* Add support for the ibm-granite/granite-embedding-{97m,311m}-multilingual-r2 embedding models:
* Added a version of the gpt4o tokenizer that has a fixed regex (better handling of marks), and different token merging setting for the 97m model
* Reused gemma4 tokenizer for the 311m model
* granite-embedding-*-multilingual-r2 : add support SwiGLU FFN for Granite Embedding Multilingual R2
* added new GGUF key <arch>.hidden_activation (LLM_KV_HIDDEN_ACT) + writer
* added a forward declaration of llm_ffn_op_type to llama-hparams.h
* added llm_ffn_op in hparams
* added LLM_FFN_NONE = 0 sentinel to llm_ffn_op_type (value-initialization), modern-bert: explicitly assigns LLM_FFN_GEGLU before reading GGUF (unchanged).
* centralized hidden_act mapping in llama-model.cpp, added llm_ffn_op_type_from_string() helper, mirroring rope_scaling_type/llama_rope_scaling_type_from_string()
* modern-bert reads the GGUF key (when present) and uses the resulting op in its FFN graph
* Added granite-embedding-{97m,311m}-multilingual-r2 to the converter code
* Added the hashes for the granite embedding multilingual R2 models
* Set the hidden_activation in the GGUF if the field is present in config.json (such as for the granite embedding models)
* speculative : add common_speculative_n_max helper function
Extract the speculative max-draft-size logic from server_n_outputs_max
into a reusable common_speculative_n_max() function in common/speculative.
Assisted-by: llama.cpp:local pi
* cont : draft context always has n_parallel outputs
* llama : log n_outputs_max
* speculative : remove draft-simple auto-enable
* ci : enable server tests on PRs
* llama: save more VRAM by reserving n_outputs == n_seqs when possible
* add n_outputs_per_seq
* move n_outputs_max to server-context
* change ubatch to batch everywhere
After #23007 reclassified integrated CUDA/HIP devices as IGPU, the device
selection logic dropped the local iGPU whenever any RPC server was added,
because RPC devices made `model->devices` non-empty. On systems where the
"iGPU" is the main compute device (e.g. Strix Halo with 128 GiB of unified
memory), this caused all tensors to be allocated on the RPC peer alone and
model loading to fail.
Gate the iGPU inclusion on `gpus.empty()` instead, so RPC peers no longer
suppress the local iGPU.
closes: #23858
Add minicpm5 pre-tokenizer hash via convert_hf_to_gguf_update.py and
implement hardcoded regex handling in llama-vocab.cpp, consistent with
other BPE pre-tokenizers.
Co-authored-by: zhangtao <zhangtao2@modelbest.cn>
* initial talkie support, coherent
* reorder to follow convention
* absorb inverse rope
* stop folding scalars to improve quantization
* use broadcasting instead of duplication
* style cleanup
* add scaling support to LoraTorchTensor; use that path in conversion
* use layer_out_scale instead of embd_skip_scale
ffn_latent_down/up are declared GGML_OP_MUL in LLM_TENSOR_INFOS but
nemotron-h feeds them through ggml_mul_mat. The loader buft probe asks
the backend about the declared op, so it tested an elementwise MUL on a
q8_0 weight. That used to return true unconditionally and the weight
stayed on GPU by luck. Once supports_op told the truth, the probe got a
no and the loader pushed the weight and its matmul to CPU, splitting the
graph. Tagging it MUL_MAT asks the real question, the math is unchanged.
Verified on Nemotron 3 Super 120B Q5_K_M: from 64.9 back to 103.22 t/s.
* vocab : add Carbon-3B (HybridDNATokenizer) support
Adds a new BPE pre-type LLAMA_VOCAB_PRE_TYPE_CARBON for the
HybridDNATokenizer used by HuggingFaceBio/Carbon-{500M,3B,8B}.
The base BPE is Qwen3-4B-Base's; what differs is that text inside
<dna>...</dna> regions is chunked into fixed 6-mers (right-padded
with 'A' on the trailing partial), and any base outside ACGT maps
to <oov>.
* src/llama-vocab.{h,cpp}: new pre-type, dispatched from
llm_tokenizer_bpe_session::tokenize.
* src/llama-vocab-carbon.h: pure helpers (tokenize_carbon,
emit_dna_kmers) factored out for unit testing — no llama_vocab
dependency, vocab access goes through a std::function.
* conversion/base.py: detect HybridDNATokenizer by class name in
get_vocab_base_pre (chktxt collides with Qwen3 base since it
has no <dna>), and pass trust_remote_code=True in get_vocab_base
so the custom tokenizer class can load.
