ggml-org/llama.cpp b10660 on GitHub

Details

model: add Qwen3.8-Flash-Next (qwen4exp) (#27742)

gguf: add qwen4exp (Qwen3.8-Flash-Next) arch and converter

Adds the GGUF-side plumbing for HF model_type qwen4_exp:

MODEL_ARCH.QWEN4EXP plus tensors for the low-rank hyper-connection
variant (hc_norm/down/up/inject) and the PLE n-gram hash embeddings.
The DeepSeek-V4 hc_fn/base/scale tensors are a different
parameterisation, so these are separate entries rather than reuse.
Reuses the existing indexer, per_layer_token_embd, SSM and
compress_ratios keys unchanged.
conversion/qwen4exp.py inherits the Qwen3.5 linear-attention V-head
reorder and interleaved mrope, concatenates the 128 PLE embedding
shards, and splits index_qk_proj into separate indexer q/k tensors.

The PLE hash multipliers reach ~2.4e13. prepare_tensors() casts every
non-float dtype to float32 before modify_tensors() runs, and GGUF array
writes infer INT32 from Python ints, so both paths are bypassed: the
constants are read from the pre-cast lazy tensors and written as
explicit UINT64 arrays.

Additive only; no existing arch changes behaviour.

llama: load qwen4exp (Qwen3.8-Flash-Next) hparams and tensors

Adds LLM_ARCH_QWEN4EXP with its hparams and tensor loading. The graph
comes in the next commit; this makes the model load and report correct
metadata.

hyper-connections set n_embd_out_impl = hc_count * n_embd, so the
residual stream is 4x wide and there is no output_norm: the final
mixer's hc_norm is the last norm in the model.
registered as hybrid and given the same recurrent/attention memory
filters as Qwen3-Next and Qwen3.5.
reuses the existing indexer, per_layer_token_embd, SSM and
compress_ratios keys as-is.
the PLE table row count is read back from the file rather than
recomputing the vocab padding rule.

llama-model-loader gains UINT64 array support. That branch previously
threw, so no existing caller changes behaviour; it is needed because the
PLE hash multipliers do not fit in int32.

qwen4exp: shorten comments
llama: qwen4exp text graph with hyper-connections, GDN and MoE

Implements the decode graph for Qwen3.8-Flash-Next: the hyper-connection
residual stream, gated delta net layers, the MoE block with its gated shared
expert, and dense full attention. The QSA indexer and the PLE n-gram embedding
are not wired up yet and land in later commits.

Hyper-connections are implemented here rather than shared with deepseek4.cpp.
The two formulations agree on the [n_embd, hc, n_tokens] layout and little
else: DeepSeek-V4 mixes with a full-rank projection and Sinkhorn-normalises
it, whereas this model uses a low-rank down/silu/up sigmoid gate and collapses
by a plain mean. Only the ~10 line stream mean is genuinely common, so sharing
would mean touching DSV4's hot path and its three fused CUDA ops to reuse very
little. What is reused is the substantive part: the LLM_KV_HYPER_CONNECTION_*
keys, the n_embd_out_impl wide-residual support already in the loader, and the
layout convention.

Also allows a checkpoint to carry no PLE layers at all, which makes it
possible to bring the graph up and validate it in stages.

Validated against vLLM, the only working reference implementation. On a
scaled-down model with an init scale large enough to give non-uniform logits,
agreement with vLLM sits at the numerical noise floor: llama.cpp f32 against
its own bf16 gives 84.3% top-1 agreement over 255 positions, and this graph
against vLLM gives 85.1%. The comparison was calibrated by seeding three
deliberate bugs (silu instead of sigmoid on the delta net gate, dropping the
1/hc scale in the mix, dropping the 2x in the combine); each drops top-1 to
between 0% and 11%, an order of magnitude below the floor.

llama: qwen4exp PLE n-gram hash embedding

Adds the per-layer embedding: a custom I32 graph input hashes each token with
its ngram_size-1 predecessors host-side and the result is a plain row gather
over the shared table, the same shape gemma3n's per-layer embedding uses. The
hash has to run on the host because the splitmix64-derived multipliers reach
2^45, so the products need 64-bit integers and an xor, neither of which ggml
has.

Predecessors that fall outside the ubatch come from a small per-sequence
history on the model, mirroring the per-request ngram_context the reference
carries. It is only trusted when contiguous with the incoming position, so a
fresh prompt or a rewound cache falls back to EOS padding rather than hashing
against stale tokens.

The depthwise conv is written out as a sum of shifted, per-channel-scaled
copies rather than through ggml_conv_1d_dw, which carries a correctness
warning upstream.

Verified two ways. The row indices match a transcription of the reference's
tensor formulation exactly, 1024 of 1024 rows, including sequences with EOS
tokens sprinkled through them to exercise the segment reset. Separately, with
PLE placed on layer 0 so its input is just the token embedding, ple_embd and
ple_gated_value match a PyTorch computation from the same checkpoint to every
printed digit.

End to end over 1023 scored positions the port sits the same distance from
vLLM with PLE as without it, 6.3 points of top-1 against 6.0, so PLE costs no
accuracy relative to the rest of the model. That common offset is vLLM's bf16
activations, which cannot be removed: its QSA kernel refuses float32.

Two bugs found along the way, both caught by the row-index check. The history
was read and updated in the same pass, so a token early in a ubatch could pick
up an earlier token of that same ubatch as prior context; it is now snapshotted
first. And an EOS token was cutting its own context, where the reference takes
the last EOS strictly before the position, so a boundary only hides tokens from
the positions after it.

