This commit removes the precompiled headers that I added in Commit
3bcfeb700 ("cmake : add PCH and unity build to improve build times
(#28091)").
The motivation for this is that this looked good when developing this
but has caused multiple issues that I had taken into consideration and
we have decided to remove it and only keep the unity builds from the
above commit.
Refs: https://github.com/ggml-org/llama.cpp/pull/28882#issuecomment-5662272126
* gguf-py: add Maple tensor constants
Add MODEL_ARCH.MAPLE, its "maple" name, and the tensor list for the
Maple 20B-A1B ternary MoE architecture: token embeddings, output,
attention with Q/K RMS norms, and per-expert FFN tensors.
* convert: add Maple HF->GGUF converter
Register MapleForCausalLM in the HF architecture map and add the
converter for the Maple 20B-A1B ternary MoE model: 24 layers, 256
experts with 8 active, sliding-window attention (SWA-512) interleaved
with global attention at a 3:1 ratio, partial rotary factor 0.5, and
per-expert weight stacking into merged 3D tensors.
* llama: add Maple architecture (20B-A1B ternary MoE)
Add the Maple 20B-A1B ternary MoE architecture: 24 layers, 256
experts with 8 active, sliding-window attention (SWA-512) interleaved
with global attention at a 3:1 ratio, and ternary TQ1_0/TQ2_0
quantization support.
- register LLM_ARCH_MAPLE between MAMBA2 and JAMBA
- implement llama_model_maple: Q/K RMS norms after projection (GEMMA4
style), rope applied only on SWA layers (nope_on_global_attention),
ISWA KV cache, and MoE FFN with swiglu gate clamp at +7 (DEEPSEEK4
style)
- mark MAPLE as unsupported by the model saver (roundtrip skipped)
* tests: mark Maple as MoE-mandatory
Maple is always-MoE: the model throws when n_expert == 0, so the
test harness must only run the MoE config for LLM_ARCH_MAPLE.
* maple: apply review feedback (n_ff_exp_arr, get_arr, rope params)
- load_arch_hparams: use n_ff_exp_arr + n_ff_exp() accessor (upstream
changed these from a scalar member during the rebase)
- sliding_window_pattern: get_arr, the pattern is mandatory for this arch
- partial_rotary_factor: read only from rope_parameters (base.py mirrors
the top-level key automatically)
- document why TOKEN_EMBD/OUTPUT are forced to F16 (they are the two
dense tensors in Maple, and the reference GGUFs ship them as F16)
- add @ModelBase.example("deepgrove/maple-preview")
* tests: add Maple to the SWA pattern array list
get_arr for maple.attention.sliding_window_pattern requires an array, but
the harness only emitted a per-layer array for the arches in its list, so
test-llama-archs -a maple failed to load the model.
Assisted-by: DeepSeek Harness
* maple: move swiglu_clamp_exp to the converter
The loader prefilled 7.0 and read the key optionally. The converter now
writes it and the loader reads it as required, because llama-graph.cpp
skips the clamp when the limit is 0 and an optional read would silently
run unclamped. The test harness provides the key for the same reason.
Also drops tensor_force_quant: base.py already forces FFN_GATE_INP to F32
and TOKEN_EMBD/OUTPUT to F16 for ternary file types.
Assisted-by: DeepSeek Harness
* convert: fix the LazyBase func signature in the Maple converter
ty flagged the stack() closure: it takes no argument, while LazyBase is
annotated with func: Callable[[Any], Any]. Pass the tensor list through
args instead of closing over it, the same way kimi_k3 does, so the
callable shape matches.
Assisted-by: DeepSeek Harness
1) Combine two consecutive lookups (find + insert) into a single insert-attempt/lookup routine so that we don't per
form two O(log(n)) lookup operations in a row anymore -- we only need to do it once and then see if the insert succeeded.
