84075273c8
* spec: add DSpark speculative decoding DSpark (DeepSpec, 2026) on top of the merged DFlash drafter. It reuses the DFlash encoder/decoder graph, target feature extraction and KV-cache injection, and the verify/accept path unchanged; the draft model is a new "dspark" arch adding a low-rank Markov head (markov_w1/w2) and an optional (unused here) confidence head. No new public APIs. The proposal is the only change: the block is anchor-first (position 0 already predicts the first draft) and the decoder graph applies a semi-autoregressive, previous-token conditioned logit bias in-graph, chained per block position: logits'(i) = logits(i) + markov_w2 . markov_w1[prev(i)] prev(0) = the block's anchor token, prev(i>0) = argmax(logits'(i-1)) vectorized across all blocks in the batch; the anchors are fed through a dedicated graph input (token 0 of every block). Greedy stays lossless (verify unchanged, same as DFlash). - new arch "dspark" (llama_model_dspark : llama_model_dflash, reuses the graph, loads the markov/confidence tensors; shares the target's embed/lm_head). - Qwen3DSparkModel converter. - new spec type "draft-dspark" (common_speculative_impl_draft_dspark : common_speculative_impl_draft_dflash, overrides draft() only: submits whole anchor-first blocks and greedily reads back the biased logits). * spec: read draft block size in the dflash impl * docs: add DSpark section to speculative.md * spec: keep dspark block size read in the dspark impl * dspark : add TODOs for incomplete parts - confidence head is loaded but not used yet - confidence-scheduled prefix pruning is not implemented - the in-graph Markov chain is greedy-only - only Qwen3 backbones are supported for now (also noted in docs) * spec: fold DSpark into the DFlash arch Address review: drop LLM_ARCH_DSPARK and the dspark.block_size / markov_rank GGUF keys. A DSpark draft now converts to a DFlash GGUF; the Markov head tensors are detected by presence (like eagle3 d2t), block_size is read from the existing dflash.block_size key, and the block anchors are taken as a strided view of the decoder's token input instead of a separate graph input. * spec: add confidence-based draft pruning for DSpark The DSpark confidence head predicts per-position acceptance of the drafted block. --spec-draft-conf-min truncates the block at the first position below the threshold (default 0 = disabled). * fold the dspark impl into dflash, selected by spec type * address review comments * dspark: clean up and improve naming * update readme * remove trailing whitespace * dflash: draft full n_max blocks, defer dp.n_max to the central truncation The DSpark markov head views the draft batch as a uniform [n_seqs x block] grid, but the per-seq dp.n_max clamp could produce blocks of different sizes, silently corrupting the strided views and the resulting logits. Drop the clamp and always draft the full n_max block for every sequence: dp.n_max is already enforced by the central truncation in common_speculative_draft(), the same way eagle3 handles it. Co-authored-by: Zaire404 <3147879462@qq.com> * dflash: assert the markov head block-uniformity invariant, require the conf head With the draft batch always submitting equal-size n_max blocks, a non-divisible token count can only mean the batch was split across ubatches or a caller broke the layout - fail loudly instead of silently dropping the markov bias. The block_drafts > block_size early return stays: worst-case graph reserve passes legitimately build with n_seq_tokens > block_size. Also make conf_proj required when the markov head is present: the confidence head is part of the DSpark checkpoint format, and a missing head would otherwise leave --spec-draft-conf-min silently reading stale embeddings instead of confidences. Co-authored-by: Zaire404 <3147879462@qq.com> * dspark: fold conf_min into p_min p_min and conf_min express the same thing - the minimum predicted survival probability for a drafted position - differing only in how the estimate is obtained: token probability for regular drafters, the trained confidence head for DSpark. The DSpark readback never used p_min, so reuse it for the confidence threshold and drop the separate --spec-draft-conf-min flag. Both defaulted to 0 (disabled), so behavior is unchanged. Co-authored-by: Zaire404 <3147879462@qq.com> * dflash: note the confidence broadcast workaround Requested in review: the ggml_repeat only adapts the [1, n_tok] confidences to the n_embd-wide embd_nextn transport so that llama_get_embeddings_nextn can be reused - not a placeholder. Co-authored-by: Zaire404 <3147879462@qq.com> * cont : clarify [no ci] --------- Co-authored-by: Ruixiang Wang <wangruixiang07@outlook.com> Co-authored-by: Zaire404 <3147879462@qq.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
22 KiB
Speculative Decoding
llama.cpp supports speculative decoding, a technique that can significantly accelerate token generation by predicting multiple tokens ahead of the main model.
