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
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%.
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.
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)
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 was missing from the gated delta net branch of get_split_segments,
so its attn_qkv.weight, shaped {n_embd, 2*key_dim + value_dim}, fell through
to the generic fused QKV rule and tripped
GGML_ASSERT(tensor->ne[axis] == n_embd + 2*n_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)
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)
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.
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.
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.
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.
* DSV4: sm tensor
* set coarser granularity for head splits
* fix dspark
* add model saving for dsv4 + allow dflash to return on specific device
* add comment about dsv4 seq_rm
* simplify
* add shared expert delayed allreduce
* remove special test for dsv4
* feat(convert): Add conversion for GraniteSWAForCausalLM
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob, OpenCode + Qwen3.6-35b)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat(llama): Add granite_swa support
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob, OpenCode + Qwen3.6-35b)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat(conversion): Add conversion infra for rope_pattern array
NOTE: There is other work also targeting this, so this may be
removed depending on merge order.
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob)
Signed-off-by: Gabe Goodhart <[email protected]>
* fix(conversion): Fix SWA pattern logic and support for non-rope layers
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat(conversion): Add support for GraniteMoeSWA
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: Add llama_hparams::has_rope and arch constants
NOTE: This shadows the work done for Granite Speech
https://github.com/ggml-org/llama.cpp/pull/25107
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: Add support for per-layer rope determination
Branch: GraniteSWAForCausalLM
AI-usage: full (Bob)
Signed-off-by: Gabe Goodhart <[email protected]>
* style: Fix failing flake8 for extra newlines
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver
Branch: GraniteSWAForCausalLM
AI-usage: full (OpenCode + Qwen3.6-35b)
Signed-off-by: Gabe Goodhart <[email protected]>
* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Load MoE params as optional
Branch: GraniteSWAForCausalLM
AI-usage: draft (OpenCode + Qwen3.6-35b)
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: Handle MoE params in conversion
branch: GraniteSWAForCausalLM
AI-usage: full (OpenCode + Qwen3.6-35b)
Signed-off-by: Gabe Goodhart <[email protected]>
* style: Remove unnecessary newline
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Remove unnecessary tensor additions to GRANITE architecture
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Correctly handle naming for ffn gate inp
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Always default hparams.rope_pattern to 1s
This isn't strictly necessary, but it will allow other models to rely on
hparams.has_rope(il) without needting to prepopulate.
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: Move to has_rope for all granite model architectures
Now that we have a proper hparam for this, it's better to use it and not
require a hacky fallback in the hparam method itself.
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: No hacky rope_finetuned fallback in has_rope
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Fully remove rope hparam filling in granitemoe
There are no granitemoe models that use NoPE (it's not actually used in the
layer building below), so this was just dead code.
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Save out rope_pattern in model-saver
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Set hparams.rope_finetuned for round trip
Since the value is _read_ from rope_finetuned, we need to persist it when
the model is saved with the saver.
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Code review cleanup
Signed-off-by: Gabe Goodhart <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* refactor: Keep gate/up fused for MoE path
Branch: GraniteSWAForCausalLM
AI-usage: full (Claude + Sonnet 5)
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Skip GRANITE_SWA in model saver
https://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651
Keeping is_swa_impl in the saver can break other models.
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* add sliding window pattern for model in test
* style: Fix indentation
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* fix: Fix \r\n
Thanks Claude!
Branch: GraniteSWAForCausalLM
AI-usage: none
Signed-off-by: Gabe Goodhart <[email protected]>
* feat: Keep shared expert fused
Branch: GraniteSWAForCausalLM
AI-usage: full (Claude + Sonnet 5)
Signed-off-by: Gabe Goodhart <[email protected]>
* style: More indentation fixes
Signed-off-by: Gabe Goodhart <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
---------
Signed-off-by: Gabe Goodhart <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Adding support for bailingmoe3
* Adds speculative decoding support
* Make BailingMoE3 safe gate metadata optional
* bailingmoe3: apply trained SwiGLU clamps
* common: fix Bailing V3 tool argument parsing
* llama-model-saver, instantiate float vector metadata writer
* bailingmoe3: support Q-LoRA (Ling-3.0-tiny)
Ling-3.0-flash sets q_lora_rank: None and projects Q directly, so the current
implementation loads a single ATTN_Q tensor. Ling-3.0-tiny sets q_lora_rank: 256
and routes Q through a LoRA bottleneck instead:
q_a_proj -> q_a_layernorm -> q_b_proj
Conversion therefore failed with:
ValueError: Can not map tensor 'model.layers.3.attention.q_a_layernorm.weight'
Add the missing path, mirroring the existing deepseek2 MLA implementation:
* constants.py - add ATTN_Q_A / ATTN_Q_B / ATTN_Q_A_NORM to BAILINGMOE3
* tensor_mapping.py - map model.layers.{bid}.attention.q_{a,b}_proj and
q_a_layernorm
* conversion - emit attention.q_lora_rank when the config has it
* bailingmoe3.cpp - read n_lora_q; create the Q-LoRA tensors and build Q
through the bottleneck when q_lora_rank > 0
Everything is gated on q_lora_rank > 0. Ling-3.0-flash's config has no
q_lora_rank, the converter only emits the key when present, hparams.n_lora_q
defaults to 0, and get_key(..., required=false) leaves the target untouched when
the key is absent - so flash keeps taking the existing direct-Q branch.