2) Instead of copying every potential stack (expensive) and then moving it (cheap) to new_stacks when it's a final output state, we switch the order so that we move every potential stack (cheap), and then only copy it (expensive) to new stacks when it's a final output state. There are a LOT of intermediate states that get generated, and unless they become final output states, then all of these expensive intermediate copies are wasted.
Before: lookup -> lookup/insert + copy -> optional move to output
New: lookup/insert + move -> optional copy to output
The NextN/MTP tail loop derives the expert FFN size as n_ff/n_expert_used
when expert_feed_forward_length gives nothing for the layer. Both values come
from per-layer arrays that legitimately hold 0 on layers that are not MoE, so
a checkpoint whose predict layers hold 0 in both divides by zero and dies with
SIGFPE at load time, with no error message. Report the malformed metadata
instead.
* scripts : add initial profiling script (wip)
* src : add precompile headers (PCH) for models.h
* common : add common.h as PCH
* ggml : add PCH for ggml-impl.h
* mtmd : use PCH for models.h
* scripts : add script to build with Server/Tools/Tests
* server : add PCH for common.h
* docs: add profiling progress notes (wip)
* ggml : add exclude for GCC + SVE on ARM
Refs: https://github.com/ggml-org/llama.cpp/actions/runs/33393906061/job/99493756214?pr=28091
* ggml : attempt to fix use of std::hardware_destructive_inference_size
Refs: https://github.com/ggml-org/llama.cpp/actions/runs/33396221677/job/99501265689?pr=28091
* squash! ggml : attempt to fix use of std::hardware_destructive_inference_size
Add a version check for GCC 12 to conditionally apply the `-Winterference-size`
pragma.
* editorconfig : exclude profiling reports dir
This directory will not be included in the merge later and this commit
can be ignore at that point. Just fixing to keep CI happy.
* ggml : skip PCH for gcc on non-x86 architectures
* tests : add PCH for peg-parser/tests.h
There are 7 peg-parser tests that can share one PCH instead of then each
parsing the full tests.h.
* common : add PCH for chat.h
* docs : update linux build profiling full results
Just updating after a number of PCH additions. These are not exact
figures and will vary a bit from run to run, but they give a general idea
of the performance impact of PCH.
* cmake : introduce unity build for models
This commit introduces a unity build for the models to improve
compilation time.
The improvements were roughly the following:
```console
+------------------------+-----+------------+------------+------------+
| Build | TUs | Frontend | Backend | Total |
+------------------------+-----+------------+------------+------------+
| Full, master | 396 | 811.0 s | 692.2 s | 1,503.2 s |
| Full, with PCH | 405 | 380.0 s | 664.7 s | 1,044.7 s |
| Full, with PCH + UB | 264 | 357.7 s | 635.7 s | 993.4 s |
+------------------------+-----+------------+------------+------------+
TU = Translation Unit.
Full = includes Server, Tools, and Tests.
PCH = precompiled headers.
UB = unity build for models.
```
* docs : update linux profiling table with unitiy build results
* docs : update mac profiling results to include unity build [no ci]
* docs: remove profiling reports
* scripts : merge build profile scripts into one script
I was lazy before and just copied the first script to enable Tests,
Server, and Tools. This now merges them into a single script.
* Revert "editorconfig : exclude profiling reports dir" [no ci]
This reverts commit 2922a12118.
* src : rename ggml_view_2d_slice to gemma3n_view_2d_slice
This is to be consistent with the rename in gemma4.cpp which was
required to avoid a name clash.
* cmake : add build profile script for windows [no ci]
This commit adds a port of the scripts/build-profile.sh script to
windows powershell.
This was developed on Windows on ARM but should work on X64 as well but
needs to be tested there as well.