Known gap: the conv carries no state across ubatches, so it is exact only for a
prefill that starts at position 0. Chunked prefill and decode need the conv
state wired into the recurrent memory, and the conv branch itself is still
numerically unverified because the fixture zeroes its weights.

llama: carry the qwen4exp PLE conv state across ubatches

The PLE depthwise conv was zero-padding on the left, which is only right for a
prefill that starts at position 0. Decode and chunked prefill saw a truncated
history for the first (kernel-1)*ngram_size positions of every ubatch.

The PLE module sits on a layer that is also a delta-net layer, so both need a
conv history in the same recurrent row. Rather than plumb a per-layer state
size through build_rs and build_conv_state, the row is widened once and each
convolution addresses its own slice through a local helper. n_embd_r() gains
the extra span, which is zero for every other architecture because it is
derived from ple_n_heads.

Verified by feeding the same 1024 token sequence in chunks instead of one
shot: at 64 tokens per decode the logits are bit-identical to the single-shot
run, 1023 of 1023 top-1 and a maximum logprob deviation of exactly zero. At
one token per decode they differ slightly, but the no-PLE model differs more
under the same test (94.6% against 97.1%), so that is the usual gemv-versus-
gemm accumulation difference and not the state.

The conv branch is also no longer unverified. With non-zero conv weights the
port sits 6.3 points of top-1 below the numerical floor, the same distance as
with the weights zeroed and as the model with no PLE at all, so the branch
adds no error of its own.

test-llama-archs passes every existing architecture at 0.00e+00, including the
delta-net models that share this code path.

llama: fix the qwen4exp PLE conv state and unblock test-llama-archs

build_rs writes into the state tensor in place, zeroing one row and copying the
carried-over states, so calling it twice for the same layer let the second call
clobber the first write-back. The PLE layer is also a delta-net layer, so that
is exactly what happened: both convolutions gathered the same row. They now
share a single gather per layer.

The earlier claim that the conv state was carried correctly was tested on a
fixture whose conv weights are zero, where the branch contributes nothing and
chunking matches trivially. Re-running with non-zero conv weights showed the
divergence, growing with the number of ubatch boundaries: 97.1% top-1 at one
boundary down to 90.2% at seven. With the shared gather it is bit-identical to
the single-shot run at every chunk size tried, 512, 128 and 64, with a maximum
logprob deviation of exactly zero over 1023 positions. The delta-net-only model
stays bit-identical too, so nothing regressed there.

Also derive the delta-net conv channel count the way load_arch_tensors sizes
wqkv instead of from ssm_d_inner. The two agree for this model, but n_embd_r()
only bounds the row and the convolution has to match the tensor feeding it.

test-llama-archs previously aborted on this architecture and took every later
architecture with it. qwen4exp is marked MoE-only, given the hyper-connection
keys and an ssm_d_inner consistent with its tensor derivation, and skipped for
now: the hyper-connection keys written by get_gguf_ctx are not reaching the
synthesised file, which needs a separate look. The suite completes again, 124
architectures at 0.00e+00.

llama: optional indexer key cache in llama_memory_hybrid

Groundwork for qwen4exp's QSA sparse attention. Its indexer needs a per-token
key history for the full-attention layers, but a hybrid model cannot use
llama_kv_cache_dsa: that class derives from llama_memory_i rather than
llama_kv_cache, and llama_memory_hybrid constructs its attention cache
directly. No existing architecture pairs recurrent state with a sparse
indexer, so there was nothing to reuse wholesale.

llama_memory_hybrid therefore gains a third, optional cache, shaped the same
way llama_kv_cache_dsa shapes its lightning-indexer cache: a copy of hparams
with n_head_kv forced to 1 and n_embd_head_k_full set to indexer_head_size.
It is built only when a filter_idx callback is passed, which defaults to
nullptr, so every existing architecture gets exactly what it got before. The
per-sequence operations and the batch preparation forward to it under a null
check, matching how the DSA cache prepares its two caches over the same
ubatches.

test-llama-archs passes all 124 architectures at 0.00e+00, including the 12 in
the hybrid family that share this code. The qwen4exp fixtures are unchanged:
same logits against vLLM, and chunked evaluation still bit-identical to
single-shot.

llama: QSA sparse attention for qwen4exp

The full-attention layers of this model do not attend to everything. An
indexer scores one mean-pooled key per block of compress_ratio tokens and
keeps a budget of the best blocks, plus the tail of tokens that do not yet
form a complete block. Below indexer_top_k + compress_ratio - 1 cached
tokens every block fits in the budget, so the result is exactly dense.

What is reused rather than rebuilt:

the mask machinery. build_attn's DSA overload already turns a list of
token indices into a KQ mask via ggml_set_rows, so that block is lifted
out verbatim into build_attn_mask_top_k and shared with a new overload
on llm_graph_input_attn_kv. DSA's node sequence is unchanged; the new
overload exists because llama_kv_cache_dsa assumes MLA and cannot be
dropped into a hybrid model.
the indexer key cache, which is the optional third cache added to
llama_memory_hybrid in the previous commit. It holds raw keys, because
pooling happens before the norm and the rotation.

The graph expands block scores rather than block indices: giving every
token of a block its block's score needs only a gather, where expanding
indices would need an integer multiply-add that ggml has no op for. Since
the budget is a whole number of blocks and a block's members tie exactly,
the cut still lands on a block boundary.

Everything that depends on cache layout is computed host-side in
set_input_qsa. Blocks are cuts of the position line rather than of the cell
array, so nothing assumes the cache is contiguous.