The LoRA path produces the same shape as the direct projection, so the
nope/rope split, RoPE application and wk_b absorption downstream are unchanged.
* small mtp change
* bailingmoe3: support separate MTP GGUF and Q-LoRA MTP
* gguf: remove duplicate add_kda_gate_lower_bound definition
---------
Co-authored-by: bloomer <[email protected]>
Co-authored-by: Dyluhn <[email protected]>
* model: add Kimi-K3 text model
Hybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, plus five
things that architecture does not have:
1. cross-layer residual attention (attn_res_block_size)
2. latent MoE (routed experts run at n_expert_latent)
3. situ activation (replaces SwiGLU everywhere)
4. MLA output gate (sigmoid gate before o_proj)
5. full-rank KDA gate (single ssm_g instead of ssm_g_a/ssm_g_b)
K3's text_config reports KimiLinearForCausalLM - the older 48B architecture -
so get_model_architecture routes on the top-level name instead.
The KDA decay gate has two forms, selected by linear_attn_config's
gate_lower_bound. It is not a clamp: when set it swaps the activation entirely
(fla/ops/kda/gate.py), from -exp(A_log)*softplus(x) to
lower_bound*sigmoid(exp(A_log)*x). K3 sets it to -5.0; kimi-linear leaves it
unset, so that path is unchanged.
Cross-layer residuals reuse ggml_dsv4_hc_pre for the weighted sum. That op is
CPU + CUDA only, so Metal/Vulkan will fall back per-node until those kernels
exist.
The routed experts ship as compressed-tensors "mxfp4-pack-quantized". That is
bit-compatible with ggml's MXFP4 - same E2M1 code assignment, same E8M0 scale
byte, only the nibble positions within a block differ - so they are repacked
rather than dequantized, losslessly and without a ~5.5 TB bf16 round-trip.
The repack is built lazily because gguf_writer holds every added tensor until
the final write. DeepSeek-V4 was already doing the identical bit-shuffling, so
it now shares the helper.
Verified against Moonshot's own code path (transformers + fla's Triton KDA
kernels) on a tiny model exercising every K3-specific feature. Final-position
logits vs the fp32 reference: 6.7e-05 rel / corr 1.00000000 for both the
chunked and the recurrent delta-net path. MXFP4 blocks dequantize to the source
weights with 0.0e+00 error.
Assisted-By: Claude Opus 5 (1M context) <[email protected]>
* model: fix ty errors in the Kimi-K3 converter
- `_res_parts` buffers (kind, tensor) pairs, not bare tensors
- `get_tensors` must return an Iterator, matching ModelBase
- LazyBase's `func` takes one argument, so pass the expert loaders through
`args` instead of the closure
- borrowing KimiLinearModel.set_vocab from an unrelated TextModel is
deliberate and safe, but not expressible in the signature
No behaviour change: the MXFP4 repack still dequantizes to the source weights
with 0.0e+00 error and end-to-end logits are unchanged (8.386e-03 rel,
corr 0.99996630).
Assisted-By: Claude Opus 5 (1M context) <[email protected]>
* Update conversion/kimi_k3.py
Co-authored-by: Boris Dvorkin <[email protected]>
* Increase LLAMA_MAX_EXPERTS from 512 to 1024
* tests : support for Kimi K3 in archs test
* chat : add Kimi K3 chat format (reasoning, content, typed tool calls)
K3's assistant output is an XTML-ish tagged format built by the template's
open_tag/close_tag macros. Two properties break generic parsing:
1. The generation prompt ends with open_tag('think'), so the completion
starts inside the think section with no opening marker in the output
(thinking_forced_open).
2. Only <|open|>/<|close|>/<|sep|>/<|end_of_msg|> are special tokens; tag
names ("think", "response", "message") are ordinary text tokens.
Adds common_chat_params_init_kimi_k3 (PEG_NATIVE) with detection on the
marker trio, reasoning extraction, response unwrapping, and tool-call
parsing of the tools/call/argument tag structure with argument types
taken from the tool schema. Includes the K3 chat template fixture and 9
test-chat cases derived from real generations of the full 2.8T model.