Speculative decoding leverages the fact that computing n tokens in a batch (as in prompt processing) is more efficient than computing n sequentially (as in response generation). By generating draft tokens quickly and then verifying them with the target model in a single batch, this approach can achieve substantial speedups when the draft predictions are frequently correct.
Implementations
The llama-server application supports several implementations of speculative decoding. An implementation with draft model can be mixed with an implementation without draft model.
Draft Model (draft)
A much smaller model (called the draft model) generates drafts. A draft model is the most used approach in speculative decoding.
EAGLE-3 (draft-eagle3)
EAGLE-3 uses a small draft model that reads the target model's hidden states to predict the next tokens, so it
reaches higher acceptance than a standalone draft model of the same size. The draft is a one-layer transformer
trained for a specific target model; it shares the target model's tokenizer and, optionally, uses a reduced draft
vocabulary with its own lm_head, which is mapped back using a d2t table.
Convert the EAGLE-3 checkpoint with --target-model-dir so it inherits the target's tokenizer and the layer
indices to read. Both the SpecForge LlamaForCausalLMEagle3 and the vLLM/AngelSlim Eagle3LlamaForCausalLM
checkpoint formats are supported (for example AngelSlim/Qwen3-4B_eagle3
for Qwen/Qwen3-4B):
python convert_hf_to_gguf.py AngelSlim/Qwen3-4B_eagle3 \
--target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-eagle3.gguf
llama-server -m Qwen3-4B.gguf -md Qwen3-4B-eagle3.gguf --spec-type draft-eagle3
Supported EAGLE-3 draft models include:
- yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
- yuhuili/EAGLE3-LLaMA3.3-Instruct-70B
- RedHatAI/gemma-4-31B-it-speculator.eagle3
- RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3
- Tengyunw/qwen3_8b_eagle3
- Tengyunw/qwen3_30b_moe_eagle3
- AngelSlim/Qwen3-1.7B_eagle3
- AngelSlim/Qwen3-4B_eagle3
- AngelSlim/Qwen3-8B_eagle3
- AngelSlim/Qwen3-14B_eagle3
- AngelSlim/Qwen3-32B_eagle3
- AngelSlim/Qwen3-a3B_eagle3
- RedHatAI/gpt-oss-20b-speculator.eagle3
- lmsys/EAGLE3-gpt-oss-120b-bf16
- nvidia/gpt-oss-120b-Eagle3-long-context
For the full and up-to-date list of supported models, see #18039.
DFlash (draft-dflash)
DFlash produces an entire block of draft tokens in a single forward pass (block diffusion) and injects the target model's hidden states into the draft model's attention, instead of drafting one token at a time. This keeps the draft model small while making drafting GPU-friendly. Unlike EAGLE-3 (a single-layer autoregressive draft), the DFlash draft uses several transformer layers but emits a whole block per draft step.
The draft is a small block-diffusion model trained for a specific target (for example
z-lab/Qwen3-4B-DFlash for Qwen/Qwen3-4B). Convert it with --target-model-dir so it inherits the
target's tokenizer and token embeddings:
python convert_hf_to_gguf.py z-lab/Qwen3-4B-DFlash \
--target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-DFlash.gguf
llama-server -m Qwen3-4B.gguf -md Qwen3-4B-DFlash.gguf \
--spec-type draft-dflash --spec-draft-n-max 15 -fa on --jinja
--spec-draft-n-max is clamped to the draft model's trained block size.
See:
- #22105
DSpark (draft-dspark)
DSpark extends DFlash with a semi-autoregressive Markov head: the draft still emits a whole block per forward pass, but each block position's logits are biased by a low-rank term keyed on the previous token, chained in-graph across the block. This keeps drafting at one decode per block while recovering some of the left-to-right signal that pure block diffusion loses.
The draft is a small DeepSpec checkpoint trained for a specific target (for example
deepseek-ai/dspark_qwen3_4b_block7
for Qwen/Qwen3-4B). Convert it with --target-model-dir so it inherits the target's tokenizer
and token embeddings:
python convert_hf_to_gguf.py deepseek-ai/dspark_qwen3_4b_block7 \
--target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-DSpark.gguf
llama-server -m Qwen3-4B.gguf -md Qwen3-4B-DSpark.gguf \
--spec-type draft-dspark --spec-draft-n-max 7 -fa on --jinja
--spec-draft-n-max is clamped to the draft model's trained block size.
