriers for the second batch of kernels
* Further refinements after rebase.
* Moves pdl logic to separate function, removes some whitespace
* Strips post-hoc PDL logic
* Adds stream capture PDL setup. Enrolls quantize_q8_1 to leverage pdl to
overlap execution with previous kernels
* Enrolls mul_mat_vec_q, rms_norm_f32 and k_bin_bcast (partly) into PDL
* Enrolls mmvf, rope, set-rows and topk kernels for gpt-oss into PDL
* Introduce ggml_cuda_kernel_launch, to abstract away cudaLaunchKernelEx,
to enable hip/musa compatibility
* Enrolls cpy_scalar_contiguous, k_get_rows_float and rms_norm_f32
* Enrolls flash_attn_combine_results
* Fix: Drops needless and broken check of CUDA arch for PDL. PDL either
works or is without effect.
* Enrolls flash-attention kernels to pdl
* Fix: inlines ggml_cuda_kernel_launch, and uses perfect forwarding for
kernels args. This fixes PDL.
* Perf: Enrolls k_bin_bcast variadic template invocation into PDL, via
and template alias and template expansion
* Enrolls all remaining kernels for qwen3-coder-next into PDL
* Remove all PDL LC calls to create a baseline
* Added LC according to internal guidance and tested kernel performance.
* Enrols missing qwen3-5 kernels passively into PDL.
* Kernel optimizations (LC signals) for qwen3.5
* Enrolls ssm-scan kernels into PDL
* Adds GGML_CUDA_PDL command line option to toggle PDL.
* Fix: Ada and lower compilation by guarding PDL calls correctly
* Cleanup: Removes commented out GGML_CUDA_PDL_LC
* Cleanup: Removes experimental comments
* Adds 90-virtual to build script so that Hopper GPUs can leverage PDL.
* Adds stricter checks to enable PDL, adds env-check to disable it, and removes now superfluous compile option to enable PDL.
* Fix: Correct PDL en/disablement based on device-side arch check. Host
side check is UB. Required moving from macros to inlined functions
* Fix: default-disable PDL. Enable by setting GGML_CUDA_ENABLE_PDL=1
* Enable PDL by default for Hopper+ devices
* Enrolls softcap_f32 and two flash_attn kernels into PDL.
* Improves flash attn PDL barrier placement
* Fix: Perf regression on ada; excludes ada and below from PDL launches
* Improves some sync barrier placements
* Drops superfluous constructor
* Adds #endif guard comments
* Reverts experimental change to top-k-moe.cu, which moved expensive allocations
in front of the PDL barrier. It did not have a meaningful impact.
* Exchanges GGML_CUDA_DISABLE_PDL with GGML_CUDA_PDL. IFF GGML_CUDA_PDL=0
PDL is disabled
* Revert "Drops superfluous constructor". Adds const to remaining
arguments
This reverts commit 12b1d250da0089ae02a9bb71bbb3fd6d70f6f2f1.
* Cleanup: Removes and fixes some comments and whitespace
* Clarifies comment of sync-barrier position
* Relocates and refactors PDL launch functions and accessories
* Adds error checking to the regular kernel launch path
* Drops "auto" in favor of "ggml_cuda_kernel_params"
* Adds "const" to ggml_cuda_kernel_launch_params
* [Whitespace] Adds final newline to common.cuh to make editorconfig CI job happy
* opencl: add q4_k moe support
* opencl: add q5_k moe support
* opencl: add q6_k moe support
* opencl: adjust format
---------
Co-authored-by: Li He <lih@qti.qualcomm.com>
* sycl: add GGML_SYCL_USE_ASYNC_MEM_OP env toggle
Signed-off-by: Chun Tao <chun.tao@intel.com>
* Use async mem ops for correctness when SYCL graphs are explicitly on.
Signed-off-by: Tao, Chun <chun.tao@intel.com>
---------
Signed-off-by: Chun Tao <chun.tao@intel.com>
Signed-off-by: Tao, Chun <chun.tao@intel.com>
Co-authored-by: Chun Tao <chun.tao@intel.com>
With the introduction of MTP we can have multiple compute contexts for
the same RPC device. In this case last_graph_uid is not updated properly
when contexts are being switched. This patch fixes this by moving
last_graph_uid to the device context, making sure it is always updated.
closes: #23242
* ggml-hexagon: add PAD op HVX kernel
Implements GGML_OP_PAD on the Hexagon HTP backend using HVX vectorized
kernels. Supports zero-padding and circular padding across all 4 tensor
dimensions.