* metal : rework fusion patterns into a single table
All fusable op patterns for the Metal backend are now declared once in a
fusion table (ggml-metal-fuse.cpp) and consumed by both the graph optimizer
(ggml_metal_fuse_max, packing) and the op encoders (ggml_metal_fuse_next,
compute). The two phases share the same pattern table plus ggml_can_fuse_subgraph_ext
for the structural checks, and differ only in the mode used for the pattern
check (STRUCTURAL at optimize time, since tensors are not allocated yet, and
FULL at compute time, including Metal buffer placement). This also protects the
snake activation (MUL + SIN + SQR + MUL + ADD) from being reordered during graph
optimization, which was previously unprotected.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : fix absolute output indices in fusion patterns
ggml_can_fuse_subgraph_ext expects the outputs array to contain absolute graph
node indices (it indexes cgraph->nodes[outputs[i]]), but the fusion table query
was passing a relative index (n_ops - 1). As a result the last node of every
pattern was not recognized as an output and was subjected to the elidable
use-count check, which failed for essentially all fusions. This silently
disabled the norm/MUL fusion and caused a ~5% token-generation regression.
Pass the absolute graph index of the last node instead.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : fuse gated_delta_net with cache cpy
Add GGML_METAL_FUSE_GDN_CACHE to the fusion table: when the gated_delta_net
kernel is followed by a cpy that scatters its recurrent state snapshots into
the KV cache, the kernel writes the snapshots straight into the cache buffer
and the trailing cpy is elided.
The gdn output has other consumers (the attn scores view), so unlike the
elision-chain patterns this is not a simple chain: a 'raw' flag on the fusion
pattern skips the generic chain/shape and ggml_can_fuse_subgraph_ext checks,
making the pattern-specific check callback the sole validator. Packing
(ggml_metal_fuse_max) now matches on the same view-transparent node sequence
that the compute phase uses, so the gdn + cache cpy group is packed along with
any intermediate views and stays adjacent through the reorder.
The fused cpy is a view consumer of the gdn (it writes the cache directly),
so its mem-range is skipped in the encoder; the skip is restricted to CPY
nodes consuming the previous fused node through a view so other fusions are
unaffected.
Add test_gated_delta_net_cache_fusion and register 5 cases.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : drop is_view_consumer mem-range skip
The is_view_consumer skip was carried over from the upstream gated_delta_net
cache-fusion draft, but it is not needed: keeping the elided cpy's mem-range in
the concurrency tracker only ever adds a (conservative) memory barrier at the
fusion point. It can never remove a barrier, so it cannot introduce a race. The
worst case is one spurious barrier per gdn+cache-cpy fusion, which is within
run-to-run noise on Qwen3.5-0.8B Q8_0.
Dropping the check keeps the mem-range loop uniform for all fused groups.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : rename gated_delta_net fused state output args
Rename the fused cache-write kernel argument to match the rest of the kargs:
state_out_stride -> nb_out (and widen it to uint64_t), and the local buffer id
bid_state_out -> bid_out.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : rename raw fusion flag to unsafe
raw did not convey that the flag opts a fusion pattern out of the generic
elision-chain safety net (ggml_can_fuse_subgraph_ext + chain/shape checks).
rename it to 'unsafe' to make explicit that the pattern's check callback is the
sole validator and must re-establish the safety guarantees itself.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : tidy fusion pattern checks and table
- const-correct ggml_metal_fuse_outputs buffer
- annotate unused check-callback parameters
- drop a redundant size_t cast
- align the ops/table initializers and add blank-line separation
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* metal : add generic fusion stats via ad-hoc proc-address API
Add a device-owned fusion context that lets a test tool count how many
times each fusion pattern fires and toggle fusion. It is exposed through
the ad-hoc ggml_backend_reg_get_proc_address mechanism with generic names
so the testing tool is backend-agnostic:
- ggml_backend_fusion_stats_init: start collecting fusion stats; when a
context is created afterwards it registers the labels/counters and
encodes single-threaded (n_cb == 0) so the counters are race-free
- ggml_backend_fusion_stats_reset / _get_stats / _set_enabled
The context lives on the metal device (not on the last backend context),
so counters accumulate across contexts and reads are always consistent.
