Views
No views yet
DFlashLagunaForCausalLM (6 sliding-attention layers, block_size 16).draft_vocab_size == vocab_size (no d2t/t2d).laguna_dflash) and TRT-LLM (pytorch DFlash backend) as the draft model in a speculative config.1vllm serve --model poolside/Laguna-S-2.1 \
2 --speculative-config '{"model":"poolside/Laguna-S-2.1-DFlash","num_speculative_tokens":7,"method":"dflash"}' \
3 ...1python -m sglang.launch_server --model-path poolside/Laguna-S-2.1 \
2 --speculative-algorithm DFLASH \
3 --speculative-draft-model-path poolside/Laguna-S-2.1-DFlash \
4 ...laguna-s-2.1-DFlash-BF16.gguf) is
published in
poolside/Laguna-S-2.1-GGUF
alongside the target GGUFs. It embeds the target tokenizer and the DFlash metadata
(dflash.decoder_arch = laguna, capture layers, block size), so it works directly
as the -md draft model:1git clone --branch laguna https://github.com/poolsideai/llama.cpp
2cd llama.cpp && cmake -B build && cmake --build build -j
3
4./build/bin/llama-server \
5 -m laguna-s-2.1-Q4_K_M.gguf \
6 -md laguna-s-2.1-DFlash-BF16.gguf \
7 --spec-type draft-dflash --spec-draft-n-max 7 \
8 -fa on --jinja --port 8000--spec-draft-n-max is clamped to the trained block size (15 draft tokens + 1).[!NOTE] Requires Poolside's llama.cpp fork, branchlaguna. Upstream llama.cpp ships the generic DFlash framework but not the Laguna decoder contract this draft model needs, and upstream PR ggml-org/llama.cpp#25165 covers the target architecture only.