* DeepseekV4: fix rollback with multi-seq
* fix model loading
* make pending rollback single use
* only clear cache for seq_id for full load
* add assert for compress ratio
* make graph topology static
* pass true instead of flags in clear_compressed
* cont : clean-up + TODOs
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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 <bloomer@booper.brushtail.me>
Co-authored-by: Dyluhn <dylanranejohnston1@gmail.com>
* 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) <noreply@anthropic.com>
* 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) <noreply@anthropic.com>
* Update conversion/kimi_k3.py
Co-authored-by: Boris Dvorkin <b_dvorkin@niuitmo.ru>
* 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 <noreply@anthropic.com>
* 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 <noreply@anthropic.com>
* 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 <b_dvorkin@niuitmo.ru>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Deepankar Singh <singh.deepankar39@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Caleb DeLeeuw <caleb.deleeuw@gmail.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* 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 <qscqesze@gmail.com>
* 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 <sigbjorn.skjaeret@huggingface.co>
* model : use common functions to make MiniMax-Text-01 implementation more concise
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* 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 <sszymczy@gmail.com>
Co-authored-by: QscQ <qscqesze@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Initial changes for Recurrent state rollback for nemotron for cpu and cuda
* Removing CPU RS rollback. Will enable it in subsequent PRs
* addition of test case
* Removing assert and calling runtime API to check if op is supported
* removing extra API and updating the call sites for K
* replace static cuda detection to runtime fused_op api
* address review comments and fallback when SSM rollback not supprted
* Adding changes for supporting RS-rollback in CPU. Also added test-backend-ops for cpu and cuda
* removing memory manipulation as rs rollback is now supported in CPU
* removing the static probe which is not needed now
* correcting the format
* address review comments
* enabling test for all the backends, unsupported backends will fallback to CPU
* Apply suggestions from code review
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* choose different graph based on the result of fused_ssm_op is supported or not and also handled memory->n_rs_seq >1 case incase of op is not supported
* Support K > 1 in ssm_scan for all backends
* Fix CI Issues
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Gaurav Garg <gaugarg@nvidia.com>
* server : save serialized image chunks at the end of the llama state
* server : support multimodal slot state save/restore with packed payload
* server : refine image slot state serialization
* server : support media slot state and centralize media validation
* server : remove unnecessary comment
* server : remove defensive media checks and move the chunk type check to validate()
* Enable backend sampling with token speculation
* Clamp the mask sum before converting it into the sampled index
* Add a numeric context parameter declaring the maximum outputs one sequence
* More fixes
* Don't reuse memory for output views.
* Match dist between CPU and GPU
* Fix CPU and backend sampling mismatches
* Simpify some of the changes
* Fix tests on Vulkan
* More test fixes
* Rebase changes
* Rebase and address review comments
* Address review comments
* Address review comments
* Update src/llama-sampler.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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 : enforce the same K and V cache types for DeepSeek V4; enable FA if V cache is quantized
* llama : enforce the same K and V cache types for MLA models
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* 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 <1629204+CISC@users.noreply.github.com>
* Remove whitespaces from constants.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Tighten comment in minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* 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 <1629204+CISC@users.noreply.github.com>
* Update gguf-py/gguf/gguf_writer.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update gguf-py/gguf/tensor_mapping.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Update conversion/minimax.py
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove whitespace in src/llama-kv-cache.cpp
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove Whitespace in Update src/llama-model.h
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* Remove whitespace in src/llama-hparams.h
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
* 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 <danielhanchen@gmail.com>
Co-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* convert: add dsv4 conversion
* add basic setup
* add llm_graph_input_dsv4
* add save-load state
* add sinkhorn eps - correction by @fairydreaming
* add rope fix
* cleanup dead code
* fix bugs
* support pro model: added by @fairydreaming
* remove redundant V cache
* Chat template
* remove debugging leftovers
* Add mechanism for inlining templates based on architecture
* s/deepseek-v4-flash/deepseek4/g
* s/deepseek-v4-flash/deepseek4/g continued
* enable graph reuse
* enable FA
* fix test llama archs
* rename
* compatibility with antirez ds4 GGUFs
* simplified set_gguf_parameters() by calling super class method, replaced moe.score_func with expert_gating_func.
* reserve worst-case kv-cache
* revert max split inputs
* address review comments
* add padding to enable FA
* pad only the final value of plan.n_kv to 256
* remove built-in cpp chat template
* cont: remove cpp built-in template
* rm outdated test
* replace ggml_view_3d() with ggml_reshape_3d()
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* only support n_seq=1 for now
* remove unused var
* cont: remove unused var
* use scale bias
* use correct ptr for can_reuse
* remove gen-chat-inline-templates.py
* simplify graph reuse
* cont: cleanup
* remove unused inputs
* enable partial checkpointing
* add correct shape for kq_mask + set llama_model_n_swa to 0 for dsv4
* precompute source_idx + add comment about dummy write
* support multi-seq
* remove restored_trim_pos
* use split_equal when possible
* fix indent
* address review comments
* use LLM_KV
* fix ci
---------
Co-authored-by: Piotr Wilkin <piotr.wilkin@syndatis.com>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* 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
* 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.
* 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>
* 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>
* ggml: backend-agnostic tensor parallelism
* support for GPT-OSS, Qwen 3 MoE
* partial Vulkan fix
* add support for 4/8 GPUs
* unconditional peer access
* re-use buffers + ggml contexts
* fix output pattern
* NCCL support
* GGML: HIP: add RCCL support
* Remove shfl and AllReduce from backend interface
* move allocation workaround out of ggml-alloc.c
* 2d tensor set/get support
* Fix the seg fault without NCCL
* Apply suggestion from JohannesGaessler
* support for tensor dims % n_devs != 0
* fix view_offs scaling
* arbitrary num. of GPUs/tensor split
* fix compilation
* better granularity estimate
* Support device-specific host buffer types if all underlying backends expose the same type. This allows using pinned memory instead of pageable memory for CUDA.
