README.md

---
license: apache-2.0
pretty_name: Ternary-Bonsai-2-27B + in-file MTP reproduction (Ada/SM89)
tags:
- ternary
- speculative-decoding
- mtp
- llama.cpp
- bonsai
---

# Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)

> **This repository holds the raw data. The method (build script, harness, launch units, write-up) lives on GitHub:**
> **<https://github.com/zhaoyilun/bonsai2-27b-mtp-repro>**
>
> Both are the same piece of work: the GitHub repo has the code and the how-to, this dataset has the
> measurements it produced. Cross-linked in both directions.

Raw measurements, scripts and notes for the two discussions:

- official model repo: https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/discussions/23
- drafter repo: https://huggingface.co/ProCreations/Ternary-Bonsai-2-27B-MTP/discussions/2

**Headline**: with the MTP block packaged *inside* the target GGUF (so the draft context is created against the
target and shares its vocabulary), MTP speculative decoding gives a **median 1.338x decode speedup**
(range 1.23x–1.70x, aggregate draft acceptance **68.1%**, 803/1180) on a single 16 GB Ada card — after
working around a hard guard in the official fork that refuses the MTP graph on Hadamard-folded weights.

![A/B results](https://huggingface.co/datasets/zhaokeqi/bonsai2-27b-mtp-repro/resolve/main/chart_en.png)

## Environment

| item | value |
|---|---|
| GPU | NVIDIA RTX 4080 SUPER, 16376 MiB (Ada, SM89) |
| Driver / toolkit | CUDA UMD 13.3, CUDA toolkit 13.3.73 |
| Host | WSL2 (kernel 5.15), 32-core CPU, 46 GB RAM |
| Model | `ProCreations/Ternary-Bonsai-2-27B-MTP` → `Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf`, 7.66 GB, sha256 `3cb3f0056d2e34ee44245a64396004a21f8492573d6ce1266ec4b7222c131dd4` (matches the publisher's `SHA256SUMS`) |
| Runtime | built from `runtime/prism-dflash2-source.tar.gz`, sha256 `8c0f589673b25574f27f013bb3278824384eb35eb984034a2445af3c437b9d05`, base `d8f26ee`, `-DCMAKE_CUDA_ARCHITECTURES=89` |

The published prebuilt runtime archive is CUDA 13.3 **SM120 (Blackwell) only**, so on Ada it has to be built from
the source they ship. Build on 32 cores: 484 targets, ~9 minutes.

One gotcha worth knowing: with `nvcc` not on `PATH`, CMake fails with `CMAKE_CUDA_COMPILER-NOTFOUND` unless the
compiler is passed explicitly:

```
cmake -S llama -B llama/build -G Ninja -DGGML_CUDA=ON \
      -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
      -DCMAKE_CUDA_ARCHITECTURES=89 -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=OFF
cmake --build llama/build -j 28 --target llama-server llama-bench
```

## The blocker on the official fork, and the fix

`prism-b10683-d8f26ee` refuses to create the MTP context for this file:

```
E llama_init_from_model: failed to initialize the context: Hadamard-latent table 'token_embd.weight' is read without the inverse transform
E common_speculative_init_result: failed to create MTP context
```

The guard is `llama_verify_hadamard_graph` (both strings are visible in `libllama.so`); no flag bypasses it. The MTP
graph reads `token_embd` with a plain `ggml_get_rows`, but the table is stored in the rotated (Hadamard-latent) basis.
The fix is to apply the inverse transform to that lookup — `runtime/bonsai-mtp-embedding.patch` in the drafter repo
does exactly that in `src/models/qwen35.cpp::graph_mtp`, and the source archive already contains it.

## Results

### A/B, 12 prompts x 5 categories, n_predict = 128

Both arms identical except `--spec-type`:

```
llama-server -m Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf \
  -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -np 1 -t 16 --temp 0 \
  --spec-type none                      # arm A
  --spec-type draft-mtp --spec-draft-n-max 2   # arm B
```

Sampling: `POST /completion`, `temperature=0, top_k=1, seed=7, cache_prompt=false`. Acceptance is read from the
response `timings` object (`draft_n`, `draft_n_accepted`) — `llama-cli` does not print acceptance, only the server does.

| # | prompt (abbrev.) | baseline t/s | draft-mtp t/s | speedup | acceptance | accepted/drafted |
|---|---|---:|---:|---:|---:|---:|
| R1 | reasoning prose | 67.2 | 88.0 | 1.31x | 57.6% | 68/118 |
| R2 | reasoning prose | 65.4 | 82.0 | 1.25x | 47.7% | 62/130 |
| R3 | reasoning prose | 66.6 | 84.6 | 1.27x | 57.6% | 68/118 |
| C1 | Python continuation | 67.1 | 99.8 | 1.49x | 77.0% | 77/100 |
| C2 | Python continuation | 67.4 | 114.8 | 1.70x | 94.3% | 83/88 |
| C3 | async Python | 67.7 | 89.2 | 1.32x | 62.5% | 70/112 |
| M1 | step-by-step math | 68.1 | 97.2 | 1.43x | 72.1% | 75/104 |
| M2 | probability recursion | 68.2 | 84.0 | 1.23x | 54.5% | 66/121 |
| F1 | JSON repetition | 68.3 | 111.5 | 1.63x | 90.0% | 81/90 |
| F2 | list continuation | 68.3 | 112.3 | 1.64x | 92.1% | 82/89 |
| Z1 | Chinese rewrite | 68.5 | 91.7 | 1.34x | 64.5% | 71/110 |
| **median** | | **67.5** | **90.4** | **1.338x** | **68.1%** | **803/1180** |

By category: code 1.32–1.70x (acceptance 62–94%), format/repetitive 1.63–1.64x (90–92%), math 1.23–1.43x,
reasoning prose 1.25–1.31x (48–58%). The pattern is the usual one: speculation pays where text is predictable.