The enable/disable toggle is initialized from GGML_METAL_FUSION_DISABLE
and can be overridden by the test through set_enabled. Labels are
synthesized from the fuse table via ggml_metal_fuse_label (e.g.
"GATED_DELTA_NET+CPY").
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : add fusion count regression test with per-backend baseline
test-fusion runs every dummy model generated by test-llama-archs on a
single backend (single-threaded encoding, n_cb == 0) with fusion enabled
and disabled, and for each mode (prefill / decode) reports the per-fusion
counters and the NMSE between the fused and unfused logits, plus the NMSE
against a CPU reference.
A fusion pattern that silently stops matching (or fires when it should
not) is caught as a regression by comparing the counters against a
committed per-backend TSV baseline:
- --record writes the golden baseline, --check (default) validates it
- the unfused run doubles as a control: its counters must be all-zero
- NMSE is skipped when it is NaN or the arch is already broken on the
device (e.g. plamo2 on Metal), so the count check is the hard gate
- baseline counts depend only on graph structure, not weights (verified
stable across weight seeds)
- the fusion stats API is resolved through the ad-hoc get_proc_address
mechanism with generic names; a backend that does not export it makes
the test fail with an error
The committed MTL0.tsv baseline covers 110 dummy archs (298 rows).
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : rename fusion api helpers to match stats_init signature
Align the test with the ad-hoc fusion stats API: fusion_stats_init no
longer takes an enable bool (stats are turned on by calling it), so the
proc-address wrappers and typedefs are renamed to the api_* convention.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : rename backend to device in fusion test CLI
The fusion test operates on a compute device (e.g. MTL0), not a backend,
so rename the --backend argument to --device and the backend_name
variable to device_name. Keep "backend" where it refers to the ggml
backend interface (the ad-hoc proc-address mechanism).
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : add --model and --help to fusion test
--model FILE runs the fusion regression test over a single model file
instead of enumerating a --models DIR. --models and --model are mutually
exclusive. Also add a --help/-h option that prints the usage.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : use backend base name for fusion baseline output
The fusion test is invoked with a specific device name (e.g. MTL0), but
its output - the recorded baseline and the header it writes - should be
named after the backend base name (e.g. MTL, via ggml_backend_reg_name),
since the counters depend on the backend, not on the specific device
index. Rename the committed baseline to MTL.tsv.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : run fusion test from ci instead of ctest
The fusion test needs Metal and generates a lot of dummy models, so it
does not belong in the generic ctest suite. Move it to ci/run.sh as
gg_run_test_fusion, gated on GG_BUILD_METAL like
gg_run_test_llama_archs_tensor_split: it generates the dummy models with
test-llama-archs -o and then validates the fusion counts against the
committed baseline. test-fusion.cpp is still built (llama_build) but no
longer registered as a ctest.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : align fusion baseline TSV columns
Pad the TSV fields to fixed widths so the columns line up regardless of
the variable arch and fusion-label lengths, and trim each field on parse
so the padded file is still accepted. Regenerate the committed MTL.tsv
baseline in the padded format (data unchanged, verified identical modulo
padding).
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : widen label column and align fusion TSV header
Give the label column more room (28 chars) and fix the column header
widths so they match the data rows (moe/mode/label), keeping the header
aligned with the values. Regenerate the MTL.tsv baseline in the new
format (data unchanged).