Measured on the tiny fixture against vLLM, comparing the selected token
indices directly rather than the logits:

below the budget selection identical, and 1024-token logits are
bit-identical to the pre-QSA dense path
above the budget mean jaccard 0.975

The direct index comparison is what made this correct. The reference
rectifies each head's dot product before summing over heads, which an
earlier reading of it had missed; on logits alone the resulting port looked
fine, because on a randomly initialised fixture the known-correct dense
path already disagrees with vLLM by more than the bug did. Comparing the
indices showed 0.794, and fixing the ReLU moved it to 0.975.

llama: give the qwen4exp indexer cache the attention cache's slots

The indexer cache found its own slots, independently of the attention
cache. Both are the same size and see the same ubatches, so in a
straight-through prefill they agree, which is why every fixture and every
single-shot parity run passed. They drift once the context is being
rewritten between turns, and then the QSA top-k indices, which are applied
against the attention mask, point at the wrong cells.

The seven-turn chat test caught it on the third turn: llama-server aborted
on the assertion that the two caches report the same n_kv.

The cache is a side buffer addressed by the attention cache's cells, so it
now takes that cache's slot layout instead of computing one. Applying that
layout also marks its cells identically, so the two agree cell for cell by
construction rather than by coincidence, and the assertion can no longer
fire.

Inert where the caches already agreed: test-llama-archs green at 126 archs
and 0.00e+00, and the 4096-token tiny fixture is unchanged at max logit
delta 0.0.

tests: record what the qwen4exp arch-test skip actually observes

The old note guessed that the hyper-connection keys never reach the file.
They do: dumping the gguf_context handed to llama_model_init_from_user
shows both among its 67 KVs, and the loader still reports one missing.

tests: cover qwen4exp in test-llama-archs

The arch was skipped with a note guessing that the hyper-connection keys
never reached the synthesised file. They did. The suite builds a model, then
saves and reloads it, and llama_model_saver did not re-emit those keys, so
the failure was in the roundtrip leg rather than the first load. Three gaps,
all in shared code and all additive:

add_kv_from_model wrote no hyper-connection, compress-ratio or PLE keys.
The PLE group only means anything whole, so it is written or omitted
together; the rest follow the file's existing style of writing every key
unconditionally, since an architecture that does not read one is
unaffected by a zero.
the saver had no uint64 path at all, which the PLE hash constants need.
add_tensors_from_model enumerates model-level tensors by hand and was
missing per_layer_tok_embd and the three final-mixer tensors.

Two smaller fixes on the qwen4exp side, both found by running the test:

build_qsa_top_k divided by the compression ratio before asserting it was
non-zero, so a file without the key crashed instead of reporting.
a layer with no compression ratio now falls back to dense attention,
which is what the model computes below the budget anyway. The test then
has to write a ratio to reach QSA at all, and an indexer key length no
narrower than n_rot, since the indexer ropes with the main attention's
rotary width.

Full suite: 126 archs, qwen4exp at 0.00e+00 with roundtrip OK. The tiny
fixture is unchanged, max logit delta 0.0 against the pre-QSA dense run.

convert: stream the qwen4exp PLE table instead of concatenating it

The n-gram table arrives as 128 shards that were held in a dict and then
torch.cat-ed, so the peak was the shards plus the concatenation: around
300 GB of RSS on the real checkpoint, which rules out machines that could
otherwise convert this model.

Each shard is now written straight into a memory-mapped file at its final
row offset and dropped, so the resident set is one shard and the rest is
the page cache's problem. The temporary file sits beside the output and is
removed once the write finishes, including on failure.

Shards other than the last must be uniform for direct placement, which is
asserted rather than assumed, and a shard arriving before the stride is
known is held instead of misplaced.

Verified on the tiny fixture: the resulting GGUF is byte-identical to the
one the concatenating path produced (md5 2d274efac91ad1e9a6007efb0687e597).

quantize: fall back to F16 for 32-block types with an odd ncols

tensor_type_fallback demotes a tensor whose ncols is not a multiple of the
target's block size, but its switch only enumerates the 256-block types. A
target that is already a 32-block type (iq4_nl, q4_0, q5_0, q8_0, ...) falls
into default: and throws, even though the function already knows how to answer
that case: the ncols check right below the switch resolves an unrepresentable
shape to F16.

Route those types into that check instead of throwing. Only paths that abort
today change, so no quantization that currently succeeds is affected.

Found on a 4-wide depthwise conv kernel. llama-quantize reported nothing but
"failed to quantize model from ...", with no tensor name and no exception text,
which made a quant recipe that had simply not pinned the tensor look like a
corrupt model. It now names the tensor and continues.

quantize: let --tensor-type name per_layer_token_embd

per_layer_token_embd shares the TOKEN_EMBD category with token_embd.weight, so
--token-embedding-type is returned for it before any --tensor-type pattern is
consulted, and there is no way to give it a tier of its own.

That grouping is fine as a default and stays the default. It is a poor fit for
the size, though: on qwen4exp the table is 97.7 GiB of a 337.6 GiB BF16 file and
about 46% of a 4-bit one, roughly eighty times token_embd.weight, and it is
read by ggml_get_rows rather than a matmul so no imatrix ever covers it.

Allow an explicit --tensor-type pattern to name it, and only it. Nothing
changes unless such a pattern is passed, and token_embd.weight keeps the old
precedence in either case.

Measured on Qwen3.8-Flash-Next, Q4_K_M with an imatrix: the table lands at q8_0
(51.9 GiB, 113.5 GiB total) by following --token-embedding-type, and pinning it
q4_1 gives 30.5 GiB for 92.1 GiB total, 19% off the file.

quantize: size the output buffer exactly instead of nelements * 4

The per-tensor output buffer was sized nelements * 4, described as an upper
bound. It is a very loose one: the output is at most 2 bytes per element
(f16/bf16) and usually well under 1.1 (q8_0 and below), so between 2x and 4x of
it is never touched. The exact size is already known here, since it is what the
quantization loop writes, what new_size sums to, and what the GGUF metadata is
asserted against a few lines later.