* tests/test-tokenizer-carbon.cpp: 12 cases covering single 6-mer,
multi 6-mer, lowercase, invalid base -> <oov>, partial k-mer
right-pad, mixed text+DNA, empty <dna></dna>, unterminated <dna>,
two regions, vocab miss.
* vocab : align Carbon-3B changes with llama.cpp conventions
* Fold tokenize_carbon + emit_dna_kmers inline into
llm_tokenizer_bpe_session (drop src/llama-vocab-carbon.h),
matching how every other tokenizer keeps its helpers inside
llama-vocab.cpp.
* Replace the standalone unit test with the conventional
test-tokenizer-0 row backed by models/ggml-vocab-carbon.gguf
(vocab-only conversion) + .inp/.out fixtures covering single
6-mer, multi 6-mer, lowercase, invalid base -> <oov>, partial
right-pad, mixed text+DNA, empty <dna></dna>, unterminated <dna>,
two regions.
* Register "carbon" in convert_hf_to_gguf_update.py's model list
(pointing at HuggingFaceBio/Carbon-3B) and teach both
AutoTokenizer call sites in the updater to pass
trust_remote_code=True for it, matching how t5 is special-cased.
* vocab : move Carbon dispatch to _set_vocab_carbon + LlamaModel branch
Refactor the conversion-side changes to follow the per-tokenizer-family
convention used by _set_vocab_qwen, _set_vocab_interns1, _set_vocab_glm,
etc. instead of conditionalising the shared get_vocab_base /
get_vocab_base_pre paths.
* conversion/base.py: add _set_vocab_carbon — self-contained, loads
with trust_remote_code=True so HybridDNATokenizer's merged Qwen3 + DNA
vocab is visible, writes tokenizer.ggml.pre = "carbon" directly.
* conversion/llama.py: branch in LlamaModel.set_vocab on
tokenizer_config.json["tokenizer_class"] == "HybridDNATokenizer" and
dispatch to _set_vocab_carbon. Same precedent as conversion/bert.py
(tokenizer_class branch between BertTokenizer / RobertaTokenizer) and
conversion/phi.py.
* conversion/base.py: revert the conditional in get_vocab_base and the
class-name short-circuit in the auto-generated get_vocab_base_pre.
* tests : expand ggml-vocab-carbon.gguf fixtures with model-card examples
Add 6 cases from the Carbon-3B model card on top of the existing edge
coverage: the unterminated basic-completion prompt, the closed 33-bp
example, the metadata-conditioned prompt (with <vertebrate_mammalian>
and <protein_coding_region> which BPE-decompose since they are not in
the vocab), the documented anti-pattern of raw DNA without <dna> tags,
and the two likelihood-scoring examples. Brings the suite to 19 cases.
* vocab : promote HybridDNATokenizer to its own LLAMA_VOCAB_TYPE
Refactor per upstream review:
> This should be its own tokenizer model, ie. carbonhybriddna instead
> of gpt2 and not carbon pre-tokenizer. That way you can keep the
> correct pre-tokenizer, in case that ever changes.
Previously the tokenizer was modelled as LLAMA_VOCAB_TYPE_BPE plus a
new LLAMA_VOCAB_PRE_TYPE_CARBON, which (a) put a CARBON-specific
branch inside llm_tokenizer_bpe_session::tokenize (only existing
pre-types differ in regex, not dispatch logic), and (b) conflated
"hybrid DNA tokenization" with "Qwen3 BPE pre-tokenizer".
This change moves it to its own vocab type, peer to PLAMO2, with the
GGUF model name matching the HF tokenizer class (HybridDNATokenizer):
* include/llama.h: new LLAMA_VOCAB_TYPE_HYBRIDDNA = 7.
* src/llama-vocab.cpp: new llm_tokenizer_hybriddna + session that
owns std::unique_ptr<llm_tokenizer_bpe> for non-<dna> text and
routes raw text through a DNA-aware splitter; wired into
init_tokenizer, tokenize, type_name, byte_to_token, and the
BPE-style token_to_piece case (DNA k-mers + <dna>/</dna>/<oov>
are pure ASCII, so byte-level BPE decoding handles them).
LLAMA_VOCAB_TYPE_HYBRIDDNA gets its own branch in the vocab-type
config block alongside SPM/WPM/UGM/RWKV, where pre_type is set
to QWEN2 and the matching add_space_prefix / escape_whitespaces /
clean_spaces flags are applied — mirroring qwen2's BPE path so
byte-level BPE merging stays bit-identical to the Python
reference for non-DNA text.
