The Q8_0 reorder optimization (#21527) was missing a reorder-aware
dequantizer for the GEMM code path used during prompt processing.
After token generation reordered Q8_0 weights (via DMMV/MMVQ), the
next prompt processing pass would read them with the standard
dequantizer, producing garbage output.
Add dequantize_block_q8_0_reorder() and wire it into both
ggml_get_to_fp16_sycl() and ggml_get_to_fp32_sycl(), matching the
pattern already used by Q4_0, Q4_K, and Q6_K.
Fixes #21589
AI (Claude) was used to assist with root cause investigation and
writing the kernel code. All code was human-reviewed and tested
on real hardware.
* SYCL: fix reorder crash when device memory is full
The reorder optimization allocates a temporary buffer the full size of
the weight tensor on the device. When VRAM is nearly full (large models
on a single GPU), this allocation fails and the subsequent memcpy crashes
on a NULL pointer.
Fix: try device allocation first, fall back to host memory if device
memory is full. The reorder kernel still works correctly reading from
host memory over PCIe. This is slower for the one-time reorder (~21 t/s
vs ~38 t/s on Intel Arc Pro B70), but the optimization is preserved for
all subsequent inference. If both device and host allocation fail, skip
the reorder and fall back to the unoptimized kernel path.
Also fixes a bug where opt_for_reorder() marked tensors as reordered
even when the reorder was skipped due to allocation failure. This caused
DMMV/MMVQ kernels to read the original AoS data as if it were SoA,
producing garbage output or NaN results.
Tested on Intel Arc Pro B70 (32GB) with Q8_0, Q4_K_M models. Coding was
AI-assisted (Claude), reviewed and tested on hardware by a human.
Fixes #20478
* SYCL: add RAII temp buffer class + macro guard for host fallback
Replace sycl_ext_malloc_with_fallback/sycl_ext_free_fallback free
functions with sycl_reorder_temp_buffer RAII class. The host_fallback
bool is now a private member, and cleanup happens automatically at
scope exit.
Add GGML_SYCL_HOST_MEM_FALLBACK cmake option (default ON) to guard
the host memory fallback code path. Device access to host memory
requires Linux kernel 6.8+ (Ubuntu 26.04+); users on older kernels
can set -DGGML_SYCL_HOST_MEM_FALLBACK=OFF to disable it.
Addresses arthw's review on PR #21638.
Co-Authored-By: Claude Opus 4.6 (1M context) <redacted>
* SYCL: document GGML_SYCL_HOST_MEM_FALLBACK build option in SYCL.md
Co-Authored-By: Claude Opus 4.6 (1M context) <redacted>
* SYCL: add reorder-aware DMMV dequantizers for Q4_K and Q6_K
Q4_K and Q6_K had reorder support for MMVQ and GEMM paths but not
DMMV. When the DMMV path encountered reordered data it would abort.
Add DMMV kernels that read from the SOA reorder layout for both
types. Same math as the non-reorder versions, different memory
access pattern.
Co-Authored-By: Claude Opus 4.6 (1M context) <redacted>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <redacted>
hexagon: optimization for HMX mat_mul (llama/21554)
* hexagon: add async HMX worker
Introduce hmx-worker (dedicated thread for HMX compute) to overlap HMX
matmul with HVX dequant/DMA stages in the pipeline path, replacing the
previous synchronous HMX calls that blocked the main thread.
* hexagon: cost-based VTCM chunk search for out-stationary matmul
* hexagon: fix futex race in hmx_worker_drain
Store the boolean to local variable avoid atomic load twice
* hex-mm: hmx optimize scatter/transpose and use HMX intrinsics
* hex-vmem: drop vmem limit a touch under 3GB on v73
* hexagon: add fwd declaration of htp_context
* hex-hmx: replace hmx-worker with hmx-queue that mimics dma-queue interface
Simplifies the overall implemantion, reduces thread wakeup roundtrips.
* hex-mm: add debug log to hmx work func called from hmx-queue
* Update hmx-queue.h
Co-authored-by: Max Krasnyansky <redacted>
---------
Co-authored-by: Kim-Chyan Gan <redacted> Co-authored-by: Max Krasnyansky <redacted> Co-authored-by: Max Krasnyansky <redacted>
cmake: fix CMP0194 warning on Windows with MSVC (llama/21630)
* cmake: fix CMP0194 warning on Windows with MSVC
Set CMP0194 policy to NEW before project() call in ggml/CMakeLists.txt to suppress the "MSVC is not an assembler for language ASM" warning introduced in CMake 4.1.