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : switch fusion baseline from TSV to CSV
Use comma-separated values like the rest of the project, keeping the
padded, aligned columns. Split on ',' and trim on parse. Rename the
committed baseline to MTL.csv (data unchanged, verified identical modulo
padding/separator). Update the ci/run.sh check path accordingly.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* cont : rebase + update MTL stats
* tests : avoid graph reallocations for some archs
* metal : tidy fusion debugging context and op init
- simplify the shared fusion debugging context comments
- shorten the ggml_metal_fusion struct comment
- align the ggml_metal_fuse struct fields and comments
- move the fusion parameter of ggml_metal_op_init right after dev
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : dedup fusion baseline into any mode
prefill and decode always produce the same per-graph fusion count, so
store a single row per label with mode = "any" and the per-graph count
instead of two rows. this halves the baseline size and keeps the check
stable.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* ci : move fusion model generation to a separate step
the dummy models generated by test-llama-archs are reused by other tests,
so generate them once in their own step instead of inside test_fusion.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : bump nmse thold
* models : fix plamo2 graph
* tests : remove "skip" logic from test-fusion
* tests : set qwen3tts dummy vocab to codec head size
the dummy qwen3tts model used a vocab of 4096 while the codec head is
3072, so the graph padded the output with -inf which made the NMSE in
test-fusion produce NaN. use the exact codec head size instead so the
padding is not generated at all.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* tests : regen fusion baseline
reflect the plamo2 graph fix, which changed its fusion pattern split
(RMS_NORM+MUL 11->10, RMS_NORM+MUL+ADD 3->4; same total).
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* ci : skip dummy model generation on OpenVINO
test-llama-archs does not build on the OpenVINO platform, so do not try
to generate the dummy models there.
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
* cont : minor
* tests : enable test-llama-archs on windows
* cont : disable on windows + workaround
* metal : naming nits
* test-fusion : add instructions to update baseline
* context : fix Kimi-K3 graph reserve
* fusion : update MTL
* cont : fix naming
* metal : rework fusion info storage
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp
* metal : align fusion info API
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp
* metal : use opaque fusion handle in ad-hoc API
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp
* ci : move fusion test to dedicated workflow
Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp
* cont : run only on ggml changes
* cont : simplify
* fusion : remove multi-output stuff for now
* ci : fix typo
Some of these if statements were copypastaed in a former refactor and
never cleaned up to remove the cases that could never happen anymore. The
only thing that's shared between these relatives anymore is
llama_model_bert::graph::graph, so the rest of the code doesn't need the
conditionals.
* 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>
* model: add Tencent Hy 4 (hy_v4) preview architecture support
Adds support for the Tencent Hy 4 model (Hugging Face architecture
HYV4ForCausalLM, GGUF arch hy_v4):
Add HF -> GGUF conversion script (conversion/hy_v4.py) and wire it into the conversion registry
Register hy_v4 GGUF constants, arch enum, and writer support
Implement the hy-v4 model graph, hparams, vocab and context changes
Register the new arch in llama-arch and models registry
Extend arch tests to cover hy_v4
Assisted by Claude Opus 5
* Update convert_hf_to_gguf_update.py
Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
* Update conversion/base.py
Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
* convert : move hy_v4 entry to the same place as in convert_hf_to_gguf_update.py
* model : apply changes related to n_ff_exp becoming per-layer in Hy4-preview
* n_layer_all
---------
Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* src : add n_expert_used_max function
With Commit c61b98b875 ("model: add
NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)") it
is now possible for each layer to have a specific number of experts but
there are a few checks that need to be updated to handle this upon model
loading. For example:
```console
llama_model_load: error loading model: model has expert layers but no expert layers are used
```
And later:
```console
/llama.cpp/src/llama-model-loader.cpp:955: GGML_ASSERT(n_ids_used > 0) failed
```
This commit adds the n_expert_used_max function so that these checks
can use it.
Refs: https://github.com/ggml-org/llama.cpp/pull/25444#issuecomment-5524976031
* src : use hparams.n_expert_used_max in llama_model_base::load_hparams
* src : use 0 as initial value for n_expert_used_max
* metal : support n_kv_max sparse mask hint in flash attention vec kernel
- add kernel_flash_attn_ext_vec_idx: compacts finite mask entries into
a per-row index list (Hillis-Steele scan, one threadgroup per row)
- extend vec FA kernel with optional sparse index gathering (FC slot 5)
- add host-side gate: sparse path when n_kv_max > 0, mask present,
supported head sizes / KV types, n_kv_max <= 4096
- new buffer region extra_idx for the index list
- pipeline getter extended with has_sparse param
- add test cases: head sizes, quant types, nb>1, nr23 variants,
sinks, ALiBi, softcap, permute, v_view_of_k, no-mask fallback
Note: multi-row (nb*nr23[1] > 1) cases still failing - rid mapping
in the store phase needs revisiting for the sparse path.