* hex-ggml: remove duplicate op cases (merge conflict)
* hex-pad: fix editorconfig checks and macro alignment
---------
Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
2026-05-25 12:26:07 +03:00
Intel AI Get-to Market Customer Success and Solutions
* ci/run: set explicit SPIR-V Headers search path for macOS vulkan CI
For whatever reason, the files are under additional sub-path
`vulkan/` under the cmake directory, which does not match either
current LunarG macOS Vulkan SDK structure (`lib/cmake/SPIRV-Headers`),
nor what gets installed when you run the cmake build+install for
SPIRV-Headers itself on at least Linux (`share/cmake/SPIRV-Headers`).
This allows for SPIRV-Headers to be found, as currently the CI
runner's setup does not seem to include the relevant path in
list of search locations.
* ggml-vulkan/CMakeLists: add a check for SPIRV-Headers
This is installed by the project if it is built and installed.
Receiving an error during the configuration step is generally
preferred to receiving an error in the middle of a build.
* spec: support MTP
* fix batch size
* rename files
* cont : simplify (llama/7)
* MTP: clean-up (llama/9)
* MTP: clean-up
* review: use llama_context_type instead of llama_graph_type
* review: remove llama_model_has_mtp
* review: fix convert issues
* convert: fix pycheck
* review: formatting
* use `mtp-` for identifying mtp models
* convert: fix mtp conversion
* mtp -> draft-mtp
* remove unused llama_arch
* add need_embd in speculative
* llama: allow partial seq_rm for GDN models for speculative decoding
Currently speculative checkpoint needs to restart from a checkpoint
after some draft tokens are not accepted, this leads to some wastage in
running the target again. This PR adds the ability to rollback upto
`draft_max` by storing the GDN intermediates.
* fix pending state
* vulkan: add GDN partial rollback
* meta: extend check to axis 1
* metal: add GDN partial rollback
Extend the gated delta net kernel to store intermediate states for
partial rollback support on the Metal backend.
- Add K (snapshot slot count) as a function constant
- Read input state from slot 0 of the 3D state tensor
- Write intermediate states to different slots during token loop
- For K=1, maintain backward-compatible single-slot behavior
Ref: 8c05923630
Assisted-by: llama.cpp:local pi
* delta_net_base: use ggml_pad instead of new_tensor
* review: add need_rs_seq
* review: rename part_bounded to n_rs
* review: deslop comments
* review: rename, add asserts
* server : adjust checkpoint logic (llama/11)
* server : adjust checkpoint logic
* cont : rm asserts
* server-context: fix early exit
* spec : fix compatibility with n-gram and add TODOs (llama/13)
* metal : cleanup
* llama : fix faulty bitwise check in recurrent memory
* server : disable RS-based MTP in combination with other spec types
* spec : add TODOs
* cont : fix comment
* cont : update comment
* common : fix logic for ngram + mtp compat
* llama-memory: enable checkpointing with partial rollback
* cont: add test-case for loading into a dirty ctx
* llama-memory-recurrent: clear rs_idx in clear
* download: fix mtp path
* llama-arch: fix enorm op
* docs: update docs
* conversion: fix type annotations
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Adds RDNA3 support to the CUDA mma FA kernel. To make the RDNA3 tensor cores work with the FP16 accumulation for VKQ the tiles they need to be 32 logical units long in direction of the attention head; for head sizes 80 and 112 that are not exactly divided by 32 the regular length of 16 with FP32 accumulation is used instead. The longer tiles also enable more efficient transposition for a warp size of 32 which is why it's also used for RDNA4. However, this scrambles the data layout of the accumulators along the attention head dimension. To prevent accidental misuse I added another entry to ggml_cuda_mma::data_layout.
I also tuned the kernel parameters for RDNA3, RDNA4, and CDNA1 in general, during which I discovered that the kernel can be made to work for head sizes up to 256 for CDNA. For RDNA3/4 I was not able to get better performance that the tile kernel for head sizes > 128.
* ggml-webgpu: makes the flash attn vec path compile and size its split/reduce work from the device’s reported subgroup range instead of assuming 32 subgroup size.
* ggml-webgpu: remove the extra max_wg_size >= max_subgroup_size guard. Remove hardcoded 32 when determine the value of reduce_wg_size and vec_nwg_cap
* SYCL: fix multi-GPU system RAM exhaustion by using Level Zero allocations
Replace sycl::malloc_device with zeMemAllocDevice for GPU memory allocation
in the SYCL backend. sycl::malloc_device triggers the xe kernel driver's
DMA-buf/TTM path which mirrors every VRAM allocation 1:1 in system RAM.
zeMemAllocDevice uses the SVM/P2P path with no host staging.
On a dual Intel Arc Pro B70 system (64GB VRAM, 64GB RAM), a 15.6 GiB model
consumed 60 GiB of system RAM via sycl::malloc_device, causing OOM crashes.
With zeMemAllocDevice, the same workload uses ~6.7 GiB of system RAM with
no performance regression.
All Level Zero calls include automatic fallback to the original SYCL
allocation path if Level Zero interop is unavailable.
