* ggml : transpose fused GDN state access for coalesced memory reads (#20436)
The fused Gated Delta Net kernel accessed the [S_v, S_v] state matrix
column-wise on row-major storage, causing strided reads (stride S_v =
128 floats = 512 bytes) that waste GPU cache bandwidth. This produced a
39% regression on Qwen3.5-9B (Metal, M4 Max) compared to the unfused
path.
Transpose the state indexing so threads read contiguously:
- Metal: s_ptr[is*S_v] -> s_ptr[is] (stride 1 vs S_v)
- CUDA: curr_state[i*S_v+col] -> curr_state[col*S_v+i] (coalesced)
- CPU: restructured loops for row-wise transposed access
Also add --fused-gdn [on|off|auto] CLI flag (mirrors --flash-attn) so
users can control fused GDN independently of auto-detection.
All GATED_DELTA_NET backend-ops tests pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* ggml : use SIMD dot products in CPU GDN kernel, couple AR/chunked fused flags
- Replace scalar inner loops with ggml_vec_dot_f32 for SIMD-optimized
dot products in the CPU fused GDN kernel (delta and attention output)
- Couple fused_gdn_ar and fused_gdn_ch flags in auto-detection: if one
path lacks device support, disable both to prevent state layout mismatch
between transposed (fused) and non-transposed (unfused) formats
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* llama : rever fgdn argument changes
* graph : remove GDN state transposes
* vulkan : adapt
* cuda : remove obsolete smem code
---------
Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Oliver Simons <osimons@nvidia.com>
* llama : enable chunked fused GDN path
* models : avoid Q and K repeats when using fused GDA
* cont : fix comment
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
* cont : fix the fix
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
* cont : fix
* metal : add GDN kernel (#20361)
* metal : add Metal backend for GGML_OP_GATED_DELTA_NET
Add a fused Metal kernel for the gated delta net recurrence op
(#19504), enabling GPU-accelerated inference for DeltaNet-based
models (Qwen3.5, etc.) on Apple Silicon.
Supports both GDA (scalar gate) and KDA (per-row gate) modes
with head_size 64 and 128. Unsupported configurations (head_size
32, non-contiguous tensors) gracefully fall back to CPU.
Performance: Qwen3.5-0.8B Q4_K_M on M4 Max
tg128: 170 -> 213 t/s (+25%)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* metal : validate contiguity of all input tensors in supports_op
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* metal : add algorithm equivalence comment for GDA decay path
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* cont : unslop + optimize
* cont : clean-up
---------
Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* CUDA: AR gated delta net improvements (#20391)
* Add FastDiv to gated_delta_net_cuda
* Shard columns across warps
This reduces register pressure (avoids spill for S_v = 128) and gives
the warp-scheduler more CTAs to schedule (thus hiding data-access
latencies).
* Remove unneded include in gated_delta_net.cu
* Improve comments
* Apply code-formating
* Make sharding HIP-compatible
1. Use ggml_cuda_get_physical_warp_size() to determine warp size flexibly
2. Add test with partial warp to test sum reduction on CUDA
* Remove fastdiv_s64, as we can treat neqk1 and rq3 as uint32_t
* Rename variables
* Enable GDN also for prefill, move TODO for chunked_GDN
* Actually remove the TODO from 2068908975
* Get warp size at runtime
warp_size is not known at compile time in hip host code.
* Don't expose ggml_cuda_get_physical_warp_size on host
---------
Co-authored-by: uvos <devnull@uvos.xyz>
* llama : refactor llm_build_delta_net_base API
---------
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Oliver Simons <osimons@nvidia.com>
Co-authored-by: uvos <devnull@uvos.xyz>
* WIP: add NVFP4 quantization support
* tests
* improve NVFP4 dot product implementation performance and fix bad super call
* typo
* Use nvfp4 kvalues
* vulkan : fix NVFP4 shader compilation by including kvalues_mxfp4 lookup table
* vulcal and perf fixes
* wip
* Fix metal
* fix vulcan
* Rename threshold & fix wrong scale
* Fix MOE
* Shelf backend implementations (CUDA, Metal, Vulkan, arch-specific SIMD)
Remove NVFP4 support from GPU backends and architecture-specific
optimized dot products. These should be added in separate PRs so
backend specialists can review them independently.