The ggml project enables ASM globally for Metal (macOS) and KleidiAI (ARM) backends. On Windows/MSVC, no assembler sources are used, but CMake 4.1+ warns because cl.exe is not a valid ASM compiler.
This follows the same pattern used in ggml-vulkan (CMP0114, CMP0147).
* Add test case for dispatch to DeviceSegmentedRadixSort
We currently lack a way to force graph mode in CUDA, patch callback to
invoke ggml_backend_compare_graph_backend twice to enforce each test to
run in graph mode
Stephen Cox [Sun, 12 Apr 2026 12:15:26 +0000 (00:15 +1200)]
mtmd: add Gemma 4 audio conformer encoder support (llama/21421)
* mtmd: add Gemma 4 audio conformer encoder support
Add audio processing for Gemma 4 E2B/E4B via a USM-style Conformer.
Architecture:
- 12-layer Conformer: FFN → Self-Attention → Causal Conv1D → FFN → Norm
- Subsampling Conv Projection: 2x Conv2D(stride=2) with LayerNorm
- Full self-attention with sinusoidal RPE and sliding window mask (24)
- Logit softcapping at 50.0, ClippableLinear clamping
- Output: 1024 → 1536 → RMSNorm → multimodal embedder
Mel preprocessing (dedicated mtmd_audio_preprocessor_gemma4a):
- HTK mel scale, 128 bins, magnitude STFT, mel_floor=1e-3
- Standard periodic Hann window (320 samples), zero-padded to FFT size
- Semicausal left-padding (frame_length/2 samples)
- Frame count matched to PyTorch (unfold formula)
- No pre-emphasis, no Whisper-style normalization
- Mel cosine similarity vs PyTorch: 0.9998
Key fixes:
- Tensor loading dedup: prevent get_tensor() from creating duplicate
entries in ctx_data. Fixed with std::set guard.
- ClippableLinear clamp_info loading moved after per-layer tensors.
- Sliding window mask (24 positions) matching PyTorch context_size.
- Skip Whisper normalization for Gemma4 mel output.
Tested on E2B and E4B with CPU and Vulkan backends.
Transcribes: "Glad to see things are going well and business is starting
to pick up" (matching ground truth).
- vendors/hip.h: Add CDNA4 preprocessor define for __gfx950__
- common.cuh: Add GGML_CUDA_CC_CDNA4 and GGML_CUDA_CC_IS_CDNA4 macros
- mma.cuh: Route CDNA4 to compatible MFMA instructions:
* f32 matmul: mfma_f32_16x16x4f32 (xf32 variant unavailable on gfx950)
* bf16 matmul: mfma_f32_16x16x16bf16_1k (same as CDNA3)
* int8 matmul: mfma_i32_16x16x32_i8/32x32x16 (same as CDNA3)
- mmq.cuh: Include CDNA4 in stream-k kernel dispatch
CDNA4 is largely compatible with CDNA3 except:
- No xf32 MFMA (mfma_f32_16x16x8_xf32) — routes to f32 path
- Different FP8 format (e4m3fn vs e4m3_fnuz) — not changed here
Tested on AMD Instinct MI355X (gfx950), ROCm 7.0.1:
- Build: compiles cleanly with -DAMDGPU_TARGETS=gfx950
- llama-bench (Qwen2.5-1.5B Q4_K_M, single GPU):
* f16+FA: 40,013 tok/s prefill, 254 tok/s decode
* q8_0+FA: functional
- Flash attention: works correctly
- MMQ: works correctly with stream-k dispatch
* Remove shfl and AllReduce from backend interface
* move allocation workaround out of ggml-alloc.c
* 2d tensor set/get support
* Fix the seg fault without NCCL
* Apply suggestion from JohannesGaessler
* support for tensor dims % n_devs != 0
* fix view_offs scaling
* arbitrary num. of GPUs/tensor split
* fix compilation
* better granularity estimate
* Support device-specific host buffer types if all underlying backends expose the same type. This allows using pinned memory instead of pageable memory for CUDA.
Fix compilation errors.
* partial Qwen 3 Next support
* Fix qwen3 30b (llama/8)
* Fix crash with Qwen-30B-A3B Q4_0
Qwen-30B-A3B Q4_0 has an intermediate dimension of 768. Using a granularity of 256 forces an uneven split between GPUs, which is not supported by the current implementation.
* Decide block size based on tensor quantization type
* Fix crashes due to KV cache serialization (llama/9)
KV cache serialization requires non-zero offsets on the tensor. Add support in the meta backend to set/get a tensor with a non-zero offset.
