Pascal [Fri, 8 May 2026 09:44:09 +0000 (11:44 +0200)]
cuda: fuse snake activation (mul, sin, sqr, mul, add) (#22667)
* cuda: fuse snake activation (mul, sin, sqr, mul, add)
Add ggml_cuda_op_snake_fused with F32 / F16 / BF16 templates. 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.
Add test_snake_fuse comparing CPU naive vs CUDA fused across
F32 / F16 / BF16.
* cuda: address review feedback from @am17an
Use ggml_cuda_cast for F32/F16/BF16 conversions and rename
kernel_snake to snake_kernel to match upstream conventions.
* cuda: snake fusion fastdiv on T_len, Suggested-by: @am17an
Michał Piszczek [Fri, 8 May 2026 03:55:48 +0000 (05:55 +0200)]
convert : fix RuntimeError when stripping FP8 KV-cache scales (#22818)
* convert : fix RuntimeError when stripping FP8 KV-cache scales
In ModelBase._generate_nvfp4_tensors the final cleanup loop iterates
self.model_tensors.keys() and calls del on the same dict, which raises
RuntimeError: dictionary changed size during iteration when a ModelOpt
NVFP4 model also has FP8 KV-cache scales (e.g. mmangkad/Qwen3.6-35B-A3B-NVFP4
and any modelopt config with kv_cache_quant_algo: FP8).
Wrap the keys view in list() so the deletions happen on a snapshot.
Signed-off-by: Chun Tao <redacted> Signed-off-by: Todd Malsbary <redacted> Co-authored-by: Chun Tao <redacted> Co-authored-by: Todd Malsbary <redacted>
Daniel Bevenius [Wed, 6 May 2026 11:50:44 +0000 (13:50 +0200)]
convert : ignore non-language tensors for Gemma4Model (#22753)
* convert : ignore non-language tensors for Gemma4Model
This commit adds a check to make sure only text language tensors are
handled in filter_tensors.
The motivation is that currently when trying to convert a Gemma4 model
the following error occurs:
```console
(venv) $ ./convert-gemma.sh
INFO:hf-to-gguf:Loading model: gemma-4-E2B-it
INFO:hf-to-gguf:Model architecture: Gemma4ForConditionalGeneration
INFO:hf-to-gguf:gguf: indexing model part 'model.safetensors'
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...
INFO:hf-to-gguf:rope_freqs.weight, torch.float32 --> F32, shape = {256}
Traceback (most recent call last):
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 13752, in <module>
main()
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 13746, in main
model_instance.write()
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 945, in write
self.prepare_tensors()
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 805, in prepare_tensors
for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 7925, in modify_tensors
yield from super().modify_tensors(data_torch, name, bid)
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 7290, in modify_tensors
yield from super().modify_tensors(data_torch, name, bid)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 579, in modify_tensors
new_name = self.map_tensor_name(name)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/danbev/work/llama.cpp/./convert_hf_to_gguf.py", line 572, in map_tensor_name
raise ValueError(f"Can not map tensor {name!r}")
ValueError: Can not map tensor 'model.embed_vision.embedding_projection.weight'
```
Adrien Gallouët [Tue, 5 May 2026 10:16:25 +0000 (12:16 +0200)]
common : fix missing-noreturn warnings when compiling with clang 21 (#22702)
common/arg.cpp:3719:9: error: function 'operator()' could be declared with attribute 'noreturn' [-Werror,-Wmissing-noreturn]
3719 | [](common_params & /*params*/, int /*value*/) {
| ^
common/arg.cpp:3726:9: error: function 'operator()' could be declared with attribute 'noreturn' [-Werror,-Wmissing-noreturn]
3726 | [](common_params & /*params*/, int /*value*/) {
| ^
common/arg.cpp:3733:9: error: function 'operator()' could be declared with attribute 'noreturn' [-Werror,-Wmissing-noreturn]
3733 | [](common_params & /*params*/, int /*value*/) {
| ^
common/arg.cpp:3740:9: error: function 'operator()' could be declared with attribute 'noreturn' [-Werror,-Wmissing-noreturn]
3740 | [](common_params & /*params*/, int /*value*/) {
| ^
common/arg.cpp:3747:9: error: function 'operator()' could be declared with attribute 'noreturn' [-Werror,-Wmissing-noreturn]
3747 | [](common_params & /*params*/, int /*value*/) {
| ^
Georgi Gerganov [Tue, 5 May 2026 03:35:27 +0000 (06:35 +0300)]
server : validate --tools CLI argument against known tool names (#22538)
Previously, unknown tool names passed via --tools were silently ignored.