Verified end-to-end against Kimi-K3-Q2_K (GrEarl/Kimi-K3-GGUF) on 8x B200:
content, reasoning_content, streaming deltas, and tool_calls all correct;
finish_reason stop/tool_calls as appropriate.
Co-Authored-By: Claude Fable 5 <[email protected]>
* chat : add message_delimiters for Kimi K3
Per-role message-start markers for token-level span splitting. User and
assistant messages carry only the role attribute, so their full opener
(through <|sep|>) is used; system and tool messages continue with more
attributes (type=/tool=/index=), so those delimiters stop after the
role's closing quote. Verified against the K3 tiktoken vocabulary that
the closing quote is always a standalone token across all attribute
variants, so the token-level prefix match stays exact.
Co-Authored-By: Claude Fable 5 <[email protected]>
* fix: apply nits from @ngxson and text fixes from @danielhanchen
* tests : added missing hyperparameters and tensors for Kimi K3 in test-llama-archs
* chore : move overly verbose header file comments to Kimi K3 source file
* tests : re-enabled KIMI_K3 in test-llama-archs for WebGPU backend
* model-saver : emit kda_gate_lower_bound for Kimi K3
Quick fix. The Kimi K3 loader reads kda_gate_lower_bound and gates a graph branch on it (it scales the KDA gate when the bound is above -INFINITY), but the model
saver never wrote the key, so a save->load roundtrip silently dropped it back to the -INFINITY default and changed the model's output. The real K3 config sets gate_lower_bound = -5.0.
I propose to emit it from the saver, and set it to -5.0 in the test-llama-archs K3 case so the roundtrip check exercises it (the roundtrip fails without the saver line).
* Refactor conditional for model architecture check
* tests : re-enabled (again) KIMI_K3 and MINIMAX_M3 in test-llama-archs for WebGPU backend
* fix code comments
* add template on conversion
* move repack_mxfp4_blocks to model base
* nits
* add_value_length
* optimize res_stack construction
* nits
---------
Co-authored-by: Boris Dvorkin <[email protected]>
Co-authored-by: Stanisław Szymczyk <[email protected]>
Co-authored-by: Deepankar Singh <[email protected]>
Co-authored-by: Claude Fable 5 <[email protected]>
Co-authored-by: Caleb DeLeeuw <[email protected]>
Co-authored-by: Xuan Son Nguyen <[email protected]>
* llama : support for MiniMax-Text-01 model
* chore : renames to match the other MiniMax models
* model : add logits mask as MiniMax-Text-01 embeddings tensor has zero-valued embeddings for tokens >= 200032 that produce zero logits disrupting the token sampling process
* llama : replace hardcoded conditions with hparams.is_recr()
* model : used build_rs() for recurrent state management
* chore : code cleanup
* model : optimized MiniMax-Text-01 by removing the state tranpose operations
* chore : removed unnecessary ggml_cont() in MiniMax-Text-01 implementation
* llama : add generic logits mask graph input
* model : permuted diag_decay dimensions to avoid doing it inside MiniMax-Text-01 graph
* chore : code cleanup
* chore : code cleanup
* model : use token positions when calculating MiniMax-Text-01 decay tensors
* convert : add support for MiniMaxM1ForCausalLM as it seems to be the same as MiniMaxText01ForCausalLM
* chat : add jinja template for MiniMax-M1
Co-authored-by: QscQ <[email protected]>
* chore : code cleanup
* tests : MINIMAX_01-related fixes
* chore : silence Python lint errors
* vocab : remove unnecessary vocab type
* convert : update MiniMaxText01Model conversion to use yield when modifying tensors
* convert : suppress tokens with zero-valued embeddings during MiniMax-Text-01 conversion
* llama : removed logits mask - no longer necessary as token suppression is used instead
* model : use common functions to make MiniMax-Text-01 implementation more concise
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* model : use common functions to make MiniMax-Text-01 implementation more concise
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* convert : override non-working built-in chat template during conversion
* tests : skip arch MINIMAX_01 tests for WebGPU backend (it breaks again)
---------
Co-authored-by: Stanisław Szymczyk <[email protected]>
Co-authored-by: QscQ <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* adapt the api
* text model ok
* working impl, need verify and clean up
* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im
ggml_conv_transpose_1d has no grouped mode, so the depthwise upsample
was built as one convolution and one concat per channel, which floods
the graph with small nodes and makes kernel launches dominate the
decoder.
Fold both cases into the column form the seanet decoder already needs:
the general case reshapes the kernel to [IC, K * OC] and matmuls it
with the input, the depthwise case batches a matmul over the channels
so a step scales its own kernel. A single col2im_1d then scatter-adds
the columns back to the signal, with the same shape as before, so the
overlap-add tail, the streaming state and the bias are untouched.