Assisted-by: pi:llama.cpp/Qwen3.8-27B
* metal : fix sparse flash attention row addressing
- kernel_flash_attn_ext_vec_idx: mask param is half* but nb31 is a byte
stride, so the per-row mask offset was scaled by 2x; cast to char*
before applying the byte strides
- kernel_flash_attn_ext_vec: sparse pidx param is char* so the per-row
element offset was under-scaled by sizeof(int); scale it by sizeof(int)
to get the correct byte offset
- fixes the multi-row (nb*nr23[1] > 1) sparse flash attention failures
Assisted-by: pi:llama.cpp/DeepSeek-v4-0731
* cont : use sparse vec FA for prefill
* metal : single-pass flash attention sparse index compaction
The idx kernel previously read the mask row twice: once to count the finite
entries (for the prefix scan) and again to recover their positions. Since the
kernel is memory-bound, this doubled the mask traffic.
Keep the finite positions in a per-thread register array during the count
pass and write them out directly, avoiding the second mask read. A dense
mask with more than NLOCAL finite entries in a slice falls back to re-reading
the mask to write the remaining positions.
Assisted-by: pi:llama.cpp/DeepSeek-v4-0731
* tests : add perf cases for sparse flash attention prefill
Measure the sparse vec FA kernel across KV sizes, n_kv_max hints and batch
sizes. Run with:
./build/bin/test-backend-ops -b MTL0 -o FLASH_ATTN_EXT -p "n_kv_max=[1-9]" perf
Assisted-by: pi:llama.cpp/DeepSeek-v4-0731
* qwen4 : enable sparse attention
* cont : adjust nsg
* cont : sync test-backend-ops
* cont : disable Qwen4 for now
* cont : clean-up + tests
* 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 269a81e03d)
* 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 b1878a1017)
* 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 8f009f54be)
* 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 d1a592f278)
* 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 f824e09dc8)
* 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 189b67fc2c)
* 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 dbedc9e19c)
* 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.
get_prev_tokens() rebuilt a (seq, pos) -> token hash map on every
ubatch by walking all used cells, while llama_kv_cells already keeps
an ordered index of the positions of each sequence in seq_pos, updated
on every cell mutation to serve seq_pos_min() and seq_pos_max().
The index now stores (pos, cell) pairs in a std::set instead of a
position -> count map, so a repeated position (cache reuse via rm + add,
vision inputs with shared positions) yields distinct entries and the
removal of a cell erases its own pair. The new seq_pos_tok_le() returns
the token of the cell at the largest position <= p in logarithmic time,
which is exactly what the old window lookup and its M-RoPE gap fallback
computed together.
get_prev_tokens() shrinks to a direct lookup per (token, offset) and
for_each_token_in() goes away with its only caller. The kv-cache keeps
no n-gram logic of its own.
Measured on Qwen3.8-Flash-Next UD-Q4_K_XL at 71k context, alternating
two binaries with the first run discarded: tg 69.3 -> 72.7 t/s (+4.9%),
pp unchanged at ~2720 t/s, greedy output identical, needle retrieved.
* qwen4exp: follow up fixes
* -kvu NaN collapse fix
Assisted-by: Claude
* indexer cache ext.x/ext.y restore fix
Assisted-by: Claude
* kv-cells: rename seq_set to seq_get_all
seq_get is already taken by the single-id getter, so the suggested name
cannot be overloaded on return type alone.
Assisted-by: Claude
* memory-hybrid-idx: implement set_input_qsa on the memory class
The context held the whole implementation, where the pattern elsewhere is a
thin context forwarding to the memory class, as llama_kv_cache_context does
for set_input_kq_mask. The body reads no context state, so it moves unchanged
and the context keeps a forwarder.