Fix compilation errors.
* partial Qwen 3 Next support
* Fix qwen3 30b (#8)
* Fix crash with Qwen-30B-A3B Q4_0
Qwen-30B-A3B Q4_0 has an intermediate dimension of 768. Using a granularity of 256 forces an uneven split between GPUs, which is not supported by the current implementation.
* Decide block size based on tensor quantization type
* Fix crashes due to KV cache serialization (#9)
KV cache serialization requires non-zero offsets on the tensor. Add support in the meta backend to set/get a tensor with a non-zero offset.
* metal : fix build (#7)
* static memory allocations, fix usage count
* fix tensor granularity
* more even memory distribution
* use BF16 for allreduce
* rebase fixup
* better error message for unsupported architectures
* Fix device mismatch during scatter of allReduce. (#11)
There is a mismatch between the dst buffer device and the backend device, causing the use of sync copies
* Enable the previous allreduce implementation. It is better in both perf and stability (#12)
* delay AllReduce for Moe for less I/O
* build : clean-up compile warnings
* backend : move most of the meta backend API to ggml-backend-impl.h
* cont : hide unused public API in the implementation
* llama : use llama_device + remove ggml_backend_dev_is_meta()
* ggml-backend : remove unused alloc include
* minor : remove regex include
* ggml : introduce ggml-ext.h for staging new APIs
* rebase fixup
* fix tests
* llama : more robust logic for determining Meta devices (#16)
* llama : more robust logic for determining Meta devices
* cont : fix devs size check
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* cont : fix log type
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* disable roundtrip for meta backend
* fix arch selection
* Qwen 3.5 support
* fix Gemma 4 MoE
* fix OpenVino, SYCL
* fix test-llama-archs for CPU-only builds
* Fix Qwen 3.5 MoE
* disable meta backend tests for WebGPU
* tests : filter CPU-based devices from the Meta backend tests (#17)
* meta : formatting, naming, indentation (#18)
* formatting : llama-model.cpp
* formatting : ggml-ext.h
* formatting : ggml-backend-meta.cpp
* meta : add TODO
* add documentation
* better error messages
* fix GPT-OSS
---------
Co-authored-by: Carl Philipp Klemm <carl@uvos.xyz>
Co-authored-by: Gaurav Garg <gaugarg@nvidia.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
The MEAN/CLS/LAST pooling paths in encode() and decode() used
n_embd_inp() (16384 for qwen3vl with deepstack) to read from the
pooled embedding tensor, which only has n_embd_out() (4096) floats
per sequence. This caused a tensor read out of bounds assertion.
Fixes embedding mode for Qwen3-VL-Embedding models.
* context: zero output buffer on allocation
Address GHSA-wqq9-25mr-rw76.
The logits output buffer allocated in output_reserve() uses
posix_memalign(), which does not zero memory. The buffer is only
written during decode when needs_raw_logits() returns true. When
backend samplers cover all output sequences, needs_raw_logits()
returns false and the buffer is never written, but
llama_get_logits() still returns a pointer to it, exposing stale
heap content.
Zero the buffer after allocation to prevent information disclosure
through the public logits API.
Found-by: Pwno
* Update src/llama-context.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix pooling assertion crash in chunked GDN detection path
The chunked fused Gated Delta Net detection in sched_reserve() calls
graph_reserve(16*n_seqs, n_seqs, n_outputs, ...) where n_outputs = n_seqs.
This creates a dimension mismatch in build_pooling() for embedding models
with mean/rank pooling: build_inp_mean() creates a tensor with shape
[n_tokens=16*n_seqs, ...] while t_embd is reduced to [n_outputs=n_seqs, ...]
via out_ids, causing ggml_mul_mat to assert on ggml_can_mul_mat(a, b).
Fix: pass n_tokens as n_outputs in the chunked GDN graph reservation,
matching the pattern used by the pp/tg worst-case reservations.
Regression introduced by #20340 (d28961d).
Same class of bug as #12517, fixed by #12545.
* server : add mean pooling tests to embedding test suite
Add test_embedding_pooling_mean and test_embedding_pooling_mean_multiple
to cover the --pooling mean codepath, which was previously untested.
These tests would have caught the regression introduced by #20340 where
build_pooling() crashes with a ggml_mul_mat assertion due to mismatched
dimensions in the chunked GDN detection path.
---------
Co-authored-by: Domenico Crupi <domenico@zerovolt.it>
* tests: allow loading test-backend-ops tests from json
* add error threshold based on op
* add error when file cannot be read
* add graph operator json extraction tool
* add nb parameter for non-contiguous input tensors
* fix view check
* only use view if non-contiguous/permuted, use C++ random instead of rand()
* replace internal API calls with public llama_graph_reserve call
* reduce test description length
* fix nb[0] not getting set for view
* add name to tests
* fix inplace error
* use text file instead of json
* move llama_graph_reserve function to new llama-ext header, move export-graph-ops to tests/
* fix missing declaration
* use pragma once
* fix indent
* fix Windows build
Aman Gupta
Toby
Piotr Wilkin (ilintar)
fairydreaming
lnigam
Chipmunk
Gaurav Garg
Bar Haim
o7si
zql
Georgi Gerganov
timkhronos
Ruixiang Wang
YiChen Lv
Sigbjørn Skjæret
Johannes Gäßler
Andrei
Erik Scholz
lainon1
Tom Hillbrunner
Ruikai Peng
Andreas Obersteiner
ZeroV0LT
Ruben Ortlam