speculative: dspark block-diffusion drafter + CUDA Markov resample; Q1_0/Q2_0 dp4a fix - #55

Merged

khosravipasha merged 15 commits into

prismfrom

feat/dspark-public

Jul 13, 2026

Commits on Jul 12, 2026

speculative: dspark block-diffusion drafter + CUDA Markov resample; Q1_0/Q2_0 dp4a fix
bri-prism
committed
5eaa904
cuda: forward-declare ggml_cuda_mul_mat_q1_hopper (fix -Werror=missing-declarations)
bri-prism
committed
9d0790c
tests: fix dspark test cross-platform builds (macos -Werror, windows DLL link)
bri-prism
committed
3194b28
tests: use public vocab API for dspark n_vocab (windows DLL link)
bri-prism
committed
4ddf039
dspark: widen capture-copy size math to size_t (CodeQL overflow)
bri-prism
committed
b8ab101
llama-context: bound capture-layer writes and zero unpopulated capture rows
bri-prism
committed
e9b18b1
speculative: harden dspark drafter/target contract and recovery paths
bri-prism
committed
8ab5a01
convert: map dspark per-layer tensors under drafter./dspark. wrappers
bri-prism
committed
4f2e334
common: build the dspark Markov CUDA TU for ggml's architectures, not SM80 only
bri-prism
committed
1b1610f
tests: gate dspark tier-2 on agreement and actually run the rs-ring test
bri-prism
committed
a5c725c
docs: correct dspark Markov CUDA default and drop stale scaffolding claims
bri-prism
committed
312482c
tests: run rs-ring rotation on CPU (-ngl 0)
bri-prism
committed
2605f9e
dspark: load and run drafters with GIDD log-SNR conditioning
bri-prism
committed
44571d5
dspark: drop unused pos from llama_set_dspark_ctx; document draft-dspark contract
bri-prism
committed
fd569b3

Commits on Jul 13, 2026

speculative: validate the target's configured capture-layer count in dspark process()
bri-prism
committed
bd9ec19