Now the server validates each tool name at startup and exits with an
error if an unrecognized tool is specified, listing the available tools.
Georgi Gerganov [Mon, 4 May 2026 05:52:07 +0000 (08:52 +0300)]
docs : update speculative decoding parameters after refactor (#22397) (#22539)
* docs : update speculative decoding parameters after refactor (#22397)
Update docs/speculative.md to reflect the new parameter naming scheme
introduced in PR #22397:
- Replace --draft-max/--draft-min with --spec-draft-n-max/--spec-draft-n-min
- Replace --spec-ngram-size-n/m with per-implementation variants
- Add documentation for all new --spec-ngram-*- parameters
- Update all example commands
Assisted-by: llama.cpp:local pi
* pi : add rule to use gh CLI for GitHub resources
Assisted-by: llama.cpp:local pi
* docs : run llama-gen-docs
Llama-architecture q_proj/k_proj weights need an axis-0 row permutation
to match GGML's RoPE convention. The BF16 path applies this in
LlamaModel.modify_tensors via LlamaModel.permute, but the NVFP4 path
bypasses modify_tensors and writes weights directly through
ModelBase._repack_nvfp4. Without the permutation, attention heads end
up scrambled at inference and the model produces gibberish.
This change overrides _repack_nvfp4 on LlamaModel and applies the same
permutation to both the nibble-packed weight and the per-block scale
before delegating to ModelBase._repack_nvfp4 via super(). Reuses the
existing LlamaModel.permute static helper and respects the existing
undo_permute flag, so subclasses (Mistral, Granite, Llama4, etc.)
inherit the fix automatically.
Verified on TinyLlama-1.1B reproducer: perplexity drops from 4419
(gibberish) to 43.9, matching the BF16-dequantized baseline (44.0).
Also verified end-to-end on ALIA-40b-instruct-2601 (BSC, Llama
architecture) with multilingual generation in Spanish/Catalan/Basque/
Galician all coherent with the fix applied.
Yiwei Shao [Sat, 2 May 2026 03:29:13 +0000 (20:29 -0700)]
hexagon: hmx flash attention (#22347)
* hmx: extract shared interleave headers and unify matmul batched
* hmx: add HMX-accelerated flash attention for prefill
* hmx: replace asm wrappers with Q6_ intrinsics in hmx-utils.h
Switches three single-instruction helpers from inline asm to the matching
Q6_ intrinsics, matching the style established by aizip f8737609a and used
by the upstream PR #21554 hmx-matmul-ops.c rewrite:
hmx_load_tiles_fp16 stays on inline asm: it uses ":deep" activation
streaming, and the mixed Q6_activation_hf_mxmem_RR_deep + non-deep
Q6_weight_hf_mxmem_RR pair fails the HMX backend constraint check
("activate weight pair (1) exceeds limit (1)"). The asm bundle keeps
both halves in one VLIW packet and avoids the diagnostic.
Functionally equivalent — same instructions emitted; the Q6_ intrinsics
just give the compiler more visibility for scheduling.
* hmx: drop the duplicate interleave_fp16_weight_chunk_to_tiles
* hmx: apply upstream optimization to hmx-flash-attn-ops.c
apply restrict, __builtin_assume, and pointer accumulation to the three HMX workers (qk_dot, o_update, o_norm) and the matching inline HMX loops in op_hmx_flash_attn_ext.
* hmx: unify interleave helper
* hmx: multi-thread Q load / O store and enable prefill FA dispatch
Extract inline Q-load and O-store loops into worker_pool-parallel helpers
(fa_phase_q_load, fa_phase_o_store) so HVX threads split the F32↔F16
conversion work across row ranges. Also relax the softmax threading
gate from n_row_vec_cnt >= n_threads to >= 2, which was unnecessarily
forcing single-thread fallback when n_rows_g < 512.
On the dispatch side, remove the ne[2] != 1 guard that blocked multi-head
(prefill) FA from reaching the HTP backend — GQA is already handled
internally by both the HMX and HVX flash-attention paths.