Generation time per frame drops by 80% on CUDA and by 50% on CPU. The
output matches the previous implementation sample for sample, with a
correlation of 0.999994 and identical frame counts.
* flow_temp + frames_after_eos
* chunking
* mtmd: carry the remaining pocket-tts per-pack settings
The language packs also tune the end-of-speech padding and the padding
of short prompts, next to the temperature already carried in the
mmproj: french_24l asks for 8 tail frames instead of the guessed 3,
english_2026-01 asks for short prompts to be padded with spaces.
Write both in the mmproj as clip.gen.audio.frames_after_eos and
clip.gen.audio.pad_short_text, keyed on the pack in the conversion
script like the temperature. The loader keeps them optional, so a
mmproj without them behaves as before. Map semicolons to commas for
every pack instead, the reference only asks for it on three of them and
it costs nothing elsewhere.
Existing mmproj files must be converted again to carry the two keys.
On a long french text the port now lands within 2% of the reference:
22.96s against 23.44s, with the same peak level and the same amount of
silence.
* clip.gen.audio.model_variant
* clean up code comments
* nit: drop the dead flow_temp hparam, the pack table holds the default
* update docs
* address security problems
* less invasive base.py
* lint
* add mtmd_gen_inp_default
* add docs
* rm gen_flow_temp
---------
Co-authored-by: Pascal <[email protected]>
* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used
* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp
Co-authored-by: Max Krasnyansky <[email protected]>
* set mmap_support to false on OpenCL backend
* fix order of load modes
* use -1 for auto
* resolve load mode auto earlier to correctly pick gpu host or cpu memory
* add load mode auto to llama-bench
* bump virtgpu api version, regenerate docs
---------
Co-authored-by: Piotr Wilkin (ilintar) <[email protected]>
Co-authored-by: Max Krasnyansky <[email protected]>
Co-authored-by: Georgi Gerganov <[email protected]>
* Get started with Onyx
* Add architecture
* Skip keys handled in super()
* Loading tensors
* Shorten
* Graph
* Apply suggestion from @pcuenca
* Remove norm now embedding in transformers weights
* Add eot
* Explicit output_multiplier
* Handle post_norm_eps
* No super call; unhardcode eot.
The pattern `self._set_vocab_gpt2()` seems preferred throughout the
codebase, and it allows `set_vocab()` to be called from a different part
of the Python class hierarchy: the drafter model converter that we may
need eventually.
* Register for drafting
* DFlash: inherit rope type from the linked target.
Another option would be to store it in the gguf file itself.
* mmproj conversion
Note: some fields to be renamed after the implementation works. We are
keeping compatibility with the reference Meta gguf for testing purposes.
* "clip" header declarations
* Load mmproj
* Pre-processing
* Graph
* Go back to using delimiters.
Otherwise our generations are worse.
Transformers does not use them. We need to trace inputs to verify
whether they are equivalent.
* downsample_factor -> merge_size
* Add vision graph
lol, forgot from a previous commit
* Additional renames, align with llama.cpp / transformers
* Prefer _size instead of independent _h and _w
* Fix token layout
Co-authored-by: Young Han <[email protected]>
* onyx: bring the chat parser onto the onyx branch
common/chat.cpp on this branch has no Onyx handling, so a converted model
serves malformed chat: the assistant preamble leaks into content
("to=self<|message|>...") and tool calls fail with
HTTP 500 "The model produced output that does not match the expected
peg-native format"
common_chat_params_init_onyx exists on onyx-fair-patch, added there by
8bb73dd3d. It was never on this branch, so this is not a regression --
the two lines developed independently.
The code here is taken verbatim from that commit. It is the clean side of
`git merge origin/onyx-fair-patch`: chat.cpp is one of the files that
merges without conflict. The full merge is not viable -- it produces 13
conflicts, including add/add on conversion/onyx.py and src/models/onyx.cpp
where the q_norm-folding and metadata-scale approaches contradict each
other, and #4/#7 are stacked on this branch's side of that.
Verified on this branch: builds with 0 errors, converts an Onyx checkpoint,
and serving it gives "4" for "What is 2+2?" plus a correct
get_weather {"city":"Paris"} tool call, where the unported branch gives the
two failures above.
No converter or runtime changes are included, so this should not interact
with the q_norm work.
Co-authored-by: Beto de Paola <[email protected]>
* Less params, bilinear pos-emb interpolation as a graph op instead of CPU
* Map to symbolic V_MMPROJ instead of strings
* Make a couple params explicit
* Patchify via build_inp()
* No param for rope_theta
* Small cleanup
* Restore blank line
* Unpermute, to adapt to the latest transformers checkpoint
* Apply norm after token embeddings
This follows the latest transformers approach.