* src/llama-vocab.h: drop the short-lived LLAMA_VOCAB_PRE_TYPE_CARBON.
* conversion/base.py: _set_vocab_hybriddna writes
tokenizer.ggml.model = "hybriddna" (no separate pre).
* conversion/llama.py: dispatch on tokenizer_class ==
"HybridDNATokenizer" same as bert.py / phi.py do.
* models/ggml-vocab-hybriddna.gguf{,.inp,.out}: renamed fixture +
regenerated metadata.
* convert_hf_to_gguf_update.py: drop the stale chkhsh entry and
trust_remote_code special-case (no longer needed since dispatch
is now class-name driven, not chkhsh).
Verified end-to-end against HuggingFaceBio/Carbon-{500M,3B,8B}:
tokenization is bit-identical to the Python HybridDNATokenizer for
all 19 test fixtures plus the model-card metadata-conditioned
prompt; greedy completion produces the same DNA continuation as
the Python reference; spec-dec with 500M as draft for 8B still
works.
* vocab : relax llm_tokenizer_bpe assert to allow HYBRIDDNA
* vocab : drop llm_tokenizer_bpe vocab-type assert
* vocab : write tokenizer.ggml.pre for HYBRIDDNA, share BPE dispatch
* vocab : assert BPE or HYBRIDDNA in llm_tokenizer_bpe
* vocab : annotate #endif with PRETOKENIZERDEBUG
* vocab : drop local hybriddna fixture (moves to ggml-org/vocabs)
* deduplicate
* simplify
* simplify
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
When a model has zero non-SWA attention layers (e.g. a SWA-only slice of Gemma 4),
the base KV cache has no layer tensors. The input tensors (self_k_idxs, self_v_idxs,
self_kq_mask) are created as graph input nodes but never consumed by any compute node,
so the backend scheduler never allocates a buffer for them. Calling
mctx->get_base()->set_input_k_idxs() on an unallocated tensor then hits
GGML_ASSERT(buffer) at ggml-backend.cpp:194.
The same scenario applies symmetrically: if a model had zero SWA layers, the SWA
tensors would be unallocated.
Fix: guard both the base and SWA set_input calls with null/buffer checks, matching
the pattern already used by llm_graph_input_mem_hybrid_iswa::set_input (line ~674)
which has the comment: 'base tensors may not be allocated if there are no non-SWA
attention layers'.
Also fix can_reuse() in the same class to skip the ne[0] and kq_mask checks for
unallocated tensors, preventing a null-dereference on the reuse path.
- HunyuanOCR shares the same HF arch and vision layout as HunyuanVL butwas split into a separate path that skipped the +0.1 bilinear sampler used by the HF reference.
- Collapse OCR into the HUNYUANVL projector + HUNYUAN_VL text arch
* Move to backend sampling for MTP draft path
Run top_k(10) on the draft backend. D2H transfers happen only for the top 10 logits
Make backend sampling more robust and fallback to CPU on failure cases, such as with "-sm tensor" or when a backend doesn't support TOP_K.
* Allow sampler chains to be partially offloaded to backend
* Add --spec-draft-backend-sampling argument. Enabled by default.
This commit attempts to clarify a code comment in graph_mtp regarding
where the MTP layer is stored.
The motivation for this is that it was not obvious to me what the
original comment meant and hopefully this makes it clearer.
* spec: support MTP
* fix batch size
* rename files
* cont : simplify (#7)
* MTP: clean-up (#9)
* MTP: clean-up
* review: use llama_context_type instead of llama_graph_type
* review: remove llama_model_has_mtp
* review: fix convert issues
* convert: fix pycheck
* review: formatting
* use `mtp-` for identifying mtp models
* convert: fix mtp conversion
* mtp -> draft-mtp
* remove unused llama_arch
* add need_embd in speculative
* llama: allow partial seq_rm for GDN models for speculative decoding
Currently speculative checkpoint needs to restart from a checkpoint
after some draft tokens are not accepted, this leads to some wastage in
running the target again. This PR adds the ability to rollback upto
`draft_max` by storing the GDN intermediates.
* fix pending state
* vulkan: add GDN partial rollback
* meta: extend check to axis 1
* metal: add GDN partial rollback
Extend the gated delta net kernel to store intermediate states for
partial rollback support on the Metal backend.