On a model whose largest tensor is a few GB none of this matters. On
Qwen3.8-Flash-Next it does: per_layer_token_embd is 51.2 G elements, so the
buffer was 205 GB where 54 GB is needed at q8_0 and 32 GB at q4_1.

Measured on that model, VmHWM of a live llama-quantize was 485 GB per process.
Three of them fit in 2 TB and five did not, which is what an OOM-killed quant
ladder looks like. This removes about 150 GB of that.

Byte-identical output, verified against the same binary built at the parent
commit: q4_K, q8_0, q5_K, q6_K and IQ4_XS, over BF16 and F32 sources, with and
without a PLE table present. Six cases, six matching md5s.

qwen4exp: hash the image placeholder for multimodal batches

The PLE row indices are computed host-side from ubatch->token, and set_input
returned early when that was null. A multimodal ubatch is exactly that case:
the mtmd layer consumes the image placeholder ids and hands llama_decode
embeddings instead. The early return left the I32 index tensor uninitialised,
so ggml_get_rows indexed a 320 M row table with whatever the buffer happened to
contain, and aborted:

GGML_ASSERT(i01 >= 0 && i01 < ne01) failed
ggml_compute_forward_get_rows
mtmd_helper_decode_image_chunk -> llama_decode

Every image request crashed. Nothing caught it because the vision work had only
ever been verified by converting an mmproj, never by running one.

The reference computes the hash over input_ids, where those positions still
hold the image placeholder, so carry that id through as qwen4exp.ple.image_token_id
and hash it. The key is optional: a file converted before it existed falls back
to the PLE EOS token, which is defined and treats the image as a segment
boundary rather than crashing.

Verified end to end with llama-mtmd-cli, a Q4_K_M base and the F16 mmproj, on a
generated image with known content. The model names the red circle, the blue
square, the inverted green triangle and reads "UNSLOTH 42", each with the right
position.

qwen4exp: support a non-unified KV cache in QSA

set_input_qsa asserted n_stream == 1, so llama-server could not serve this
model with more than one slot unless -kvu was passed. With a non-unified
cache each sequence owns its own cells, and a cell index means a different
token in each stream, so a single shared mapping is wrong.

cell_blk, blk_cells and bias gain a stream dimension. At n_stream == 1
these collapse to the shapes they had, so the unified path is unchanged.
Scoring is now batched over streams. ggml_mul_mat matches ne[2] on both
operands, so stream s's queries only ever meet stream s's blocks; without
this sequences would score against each other's context.
set_input_qsa loops per stream and resolves cells through
v_cells[seq_to_stream[seq_id]], following set_input_kq_mask_impl, instead
of hardcoding v_cells[0].
llama_kv_cache_context::get_n_stream() is added, mirroring the ns that
get_k and get_v already derive from the slot info.

build_attn_mask_top_k needed no change: it already expects
[n_top_k, n_batch, 1, n_stream], so the top-k result is reshaped to meet it.

set_input_qsa has exactly one caller, so the blast radius is qwen4exp only.

Validation, UD-Q4_K_XL on one B200:

unified cache unchanged within noise: 1802.9/68.85 -> 1807.2/69.11 t/s at
batch 1, 2262.5/192.43 -> 2270.1/193.75 at batch 4.
non-unified now runs at npl 1, 4, 16 where it previously aborted, and is
22% faster than the -kvu workaround at batch 16 (1205 vs 984 t/s total),
since per-stream cells avoid the cross-sequence masking a unified cache
pays for.
no cross-stream contamination: four concurrent sequences each carrying a
distinct secret all recall their own and no other, on both cache modes.
test-llama-archs green on qwen4exp, deepseek2, gemma3n, qwen3next, llama.

Note on testing: comparing concurrent output against solo output exactly is
not a valid check. It failed 0/4 with no bug present, and the unified-cache
control failed the same way, because batch composition changes the
floating-point reduction order and near-tied tokens flip. The contamination
test above is what the exit code gates on.

llama: keep the qwen4exp top-k attention mask arch-local

The QSA graph needed a build_attn that attends only to the cells named by a
top_k tensor, and the first version got it by adding a llm_graph_input_attn_kv
overload to llm_graph_context and factoring the mask construction out of the
existing MLA sparse path into a shared build_attn_mask_top_k.

That put a new arch on the shared attention path and made the deepseek32 and
glm-dsa attention build depend on a helper introduced for qwen4exp. Build the
mask in src/models/qwen4exp.cpp instead and leave llama-graph.{h,cpp} exactly as
they were: the MLA path keeps its own copy of the same node sequence.

The nodes emitted are unchanged, so this is bit-identical.

llama: hold the qwen4exp indexer cache in a new llama_memory_hybrid_idx

The indexer key cache was added by extending llama_memory_hybrid with an
optional third cache, and the host-side cell/block mapping that drives QSA was
added as set_input_qsa on llama_kv_cache. Both are shared classes that every
hybrid and every attention model goes through.

Move both into a new memory type, llama_memory_hybrid_idx, following
llama_kv_cache_msa: the indexer cache and the pos<->cell translation live with
the sparse-attention memory rather than in the classes that serve every other
architecture. llama-kv-cache.{h,cpp} and llama-memory-hybrid.{h,cpp} are
restored to their unmodified state.

init_batch is repeated from llama_memory_hybrid because the indexer cache has to
be handed the attention cache's slot infos, and those are not reachable through
the context the base returns. Allocating them separately lets the two caches
drift, which is what pointed QSA's top-k at the wrong cells before.