* Remove duplicated function
* build_vit
* onyx: use the model rope theta on sliding-window layers
* DFlash: conversion from transformers drafter
* Revert rope_type derivation from target
NOTE: this breaks compatibility with Meta's distributed DFlash GGUFs, as
the Q/K are stored in "NEOX" (rotated half) format, like in
transformers.
* Apply suggestion from @pcuenca
* Set model type
* Remove comment that will become obsolete
* Hardcode post_norm_rms_eps instead of new param
* Derive SWA+RoPE pattern from gguf array or scalar
* Fix model type <-> number of layers
* Reorder
* Rename
* Fix typo
* DFlash: seed the draft KV cache from multimodal embedding batches
`common_speculative_impl_draft_dflash::process()` returned early on any batch carrying embeddings, so an image prefill never had its target-layer features fused through the DFlash encoder and injected into the draft's KV cache. That left a hole spanning the image's positions, and the next injection at a post-image position failed to initialize its batch:
```
decoding image batch 1/1, n_tokens_batch = 256
decode: failed to initialize batch
llama_decode: failed to decode, ret = -1
process: llama_decode(ctx_dft) failed rc=-1 (n_tokens=17, offset=0)
srv decode: failed to process speculative batch
```
Every image request with `--spec-type draft-dflash` failed with HTTP 500. Text-only was unaffected, since those batches carry token ids and were let through.
Restore the earlier condition, which admits a batch that is either tokens or embeddings and skips only the degenerate neither/both cases. The rest of `process()` is already layout-agnostic -- it gathers features via `llama_get_embeddings_layer_inp()` and indexes `batch_in.pos[]` / `batch_in.seq_id[]`, none of which assume token ids -- so this is the whole fix.
Validated against `muse-glimmer-30B-bf16.gguf` + `mmproj-muse-glimmer-30B-bf16.gguf` + a DFlash draft head, on an image describe-the-shapes request:
- before: HTTP 500, `failed to process speculative batch`
- after: HTTP 200, draft acceptance 0.34012 (167 accepted / 491 generated), mean len 3.04
Output equivalence holds, which is the property that matters: at temperature 0 the drafted response is byte-identical to the same request served with no draft attached (1213/1213 chars), so the draft is drafting correctly through the image context rather than merely not crashing.
* Conversion: prefer rewrite to mapping
* Revert "Conversion: prefer rewrite to mapping"
This reverts commit a92d0ac584.
* fix lint
* sliding_window metadata is not optional
* disable state save/load
* Apply suggestion from @pcuenca
---------
Co-authored-by: Young Han <[email protected]>
Co-authored-by: Beto de Paola <[email protected]>
Co-authored-by: Daniel Han <[email protected]>
Co-authored-by: ruanrms <[email protected]>
Co-authored-by: Xuan Son Nguyen <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Restore quantization of mmprojs
This was lost in the refactor undertaken in #22004.
* add noreturn
---------
Co-authored-by: Xuan Son Nguyen <[email protected]>
* granite-switch: add llama.cpp backend (POC, CPU)
New "granite-switch" architecture: a dense, all-attention Granite-4.1
model with N embedded LoRA adapters selected per-token by control tokens.
- gguf-py schema (arch, KV keys, stacked LoRA tensor names) + writer helpers
- conversion/granite.py: GraniteSwitchModel converter (stacks N adapters +
zero base slot into per-projection A/B tensors; emits switch metadata)
- C++ arch registration (llama-arch.{h,cpp}, llama-model.{h,cpp})
- src/models/granite_switch.cpp: load + per-token switched-LoRA graph via
ggml_mul_mat_id over stacked tensors; sticky per-token index + control-token
substitution in llm_graph_input_switch::set_input
- llm_graph_input_switch in src/models/models.h
Runs end-to-end on CPU: convert 3b checkpoint (842 tensors, stacked dim 13)
and generate on both base and control-token paths. Sticky switch state is
single-sequence (POC); full multi-sequence machinery is a follow-up.
* granite-switch: add Mac (Metal) build + mid-sequence switch demo script
Self-contained script to build llama.cpp on Apple Silicon (Metal),
convert the composed 3b checkpoint, and run the crisp mid-sequence
adapter-switch demos verified on Vela:
- answerability: <|answerability|> mid-seq -> "unanswerable"
- query_rewrite: <|query_rewrite|> mid-seq -> {"rewritten_question": ...}
Each demo runs the same prompt twice, differing only by a control token
placed before the assistant turn, so the per-token switch is visible.
* granite-switch mac demo: add -no-cnv so each run is one-shot
The composed model ships a chat template, so llama-completion auto-enables
interactive conversation mode and halts at a `>` prompt after generating,
stalling the script. -no-cnv disables conversation mode: generate once from
the raw prompt and exit (also prints special tokens, making the switch visible).
