teamblobfish/DeepSeek-V4-Pro-GGUF

README: drop multi-GPU caveats list (kept recommended config)

README: add WIP multi-GPU CUDA section with recommended layer-split config

README: CUDA build — add SM table, BUILD_TYPE=Release, multi-GPU needs both CXX+CUDA flags

README: drop #21149 gate reference, lead with fork install

README: correct -cmoe technical explanation (op_offload + peak-liveness compute buffer) per codex audit

README: document -cmoe + -ub 128 recommendation (fairydreaming's RTX PRO 6000 OOM report)

README: document -DGGML_SCHED_MAX_SPLIT_INPUTS=128 flag for multi-GPU CUDA builds

README: CUDA support landed (19/19 on RTX 5090), branch renamed to feat/v4-port-cuda

Add Q8_0 shards

Add Q4_K_M-XL shards

README: Q2_K-XL gate-tools ✓ pass; add real timing numbers

Add Q2_K-XL shards

README: explicit CUDA / non-Metal backends not supported (Metal-only ops)

Add README

Add DSML chat template

initial commit