--spec-draft-conf-min P truncates each drafted block at the first position whose predicted
acceptance (from the draft's confidence head, if present) falls below P (default 0 = disabled).
Currently only drafts with a Qwen3 backbone are supported; support for other backbones (e.g. Gemma4) is planned.
See:
- #25173
n-gram Cache (ngram-cache)
An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences. A draft is computed using probabilities derived from these statistics. External statistics can also be loaded from files for improved accuracy.
See:
- #5479, #6828, #6848
n-gram Map (ngram-simple, ngram-map-*)
These implementations search the token history for patterns and use matching sequences as draft candidates. They require no additional model but rely on patterns that have already appeared in the generated text. An example to use this approach can be the rewriting of source code by a LLM.
n-gram Map (ngram-simple)
This implementation looks for the last n-gram in history that matches the current n-gram and creates a draft using the m tokens following the matched n-gram. It is the simplest self-speculative approach with minimal overhead.
llama-server [...] --spec-type ngram-simple --spec-draft-n-max 64
n-gram Map Key (ngram-map-k)
This implementation looks for the current n-gram of size n (called the key) in the token history. If the key n-gram is followed by the same m tokens (called the mgram) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument --spec-ngram-map-k-min-hits, default is 1) before generating drafts.
The number of accepted tokens is stored for each used n-gram.
Example:
llama-server [...] --spec-type ngram-map-k --spec-draft-n-max 64
n-gram Map Key-4-Values (ngram-map-k4v)
This experimental implementation looks for the current n-gram of size n (called the key) in the token history. For each key, up to four values (n-grams of size m, called mgrams) are tracked. An internal statistic counts the occurrences of each mgram after the key n-gram. If one mgram is significantly more frequent than the others, it is used as the draft.
The number of accepted tokens is stored for each used n-gram.
Example: Server options to be used if there are a lot of longer repetitions.
llama-server [...] --spec-type ngram-map-k4v --spec-ngram-map-k4v-size-n 8 --spec-ngram-map-k4v-size-m 8 --spec-ngram-map-k4v-min-hits 2 --spec-draft-n-max 64
n-gram Mod (ngram-mod)
Add basic ngram hasher for speculative decoding:
- For each ngram, compute a hash using LCG
- For each computed hash, store the next token
- During speculation, iteratively compute the rolling hash of the last n tokens and pick the next token from the storage
Some characteristics:
- Lightweight (~16 MB)
- Constant memory and complexity
- Can generate variable draft lengths (i.e. m is not fixed)
Currently, a single hash pool is shared across all server slots, so different requests can benefit from each other.
Sample usage:
# notes:
# - small `n` are not recommended
# - MoEs require long drafts
# - dense models: can reduce `--spec-ngram-mod-n-min` and `--spec-ngram-mod-n-max`
llama-server ... --spec-type ngram-mod --spec-ngram-mod-n-match 24 --spec-ngram-mod-n-min 48 --spec-ngram-mod-n-max 64
Applications:
- Iterating over a block of text/code (e.g. in llama.vim)
- Reasoning models (when they have to repeat their thinking in the final answer)
- Summarization
Example Video:
- See #19164
Differences between ngram-simple, ngram-map and ngram-mod
- ngram-simple looks for a previous matching n-gram and inserts the following m-gram.
- ngram-map-k looks for a previous matching n-gram and inserts the following m-gram but uses an internal hash-map of n-grams in the current context window.
- ngram-mod uses a hash pool which is shared across all server slots. The hash pool is a map from n-gram hash to the next token (not the next m-gram as in ngram-map).
Command-Line Options
If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.