* granite-switch: replace global sticky index with in-graph router attention
The POC computed the per-token adapter index on the CPU and carried it
across ubatches in ONE global `mutable int32_t poc_sticky_index`, reset
only when a ubatch contained sequence position 0. That global had two
problems:
1. Concurrency: with multiple sequences in a batch it was last-writer-
wins — one sequence's adapter leaked into the others.
2. Multi-turn: an interactive `ollama run` chat continues one KV cache,
so turn 2 never saw position 0 and the index never reset — the
adapter stayed stuck on across turns.
Port the vLLM/HF backend mechanism faithfully: a single-head causal
"router" attention recovers the adapter index in-graph. Per token, only
dim 0 carries signal — Q[0]=1, K[0]=+gain for a control token / -gain
otherwise, V[0]=adapter slot / 0 — and the causal softmax over the single
visible control token recovers that adapter's slot (readback =
clamp(round(V[0]), 0, n_adapters)). gain=15 matches config.py and is
F16-safe (no F32 cache).
The router's K/V live in the model KV cache at an extra layer
R == hparams.router_layer (== n_layer). We bump n_layer_all to n_real+1
so the cache allocator gives the router its own per-sequence slot, and
set n_layer_nextn=1 so n_layer() stays n_real — the decoder loop and
tensor loading are untouched and never reference layer R. The router K is
exempted from the k-shift RoPE loop (its dim-0 value is a literal
magnitude, not a rotation).
Because the selection now lives in the per-sequence KV cache, CONCURRENT
requests are isolated for free (problem 1 fixed; verified by
scratch/concurrent_switch_test.cpp). set_input becomes stateless pure
per-token maps; the global is gone.
Single-switch contract / known limitation, identical to vLLM & HF: the
gain is flat (no recency), so within one sequence there is no mechanism to
revert to base mid-sequence — once an adapter fires it stays on until that
sequence ends (problem 2 is therefore NOT fixed by a faithful copy; vLLM/HF
avoid it only because each served request is a fresh sequence). A client
continuing one KV cache across turns must start a fresh sequence per turn,
or opt into a recency-biased router (a deliberate divergence, not done
here). Documented in granite_switch.cpp and asserted by
scratch/multiturn_leak_test.cpp.
Verified (CPU): both demos unchanged (answerability -> "unanswerable",
query_rewrite -> rewritten query); concurrent two-sequence isolation
passes; multi-turn carry-over matches the vLLM/HF contract.
* granite-switch: drop scratch tests and mac demo for upstream PR
Remove the local-only development artifacts that should not ship in the
upstream PR:
- granite-switch-mac-demo.sh (local Metal build + demo driver)
- scratch/concurrent_switch_test.cpp
- scratch/multiturn_leak_test.cpp
Also drop the now-dangling reference to the scratch tests from the
granite_switch.cpp header comment. Leaves only the core architecture
support (conversion, gguf constants, llama-arch/model/kv-cache, and the
granite_switch graph).
* granite-switch: trim comments to match native llama.cpp style
* granite-switch: trim conversion comments to match native style
* granite-switch: drop unused adapter_ranks metadata
* granite-switch: rename arch to graniteswitch and drop obid alias
* granite-switch: fix non-ASCII comments and document router gain assumption
* granite-switch: drop section comments from constants.py to match native style
* granite-switch: add functional tensor block comments matching Granite4 Vision style
* granite-switch: clarify n_expert_used comment
State the actual constraint: mul_mat_id needs n_expert_used == 1, and
since the GGUF carries expert_count = 0 the generic loader's
n_expert == 0 => n_expert_used == 0 assertion has already passed by the
time load_arch_hparams runs, so it is forced to 1 here.
* granite-switch: note n_layer_nextn reuse has no MTP
The router carving reuses n_layer_nextn, normally the MTP/next-token
count. Clarify in the comment that it is borrowed here purely as the
trailing-layers lever and that there is no MTP head, to spare readers
the double-take.
