* models: use flash-linear-attention's l2norm for gated delta net q/k
The GDN q/k normalization is defined by flash-linear-attention as
l2norm(x) = x * rsqrt(sum(x*x) + eps)
with eps inside the root. Every GDN call site in the tree uses ggml_l2_norm
instead, which is x / max(sqrt(sum(x*x)), eps), i.e.
torch.nn.functional.normalize - its CUDA kernel cites that page.
The clamp never engages at these magnitudes, so in practice llama.cpp
normalizes with no epsilon at all where the reference has one inside the
root.
transformers made the same substitution when it first added Qwen3-Next and
corrected it three days later in huggingface/transformers#40842, 'Fix the
misalignment between the l2norm in GDN of Qwen3-Next and the implementation
in the FLA library'. vLLM and SGLang vendor FLA rather than reimplementing
it, so neither ever had the clamp.
eps keeps coming from the checkpoint, exactly as every call site already
passed it. The references hardcode 1e-6 for this norm; that is a separate
question and the two agree on every GDN checkpoint in the wild.
ggml_l2_norm itself is correct and unchanged, as is rwkv7-base, its original
caller, which passes normalize's own default eps of 1e-12.
No new ggml op: rms_norm already carries eps inside the root, so
rms_norm(x, eps/n) * (1/sqrt(n)) is exactly x * rsqrt(sum(x*x) + eps).
* Update src/models/models.h
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-array loading
G1/G2 infrastructure for variable-per-layer expert FFN size and top-k routing
(required for Puzzle-75B which has 5 distinct n_ff_exp values and 7 top-k values
across its 40 MoE layers).
Design: rename scalar members to _impl suffix (following existing convention),
add LLAMA_MAX_LAYERS arrays, add n_ff_exp(il)/n_expert_used(il) accessors with
scalar fallback. No new GGUF keys: reuses existing expert_feed_forward_length and
expert_used_count keys via get_key_or_arr (scalar -> broadcast, array -> per-layer).
- llama-hparams.h: n_ff_exp -> n_ff_exp_impl, n_expert_used -> n_expert_used_impl;
add n_ff_exp_arr / n_expert_used_arr arrays; add per-layer accessor declarations.
- llama-hparams.cpp: implement n_ff_exp(il) and n_expert_used(il); out-of-range
il returns impl safely (shared code, no abort).
- llama-model.cpp: central n_expert_used load changed to get_key_or_arr; derive
impl as max-of-array for validations and backward compat; zero both new arrays;
HunyuanVL override also zeroes n_expert_used_arr.
- llama-graph.cpp: aggregation loop in build_moe_ffn uses hparams.n_expert_used(il)
so per-layer top-k bounds the ggml_view loop correctly.
- All other files: mechanical rename hparams.n_{ff_exp,expert_used} -> *_impl.
Scalar arches are unaffected (broadcast fills all array slots with the single value).
(cherry picked from commit 269a81e03d66e1c353e1a203a0c03a03eb2b1a4e)
* nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE call-sites
Load n_ff_exp via get_key_or_arr into hparams.n_ff_exp_arr in load_arch_hparams;
derive impl as max for existing uniform GGUFs.
In load_arch_tensors, compute n_ff_exp_i = hparams.n_ff_exp(i) with fallback to
n_ff(i)/n_expert_used(i) for GGUFs that omit expert_feed_forward_length.
In build_ffn_layer, pass hparams.n_expert_used(il) to build_moe_ffn so per-layer
top-k is used for expert routing selection.
All other nemotron-h behaviour (mamba2, attention, shared-exp, latent projection,
routed_scaling_factor, expert_weights_norm, sigmoid gating) is unchanged.
(cherry picked from commit b1878a101793cd4e59868ac72635a86ea694987c)
* arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_expert_used->n_expert_used_impl
All non-nemotron arch files continue using the scalar impl member directly.
