[Maphist0]

Overview

Add support for DiffuSpace's Clover-35B-A3B-Preview, including GGUF conversion, loading and hybrid attention.

Following the DFlash #22105 and DFlash2 series (#27342, #27816), this adds another speculative decoding mode:

• draft-self, for self-speculative generation through the existing speculative framework.

The same model proposes masked blocks and verifies them causally. Block construction is shared with DFlash; no separate draft model is required.

The speculative example includes opt-in entropy-threshold acceptance for unconstrained temperature-zero decoding.

Additional information

Validated with BF16 CPU and H200 CUDA inference.

Validation:

CPU compile:

cmake -S source -B build-clover-cpu \
    -DGGML_CUDA=OFF \
    -DGGML_BLAS=OFF \
    -DLLAMA_CURL=OFF \
    -DLLAMA_BUILD_TESTS=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
    -DCMAKE_BUILD_TYPE=Release

cmake --build build-clover-cpu --target llama-speculative-simple

CUDA compile:

cmake -S source -B build-clover-cuda \
    -DGGML_CUDA=ON \
    -DCMAKE_CUDA_ARCHITECTURES=90a \
    -DLLAMA_CURL=OFF \
    -DLLAMA_BUILD_TESTS=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
    -DCMAKE_BUILD_TYPE=Release

cmake --build build-clover-cuda --target llama-speculative-simple

Ckpt convert:

python source/convert_hf_to_gguf.py \
    <clover_ckpt_path> \
    --outtype bf16 \
    --outfile clover-bf16.gguf

The example llama-speculative-simple program:

build-clover-cuda/bin/llama-speculative-simple \
    -m clover-bf16.gguf \
    --spec-type draft-self \
    --spec-draft-n-max 15 \
    --spec-accept-ratio 0.9 \
    --spec-accept-delta 0.6 \
    -ctk bf16 -ctv bf16 \
    -f explanation.prompt \
    -n 64 -c 1024 -b 256 -ub 256 \
    --temp 0 --top-k 1 --seed 42 \
    -ngl 99 -t 8 -tb 8 --flash-attn on --color off

// another example

build-clover-cuda/bin/llama-speculative-simple \
    -m clover-bf16.gguf \
    -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
    --spec-type draft-self \
    --spec-draft-n-max 15 \
    --color on \
    --spec-accept-ratio 0.9 \
    --spec-accept-delta 0.6 \
    -ctk bf16 -ctv bf16 \
    -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4

The current implementation generates around 32 tokens/s on an H200, with substantial room for improvement. Further performance optimizations are planned as separate PRs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, I've used GPT-6.1 Sol for majority of code generation under my review. Code generated by AI was reviewed before adding into the codebase.