* SYCL: address review feedback - remove try/catch, check device types, deduplicate
- Remove try/catch from malloc/free/memcpy helpers, check backend and
device type upfront instead (ggml_sycl_is_level_zero, ggml_sycl_is_dgpu)
- Move shared helpers (is_level_zero, is_dgpu, free_device) to common.cpp
and declare in common.hpp to eliminate code duplication
- Use SYCL_CHECK(CHECK_TRY_ERROR()) for fallback sycl::free calls
- Guard dev2dev_memcpy L0 path to dGPU-to-dGPU only, preserving the
host-staged path for iGPU-to-dGPU transfers
- Add Windows Level Zero SDK path detection (LEVEL_ZERO_V1_SDK_PATH)
in CMakeLists.txt (co-authored with @arthw)
* SYCL: add build/runtime flags for Level Zero, address review feedback
Implements the architecture suggested by @arthw: compile-time and runtime
flags to cleanly separate Level Zero and SYCL memory API paths.
- Add GGML_SYCL_SUPPORT_LEVEL_ZERO cmake option (default ON). All Level
Zero code is wrapped in #ifdef so the build works on systems without
the Level Zero SDK installed (e.g. CPU-only CI servers). Both the
loader library and headers are checked before enabling.
- Add GGML_SYCL_ENABLE_LEVEL_ZERO runtime env var (default 1). Controls
whether Level Zero or SYCL memory APIs are used. Only one API style is
used per session, no mixing. If Level Zero is enabled but the devices
don't support the Level Zero backend, it auto-disables with a warning.
- Remove Level Zero code from dpct_malloc. It was unused (dpct::device_memory
is not called anywhere in the backend) and used try/catch for flow control.
- Update SYCL.md with documentation for both new parameters.
Tested on Intel Arc Pro B70 (32GB), single-GPU and dual-GPU, with both
GGML_SYCL_SUPPORT_LEVEL_ZERO=ON and OFF builds. AI-assisted development
(Claude). Code reviewed and tested on my hardware.
* SYCL: unify Level Zero malloc/free call sites, address review feedback
Move ggml_sycl_malloc_device to common.cpp alongside ggml_sycl_free_device.
Both functions are now unconditionally available — Level Zero code is
#ifdef'd inside the functions, not at call sites. All call sites use
uniform SYCL_CHECK(CHECK_TRY_ERROR()) wrapping with no #ifdef blocks.
Addresses arthw's review: wrap all malloc/free in SYCL_CHECK for stack
traces on failure, eliminate duplicated #ifdef/else patterns at 6 call
sites (-29 lines net).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* SYCL: add Level Zero SDK to CI, fix device check and missed alloc paths
Add Level Zero SDK installation to Ubuntu and Windows SYCL CI jobs
so the Level Zero code path is compiled and tested in CI.
Fix two bugs found during extended dual-GPU testing (no
ONEAPI_DEVICE_SELECTOR set):
- The Level Zero backend check was iterating all SYCL devices
including CPU. The OpenCL CPU device caused Level Zero to be
disabled for the GPUs, defeating the fix on multi-GPU systems.
Added is_gpu() filter so only GPU devices are checked.
- sycl_ext_malloc_device/sycl_ext_free (tensor reorder temp buffers)
were still calling sycl::malloc/sycl::free directly, bypassing the
Level Zero path. Routed through ggml_sycl_malloc_device/free_device
for consistency with the other device memory call sites.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* SYCL: address arthw review feedback on Level Zero memory API structure
- Move ggml_sycl_malloc_device to static function in ggml-sycl.cpp;
only ggml_sycl_free_device (used by common.cpp) stays in common.cpp
- Switch both helpers to use g_ggml_sycl_enable_level_zero global
instead of per-call queue backend checks
- Remove #ifdef wrapper from global definition; always declare at 0,
add #else branch in init block so it stays 0 when L0 not compiled in
- Update init loop comment to explain GPU-only device check
- CMakeLists: message(STATUS) before the if block; align option wording
AI-assisted implementation. Reviewed and tested on dual Intel Arc Pro
B70 (32 GB each): test-backend-ops OK on both GPUs, single/dual-GPU
Q4_K_M and Q8_0 bench correct, zeMemAllocDevice GTT delta confirmed
<5 MiB per 4 GiB allocation (vs ~4 GiB shadow with sycl::malloc_device).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* SYCL: remove unused cstdio/cstdlib includes from common.cpp
Leftover from the deleted ggml_sycl_queue_supports_level_zero helper.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* Apply suggestions from code review