Also shortens the seq_get_all comment as suggested.
* tests: check that a sequence state survives a save/restore round-trip
Saves seq 0, erases it, restores the blob and saves again, requiring the two
blobs to match. Compares blobs rather than generated text, which cannot see a
field dropped on the way back in.
Note this passes on master for qwen4exp, so it does not demonstrate the
ext.x/ext.y drop this PR fixes; reaching that needs 2D mrope content.
* tests: give the synthetic qwen4exp a PLE so the state test bites
has_cell_ext() is n_pos_per_embd() > 1 || ple_n_heads > 0, and the indexer
cache sets rope_type = NONE, so without a PLE it serializes no cell ext at
all and the round-trip test cannot see a dropped ext.x/ext.y. With one,
removing the ext_set restore in state_read_meta fails the test: 198 of
335692 bytes differ, first at offset 282092.
Loading such a model needed two fixes:
- the row count of per_layer_token_embd came from require_weight(), which a
model synthesised from metadata alone has no file to answer. Derive it
from the head ranges and prefer the file's padded count where there is one.
- the PLE conv history is a row of the recurrent cache, so a PLE on a full
attention layer dereferenced a null p_l. Reject it at load time instead.
The meta mirror is skipped for qwen4exp. It returned NaN logits before this
fixture carried a PLE, which the nmse check passes since a NaN comparison is
false, and aborts with one. -sm tensor on real devices works.
Assisted-by: Claude
* llama: disable -sm tensor for qwen4exp
test-llama-archs skipped the tensor split for this arch from inside the
test, so the arch still advertised support it does not have. Declare it in
llm_arch_supports_sm_tensor instead and drop the test-side exception; the
existing llm_arch_supports_sm_tensor branch then does the skipping.
Assisted-by: Claude
MTP speculative decoding needs the target state to move back by the
number of rejected draft tokens. Without rollback support the context
is classified as SEQ_RM_TYPE_FULL and the server serializes the whole
recurrent state to host memory on every round, which costs more than
the drafting saves.
The recurrent cache already holds n_rs_seq + 1 snapshot planes and the
delta net writes its SSM state into them, but build_conv_state_at wrote
a single plane, so a rollback restored a convolution history that was
never captured. It now writes one snapshot per slot, each ending one
token earlier, for the delta net QKV convolution and for the PLE
convolution alike.
Measured on Qwen3.8-Flash-Next UD-Q4_K_XL with the standalone MTP
draft, n-max 3 and a single slot: decoding reaches 183 tok/s on code
and 144 tok/s on prose. The same branch before this change, where the
server falls back to checkpointing the state to host memory, reaches
123 and 83 tok/s, for 108 tok/s without a draft.
* qwen4exp: sum the indexer heads by slices
The head reduction went through a transpose and a sum_rows over ne[1],
which left sum_rows with ne0 = 4, one block per row for a four element
reduction, and the transpose copied the whole block by token surface
twice on the way in.
The heads are adjacent on ne[1], so each one is a strided view and the
sum is a short chain of adds.
RTX PRO 6000, Qwen3.8-Flash-Next UD-Q4_K_XL, fa on, 55k context, warm
runs on top of #28011:
prompt processing 2170 -> 2366 t/s
Generation is unaffected. The removed work scales with n_blocks by
n_tokens, so the gain grows with context and with ubatch size.
* qwen4exp: drop the redundant cont on the indexer query
rope returns a freshly allocated, contiguous tensor, so the reshape that
feeds the matmul does not need a copy. ggml_reshape_3d asserts
contiguity, so a layout that would need the cont cannot slip through
silently.
Greedy output is unchanged token for token.
Address review from @ggerganov
* kv cache : batch state restore scatter reads per contiguous run
When restoring state into non-contiguous destination cells (e.g. a
prompt-cache snapshot into a fragmented ring), state_read_data issued
one small copy per KV cell - ~1.4M copies of a few KiB each for a
40k+ token restore, taking 25-63 s on the CUDA backend.