* metal : fix build (llama/7)
* static memory allocations, fix usage count
* fix tensor granularity
* more even memory distribution
* use BF16 for allreduce
* rebase fixup
* better error message for unsupported architectures
* Fix device mismatch during scatter of allReduce. (llama/11)
There is a mismatch between the dst buffer device and the backend device, causing the use of sync copies
* Enable the previous allreduce implementation. It is better in both perf and stability (llama/12)
* delay AllReduce for Moe for less I/O
* build : clean-up compile warnings
* backend : move most of the meta backend API to ggml-backend-impl.h
* cont : hide unused public API in the implementation
* llama : use llama_device + remove ggml_backend_dev_is_meta()
* ggml-backend : remove unused alloc include
* minor : remove regex include
* ggml : introduce ggml-ext.h for staging new APIs
* rebase fixup
* fix tests
* llama : more robust logic for determining Meta devices (llama/16)
* llama : more robust logic for determining Meta devices
* cont : fix devs size check
Co-authored-by: Johannes Gäßler <redacted>
* cont : fix log type
Co-authored-by: Johannes Gäßler <redacted>
---------
Co-authored-by: Johannes Gäßler <redacted>
* disable roundtrip for meta backend
* fix arch selection
* Qwen 3.5 support
* fix Gemma 4 MoE
* fix OpenVino, SYCL
* fix test-llama-archs for CPU-only builds
* Fix Qwen 3.5 MoE
* disable meta backend tests for WebGPU
* tests : filter CPU-based devices from the Meta backend tests (llama/17)
* meta : formatting, naming, indentation (llama/18)
* formatting : llama-model.cpp
* formatting : ggml-ext.h
* formatting : ggml-backend-meta.cpp
* meta : add TODO
* add documentation
* better error messages
* fix GPT-OSS
---------
Co-authored-by: Carl Philipp Klemm <redacted> Co-authored-by: Gaurav Garg <redacted> Co-authored-by: Georgi Gerganov <redacted>
sycl : add flash-attn support for head size 512 (llama/21654)
* sycl : add flash-attn support for head size 512
This patch extends the SYCL Flash Attention implementation to support head sizes (DKQ/DV) of 512.
Changes:
- Added DKQ/DV 512 cases to both tile and vector Flash Attention kernels.
- Updated kernel selection logic to allow vector kernels for head sizes up to 512 (previously 256).
- Removed unused/redundant AMD and RDNA-specific configuration functions in `fattn-tile.hpp`.
- Refactored `ggml_backend_sycl_buffer_init_tensor` to use a switch statement for clearer tensor extra buffer initialization.
- Added necessary template instances for the new 512 head size across various quantization types.
* remove defunct mxfp4 reorder from setting buffer type
fix: free ctx_copy in ggml_opt_free to plug per-training-session leak (llama/21592)
* fix: free ctx_copy in ggml_opt_free to plug per-training-session leak
ggml_opt_alloc populates opt_ctx->ctx_copy via a free+init pair every
time the allocated graph shape changes. The last ctx_copy from the
final ggml_opt_alloc call survives until ggml_opt_free is invoked,
but ggml_opt_free was only freeing ctx_static and ctx_cpu, never
ctx_copy. Each opt_ctx lifetime therefore leaks the final per-batch
context — ~900 KB for a typical GNN training session in
sindarin-pkg-tensor, surfaced via AddressSanitizer.
ctx_copy is nullptr-initialized and ggml_free() handles NULL safely,
so the new release is guard-free.
* Update ggml/src/ggml-opt.cpp
Co-authored-by: Johannes Gäßler <redacted>
---------
Co-authored-by: realorko <redacted> Co-authored-by: Johannes Gäßler <redacted>
ggml-cuda: ds_read_b128 for q4_0 and q4_1 mmq kernels (llama/21168)
* ds_read_b128 for q4_0 and q4_1 mmq kernels
Current for loop generates ds_read_b32 instructions with hip compiler, the new solution generates ds_read_b128 instructions for the same operation, saving some LDS bandwidth. Tested on MI50 and RX6800XT, its faster on both.