* granite-switch: rename source file and apply review nits
* granite-switch: don't force LoRA tensors to F16, follow --outtype instead
* granite-switch: drop redundant _permute_qk wrapper, call LlamaModel.permute directly
* granite-switch: read router gain from GGUF (control_token_gain) instead of hardcoding 15.0
* granite-switch: derive n_slots()
* granite-switch: move llm_graph_input_switch into granite-switch.cpp
* granite-switch: cut AI-style narration comments
* granite-switch: collapse multi-line comments
* granite-switch: rename control_token_* maps to adapter_token_*
* granite-switch: cut noise comments
* granite-switch: rename embedded LoRA tensors to <base>.lora_a/lora_b
* granite-switch: GGML_ASSERT token input to avoid UB on embeddings
* granite-switch: TODO for raw embedding input support
* granite-switch: collapse LoRA tensor constants to .lora_a/.lora_b suffix
* granite-switch: drop n_expert_used hack, guard mul_mat_id buft probe
* granite-switch: stop forcing dense expert counts, read from config
* granite-switch: renamed control_token_gain metadata key to router_gain
* granite-switch: trim header comments to match native style
* granite-switch: collapse LoRA tensors to base name + suffix
* granite-switch: inline suffix checks in tensor op resolution
* granite-switch: drop switch-lora struct comment
* granite-switch: guard router layer index and inline n_slots
* granite-switch: group adapter metadata under {arch}.adapters.* namespace
* granite-switch: add hparams.has_rope(il) for KV-shift rope skipping
* granite-switch: skip arch in test-llama-archs (adapter fixture missing, TODO)
* granite-switch: Keys.Adapters namespace + simplify n_slots
* granite-switch: validate substitute token ids against n_vocab
* granite-switch: bound adapter count and lora rank from GGUF
* granite-switch: reject MTP context type when router_layer is set
* granite-switch: throw on bad adapter metadata instead of GGML_ASSERT
* granite-switch: use ASCII +/- in router K signal comment
* granite-switch: document n_layer_nextn repurpose and its leak points
* granite-switch: gate lora_a/lora_b op mapping on router_layer
* granite-switch: label all three preview model sizes
* llama : MTP support for DeepSeek V3.2
* model : no need to include MTP layers during DeepSeek V3.2 model type discovery
---------
Co-authored-by: Stanisław Szymczyk <[email protected]>
* llama : load MTP tensors only if they are really used
* llama : skip loading MTP (if not used) in remaining models that support MTP
---------
Co-authored-by: Stanisław Szymczyk <[email protected]>
* model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2)
Adds GLM-5.2 NextN/MTP as a --spec-type draft-mtp target: nextn tensor
loading via the qwen35moe/step35-style presence probe, a graph_mtp
builder (enorm/hnorm/eh_proj + dense MLA + sigmoid-gated MoE with
shared expert + shared head with fallbacks, _s scale tensors passed
for NVFP4), t_h_nextn extraction in the trunk graph, and MTP-context
KV setup: the draft head runs dense MLA, so the MTP context uses a
plain attention KV cache holding only the nextn layer(s) (same
pattern as the hybrid Qwen3.5 MTP context) while the main context
keeps the DSA cache, now filtered to trunk layers only.
Co-Authored-By: Claude Fable 5 <[email protected]>
* convert : support --mtp/--no-mtp export for GlmMoeDsaForCausalLM (GLM-5.2)
Opt GLM-5.2 into the supports_mtp_export contract (post-#25641 shape,
mirroring HYV3Model/Step35Model): --no-mtp drops the appended NextN
block (blk.78) and its nextn_predict_layers KV; --mtp keeps only the
NextN block plus shared embeddings/norm/lm_head. Default (bundled)
output is unchanged.
Co-Authored-By: Claude Fable 5 <[email protected]>
---------
Co-authored-by: Claude Fable 5 <[email protected]>
* Add preliminary MiniMax-M3 support
Text-only port that re-uses existing components: MiniMax-M2 style GQA with
per-head QK-norm and partial rotary, DeepSeek-V3 style leading-dense and
routed/shared experts, and swigluoai activation. Sparse attention is not
yet supported (dense fallback); vision tower and MTP heads are dropped.
* MiniMax-M3 vision tower (mmproj + clip graph)
* Delete m3_vision_ref.py
* Update clip.cpp
* MSA
* Update constants.py
* Update minimax.py
* Cache creation. Working withotu flash attention
* Added flash attention for sparse layers
* Decomposed slow cpu OP into GPU + CPU ops. Massive speedup over long ctx
* Rewrote indexer op to be cuda native. Modified flash attention to match per group block picking
* Implement sparse attention calc out of stock ops.
* Fix a cache allocation and cont issue
* Fixed -fa auto crash, flagged debug spots
* Delete vocab.json
* Delete model.safetensors.index.json
* Delete generation_config.json
* Delete Minimax directory
* Handled multi stream case to fall back on Dense Attention
* Development scaffolding cleanup. No functional change to the decode or
4-way paths. Full debug harness remains at <8136a9c68ed7a5eb009aa67bba3fda8062f4648f> for reproducing the
selection-parity validation.
* Remove redundant comment from minimax-m3.cpp
* Changed 3 Gelu Ops for vision into Gelu_erf ops
* Assert that n_kv is multiple of 128
* Rename MSA index tensors to indexer convention
Note: All GGUFs generated before this change will need to be regenerated.