General Speculative Parameters
--spec-type [none|draft-simple|draft-eagle3|draft-dflash|draft-dspark|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
comma-separated list of types of speculative decoding to use
(default: none)
(env: LLAMA_ARG_SPEC_TYPE)
--spec-default use default speculative decoding config
(enables ngram-mod)
Draft Model Parameters
--spec-draft-model, -md, --model-draft FNAME
draft model for speculative decoding (default: unused)
(env: LLAMA_ARG_SPEC_DRAFT_MODEL)
--spec-draft-hf, -hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]
HuggingFace repository for the draft model
(env: LLAMA_ARG_SPEC_DRAFT_HF_REPO)
--spec-draft-n-max N
number of tokens to draft for speculative decoding (default: 3)
(env: LLAMA_ARG_SPEC_DRAFT_N_MAX)
--spec-draft-n-min N
minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_SPEC_DRAFT_N_MIN)
--spec-draft-p-split, --draft-p-split P
speculative decoding split probability (default: 0.10)
(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT)
--spec-draft-p-min, --draft-p-min P
minimum speculative decoding probability (greedy) (default: 0.00)
(env: LLAMA_ARG_SPEC_DRAFT_P_MIN)
--spec-draft-ngl, -ngld, --gpu-layers-draft, --n-gpu-layers-draft N
max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT)
--spec-draft-device, -devd, --device-draft <dev1,dev2,..>
comma-separated list of devices to use for offloading the draft model
(use --list-devices to see available devices)
Draft Model CPU Scheduling Parameters
--spec-draft-threads, -td, --threads-draft N
number of CPU threads to use during generation
--spec-draft-threads-batch, -tbd, --threads-batch-draft N
number of threads to use during batch and prompt processing (default: same as --threads-draft)
--spec-draft-cpu-mask, -Cd, --cpu-mask-draft M
Draft model CPU affinity mask. Complements cpu-range-draft
--spec-draft-cpu-range, -Crd, --cpu-range-draft lo-hi
Ranges of CPUs for affinity. Complements --cpu-mask-draft
--spec-draft-cpu-strict, --cpu-strict-draft <0|1>
Use strict CPU placement for draft model (default: same as --cpu-strict)
--spec-draft-prio, --prio-draft N
set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime
--spec-draft-poll, --poll-draft <0|1>
Use polling to wait for draft model work (default: same as --poll)
--spec-draft-cpu-mask-batch, -Cbd, --cpu-mask-batch-draft M
Draft model CPU affinity mask for batch. Complements cpu-range-batch-draft
--spec-draft-cpu-range-batch, -Crbd, --cpu-range-batch-draft lo-hi
Ranges of CPUs for affinity for batch. Complements --cpu-mask-batch-draft
--spec-draft-cpu-strict-batch, --cpu-strict-batch-draft <0|1>
Use strict CPU placement for draft model batch (default: --cpu-strict-draft)
--spec-draft-prio-batch, --prio-batch-draft N
set draft process/thread priority for batch : 0-normal, 1-medium, 2-high, 3-realtime
--spec-draft-poll-batch, --poll-batch-draft <0|1>
Use polling to wait for draft model work for batch (default: --poll-draft)
Draft Model KV Cache and Tensor Override Parameters
--spec-draft-type-k, -ctkd, --cache-type-k-draft TYPE
KV cache data type for K for the draft model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(env: LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_K)
--spec-draft-type-v, -ctvd, --cache-type-v-draft TYPE
KV cache data type for V for the draft model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(env: LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V)
--spec-draft-override-tensor, -otd, --override-tensor-draft <tensor name pattern>=<buffer type>,...
override tensor buffer type for draft model
--spec-draft-cpu-moe, -cmoed, --cpu-moe-draft
keep all Mixture of Experts (MoE) weights in the CPU for the draft model
(env: LLAMA_ARG_SPEC_DRAFT_CPU_MOE)
--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft N
keep the MoE weights of the first N layers in the CPU for the draft model
(env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE)
n-gram Mod Parameters
--spec-ngram-mod-n-match N
ngram-mod lookup length (default: 24)
--spec-ngram-mod-n-min N
minimum number of ngram tokens to use for ngram-based speculative decoding (default: 48)
--spec-ngram-mod-n-max N
maximum number of ngram tokens to use for ngram-based speculative decoding (default: 64)
n-gram Simple Parameters
--spec-ngram-simple-size-n N
ngram size N for ngram-simple speculative decoding, length of lookup n-gram (default: 12)
--spec-ngram-simple-size-m N
ngram size M for ngram-simple speculative decoding, length of draft m-gram (default: 48)
--spec-ngram-simple-min-hits N
minimum hits for ngram-simple speculative decoding (default: 1)
n-gram Map Key Parameters
--spec-ngram-map-k-size-n N
ngram size N for ngram-map-k speculative decoding, length of lookup n-gram (default: 12)
--spec-ngram-map-k-size-m N
ngram size M for ngram-map-k speculative decoding, length of draft m-gram (default: 48)
--spec-ngram-map-k-min-hits N
minimum hits for ngram-map-k speculative decoding (default: 1)
n-gram Map Key-4-Values Parameters
--spec-ngram-map-k4v-size-n N
ngram size N for ngram-map-k4v speculative decoding, length of lookup n-gram (default: 12)
--spec-ngram-map-k4v-size-m N
ngram size M for ngram-map-k4v speculative decoding, length of draft m-gram (default: 48)
--spec-ngram-map-k4v-min-hits N
minimum hits for ngram-map-k4v speculative decoding (default: 1)
--spec-type TYPE
Specifies a comma-separated list of speculative decoding types to use.