The context derives from llama_memory_hybrid_context so build_inp_mem_hybrid
keeps working unchanged, and get_n_stream is computed from the slot infos
exactly as llama_kv_cache_context did.

Behaviour is unchanged: logits over an 8192-token sequence are bit-identical to
the previous implementation, sparse and dense alike.

llama: save and restore the qwen4exp indexer KV cache

llama_memory_hybrid_idx forwarded clear, seq_rm, seq_cp, seq_keep, seq_add and
seq_div to the indexer cache but not state_write / state_read, so a saved
session dropped the indexer keys and a restored one selected QSA top-k against
an empty cache. The effect is invisible until the context passes
indexer_top_k + compress_ratio - 1 cells, because QSA is exactly dense below
that and the indexer contents cannot change the result.

The indexer section is written last rather than next to the attention cache it
mirrors. As a suffix, a reader that does not expect it stops early and the
trailing bytes are caught by the size check in state_load_file; placed between
the attention and recurrent sections it would instead be parsed as recurrent
state, which can succeed and restore silent garbage. It follows the same
LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY gate as the attention cache, since a partial
checkpoint deliberately skips the token-level attention caches.

The indexer restores its own cells instead of taking the attention cache's
restored slots. The two caches share size, padding and every sequence
operation, and init_batch hands the indexer the attention cache's slot infos,
so both state_read_meta calls run find_slot over identical occupancy and land
on identical cells.

The overrides live on llama_memory_hybrid_idx, the only memory type that owns
an indexer cache, so llama_memory_hybrid and every architecture that uses it
write and read exactly the bytes they did before.

The session and sequence state versions are bumped because the qwen4exp state
layout changed. The session path already rejects a short read via its size
check, but llama_state_seq_load_file accepts one silently, so only the version
check stops a pre-fix blob from being half-restored by a fixed build.

(cherry picked from commit 2721542354f8e158c3217625f4e2e7b83e51e3fe)

llama: make the qwen4exp PLE n-gram history per context and serialise it

The PLE hash of a token mixes in the ple_ngram_size - 1 tokens before it, which
a decode ubatch does not carry, so they were remembered in a map on
llama_model_qwen4exp. That is the wrong owner twice over.

A llama_model is shared by every context that loads it, and the map was keyed
only by llama_seq_id, so two contexts running the same sequence id - two server
instances on one model, or a draft/target pair - overwrote each other's window.
The next_pos guard turned that into EOS padding instead of a crash, so it
degraded quality silently.

The map was also in no state blob: grep found ple_hist in neither
llama-kv-cache.cpp nor llama-memory-*.cpp nor llama-context.cpp. A restored
context therefore failed the next_pos check on its first ubatch and hashed the
first tokens after the restore against EOS padding. This is why a session blob
round-tripped byte for byte while the restored context computed different
logits: the state was never in the bytes.

It moves to llama_memory_hybrid_idx, which is per context, is the memory type
qwen4exp always builds, and already does the per-sequence bookkeeping this
needs. Every sequence operation now carries the window with it:

seq_rm a rewind (p1 < 0) truncates the window to the surviving prefix and
moves next_pos to p0, so a rollback keeps exact context; a hole
punched in the middle leaves the window non-contiguous, so it is
dropped
seq_cp the destination inherits the source's window, truncated to the
copied position range - a copied sequence continues with the same
n-grams the source would have used
seq_keep every other sequence's window is dropped, like its cells
seq_add a shift that moves the whole window keeps it and moves next_pos with
it, which is the context-shift case; one that cuts through it drops
it
seq_div positions stop being consecutive, so an overlapping window is
dropped
clear everything is dropped

Dropping means next_pos = -1, which set_input turns into full EOS padding: the
same thing a fresh sequence gets, and the same thing this code did before it
followed the sequence operations at all, so no case is worse than before.

The state payload is a self-delimiting list, u32 count then per entry
{ i32 seq_id, i32 next_pos, u32 n_toks, i32 toks[n_toks] }, so a whole-context
save and a single-sequence save share one format and a single-sequence restore
can retarget the window at its destination seq_id. It is written after the
indexer section, last, for the same reason that one is: as a pure suffix an
older reader stops early instead of parsing these bytes as something else.

Unlike the indexer section it is not under LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY.
The window is recurrent state - it is the input the PLE convolution's own
recurrent state is derived from - and the recurrent cache beside it is written
for partial checkpoints too. Gating it would leave the server's speculative
decoding checkpoints restoring the conv state without the window that produced
it.

No further version bump: LLAMA_SESSION_VERSION 10 and LLAMA_STATE_SEQ_VERSION 3
were introduced for the indexer section in the same unreleased series, and both
changes are qwen4exp-only additions to the same blob layout.

Also fixes the padding of a short window. set_input pads a window shorter than
ngram_size - 1 up to that length, but prev() indexes the snapshot with the most
recent token last, and resize() pads at the back, so the filler EOS landed where
the immediately preceding token belongs. It now pads at the front. A window is
short at a sequence start after a one-token prefill, and after a seq_rm rewind,
which the new bookkeeping makes common.

Every architecture other than qwen4exp builds llama_memory_hybrid rather than
llama_memory_hybrid_idx, has no PLE table and never asks for a history, so
nothing about its graph, its sequence operations or its state bytes changes.

(cherry picked from commit de170364c052c68fcf63285cc0028095edb9f23c)

qwen4exp: tidy comments and simplify image token read

Rewrite the comments this series adds to the AGENTS.md rules: one or two lines,
no prose hard-wrapped mid-sentence, no narrative or history, and no comment that
only restates the code. Net 146 fewer comment lines, no code change.