* Fix incorrect Assert
* Review driven changes (#3)
* Remove comment from conversion minimax.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Remove whitespaces from constants.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Tighten comment in minimax.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* inherit MiniMax-M3 from MiniMax-M2
* drop dead text_config fallbacks
* Add indexer writer methods
* Reuse LLM_FFN_SWIGLU_OAI_MOE
* Remove duplicate indexer setters, add only block_size/local_blocks, follow value naming convention
* Fix conversion error /gguf_writer.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Update gguf-py/gguf/gguf_writer.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Update gguf-py/gguf/tensor_mapping.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Remove whitespace in src/llama-kv-cache.cpp
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Remove Whitespace in Update src/llama-model.h
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* Remove whitespace in src/llama-hparams.h
Co-authored-by: Sigbjørn Skjæret <[email protected]>
* remove multimodal code upon maintainer request. Will be made as a separate PR
* Whitespace clean in tensor_mapping.py
* Log cache size on launch, block ctx shift, support prompt caching
Log indexer cache size on launch
Disallow ctx shift
Support prompt caching
* Update minimax-m3.cpp
* Optimize implementation, add multi stream support.
Fully rewrote minimax-m3.cpp for speed and buffer size gains:
Unified the 4-way + decode, 1 FA call per layer instead of 4, with the groups mapped onto ne[3]
Custom CPU op now emits block-level mask, expanded on GPU, which causes CPU to GPU transfer to shrinks at prefill
Decode: ~25 nodes/layer vs ~50, no per-group concats/conts
Unified selection semantics, so both regimes rank bs + local bias (position-anchored local force), which means prefill/decode can no longer disagree on selection
can_reuse on the MSA bias input. Graph reuse at decode restored (was rebuilding the full graph every token)
In-place mask adds, shrinking compute buffer ~6.8 to ~4.2 GiB at ub2048/62k
Multi-stream: MSA now runs with -np N when kv_unified=false. Decode stays batched across streams (still 1 FA call), prefill loops per stream. dense fallback only for --kv-unified + multi-seq
Measured effect on expert offload bound setup: decode 6.2(4WAY)–7.15(MSA_decode) -> 7.7~7.8 t/s, flat from 5k to 60k+. prefill around 10% faster. buffer about 20% smaller, multi-user support.
* set default cache type to F32
* Fix potential DSA double indexer cache allocation bug, only allocate in-cache k_idx for archs that opt in
* remove F16 downcasts in MSA attention, force F32 indexer score accum
* Add Minimax eos to llama vocab
* Guard edge case where idx cache can become stale after a tail trim
* Update llama-kv-cache.h
* Update llama-kv-cache.cpp
* Update llama-kv-cache.cpp
* Update llama-kv-cache.h
* Update llama-kv-cache.cpp
* Review driven changes
* style fix
* indexer hparams are required
* fix tests
* fix lint
---------
Co-authored-by: Daniel Han <[email protected]>
Co-authored-by: Sigbjørn Skjæret <[email protected]>
Co-authored-by: Xuan Son Nguyen <[email protected]>
* Start building graph - reuse deepseek32
* Enable kv cache and rotation for glm_dsa architecture
Just follow Deepseek 3.2 for now.
* Reuse prev_top_k for "shared" indexer layers
* GLM 5.2 uses LLAMA_ROPE_TYPE_NORM for the indexer.
This is transformers' `apply_rotary_pos_emb_interleave`
* Default indexer types to GLM pattern
Previous converted GGUFs like https://huggingface.co/unsloth/GLM-5.2-GGUF write indexer weights to _all_ layers, even if they are only required for "full" types. This PR relies on a new key "%s.attention.indexer.types"; if absent, it will use the default GLM 5.2 schedule as defined in https://huggingface.co/zai-org/GLM-5.2/blob/main/config.json#L26.
Note that conversion is not saving this key yet.
* Save indexer types to gguf, restore on load
* Use ggml_lightning_indexer when cparams.fused_lid
Co-authored-by: fairydreaming <[email protected]>
* GLM 5 and 5.1 use full indexers
Co-authored-by: fairydreaming <[email protected]>
* Fix indentation
* Ensure array is zero-filled
* Prefer explicit std::fill
* Assert prev_top_k exists for shared indexer
---------
Co-authored-by: fairydreaming <[email protected]>
Xuan Son Nguyen
danielhanchen
Daniel Han
Aman Gupta
Gaurav Garg
Chris Danis
Gabe GoodhartandSigbjørn Skjæret
Oğuzhan AkkayaandSigbjørn Skjæret
Ruixiang Wang
Bar Haim
Sergey Malinin
timkhronosandGeorgi Gerganov
Masashi YoshimuraandSigbjørn Skjæret
Tunahanandtnhnyc
Satinder GrewalandClaude Fable 5
Guido Imperiale
zqlandroot
Aaron Teo
Joe Rowell
Johannes Gäßler