A llama.cpp fork that verifies 8 draft tokens in 33 ms instead of 54.
Qwen3.8-27B + DFlash2 speculative decoding on one RTX 3090. Kernel changes keep the output bit-identical.
BASED ON 8DF332DE1 CUDA / SM_86 OPT-IN FEATURES MIT
What's different · Tested with · Build · Switches · Upstream llama.cpp
This is llama.cpp at upstream 8df332de1 (tag
upstream-base) plus a linear series of commits on opt/main. They target one setup:
a 27B hybrid Gated-DeltaNet/attention model (Qwen3.8-27B, IQ4_XS) drafting with DFlash2 on an
RTX 3090, verifying 8 tokens per step at long context. On that setup a verify step drops from
~54 ms to ~33 ms.
Everything else is upstream llama.cpp. Its documentation lives in docs/, with build instructions and the server README.
| Change | Effect | |
|---|---|---|
| MMSQ | Small-batch (N <= 16) quantized GEMM on int8 tensor cores for IQ4_XS, Q4_K, Q5_K, Q6_K: split-K, activation reuse, occupancy-aware dispatch | Weights stream at 750-830 GB/s of ~840; most of the speedup |
| Fused norm | Residual add + RMS norm + weight + MMSQ activation quantization as one kernel | 127 fewer kernels per step, bit-exact |
| Q4 MMA attention | q4_0 K/V decoded straight into the 8-query MMA flash-attention tiles | 95.8 -> 109.7 tok/s at 91K context |
| GPU grammar | Tool-call grammars compiled to a DFA, masked and sampled on the GPU (Philox RNG) | No logits copy to the host on tool requests |
| DFlash features | Target features for the DFlash2 draft stay in a CUDA staging tensor | No host round trip per step |
| Small kernels | Small top-k, conv-state snapshot fusion, small F32 matmul | Fewer, cheaper launches |
| Sleep cache | Server snapshots KV and recurrent state to disk on idle unload and restores it on wake | Long contexts survive idle sleep |
| Race fix | Write-after-read race on KQ in flash_attn_ext_vec (fixed upstream in #30573, after upstream-base) |
racecheck 2.2M hazards -> 0 |
GPU sampling falls back to standard sampling for any request it can't serve (penalties, logprobs, reasoning budget, regex triggers), so no request is rejected.
| Target | Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-IQ4_XS.gguf (HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF @ 993a5971) |
| Draft | stock Qwen3.8-27B-DFlash2-Q4_K_M.gguf (z-lab/Qwen3.8-27B-DFlash2-GGUF @ 2d9571f8), 7 drafted + 1 = 8 tokens verified per step |
| KV cache | target q4_0 for K and V (3.6x smaller than f16, 1.7 GiB at 98K), draft f16 |
| Context | 98,304 tokens |
| GPU | RTX 3090 24 GB (sm_86), eGPU on PCIe 4.0 x4 (OCuLink) |
| ms per verify step | explanation | python100 (tools) | story |
|---|---|---|---|
upstream 8df332de1 |
53.3 | 55.1 | 53.8 |
| llama.cpp-opt | 32.7 | 33.4 | 33.3 |
Same files, flags and sampling, run back to back. Every fork change since 2026-10-04 keeps the output bit-identical, checked by output hashes.
Model SHA-256 and revisions
target 034b4c6b7c5cd9f4a2f99570d5097698f9d1e265a0fb000207223cbf9f9bd0b0 rev 993a5971fda8f30dd1b7eb2654792ba4415c7460
draft 1a25c56858e1ebe93f2718ac1d49d1151f9323325c1bbfd6209370f4db131ebd rev 2d9571f8ce46e151f61c6499c99dee6079e1d610
git clone -b opt/main https://github.com/invaliddayta/llama.cpp-opt && cd llama.cpp-opt
cmake -S . -B build -G Ninja -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86 -DCMAKE_BUILD_TYPE=Release \
-DGGML_CUDA_FA_QUANTS="q4_0-q4_0;q8_0-q8_0;f16-f16;bf16-bf16"
ninja -C build llama-serverExample server for the target setup:
GGML_CUDA_FATTN_Q4_MMA=1 LLAMA_GPU_SAMPLING=1 LLAMA_DFLASH_GPU_FEATURES=1 \
build/bin/llama-server -m target-IQ4_XS.gguf --ctx-size 98304 --gpu-layers 999 --flash-attn on \
--cache-type-k q4_0 --cache-type-v q4_0 \
--spec-draft-model DFlash2-Q4_K_M.gguf --spec-type draft-dflash --spec-draft-ngl 999 \
--spec-draft-n-max 7 --spec-draft-p-min 0 --backend-sampling --jinjaThe benchmarks also used --parallel 1 --split-mode none --spec-draft-device CUDA0 --spec-draft-poll 1 --spec-draft-type-k f16 --spec-draft-type-v f16 --threads 12 --threads-batch 16 --batch-size 2048 --ubatch-size 512 --reasoning-format deepseek --poll 10.
| Env var | Default | Effect |
|---|---|---|
GGML_CUDA_FATTN_Q4_MMA=1 |
off | Fused q4_0 KV loader in MMA flash attention (sm86, head 256, 24/4 heads, 8 queries) |
LLAMA_GPU_SAMPLING=1 |
off | Sampling and tool grammar on the GPU, per-request fallback |
LLAMA_DFLASH_GPU_FEATURES=1 |
off | DFlash target features stay on the GPU |
GGML_CUDA_MMSQ=0 |
on | Turns MMSQ (and its fusions) off |
GGML_CUDA_MMSQ_FUSE_NORM=0 |
on | Turns only the fused norm/quantization off |
LLAMA_SLEEP_CACHE_DIR |
unset | Sleep cache: with --sleep-idle-seconds N, slots are written here on sleep and restored on wake |
LLAMA_SLEEP_CACHE_KEY |
required with DIR | Cache identity; change it whenever model, build or settings change |
LLAMA_SLEEP_CACHE_MAX_MIB |
16384 | Size limit of the snapshot |
MMSQ needs Ampere or newer. Shapes, types and batch sizes outside its range use upstream's kernels, and every opt-in leaves the upstream path untouched when off.
opt/main is a linear series on top of upstream-base. The whole difference as one patch:
git diff upstream-base opt/main -- . ':!README.md' ':!media/llama-cpp-opt.svg'MIT, like upstream llama.cpp (LICENSE). Not affiliated with ggml-org.