- Add K (snapshot slot count) as a function constant
- Read input state from slot 0 of the 3D state tensor
- Write intermediate states to different slots during token loop
- For K=1, maintain backward-compatible single-slot behavior
Ref: https://github.com/ggml-org/llama.cpp/commit/8c05923630110223669f069af2000e9cf10c02bc
Assisted-by: llama.cpp:local pi
* delta_net_base: use ggml_pad instead of new_tensor
* review: add need_rs_seq
* review: rename part_bounded to n_rs
* review: deslop comments
* review: rename, add asserts
* server : adjust checkpoint logic (#11)
* server : adjust checkpoint logic
* cont : rm asserts
* server-context: fix early exit
* spec : fix compatibility with n-gram and add TODOs (#13)
* metal : cleanup
* llama : fix faulty bitwise check in recurrent memory
* server : disable RS-based MTP in combination with other spec types
* spec : add TODOs
* cont : fix comment
* cont : update comment
* common : fix logic for ngram + mtp compat
* llama-memory: enable checkpointing with partial rollback
* cont: add test-case for loading into a dirty ctx
* llama-memory-recurrent: clear rs_idx in clear
* download: fix mtp path
* llama-arch: fix enorm op
* docs: update docs
* conversion: fix type annotations
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* unicode,test: add Qwen3.5 non-backtracking tokenizer handler and regression tests
- Add unicode_regex_split_custom_qwen35() to [src/unicode.cpp](src/unicode.cpp), a non-backtracking handler for Qwen3.5's [\p{L}\p{M}]+ regex (letters + combining marks).
- Register the handler in the custom tokenizer dispatch table to prevent stack overflows on long inputs (fixes#21919).
- Add [models/ggml-vocab-qwen35.gguf](models/ggml-vocab-qwen35.gguf) (test vocab), [models/ggml-vocab-qwen35.gguf.inp](models/ggml-vocab-qwen35.gguf.inp) (test cases), and [models/ggml-vocab-qwen35.gguf.out](models/ggml-vocab-qwen35.gguf.out) (expected output) for regression testing.
- Update [tests/CMakeLists.txt](tests/CMakeLists.txt) to include the new test entry.
This mirrors the Qwen2 fix (commit 0d049d6), but adapts for Qwen3.5's regex. Ensures robust Unicode tokenization and prevents std::regex stack overflows.
Closes#21919.
* fix: enhance regex handling for Qwen3.5 tokenizer to include accent marks
* cont : remove trailing whitespace
---------
Co-authored-by: Kabir <kabir@example.com>
Co-authored-by: Alde Rojas <hello@alde.dev>
* spec : refactor
* spec : drop support for incompatible vocabs
* spec : update common_speculative_init()
* cont : pass seq_id
* cont : dedup ctx_seq_rm_type
* server : sketch the ctx_dft decode loop
* server : draft prompt cache and checkpoints
* server : improve ctx names
* server, spec : transition to unified spec context
* cont : sync main and drft contexts
* cont : async drft eval when possible
* cont : handle non-ckpt models
* cont : pass correct n_past for drafting
* cont : process images throught the draft context
* spec : handle draft running out of context
* server : fix mtmd draft processing
* server : fix URL for draft model
* server : add comment
* server : clean-up + dry
* speculative-simple : update
* spec : fix n_past type
* server : fix slot ctx_drft ptr
* tools : update readme
* naming : improve consistency
* spec : refactor for multi-sequence speculative context
* cont : prepare params
* cont : prepare params
* spec : support parallel drafts
* server : support parallel drafting
* llama : reuse device buffers when possible
* server, spec : clean-up
* cont : clean-up
* cont : minor
* spec : reset `drafting` flag at the end
* spec : introduce `common_speculative_process()`
* spec : allow for multiple spec types (chain of speculators)
* replace old type field of type common_speculative_type in the
common_params_speculative struct with a vector to allow multiple
types to be specified
* introduce common_get_enabled_speculative_impls(const std::vector<enum common_speculative_type>)
to figure out which implementations the user has enabled
* introduce common_speculative_type_from_names(const std::vector<std::string> & names)
to parse the already user provided spec types
* all speculators run sequentially, best one wins (we verify its drafted tokens)
* maximize expected accepted tokens for current round by calculating the
product between the probability of accepting current token (n_acc_tokens / n_gen_drafts)
and the draft's length
---------
Co-authored-by: Petros Sideris <petros.sideris@nokia.com>
Mikhail Podvitskii