Reverted files:
- ggml-cuda: common.cuh, convert.cu, mmq.cu/cuh, mmvq.cu, vecdotq.cuh,
quantize.cu/cuh, mma.cuh, ggml-cuda.cu, fattn-tile.cuh
- ggml-metal: ggml-metal.metal, ggml-metal-device.cpp, ggml-metal-impl.h,
ggml-metal-ops.cpp
- ggml-vulkan: ggml-vulkan.cpp, all vulkan-shaders/*
- ggml-cpu arch: arm/quants.c, x86/quants.c, powerpc/quants.c, s390/quants.c
Core NVFP4 support (type definition, CPU fallback dot product,
quantization, dequantization, conversion) is retained.
* Fix arch-fallback.h: add NVFP4 generic fallback for all platforms
After shelving backend-specific SIMD implementations, the generic
CPU dot product needs to be aliased on ARM, x86, PowerPC, and s390
platforms that previously relied on arch-specific versions.
* quantize: add NVFP4 as a quantization type option
* Fix ggml_fp32_to_ue4m3: handle subnormal values
Previously, values with ue4m3_exp <= 0 were clamped to 0, causing
all small scales to underflow. This made NVFP4 quantization via
llama-quantize produce garbage (PPL = 5.8M) since typical transformer
weights have amax/6.0 in the range 0.001-0.01, which falls in the
UE4M3 subnormal range.
Now subnormals are properly encoded as man * 2^-9 (exp=0, man=1..7),
matching the decode path in ggml_ue4m3_to_fp32.
Result: NVFP4 requantization now produces PPL = 15.25 (vs F16 = 14.33),
comparable to Q4_1 (PPL = 15.81) at slightly lower BPW (4.70 vs 5.15).
* Restore ARM NEON NVFP4 dot product implementation
Restores the optimized ggml_vec_dot_nvfp4_q8_0 for ARM NEON using
vqtbl1q_s8 lookup and ggml_vdotq_s32 dot products.
tg128 performance: 4.37 t/s (generic) -> 13.66 t/s (NEON) = 3.1x speedup
* Optimize ARM NEON NVFP4 dot product: LUT + vpaddq + vfmaq
- Add ue4m3_scale_lut[128] to ggml-common.h replacing branch-heavy
ggml_ue4m3_to_fp32() in the hot loop
- Use vpaddq_s32 for pairwise int32 reduction instead of vaddvq_s32
- Accumulate with vfmaq_f32 into float32x4_t vector accumulators
tg128: 8.1 -> 31.0 t/s (3.8x speedup, 77% of Q4_1 speed)
* ARM NEON NVFP4: rearrange q8 to match nibble layout
Alternative approach: rearrange q8 data to match the NVFP4 lo/hi
nibble layout instead of rearranging the looked-up NVFP4 values.
Eliminates vcombine_s8(vget_low, vget_low) shuffles.
Performance is equivalent (~18.5 t/s) - the bottleneck is the 2x
block overhead from QK=16 vs QK=32, not the shuffle instructions.