Correct the PLE image comment: mtmd does not consume the placeholder ids. An
image is decoded as an embeddings-only batch, so ubatch->token is null and the
per-position ids never exist here. gemma3n and gemma4 hit the same case and
stand in row 0 of per_layer_token_embd; qwen4exp stands in the configured image
token id instead.

Read image_token_id straight from self.hparams in the converter. base.py merges
text_config into the root of hparams, and the key sits at the root of
config.json, so the config.json re-read was redundant.

(cherry picked from commit 205840c12169057da3e8d2f65ec4ceec3e18b980)

qwen4exp: support a quantized KV cache in the QSA attention path

(cherry picked from commit 4c30574f81dc1115d08078c47b6cf8c789c0a842)

llama: give qwen4exp a large-graph node budget

(cherry picked from commit 37c8c194e6a30e4c46ac29bee3fb264f091596ef)

qwen4exp: drop an unused variable that breaks -Werror builds

(cherry picked from commit 528d032b51fa3cf935ed3ef6e0fb1c7401df53b5)

quantize: dequantize and quantize large tensors in row bands

f32_conv_buf held the whole dequantized tensor, which is 204.8 GB for
per_layer_token_embd alone and dies with std::bad_alloc long before the
work buffer is reached. Dequantize and quantize in bands of whole rows
instead, capping the f32 staging at 1 GiB per band.

Rows are independent and the imatrix is indexed by column, so band
boundaries cannot change any output byte. Bands nest inside the existing
per-expert loop so each expert slice keeps its own imatrix, and a band is
kept to at least one quantization chunk per worker thread so the existing
multithreading still has work. F32 sources still stage nothing and are
banded by pointer arithmetic into the tensor.

llama_tensor_dequantize_impl now takes a first element offset; the single
caller is updated.

(cherry picked from commit 658c22549613555dbce57a772be4de8509eba3ee)

llama: segment the qwen4exp fused QKV for tensor split

qwen4exp was missing from the gated delta net branch of get_split_segments,
so its attn_qkv.weight, shaped {n_embd, 2key_dim + value_dim}, fell through
to the generic fused QKV rule and tripped
GGML_ASSERT(tensor->ne[axis] == n_embd + 2n_embd_gqa) while loading with
--split-mode tensor. --split-mode layer was unaffected.

qwen4exp broadcasts K to the V heads by tiling, k_conv is grown with a plain
ggml_repeat_4d over the head axis so that v head j pairs with k head
j % n_k_heads. That is the Qwen 3.5 pattern, not the repeat interleave that
Qwen 3 Next builds explicitly, so qwen4exp takes the else branch and its V is
segmented on the scale of K.

Reported by benklop.

(cherry picked from commit 353d753f595dc81634ae6130188b31f06018f5ae)

llama: fix the qwen4exp PLE history seq_rm(-1) iterator invalidation and the fatal-warning build

ple_hist_rm recursed over ple_hist with a range-based for and the recursive call
erases the entry it is iterating when the whole sequence is removed (p0 <= 0,
p1 < 0), so the loop then increments an invalidated iterator. It is unreachable
today only because llama_memory_recurrent::seq_rm rejects seq_id < 0 before
llama_memory_hybrid_idx::seq_rm reaches the history, which is a guard in another
class. Advance past the entry before recursing.

Two smaller things in the same area:

the n_toks sanity bound in ple_hist_state_read was the literal 64, which is
the value of LLAMA_MAX_PLE_HEADS, not of the quantity being checked. The
window is at most ple_ngram_size - 1 tokens, so the bound is
LLAMA_MAX_PLE_NGRAM - 1, eight times tighter.
build_conv_state_at left mem_size unused, so -DLLAMA_FATAL_WARNINGS=ON does
not compile. Predates this series; drop the line.

(cherry picked from commit 6eba44a89d5f328eb4859b844e1d28fb564cbe3e)

qwen4exp: include llama-impl.h explicitly for llama_mul_mat_hadamard

(cherry picked from commit b634fd4d250d181ef82bf78bd00c1ae3b96a7af6)

convert: fix the qwen4exp lint and type-check failures

flake8 flagged an unused MmprojModel import, and ty flagged seven errors in
the PLE streaming path: eos_token_id can be absent, and _ple_map, _ple_path,
_ple_row_dim and _ple_rows_per_shard are all Optional at the declaration but
were dereferenced without narrowing.

The map is opened and the stride fixed before the first shard is written, and
_finish_ple_table only runs once every shard has landed, so the invariants
hold. Assert them so the checker can see it. A missing eos_token_id now raises
with the reason instead of a TypeError from int(None).

llama: give the qwen4exp indexer cache its own tensor names

The indexer KV cache and the attention KV cache both named their tensors
cache_k_l%d, so the Meta backend matched the indexer cache against the
attention split pattern and aborted in handle_set_rows. Tag the names
instead, and mirror the indexer cache: it has one key head and its
projections are mirrored.

(cherry picked from commit a1cdc8181134659766763a17762545a1f0e5db7b)

qwen4exp: double the Q split granularity for tensor parallelism

qwen4exp fuses the attention gate into attn_q.weight the same way qwen3next
and qwen 3.5 do, so a device boundary must fall on a whole q+gate pair or the
Q heads stop lining up with the K/V heads and attn_output rows.

(cherry picked from commit 6c9a592f0a425a459ab6efae3b897cf68460e244)

qwen4exp: keep the indexer cache in step across server slots

The QSA indexer keeps a side cache addressed by the cells of the attention
cache, so cell j has to hold the same token in both: the top-k indices it
produces are applied to the attention KQ mask. init_batch already hands the
indexer the attention cache's slot layout rather than letting it look for its
own, but the restore path did not. state_read called llama_kv_cache::state_read
on the two caches in turn and each ran its own find_slot over its own occupancy.
That agrees only for as long as nothing has already pushed the two caches apart,
which is the property a restore is supposed to re-establish rather than one it
can lean on.