* Vectorized lds load update: used ggml_cuda_get_max_cpy_bytes and ggml_cuda_memcpy_1 functions for generic implementation
* Explicit for loop in mmq, renamed vec into tmp
* Fixed max_cpy usage in the loading loop
* Fixed typo in q4_1 kernel
* Update ggml/src/ggml-cuda/mmq.cuh
Co-authored-by: Johannes Gäßler <redacted>
* Update ggml/src/ggml-cuda/mmq.cuh
Co-authored-by: Johannes Gäßler <redacted>
* Update ggml/src/ggml-cuda/mmq.cuh
Co-authored-by: Johannes Gäßler <redacted>
* Renoved trailing white line 500
* Update mmq.cuh removed other whitelines
* Remove trailing whitespaces
---------
Co-authored-by: iacopPBK <redacted> Co-authored-by: Johannes Gäßler <redacted> Co-authored-by: iacopPBK <redacted>
vulkan: add FA dequant for q4_1, q5_0, q5_1, iq4_nl (llama/21029)
Add dequantize4() implementations for Q4_1, Q5_0, Q5_1, and IQ4_NL
in the flash attention base shader. Register them in the shader
generator, pipeline creation, and enable in the scalar/coopmat1 FA
support check.
Extend the existing reorder optimization to Q8_0. The reorder
separates scale factors from weight data for coalesced memory
access -- was implemented for Q4_0/Q4_K/Q6_K but Q8_0 was missing.
On Arc Pro B70 (Xe2), Q8_0 tg goes from 4.88 to 15.24 t/s (3.1x)
on Qwen3.5-27B. BW utilization: 21% -> 66%.
The key fix beyond the kernels: Q8_0 was missing from the type
check in ggml_backend_sycl_buffer_init_tensor() that allocates
the extra struct carrying the reorder flag -- so the optimization
was silently skipped.
AI (Claude) was used to assist with root cause investigation and
writing the kernel code. All code was human-reviewed and tested
on real hardware.
Write an optimized flash_attn_stream_k_fixup kernel (llama/21159)
* Write an optimized flash_attn_stream_k_fixup kernel
Write a specialized and more optimized kernel for cases where nblocks_stream_k is multiple of ntiles_dst.
Make nblocks_stream_k to multiple of ntiles_dst if nblocks_stream_k > 2 * ntiles_dst
* Use the new kernel only for nblocks_stream_k_raw > 4 * ntiles_dst to make sure we have enough concurrency on GPUs
ggml-zendnn : add MUL_MAT_ID op support for MoE models (llama/21315)
* ggml-zendnn : add MUL_MAT_ID op support for MoE models
- Add MUL_MAT_ID op acceleration for Mixture-of-Experts models
- MUL_MAT_ID op fallback to CPU backend if total experts > 32
- Point ZenDNN lib to latest bits ZenDNN-2026-WW13
* ggml-zendnn : add braces to sgemm failure condition for consistency
Reuse the buffer for the ggml context which is used for creating the
compute graph on the server side. This partially addresses a memory leak
created by the CUDA backend due to using buffer addresses as cache
keys.
CUDA/HIP: Fix kernel slection for mmvq mmid kernel to align host selection with device launch bounds (llama/21238)
The conditions cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE and cc >= GGML_CUDA_CC_TURING match all non-nvidia devices. This causes us to attempt to launch the kernel for batch sizes with larger configurations than our launch bounds on HIP devices. This pr fixes the conditionals in get_mmvq_mmid_max_batch.
ggml-webgpu: port all AOT operators to JIT (llama/20728)
* port cpy pipeline to shader lib with JIT compilation
* port glu pipeline to shader lib with JIT compilation
* port rope pipeline to shader lib with JIT compilation
* port soft_max pipeline to shader lib with JIT compilation
* removed unused functions from embed_wgsl.py which were used for
old AOT template expansion
When ollama calls ggml_backend_tensor_set from multiple threads (each
writing a different chunk of the same tensor), the CANN backend had
three concurrency issues:
1. Quantized tensors (Q4_0/Q8_0) require a full-tensor format transform
before uploading to device. Per-chunk transforms produced corrupt data.
2. ND-to-NZ weight conversion requires complete tensor data on device.
Per-chunk conversion operated on incomplete data.
3. The global g_nz_workspaces array had unprotected concurrent access.
Fix by introducing a TensorSetTracker that accumulates write progress
per tensor. For quantized tensors, raw data is staged in a host buffer
and the transform + upload is deferred until all chunks arrive. For NZ
weights, chunks are uploaded directly but conversion is deferred. The
tracker and its staging buffer are released immediately after
post-processing completes.
Add per-device mutex to g_nz_workspaces to prevent data races.
* CANN: fix L2_NORM ignoring eps parameter
The L2_NORM implementation was not using the eps parameter from
op_params, causing incorrect results when eps is large (e.g. 10.0).
The CPU reference computes scale = 1/fmaxf(norm, eps), so add a
Clamp step to clamp the norm to at least eps before dividing.