* CPU only backend 64 super-block layout
* cleanup
* Remove unused LUT
* int
* exclude NVFP4 from unsupported ops in metal build
* remove quantization for now
* store scales as native UE4M3, preserve original model bits when possible
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* correct comment
* format
* reduce duplication and cleanup
* Address comments
* move detection to prepare_tensors
* Use math instead of const
* Move
* fix comment
* Shelf quantize tests
* Rebase and move check
* cleanup
* lint
* Update gguf-py/gguf/scripts/gguf_convert_endian.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Use fallback quant config
* Simplify
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* organize
* Refactor
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* add quantize_nvfp4 (required for test_quants.py)
* add quantize_nvfp4 (required for test_quants.py)
* add quantize_nvfp4 (required for test_quants.py)
* fix return type
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* llama : add support for Nemotron 3 Super
This commit adds support for the Nemotron 3 Super model (120B.A12B)
enabling this model to be converted to GGUF format and run in llama.cpp.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Matt Clayton <156335168+mattjcly@users.noreply.github.com>
* tests: add end-to-end tests per model architecture
* fixup for rebase
* fix use-after-free in llama-model-loader.cpp
* fix CI
* fix WebGPU
* fix CI
* disable CI for macOS-latest-cmake-arm64
* use expert_weights_scale only if != 0.0f
* comments
* models : add llm_build_delta_net_base
* cont : keep qwen35 and qwen35moe graphs intact
* cont : add comments [no ci]
* add kimi linear to delta-net-base
* removed unnecessary ggml_cont from g_exp_t
* removed ggml_cont from g_diff_exp_t. moved ggml_cont for o to kimi-linear.cpp
* removed unnecessary diag mask
* cont : simplify
* cont : avoid graph splits
* scale q after mul instead of beginning
* scale q after mul instead of beginning
* identical ppl
* cont : fix scale and decay mask
* minor : remove TODO
* block implementation for kda
* remove space at the end of line 101
* concat+pad
* pad+binary row concat
* chunk size 16 for kda
* removed minor differences to master
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* WIP: Add EuroBERT support with autoformatting changes
This commit includes:
- EuroBERT model implementation for GGUF conversion
- C++ backend support for EuroBERT architecture
- Unintended autoformatting changes to Python files
Saving before reverting formatting-only changes.
* feat: add back eos assert when not last token pooling
* feat: removed duplicated code and cleanup
* feat: removed not working architectures and unnecessary check
* fix: typo
* fix: dynamic pooling config
* feat: added an example model for eurobert
* feat: proper llama-vocab implementation for jina-v5
* fix: removed unnecessary comments
* model: add JAIS-2 architecture support
Add support for the JAIS-2 family of Arabic-English bilingual models
from Inception AI (https://huggingface.co/inceptionai/Jais-2-8B-Chat).
Architecture characteristics:
- LayerNorm (not RMSNorm) with biases
- ReLU² (ReLU squared) activation function
- Separate Q/K/V projections with biases
- Simple MLP without gate projection (up -> act -> down)
- RoPE positional embeddings
- GPT-2 BPE tokenizer
Supported model sizes:
- Jais-2-8B (32 layers, 26 heads, 3328 hidden)
- Jais-2-70B (68 layers, 56 heads, 7168 hidden)
Tested with quantizations: BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, Q4_K_M, Q4_0, Q3_K_M, Q2_K
Note: JAIS-2 requires F32 precision accumulators for numerical stability
and uses standard attention (not flash attention) on CUDA backends.
* fix: run convert_hf_to_gguf_update.py for jais-2 tokenizer hash
* fix: use NEOX RoPE type for JAIS2
* fix: remove Q/K permutation (NEOX RoPE doesn't need it)
* fix: enable flash attention for JAIS2 (fixed by #19115)
* fix: add dedicated JAIS2 pre-tokenizer type and control vector support
- Add LLAMA_VOCAB_PRE_TYPE_JAIS2 with cascading whitespace regex
- Include original regex from tokenizer.json as comment
- Add build_cvec call for control vector support
* no longer necessary to override set_vocab
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* model : Add tokenizer from LFM2.5-Audio-1.5B
[LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) introduced lightweight audio tokenizer.
Tokenizer based on LFM2 architecture and acts as "embedding" model with
different input `n_embd` and output `n_embd_out`.
To be used in https://github.com/ggml-org/llama.cpp/pull/18641.
To convert use
```shell
python3 convert_hf_to_gguf.py /path/to/LFM2.5-Audio-1.5B/audio_detokenizer
```
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Formatting
* Rework check for attention layers
* Add LFM2 SWA model support
* Address PR feedback
* Set vocab to none
* Move helper function definitions to cpp file
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* full modern bert support
* added gelu op in rank pooling for modern bert
* still working on stuff, added mean calculation before classifier head
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* first layer is dense, as per modern bert research paper
* Update src/llama-graph.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* fixed set input for mean pooling to check if pooling type is ranking since modern bert does mean & rank
* Update src/llama-graph.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* models : add llm_build_delta_net_base
* cont : keep qwen35 and qwen35moe graphs intact
* cont : add comments [no ci]
* add kimi linear to delta-net-base
* removed unnecessary ggml_cont from g_exp_t
* removed ggml_cont from g_diff_exp_t. moved ggml_cont for o to kimi-linear.cpp
* removed unnecessary diag mask
* cont : simplify
* cont : avoid graph splits
* scale q after mul instead of beginning
* scale q after mul instead of beginning
* identical ppl
* cont : fix scale and decay mask
* minor : remove TODO
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* support qwen3.5 series
* remove deepstack for now, and some code clean
* code clean
* add FULL_ATTENTION_INTERVAL metadata
* code clean
* reorder v heads for linear attention to avoid expensive interleaved repeat
* Unified delta net handling
* Remove old methods.