| Type | Description |
|---|---|
none |
No speculative decoding (default) |
draft-simple |
Use a simple draft model for speculation |
draft-eagle3 |
Use an EAGLE-3 draft model that reads the target's hidden states |
draft-dflash |
Use a DFlash block-diffusion draft model that emits a block per step |
draft-dspark |
Use a DSpark draft model (DFlash backbone + semi-autoregressive Markov head) |
draft-mtp |
Use Multi Token Prediction (MTP) heads from the main model |
ngram-cache |
Use n-gram cache lookup |
ngram-simple |
Use simple n-gram pattern matching |
ngram-map-k |
Use n-gram pattern matching with n-gram-keys |
ngram-map-k4v |
Use n-gram pattern matching with n-gram-keys and up to four m-gram values (experimental) |
ngram-mod |
Use basic ngram hasher for speculative decoding with shared pool |
Example: Server-instance used to refactor source code.
./llama-server [...] --spec-type ngram-simple
Example: Multiple speculative implementations.
./llama-server [...] --spec-type ngram-mod,ngram-map-k4v
--spec-ngram-*-size-n N
Sets the size N of the lookup n-gram for n-gram map based speculative decoding. The n-gram size N determines how many tokens in a row to look back when searching for matching patterns.
Each n-gram implementation has its own parameter:
--spec-ngram-simple-size-nforngram-simple--spec-ngram-map-k-size-nforngram-map-k--spec-ngram-map-k4v-size-nforngram-map-k4v--spec-ngram-mod-n-matchforngram-mod
--spec-ngram-*-size-m M
Sets the size M of the draft m-gram for n-gram map based speculative decoding. The m-gram size determines how many tokens to draft when a match is found. Larger values can provide more speedup but may reduce acceptance rate.
Each n-gram implementation has its own parameter:
--spec-ngram-simple-size-mforngram-simple--spec-ngram-map-k-size-mforngram-map-k--spec-ngram-map-k4v-size-mforngram-map-k4v
--spec-ngram-*-min-hits H
This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
Each n-gram implementation has its own parameter:
--spec-ngram-simple-min-hitsforngram-simple--spec-ngram-map-k-min-hitsforngram-map-k--spec-ngram-map-k4v-min-hitsforngram-map-k4v
Statistics
Each speculative decoding implementation prints statistics.
draft acceptance rate = 0.57576 ( 171 accepted / 297 generated)
statistics ngram_simple: #calls = 15, #gen drafts = 5, #acc drafts = 5, #gen tokens = 187, #acc tokens = 73
statistics draft: #calls = 10, #gen drafts = 10, #acc drafts = 10, #gen tokens = 110, #acc tokens = 98
draft acceptance rate = 0.70312 ( 90 accepted / 128 generated)
statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
#calls(b,g,a): number of calls of begin (new prompt), generation and accumulation of this implementations#gen drafts: number of drafts generated by this implementation#acc drafts: number of drafts accepted (partially) by the main model#gen tokens: number of tokens generated by this implementation (including rejected tokens)#acc tokens: number of tokens accepted by the main model- `dur(b,g,a): durations of begin (new prompt), generation and accumulation (process acceptance).
Benchmarking
To measure the end-to-end effect of speculative decoding (throughput, latency, and draft acceptance) across diverse prompts, see the SPEED-Bench client in tools/server/bench/speed-bench.
It runs against a running llama-server and can compare a baseline run against a speculative-decoding run.