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* SYCL: preserve Level Zero allocation path during early malloc
* ci: fix Level Zero package conflict in Intel Docker build
* ci: find Level Zero loader in oneAPI package step
* ci: allow Windows SYCL package without Level Zero DLL
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* opencl: add q5_0 moe support
* opencl: add q5_1 moe support
* opencl: avoid potential leak
* opencl: suppress unused var warning when building for non-Adreno
---------
Co-authored-by: Li He <lih@qti.qualcomm.com>
* ggml-zendnn : add runtime env var GGML_ZENDNN_ADAPTIVE_FALLBACK to control adaptive fallback (default: enabled)
* ggml-zendnn : restore original fallback logic when adaptive fallback is disabled
* hexagon: add hvx_vec_repl helpers and use those for splat-from-vtcm usecase
* hmx-mm: optimize per-group scale handling
* hmx-fa: optimize slope load from vtcm
* hmx-fa: use aligned access where possible in hmx-utils
* hexagon: add hvx_vec_repl_2x_f16 helper and consolidate repl helpers
---------
Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
* fix(mixed-types): use f32 for precision and update the shared memory calculation logic for f32
* fix(unary): correct the gelu, gelu quick and gelu erf functions
* fix(flash-attn-tile): fix the hardcode v type
* fix(flash_attn): fix tile path
* fix: pass editorconfig and address the type conflicts
* fix: remove reduant pipeline keys
* fix: remove inline min/max group size functions and revert the flash attn path order
* fix: use clamp to avoid NaN for GELU
* fix: use the right range for exp, 80 is safer for f32 exp
* Q4_1 MoE CLC pass sanity check
* remove unnecessary code
* opencl: remove unnecessary asserts and reformat
* opencl: fix supports_op for q4_1 moe
* q4_1 moe is supported by Adreno with certain shapes
---------
Co-authored-by: Li He <lih@qti.qualcomm.com>
`im2col_cuda` and `im2col_3d_cuda` both dispatch with
`block_nums.y = OW`. CUDA caps grid Y at 65535. Conv1d encoders on
raw 16 kHz audio with T > 65535 (~ 4 s) trip the limit -- e.g. SEANet
at 11 s lands at OW = 176000 -- and the launch returns
`invalid configuration argument`.
Clamp `block_nums.y` to `MIN(OW, MAX_GRIDDIM_Y)` and loop inside the
kernel with stride `MAX_GRIDDIM_Y`. Same in-kernel stride pattern
already used for the z axis (`MAX_GRIDDIM_Z`). Both 2D `im2col_kernel`
and 3D `im2col_3d_kernel` need the same fix. Bit-identical for
OW <= 65535 (single iteration of the new outer loop).
Tested on T4 / Jetson Orin with a SEANet encoder running on 11 s /
16 kHz audio (im2col reaching OW ~ 176000); pre-fix launch returns
`invalid configuration argument`, post-fix runs to completion.
Existing test-backend-ops im2col cases unchanged.
* cuda: tighten snake fusion type checks for all operands (defensive, sync vulkan)
* cuda: reject snake fusion when ne[2] or ne[3] > 1 (mirror vulkan PR review)
* cuda: merge type_ok and types_ok into a single types_ok (address am17an review)
* cuda: filter ADD/SUB/MUL/DIV in supports_op to F32/F16
bin_bcast only dispatches F32/F16 type triplets, mirror the
vulkan filter so unsupported types fall back through cpy
instead of aborting.
* test-backend-ops: extend snake_fuse to rank-4 with ne[2]/ne[3] > 1 cases
Georgi Gerganov
Neo Zhang
Winston Ma
Ruben Ortlam
kononnable
lhez
Zijun Yu
Alexey Kopytko
Francois Dugast
Jeff Bolz
Frosty40
Pascal
Andrei
Katostrofik
someoneinjd
leonardHONG
Masashi Yoshimura
ZihaoMu
shaofeiqi
Kevin Liu
Gaurav Garg
Oliver Simons
Yash Raj Pandey
ravel7524
Reese Levine
Nikhil Jain
Harkirat Gill
Xuan-Son Nguyen
Charles Xu
Mason Milburn
Kartik Sirohi
rehan-10xengineer
Andreas Kieslinger
Aman Gupta
Max Krasnyansky
Todor Boinovski
Anav Prasad
Shrivas Shankar
Johannes Gäßler
Matt Corallo
Jinyang He
fairydreaming
fl0rianr
redfox
Jaden_Mach
Martin Klacer
ymcki
l8bloom
Vladislav
forforever73
Gilad S
Yiwei Shao
dskwe
Shawn Gu
karavayev
Sachin Sharma
Chen Yuan
Daniele
Aparna M P
Intel AI Get-to Market Customer Success and Solutions
Radoslav Gerganov
Pranav Dhinakar
Gabe Goodhart
Jan Ekström
Ori Pekelman
Dev-X25874
Pranav Dhinakar
Zheyuan Chen
alex-spacemit
scutler-nv
Trivikram Reddy
yzyyzyhhh
Masato Nakasaka
guyfischman
CrispStrobe
Oliver Walsh