* Refactor and optimize
* Adapt autoregressive version from @ymcki
* Change to decay mask approach
* Fix bad permute
* Qwen 3.5 support
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Further fixes
* Use inheritance, remove unneeded conts
* Not like this!
* Remove ggml.h explicit import
* Remove transformers, fix the views
* ACTUALLY fix views, make super calls explicit in conversion.
* Fix conversion again
* Remove extra ggml.h imports
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* kimi linear model implementation
* kimi linear convert_hf_to_gguf
* kimi linear constants.py tensor_mapping.py
* Kimi Linear ggml.h
* kimi linear ggml-cpu
* Kimi Linear ggml-cuda
* Kimi Linear ggml.c
* kimi linear src/llama
* remove "const int64_t n_seq_tokens = q->ne[2];" to get rid of unused variable warning
* remove type mismatch warning
* read MoE params
* removed some hard coded code
* removed all hard code
* use DeepseekV2 tokenizer
* removed unnecessary internal methods called by the old set_vocab of KimiLinear
* rewrite get_vocab for KimiLinear. Removed all kda_scan code
* removed all traces of kda_scan
* reduce OP count by 1 due to removal of kda_scan
* Move KIMI_LINEAR to llm_arch_is_hybrid to enable KV cache
* set n_embd_head_k/v to ensure kv cache works
* don't quantize conv1d of Kimi Linear
* Kimi Linear backend agnostic
* removed LOG_INFO
* naive chunking form implemented
* fixed some comments
* add Kimi-K2 specific tokens to be recognized as EOG
* build_kda_autoregressive is implemented to replace build_kda_recurrent for faster inference. sync'd to b7682
* replaced Akk and Aqk with mul_mat and clamp
* no clamp version
* Moved Aqk computation out of the loop
* fixed typo and split wkv_b into wk_b and wv_b
* MLA KV cache support
* fix trailing spaces
* moved const llama_model & model; around to follow qwen3next format and see if it cna pass the -Wunused-private-field error
* fix trailing whitespace
* removed traling whitespaces in empty line + make sure indentation is multiple of 4
* try to make lint happy
* remove blank lines to make lint happy
* removed at least blank line containing white space
* fixed flake8 complaints locally
* return ggml_tensor * pair in kda_autoregressive and kda_chunking as in ngxson's Qwen3Next improvement
* removed Kimi-Linear specific change that causes failure at server-windows
* removed private: from kimi_linear to make build checks happy
* removed unnecessary ggml_cont before ggml_reshape
* created static function causal_conv1d to abtract similar code for q/k/v
* merged dt_bias to SSM_DT. Do -exp(log_A) in convert_hf_to_gguf.py.
* reverted to original
* fixed find_hparam calls. Fixed e_score_correction_bias to use bias instead of weight. Removed all ssm_conv bias terms.