> A 12th prompt (a second Chinese prompt) stopped after 1 token with `stop_type=eos` and empty content in **both**
> arms — an artifact of raw `/completion` without a chat template, not a model defect. It is recorded in
> `ab_base.json` / `ab_mtp.json` and excluded from the statistics above.

### Draft length

`--spec-draft-n-max 2` vs `3` on a 3-prompt quick set: 82.4 vs 83.0 t/s total (acceptance 69.9% vs 59.0%).
Roughly a wash; lower draft length favors repetitive text, higher favors code.

### Independent verification (third party)

Both the failure and the fix were reproduced on entirely different hardware — **RDNA3 / ROCm 6.4 / Windows 11**
(RX 7900 XTX, gfx1100) — reported on [PR #205](https://github.com/PrismML-Eng/llama.cpp/pull/205#issuecomment-5742011541):
identical `latent lookup 'mtp_tok_embd-64'` failure without the patch, then **decode 41.6 -> 75.9 t/s (1.82x)** with
it at **0.82 acceptance** and 2.64 tokens committed per round. On this machine (Ada/CUDA) the same fix gives
~67 -> ~90 t/s (1.34x) at 0.68 aggregate acceptance.

The multiples differ because the **baselines** differ: speculation pays more where the target decodes slower, since
the draft's roughly fixed per-round cost is then a smaller fraction of a round. The depth sweep below shows the same
effect along the context axis — decode falls 87 -> 35 t/s while acceptance rises 66% -> 84%.

### Long context (chat path, natural text, exact depths via /tokenize)

| depth (tokens) | prefill t/s (MTP) | decode t/s (MTP) | acceptance | decode t/s (no spec) | MTP speedup |
|---:|---:|---:|---:|---:|---:|
| 8,020 | 1855 | 86.9 | 65.8% | 63.0 | 1.38x |
| 31,939 | 1674 | 62.4 | 52.2% | 53.7 | 1.16x |
| 64,084 | 1332 | 55.4 | 67.9% | 43.4 | 1.28x |
| 127,870 | 974 | 46.1 | 82.3% | 33.3 | 1.38x |
| 191,099 | 722 | 35.0 | 84.1% | 26.5 | 1.32x |

Decode falls hard with depth (87 -> 35 t/s with MTP; 63 -> 26.5 without) — that is KV cache traffic, and it
dominates long-context interactivity far more than the weight packing does. MTP's edge does **not** decay with
depth (1.16-1.38x); acceptance actually rises (66% -> 84%) because a long natural-text context constrains the
continuation while the draft's own cost stays flat.

Gotcha for anyone measuring this: `timings.prompt_n` reports only the **newly evaluated** tokens. A
191k-token prompt can come back as `prompt_n = 63745` because llama-server reuses the cached prefix; the server
log shows `n_tokens = 191139, truncated = 0`, so nothing is being dropped. Interleaving a short "cache buster"
request does not evict the checkpoints — use `POST /tokenize` for the true length, or sum the incremental
`prompt_ms` across a cumulative ladder.

### Prefix reuse (same config, no speculation)

Re-sending an **identical** 12,485-token prompt evaluates only `prompt_n = 4` tokens in 0.28 s (0.4 s wall). So a
continuing conversation does not repay prefill; switching conversations does (single KV slot, `-np 1`).

### 16 GB VRAM: KV precision is the trap, context size is not

| config (`-c 262144`) | prefill (12.5k prompt) | decode | VRAM |
|---|---:|---:|---:|
| PQ2_0 + MTP, `-ctk q8_0 -ctv q8_0` | **101 → 35 t/s** (10,240 tokens took 293 s) | — | 15.7 GiB |
| PQ2_0 + MTP, `-ctk q4_0 -ctv q4_0` | **1726 t/s** | 90.4 t/s (median) | 15.7 GiB |
| PTQ1_0 (no MTP), `-ctk q4_0 -ctv q4_0` | 1073 t/s | ~76 t/s | 12.4 GiB |

The `common_fit_params: failed to fit params to free device memory` warning appears in the q4_0/262144 state too,
but performance is unaffected — the prefill collapse tracks **KV cache size**, not the warning. Also note PQ2_0
prefill here is 1.6x faster than PTQ1_0 at the same context, consistent with the model card.

## Files

| file | what |
|---|---|
| `mtp_ab.py` | the 12-prompt A/B harness (reads `draft_n`/`draft_n_accepted` from timings) |
| `ab_driver.sh` | launches one arm as a transient unit, runs the harness, tears down |
| `mtp_bench.py` | smaller 3-prompt quick bench |
| `run_mtp_ab2.sh` | single-arm driver |
| `ab_base.json`, `ab_mtp.json` | raw per-prompt results and timings for both arms |
| `mtp_n3.json` | `--spec-draft-n-max 3` arm |

## What was NOT measured

The official 14-benchmark suite; multi-run means per prompt (each prompt was run once); contexts beyond 262,144;
CPU-only runs; DFlash2 head comparison. These are single-machine measurements and should be read as such.

## Credits

Model: `prism-ml/Ternary-Bonsai-2-27B-gguf` (Apache 2.0). MTP head, patch and patched runtime source:
`ProCreations/Ternary-Bonsai-2-27B-MTP`. This bundle only records an independent reproduction on Ada/SM89 plus
the measurement protocol.