* hmx: relax matmul pipeline gate to cover k > n shapes (e.g. FFN_down)
* hmx: optimize FA softmax mask phase (no-ALiBi fast path + GQA dedup)
* hmx: Add an asm memory clobber at the phase boundary to prevent reorder bug
* [experimental]: fp16 softmax (EXP2_HF) to accelerate fa
Bake log2(e) into qk_scale and use hvx_exp2_hf directly for P and m_diff
(base-2 consistent, matches htp-ops-lib). ~22 ALU ops for 64 lanes vs
~44 for the F32 round-trip path.
* hmx flash-attn: refine cost model coefficients based on profiling data
* hmx flash-attn: replace asm clobber with targeted volatile reads on vtcm_d_tiles
* hmx flash-attn: preserve additive mask bias in no-ALiBi fast path
The no-ALiBi fast path (max_bias==0) was skipping mask add entirely on
the assumption that mask values are only {0, -inf}. This is wrong when
the mask carries additive positional bias — those terms were silently
dropped. Keep the slope-mul skip (slope≡1.0) but add mask back so the
bias survives; vmux still clamps below -16 to -inf.
Also add HMX FA coverage to test-backend-ops: prefill shapes (nb=64,
nb=32) × {mask on/off} × {ALiBi on/off} × {softcap on/off}, F16 KV,
hs ∈ {64, 128}.
- flash-attn: when EXP2_HF is on AND logit_softcap is active, fold
log2(e) into the post-tanh multiplier (v_cap) instead of pre-baking
it into qk_scale. Pre-baking shifted the tanh knee from x≈c to
x≈c/log2(e) and produced numerically wrong softcapped outputs
whenever both knobs were enabled.
- flash-attn softmax (fa_softmax_thread): replace the union+memcpy
scalar extract pattern with HVX vmux-based per-row accumulators on
rowmax/rowsum. Add hvx_vec_get_f16 helper in hvx-base.h. Functional
parity, less scalar code, clearer hf/qf16 lane-format contract.
- matmul (hmx_mat_mul_permuted_qk_0_d16a32): pick pipeline vs sequential
layout based on whether the chunker actually yields >=2 n-chunks,
instead of the static (m>=128 && n>=256) gate. Avoids paying for
output double-buffer + worker dispatch when there is no HMX/HVX
overlap to gain (e.g. shapes that collapse to one n-chunk).
- tests: add HMX flash-attention coverage over the
{mask, ALiBi (max_bias), logit_softcap} cross-product for the prefill
path — head_dim 64/128, GQA 4×4, kv=512/nb=64 plus a kv=113/nb=32
non-aligned case.
* [Help Wanted]: refactor D matrix computation into separate function for clarity and maintainability
* format code
* hexagon: looks like -O3 is causing issues with the large code base, switch to -O2 and -flto instead
* hexagon: use hex_ prefix for swap_ptr
* hexagon: move vtcm_seq_alloc into vtcm-utils.h
More vtcm allocator updates are coming so it makes sense to start the separate hdr for it.
* hmx-utils: add hmx_prefix for layout converters
* hmx-mm: move main hmx_mm functions to the end, remove unused fwd decls, etc
* hmx-mm: remove unused qweight_fetch_task_state_t and minor alignment fixes
* hmx-fa: minor alignment fixes
* hmx-fa: move hmx_flash_atten into hmx-ops.h
* hmx-fa: remove redundant workpool pointer in the hmx_fa_ctx, plus minor alignment updates
* hmx-fa: minor alignment and simplifications
* hexagon: move FA_EXP_F16 option to hostside CMake file
* hmx-fa: use hvx_vec_splat_f16 instead of fp16_to_bits
* hmx-fa: add hvx_splat_u16/u8 and use that in the fa instead custom hvx_fill
* hmx-fa: some more alignment updates in the core fa function
* hmx-fa: keep slopes in vtcm in fp16
Saves malloc/free and removes the need for float -> fp16 downcast on every use.
Jeff Bolz [Fri, 1 May 2026 13:28:32 +0000 (15:28 +0200)]
vulkan: Support asymmetric FA in coopmat2 path (#21753)
* vulkan: Support asymmetric FA in coopmat2 path
There has been some recent interest/experimentation with mixed quantization
types for FA. I had originally designed the cm2 FA shader with this in mind
(because I didn't realize it wasn't supported at the time!), this change
adds the missing pieces and enables it.
Also support Q1_0 since people have been trying that out (seems crazy, but
who knows).
We should be able to do similar things in the coopmat1/scalar path, but
there's another change open against the scalar path and I don't want to
conflict.
overview / pkg / ggml / sources / llama.cpp / log