The failure path was the worse half, and it is reachable from the public API
with nothing more than a short buffer. Truncating a good blob at 35 offsets and
feeding it to llama_state_seq_set_data left the two caches disagreeing at 5 of
them, and every one of 23 truncations of a whole-context blob did. Four of those
five land inside the attention section, so the attention cache drops the
sequence and the indexer keeps it; only the cut that lands in the indexer
section gives the opposite direction. llama_kv_cache::state_read cleans up its
own cache and rethrows, so whichever way it falls, nothing is left to bring the
two back together. The server papers over this by clearing the slot when a
prompt cache load fails; a caller of llama_state_seq_set_data that does not is
left with an indexer addressing cells that no longer mean what it thinks.

llama_kv_cache::state_read_sinfo reports the cells a restore landed in, or takes
a copy of them, and state_read_meta uses a supplied layout in place of find_slot
once it has checked that those cells are free here too. The indexer now adopts
the attention cache's restored layout by construction instead of reproducing it
by coincidence, and a layout that does not fit fails the read rather than being
applied over cells that already drifted. The hybrid restore is wrapped so that
any failure drops the sequence, or for a whole-context restore the context, from
all three caches at once, which is a state they do agree on.

kv-cache: clear the cache once when restoring a whole context

state_read walks the streams of the cache in turn, and for a whole-context restore
each stream went through state_read_meta, which starts by calling clear(). clear()
resets every stream at once, so each stream after the first threw away the streams
already restored, and the K/V buffers with them. A non-unified cache holds one
stream per sequence, so a context saved with N sequences in it came back with only
the sequence in the last stream that carried any cells - the highest sequence id.
A unified cache has one stream and never showed it.

The cache is now emptied once, before the loop, which is what a whole-context
restore means. A blob whose streams are all empty now empties the cache as well,
where before it left the old contents in place.

kv-cache: check the mirrored slot layout on a whole-context restore too

state_read_meta only looked at the layout it was given on the single-sequence path.
A whole-context restore lays the cells out from 0 in both caches, so they agree as
long as they restore the same number of cells, but nothing checked that they did: an
indexer section belonging to some other context was read over cells the attention
cache had filled from a different one, which is the state the indexer must never be
left in.

qwen4exp: give the PLE conv history its own mirrored recurrent row

n_embd_r() reserved n_conv + ple_conv_state() so that one cache_r_l row could
carry both the delta-net conv state and the PLE dilated conv history, but the
QWEN4EXP arm of get_split_segments only described n_conv. Under -sm tensor the
segment sum came up short by ple_conv_state() and llama_memory_recurrent
construction aborted in ggml_backend_meta_alloc_ctx_tensors_from_buft.

Widening the segment list is not the fix. The Meta backend propagates a view's
split descriptor from its parent unchanged, so a view of one sub-range of a
split axis is sized as the whole row on every device; declaring the PLE tail as
a second segment merely moves the abort to "shape mismatch for VIEW" at graph
allocation. The two histories also want opposite policies: the delta-net state
is split by head to match wqkv and ssm_conv1d, while per_layer_tok_embd,
ple_conv1d and ple_norm_conv are all mirrored, so every device computes the
whole dilated conv and needs the whole history. One tensor cannot be both, and
the split state has no per-segment mirroring.

Move the PLE history into its own cache_ple_r_l%d row, mark it MIRRORED, and
return n_embd_r() to n_conv. The row is allocated only on layers where is_ple
holds, so mirroring one 92160-element row per device replaces a 92160-element
tail on all 36 recurrent rows: the recurrent R footprint drops rather than
grows. build_conv_state_at now takes its width from the tensor it was handed
and keys its gather on that tensor, which also drops a cont of a strided view.

no more ple_hist (use master version)
llama: give the qwen4exp full memory context its indexer cache

graph_reserve() walks a full memory context, and qwen4exp builds its
sparse attention only when the context exposes an indexer cache. the
full-context constructor left ctx_idx null, so the reserved worst case
was the dense fallback: a smaller graph than the one decode executes.
ggml-alloc then had to grow the compute buffer on the first decode,
past the size reported at load.

with -np 4 -c 32768 -fa on -ctk q8_0 -ctv q8_0 on an IQ1_S qwen4exp,
the reserved CUDA0 buffer was 217.00 MiB against 275.71 MiB actually
used, and CUDA_Host 42.31 MiB against 191.14 MiB. reserving the sparse
graph makes both match exactly, in unified and non-unified cache mode.

Co-authored-by: Pascal admin@serveurperso.com
Assisted-by: Claude

qwen4exp: shrink the PLE hparams storage

llama_hparams is held by value inside llm_graph_params and every llm_graph_input_*,
and llm_graph_params is a stack local in graph_reserve and process_ubatch, so its
width is paid on every worker thread stack.

is_ple_impl spent 2048 bytes carrying 512 bits. It is the one per-layer flag that is
not moved through the loader's uint32 array templates, so a bitset costs nothing in
call sites and also removes the uninitialized read that non-qwen4exp archs had, since
nothing filled the array for them.

The PLE head offsets and vocab sizes are token-space indices; the gather that consumes
them already truncates to int32, so 64-bit storage was never reachable. The gguf arrays
stay uint64 for file compatibility and are narrowed on load.

sizeof(llama_hparams) 34440 -> 31944, sizeof(llm_graph_params) 34872 -> 32376.

llama: opt-in random-access mmap advice for host-resident gather tables

qwen4exp keeps per_layer_token_embd on the host: 26.8 GiB at IQ4_NL, read
by ggml_get_rows as 16 gathers of ~90-170 bytes per token, spread across
16 head regions ~20M rows apart. Measured over 4.75M gathers, no two
consecutive gathers land on the same 4 KiB page, so the readahead the
loader asks for buys nothing here and the whole table ends up cached to
serve about 4% of itself.

llama_mmap applies POSIX_FADV_SEQUENTIAL, MAP_POPULATE and a whole-file
POSIX_MADV_WILLNEED unconditionally. Those are right for streaming the
file once into buffers and wrong for whatever stays mapped afterwards.