The snapshot stores cell rows in cell order, so a maximal run of
consecutive destination indices maps to one contiguous block and can
be restored with a single copy. Precompute the runs once and use them
in all three scatter loops (K, V, transposed V). Byte-identical.
The on-device reader copies with a byte cursor when the read and
write chunking differs, so the batched reads are safe for it as well.
Batching makes equal tensor counts with a different split reachable
(save ranges [2,1] vs restore runs [1,2]); the next commit teaches the
reader's 1:1 path to fall back to the byte cursor in that case.
Verified in a production setup: 1,363,616 copies / 25-63 s -> 224
copies / 221-424 ms for the same restores (42,603 cells, 4 runs).
Assisted-by: Claude Code (unsloth/qwen3.8-27b)
* context : fall back to the byte cursor when read and write chunking differ
the on-device reader copies saved state back with a 1:1 copy by tensor
index whenever the write and read sides recorded the same number of
tensors, guarded by a per-tensor size assert.
equal tensor counts do not imply equal chunking: a state restore may
batch its reads per contiguous run of destination cells while the save
used per-range reads, so both sides can record two tensors that split
the same data differently, and the assert aborts in all builds.
compare the per-tensor sizes and only take the 1:1 path when the
chunking actually matches, otherwise fall through to the existing
byte-cursor copy. both sides enumerate the same logical data in the
same order, so the cursor copy is well-defined across tensor
boundaries.
Assisted-by: Claude Code (unsloth/qwen3.8-27b)
* tests : cover state restore scatter reads on host and on-device paths
decode the same prefix on two sequences, interleaving the seq 0 cells
between the seq 1 cells, so the seq 1 cells are isolated from each
other in the kv cache (three cells, two saved ranges). save the seq 1
state, free the interleaved seq 0 cells, and restore: the destination
is then non-contiguous (two runs), and the restore-side chunking has
the same tensor count as the save-side with a different split, so the
scatter path is batched per contiguous run and the on-device reader's
byte-cursor fallback is exercised.
the restored state is saved again on the host and compared byte for
byte with the first save: the blob is serialized in sequence cell
order, so the two saves are identical if and only if the scatter
restore wrote exactly the same KV content. this documents the
byte-identical guarantee of the run-batched scatter reads.
one test per io backend: the host (CPU) path and the on-device path.
Assisted-by: Claude Code (unsloth/qwen3.8-27b)
* dflash : fuse the encoder into the KV injection decode
The encoder is a single fc + norm, but running it as a separate
llama_encode forced a device-to-host round trip of its output before the
injection decode could re-upload it, plus a second graph build per
round. Fold the encoder into the decoder's embd branch and feed the
target features directly to one llama_decode.
Assisted-by: Claude Fable
* nit
* Apply batched suggestions from code review
Co-authored-by: Ruixiang Wang <wangruixiang07@outlook.com>
* Fix missing references from renaming
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
Co-authored-by: Ruixiang Wang <wangruixiang07@outlook.com>
for_each_token_in tested all LLAMA_MAX_SEQ sequences for every used cell,
while a cell almost always belongs to one. The scan now stops once the
cell's own sequences have been seen. Same visit order, same callback
arguments, so behaviour is unchanged.
get_prev_tokens is the only caller, so this affects the n-gram path.
RTX PRO 6000, Qwen3.8-Flash-Next UD-Q4_K_XL, fa on, warm runs:
55k context generation 56.3 -> 74.3 t/s
132k context generation 33.6 -> 50.9 t/s
Prompt processing is unchanged, the scan is amortised over the ubatch
there. The gain follows the number of used cells, so it grows with
context and is invisible on short prompts.
- DFlash2 NVFP4 draft models produced almost no accepted speculative
tokens because the Q, K, V, and output projection scales were not
passed to the corresponding graph operations.