* ggml/cann: compare op_params for POOL_2D in ACL graph cache matching
When ACL graph mode is enabled, the graph LRU cache checks whether a
cached graph matches the current computation graph. Previously,
GGML_OP_POOL_2D was not included in the op_params comparison, so two
POOL_2D nodes with different pooling parameters (kernel size, stride,
padding) but identical tensor shapes and addresses could incorrectly
reuse a cached graph, leading to wrong results or aclnn errors.
Add GGML_OP_POOL_2D to the list of ops that require op_params matching
in ggml_graph_node_properties::has_matching_properties().
* cann: fix ACL graph cache matching by adding tensor type and unconditional op_params comparison
The ACL graph LRU cache was incorrectly reusing cached graphs for
operations with different tensor types or op_params, causing test
failures for CPY (f16 vs bf16), POOL_2D, L2_NORM, NORM_MUL_ADD,
RMS_NORM_MUL_ADD, and ADD_RMS_NORM.
Changes:
- Add node_type and src_type[] fields to ggml_graph_node_properties
so the cache can distinguish tensors with different types but
identical ne/nb (e.g. f16 and bf16 both have 2-byte elements)
- Compare op_params unconditionally for all ops instead of only for
SCALE/UNARY/GLU/ROPE/POOL_2D
We wrongly calculated offset_grid as `ceildiv(nrows, block_size)`,
while it must be `ceildiv(nrows + 1, block_size)`. As a consequence, we
had uninitialized values in `offset_iterator[nrows]` for the case when
`nrows % block_size == 0`.
Fixes #21162
* Reduce nrows in test case to 256, don't need 768
Radoslav Gerganov [Mon, 30 Mar 2026 14:05:11 +0000 (17:05 +0300)]
rpc : fix misleading error log (llama/21184)
When RPC is running with a remote backend which doesn't have init_tensor
function (like CPU and Metal), the server log gets full with error
messages saying that init_tensor is being called with null buffer which
is incorrect. This patch fixes this.
Gaurav Garg [Sun, 29 Mar 2026 16:35:18 +0000 (22:05 +0530)]
Optimize MOE GEMV kernel for BS > 1. (llama/20905)
* Optimize MOE GEMV kernel for BS > 1.
The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row.
New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync).
This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization.
* Remove em-dashes
* Cherry-pick changes from @am17an PR https://github.com/ggml-org/llama.cpp/pull/20885 to enable small_k optimization only for cases where it benefits
Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8
* Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype
I noticed that we were refetch the mask rows over and over.
This simple cache avoids that.
* hex-dma: unset in-order desc bit which caused signficant perf regression
We don't rely on true in order processing of the DMA descriptors anywhere.
Turns out this mode caused significant regression of around 3-4 TPS during token gen.
* hex-rope: update comment to clarify that we don't need in-order DMA completions
bench : sync submit-results URL to ggml-org (#3769)
The project moved from ggerganov/ to ggml-org/ and the README already
references the new URL in both places it mentions issue #89 (README.md
and examples/bench/README.md). Syncing the two remaining hardcoded URLs
in examples/bench/bench.cpp and examples/bench.wasm/emscripten.cpp.
Daniel Worthington-Bodart [Fri, 17 Apr 2026 11:36:27 +0000 (12:36 +0100)]
whisper : add stateless VAD detect + explicit state reset for streaming (#3677)
whisper_vad_detect_speech resets LSTM state on every call, which is
correct for batch processing but prevents temporal continuity when
calling per-chunk in a streaming loop.
Add whisper_vad_detect_speech_no_reset (skips buffer clear) and
whisper_vad_reset_state (explicit clear between utterances).
Existing whisper_vad_detect_speech is now a thin wrapper — zero
behavior change for current callers.
Co-authored-by: Claude Opus 4.6 (1M context) <redacted>
Radoslav Gerganov [Fri, 27 Mar 2026 08:59:35 +0000 (10:59 +0200)]
rpc : proper handling of data pointers to CPU buffers (llama/21030)
The compute graph may contain tensors pointing to CPU buffers. In these
cases the buffer address is serialized as 0 and sent over the wire.
However, the data pointer is serialized as-is and this prevents proper
validation on the server side. This patches fixes this by serializing
the data pointer as 0 for non-RPC buffers and doing proper validation on
the server side.
ren [Fri, 27 Mar 2026 07:05:21 +0000 (00:05 -0700)]
metal : Fix dimension constraint violation in matmul2d descriptor (llama/21048)
Updates Metal tensor API test probe to fix the dimension constraint violation in the matmul2d descriptor (at least one value must be a multiple of 16).
overview / pkg / ggml / sources / whisper.cpp / log