Under LLAMA_MMAP_RANDOM the eager pull-in is skipped and the mapping is
advised random once every tensor has been read, so the load itself keeps
its sequential readahead. That alone drops the table to 4.4% resident but
serializes one NVMe latency per gather.

The second half is what pays for it: the PLE input already computes every
row index for the ubatch before the graph runs, so the pages those rows
fall on are handed to the kernel in one batch and the reads overlap.
POSIX_MADV_WILLNEED on POSIX, PrefetchVirtualMemory on Windows, which
takes the discontiguous ranges in a single call.

Off by default and off for every other model: the batched prefetch keys
off "this mapping was advised random", which nothing sets unless the user
opts in.

-c 512 --chunks 60, cold, IQ1_S, mean of 3:

default            35.3 s   26.82 GiB resident (100%)
advice only       104.5 s    1.19 GiB resident (4.4%)
advice + prefetch  34.2 s    1.19 GiB resident (4.4%)

PPL 4.2346 +/- 0.07862 in all three. IQ1_S KLD is unchanged in every
field, including Mean KLD 0.396070 +/- 0.001931 and Same top p 77.325%.

llama: narrow the random-access mmap advice to the gather table

The advice was applied per mapping: every mapping the model kept got
POSIX_MADV_RANDOM plus a whole-file POSIX_FADV_RANDOM, and the eager
pull-in was skipped for every file. On qwen4exp that also hit
token_embd.weight, which sits 0.33 GiB past the PLE table in the same
shard and is read densely, not by sparse gathers. Measured over
-c 512 --chunks 60 on IQ1_S it fell to 8.45% resident, against 100% with
the feature off.

A model now nominates its gather tables (qwen4exp: per_layer_tok_embd)
and only those byte ranges are advised. The range is rounded out to
whole pages, which on this model takes in 832 bytes before and 192
after. token_embd goes back to 86.55% resident and the PLE table still
drops to 4.44%; smaps shows one VM_RAND_READ VMA of exactly the table
instead of one over all 27.16 GiB that stays mapped.

posix_fadvise is dropped from the narrowed path. POSIX_FADV_RANDOM
ignores its offset and length and marks the whole open file, and the
FMODE_RANDOM it sets is only read by page_cache_sync_ra() on the read()
path, which a fault on a MADV_RANDOM vma never reaches. POSIX_FADV_
DONTNEED does take a range, so the drop mode keeps it.

The eager pull-in is now skipped only for the files holding a nominated
table, and re-issued as WILLNEED over the rest of such a file, so other
shards load exactly as before.

prefetch_rows() keys off the tensor being nominated rather than off a
mapping-level flag, so the batched readahead lands only where the advice
did.

-c 512 --chunks 60, cold, IQ1_S, mean of 3, total wall:

default              32.50 s
whole mapping        30.05 s
narrowed             30.35 s

PPL 4.2061 in all three. IQ1_S KLD is bit-identical with the feature on
and off, including Mean KLD 0.396070 +/- 0.001931 and Same top p
77.325%. tg128 73.65 +/- 0.33 narrowed against 73.49 +/- 0.34 whole.

Assisted-by: Claude

llama: fold the random-access prefetch into its own feature flag

LLAMA_MMAP_RANDOM_PREFETCH existed to measure the two halves of the feature
apart, and the measurement is done: on a cold cache over the same wikitext
run, MADV_RANDOM without the batched readahead takes 94.4 s against 36.7 s
for an untouched mapping, while the pair together take 34.1 s. Suppressing
the kernel's readahead only pays if we replace it, so the split let a user
select a 2.6x regression through a documented switch.

Keep the accessor, since the call site reads better than a mode comparison,
but derive it from the mode alone.

FACP (Fewer Acronym Classes Please)
qwen4exp: bias the QSA selection per block, not per cell

The QSA bias is a graph input, so it is pinned on the host and uploaded every
decode, and at -c 32768 -np 4 its twelve copies were 768 of the 815 MiB of
reserved host compute buffer.

Only one half of it needs a cell: whether the cell sits in the always-visible
tail, and whether its block was pooled. Both are properties of the block. The
other half - empty, other sequence, or in the future - is the plain visible/not
test the attention mask already carries over the same cells, so add that mask
instead of repeating it. The bias then holds one value per block.

A block sits wholly inside or wholly outside the tail because the tail starts on
a block boundary, so one value per block is exact. Cells no block covers keep
their -inf from the mask.

The mask is F16 and the bias F32, and a mixed ggml_add reinterprets the F16
buffer as float rather than converting it, so the cast is required.

reserved host compute buffer at -c 32768 -np 4:
--kv-unified 814.86 -> 238.86 MiB, CUDA0 721.07 -> 421.07 MiB
--no-kv-unified 214.86 -> 70.86 MiB, CUDA0 317.07 -> 265.07 MiB

Selection is unchanged: over 8192 tokens, four times the budget, every QSA
layer returns identical top-k indices and the logprobs are bitwise equal.

Two things a reviewer should know. A cell whose position divides past the last
block is guarded by an assert rather than handled, because no run reached it.
And the mask's same-position M-RoPE rule cannot fire for text and was never
exercised for images, so the 2D case is unverified.