* remove DT_B from constants.py. remove one comment line in llama-model.cpp
* new class llm_graph_input_mem_hybrid_k to get around the new MLA change. switch the concat order of ggml_concat calls in kimi-linear.cpp to accommodate MLA changes. Removed support for exp_probs_b.weight
* remove ssm_o_norm_b
* remove ssm_o_norm_b
* changed hparams.kda_head_dim to hparams.n_embd_head_kda. added TODO comment for class llama_graph_mem_hybrid_k
* removed all ggml_cont b4 ggml_reshape_4d
* Whitespace
* replaced all hparams.get with find_hparams
* added new names for n_experts, n_experts_used and score_func in TextModel and removed their code in KimiLinear in convert_hf_to_gguf.py. Removed unnecessary ggml_cont and GGML_ASSERT in kimi-linear.cpp
* use is_mla to switch between different mem_hybrid types
* fixed logical errors in convert_hf_to_gguf.py pointed out by CISC
* removed if else for required parameters kv_lora_rank and qk_rope_head_dim
* add back ggml_cont for Vcur
* minor changes
* removed extra line in llama-vocab.cpp. Added back the comment in llama-graph.cpp
* f16 gguf cannot run without context length
* made a mistake of adding back n_ctx parsing
---------
Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
* graph : avoid branches between embedding and token inputs
* models : make deepstack graphs (e.g. Qwen3 VL) have constant topology
* ci : enable -DGGML_SCHED_NO_REALLOC=ON for server CI
* cont : pad token embeddings to n_embd_inp
* mla : pass V as a view of K to the FA op
* cuda : adjust mla logic to new layout
* kv-cache : fix rope shift
* tests : remove comment
* cuda : fix reusable_cutoff
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* qwen3next: simplify qkvz projection
* use ggml_swiglu_split
* revert swiglu_split, but remove redundant repeat()
* fix missing reshape
* rm 2 redundant transposes
* move mul_mat(k,q) to outside of chunking
* rm redundant cont
* improve g_cs_chunk
* add comments about no cont
* use std::pair instead of ggml_concat
* vectorize key_gdiff calculation
* rm unused tensor
* avoid ggml_concat inside loop
* bring back ggml_concat as it may not work on other backend
* nits
* Add Gemma3nVisionModel - MobileNetV5 vision encoder convertor to convert_hf_to_gguf.py. Add gemma3n to vision projectors in gguf-py/gguf/constants.py.
* Add mobilenetv5 impl
* Fix comments, remove unused vars
* Fix permute and remove transpose of projection weights
* Fix comments, remove debugging prints from hf_to_gguf
* 1. Hard-code image_mean = 0 and image_std = 1
2. Use available tensor mapping logic
3. Remove redundant chat template replacement of soft tokens placeholder with media placeholder
* 1. Move mobilenetv5 helpers declarations to `clip_graph_mobilenetv5` struct and definitions to mobilenetv5.cpp
2.Remove unused `clip_is_gemma3n` func declarations and definitions
3. Remove redundant `rescale_image_u8_to_f32` func and use `normalize_image_u8_to_f32` with zero mean and unit std
4. Calculate n_patches using image_size / patch_size
* Remove obsolete comments
* - convert_hf_to_gguf.py & constants.py & tensor_mapping.py: Use explicit mapping: Custom map for double indexed blocks and tensor_mapping.py for rest
- convert_hf_to_gguf.py: Unsqueeze Stem Bias and Layer scale tensors to correct shape while converting to gguf
- mobilenetv5.cpp: Remove explicit reshaping of Stem Bias and Layer scale which are now handled while converting to gguf, replace fprintf with LOG_*
- clip.cpp: Remove unused embedding and hard_emb_norm tensor loading
* - Rename tensors to v.conv..., v.blk..., v.msfa... to better align with already existing terminology
* Fix stem conv bias name
* Remove explicit handling of bias term for stem conv
* - Change order of addition in "project_per_layer_inputs" to support broadcasting of vision inp_per_layer
- Simplify the vision embeddings path of "get_per_layer_inputs" to output [n_embd_altup, n_layer, 1], broadcastable
* clean up conversion script
* fix code style
* also preserve audio tensors
* trailing space
* split arch A and V
* rm unused gemma3 func
* fix alignment
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* Add Maincoder model support
* Removed SPM model vocabulary setting and MOE related GGUF parameters
Removed trailing spaces from maincoder.cpp
* removed set_vocab
* added new line
* Fix formatting
* Add a new line for PEP8
Georgi Gerganov
Richard Davison
Daniel Bevenius
Xuan-Son Nguyen
Sigbjørn Skjæret
Johannes Gäßler
Aman Gupta
ymcki
Marcel Petrick
Maximilian Werk
megemini
3 a l i
Tarek Dakhran
Ryan Mangeno
JJJYmmm
Piotr Wilkin (ilintar)
forforever73
Christian Kastner
Junwon Hwang
Simranjeet Singh
Prabod