Behaviour is identical: the impl value is the broadcast value from the GGUF scalar.
gguf_writer: add_expert_feed_forward_length and add_expert_used_count now accept
int | Sequence[int], mirroring add_feed_forward_length, so converters can write
per-layer arrays with the same existing GGUF keys.
(cherry picked from commit 8f009f54bea5ef9a6a354123bd25e9d5ea2d5e03)
* convert: support NemotronHPuzzleForCausalLM (per-block MoE config)
Parse block_configs/mtp_block_configs into per-layer arrays (scalar-or-array
keys), append the MTP [attention, moe] sub-blocks as blk.88/blk.89 with
nextn tensors, accept the backbone.* prefix, and register the arch.
Also fix a pre-existing undeclared _experts attribute on NemotronHModel.
(cherry picked from commit d1a592f278336e78457454eb6c96bca917135f10)
* nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.A12B)
Both have 88 layers; the per-layer expert_used_count array (heterogeneous
for Puzzle, broadcast-uniform for Super) is the discriminator.
(cherry picked from commit f824e09dc812169589cf5662c92d149a4c18c30a)
* convert: accept the official Puzzle BF16 checkpoint's tensor naming
The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-
75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings,
model.norm_f) where the original release used the NemotronH-style
backbone.*, and spells the router bias e_score_correction_bias instead of
e_score_correction.bias. Normalize both at the top of
NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every
tensor name in the official index (42683 keys, MTP head included) resolves
through the tensor map after normalization.
(cherry picked from commit 189b67fc2c9d50970416c94b3317a6e7baa49b03)
* laguna: use n_ff_exp_impl for the uniform-MoE FFN size
Laguna landed after this branch was cut and reads hparams.n_ff_exp as a
scalar. This series turns it into a per-layer array with an n_ff_exp(il)
accessor, so the three scalar reads no longer compile. Laguna is a
uniform MoE, so point them at the scalar fallback n_ff_exp_impl, same as
deepseek2/qwen3moe/gemma4 in this series. No behaviour change.
(cherry picked from commit dbedc9e19c50dca0acdfb402362e2707bee424ae)
* arch: extend the n_ff_exp/n_expert_used rename to archs added upstream
kimi-k3, dflash, bailingmoe3, deepseek4, granite-swa and the nemotron-h MTP
block still referenced the scalar fields by their old names. n_ff_exp and
n_expert_used are accessors now, so those reads no longer compile; point the
non-per-layer archs at the _impl scalars and use the indexed form where the
call site is per-layer.
* convert: keep Puzzle opted out of the NemotronH MTP export path
#26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers.
Puzzle's config carries that key, but NemotronHPuzzleModel bypasses
NemotronHModel.__init__ (its per-block config needs a different setup), so
_mtp_bid was never assigned and modify_tensors raised AttributeError on any
mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the
mtp.layers.* form the base maps.
Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare
supports_mtp_export = False so --mtp / --no-mtp fail at the CLI.
* llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors
Follow-up to review feedback: the previous revision kept the scalar
hparams fields alongside the new per-layer arrays, which duplicated
state that get_key_or_arr already handles by broadcasting a scalar
value over every layer.
Drop both scalars and expose n_ff_exp(il) / n_expert_used(il) built
exactly like the existing n_head_kv(il) and n_ff(il) accessors: they
index the array and GGML_ABORT out of range, with il defaulting to 0
so genuinely uniform call sites stay a plain n_ff_exp().
Arch loaders now read both keys through get_key_or_arr over
n_layer_all, and the n_expert_used validation checks the maximum
across layers instead of a single field.
* llama: restore per-key required flags on the expert hparam reads
The scalar-to-array conversion passed required=false at every call site,
which silently made mandatory keys optional. Each read now carries the
same required flag it had before the conversion.
* Add control vector functions to qwen3.5 and qwen-next models
* Add missing cvec compatibility to the rest of the models
* Adjust comments and formatting
* cleanup
* whitespace
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.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 206890897546bd16602c3b79394fd5ea09ef199f
* 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>
* 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
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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>
