# b9743 Upgrade — 2026-06-20

## What Changed

Both binaries rebuilt from llama.cpp master **b9743 (c57607016)**. Replaces:
- Old sm75 (b9341, mtp-turbo fork, 9 MB)
- Old upstream (ac4cdde, 213 MB)

One unified binary (`llama-server-sm75`) for all 7 llama-swap profiles.

## Flag Changes

| Old Flag | New Flag | Affected Profiles |
|----------|----------|-------------------|
| `--spec-type mtp` | `--spec-type draft-mtp` | qwen36-35b-mtp |
| `--cache-type-k turbo3/4` | **Not available** — use `q8_0` | All (was never used in config) |

**Preserved aliases:** `--model-draft` = `--spec-draft-model`, `--fa` = `--flash-attn`, `-nkvo` = `--no-kv-offload`

## Benchmarks (b9743 vs Old)

| Profile | Old t/s | New t/s | Old VRAM | New VRAM | Notes |
|---------|---------|---------|----------|----------|-------|
| qwen36-35b-mtp | 45.8 | **51.4** (+12%) | 7.3 GB | **6.8 GB** | Faster, less VRAM |
| gemma-12b | 117.8 | **96.2** (-18%) | 8.8 GB | 9.0 GB | Regression — see below |
| gemma4-v2 (16K) | 54.0 | 52.7 | 8.0 GB | 7.7 GB | Same speed |
| gemma4-v2 (131K) | — | ~40 (est) | — | 9.9 GB | New config |

## Gemma-12b Regression

**Known upstream bugs, not fixable from our side:**
- **#24514:** 25% CUDA perf drop between b9301→b9305 affecting MTP spec decoding. Prompt processing dropped from 584→0.4 t/s in some cases.
- **#24670:** `draft-mtp` not fully activating on **Turing (sm_75)** — our RTX 2080 Ti.

Both are open issues in ggml-org/llama.cpp as of 2026-06-20. Workaround: none from our side. Accept 96 t/s as current baseline.

## New Models Downloaded

| Model | File | Size | Arch |
|-------|------|------|------|
| Qwen3-Coder-30B-A3B | `Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf` | 18 GB | qwen35moe MoE |
| Qwen3-Coder-Next | `Qwen3-Coder-Next-Q4_K_M.gguf` | 45 GB | Hybrid DeltaNet+Attn+MoE |
| Qwen3-Coder-Next-REAP-48B | `Qwen3-Coder-Next-REAP-48B-A3B-Q4_K_XL.gguf` | 31 GB | REAP-pruned hybrid MoE |

**Note:** REAP-48B had no Q4_K_M variant — used Q4_K_XL instead (only option in lovedheart's repo).

## Build Notes

- **g++-13 required** — CUDA 12.6 nvcc can't handle g++ 14 headers
- **cmake:** `-DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=75`
- **Dynamic linking:** Binary is 18 KB + .so files. Must copy `.so*` to `/usr/local/lib/` and run `ldconfig`.
- **Build time:** ~7 min for CUDA compilation on 16C/32T Xeon
- **Disk space:** 221 GB NVMe root + 458 GB SATA SSD for models