Hugging Face's logo

Datasets:
zhaokeqi
/
bonsai2-27b-mtp-repro

Tags:
ternary
speculative-decoding
mtp
llama.cpp
bonsai
License:
Dataset card Files
xet
 Community
zhaokeqi commited on
Commit
ee9ab8d
·
verified ·
1 Parent(s): bbce209

Link the GitHub repo (method/code) and add long-context results

Browse files
Files changed (1) hide show
  1. README.md +170 -143
README.md CHANGED
@@ -1,143 +1,170 @@
1
- ---
2
- license: apache-2.0
3
- pretty_name: Ternary-Bonsai-2-27B + in-file MTP reproduction (Ada/SM89)
4
- tags:
5
- - ternary
6
- - speculative-decoding
7
- - mtp
8
- - llama.cpp
9
- - bonsai
10
- ---
11
-
12
- # Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)
13
-
14
- Raw measurements, scripts and notes for the two discussions:
15
-
16
- - official model repo: https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/discussions/23
17
- - drafter repo: https://huggingface.co/ProCreations/Ternary-Bonsai-2-27B-MTP/discussions/2
18
-
19
- **Headline**: with the MTP block packaged *inside* the target GGUF (so the draft context is created against the
20
- target and shares its vocabulary), MTP speculative decoding gives a **median 1.338x decode speedup**
21
- (range 1.23x–1.70x, aggregate draft acceptance **68.1%**, 803/1180) on a single 16 GB Ada card — after
22
- working around a hard guard in the official fork that refuses the MTP graph on Hadamard-folded weights.
23
-
24
- ![A/B results](https://huggingface.co/datasets/zhaokeqi/bonsai2-27b-mtp-repro/resolve/main/chart_en.png)
25
-
26
- ## Environment
27
-
28
- | item | value |
29
- |---|---|
30
- | GPU | NVIDIA RTX 4080 SUPER, 16376 MiB (Ada, SM89) |
31
- | Driver / toolkit | CUDA UMD 13.3, CUDA toolkit 13.3.73 |
32
- | Host | WSL2 (kernel 5.15), 32-core CPU, 46 GB RAM |
33
- | 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`) |
34
- | Runtime | built from `runtime/prism-dflash2-source.tar.gz`, sha256 `8c0f589673b25574f27f013bb3278824384eb35eb984034a2445af3c437b9d05`, base `d8f26ee`, `-DCMAKE_CUDA_ARCHITECTURES=89` |
35
-
36
- The published prebuilt runtime archive is CUDA 13.3 **SM120 (Blackwell) only**, so on Ada it has to be built from
37
- the source they ship. Build on 32 cores: 484 targets, ~9 minutes.
38
-
39
- One gotcha worth knowing: with `nvcc` not on `PATH`, CMake fails with `CMAKE_CUDA_COMPILER-NOTFOUND` unless the
40
- compiler is passed explicitly:
41
-
42
- ```
43
- cmake -S llama -B llama/build -G Ninja -DGGML_CUDA=ON \
44
- -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
45
- -DCMAKE_CUDA_ARCHITECTURES=89 -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=OFF
46
- cmake --build llama/build -j 28 --target llama-server llama-bench
47
- ```
48
-
49
- ## The blocker on the official fork, and the fix
50
-
51
- `prism-b10683-d8f26ee` refuses to create the MTP context for this file:
52
-
53
- ```
54
- E llama_init_from_model: failed to initialize the context: Hadamard-latent table 'token_embd.weight' is read without the inverse transform
55
- E common_speculative_init_result: failed to create MTP context
56
- ```
57
-
58
- The guard is `llama_verify_hadamard_graph` (both strings are visible in `libllama.so`); no flag bypasses it. The MTP
59
- graph reads `token_embd` with a plain `ggml_get_rows`, but the table is stored in the rotated (Hadamard-latent) basis.
60
- The fix is to apply the inverse transform to that lookup — `runtime/bonsai-mtp-embedding.patch` in the drafter repo
61
- does exactly that in `src/models/qwen35.cpp::graph_mtp`, and the source archive already contains it.
62
-
63
- ## Results
64
-
65
- ### A/B, 12 prompts x 5 categories, n_predict = 128
66
-
67
- Both arms identical except `--spec-type`:
68
-
69
- ```
70
- llama-server -m Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf \
71
- -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -np 1 -t 16 --temp 0 \
72
- --spec-type none # arm A
73
- --spec-type draft-mtp --spec-draft-n-max 2 # arm B
74
- ```
75
-
76
- Sampling: `POST /completion`, `temperature=0, top_k=1, seed=7, cache_prompt=false`. Acceptance is read from the
77
- response `timings` object (`draft_n`, `draft_n_accepted`) — `llama-cli` does not print acceptance, only the server does.
78
-
79
- | # | prompt (abbrev.) | baseline t/s | draft-mtp t/s | speedup | acceptance | accepted/drafted |
80
- |---|---|---:|---:|---:|---:|---:|
81
- | R1 | reasoning prose | 67.2 | 88.0 | 1.31x | 57.6% | 68/118 |
82
- | R2 | reasoning prose | 65.4 | 82.0 | 1.25x | 47.7% | 62/130 |
83
- | R3 | reasoning prose | 66.6 | 84.6 | 1.27x | 57.6% | 68/118 |
84
- | C1 | Python continuation | 67.1 | 99.8 | 1.49x | 77.0% | 77/100 |
85
- | C2 | Python continuation | 67.4 | 114.8 | 1.70x | 94.3% | 83/88 |
86
- | C3 | async Python | 67.7 | 89.2 | 1.32x | 62.5% | 70/112 |
87
- | M1 | step-by-step math | 68.1 | 97.2 | 1.43x | 72.1% | 75/104 |
88
- | M2 | probability recursion | 68.2 | 84.0 | 1.23x | 54.5% | 66/121 |
89
- | F1 | JSON repetition | 68.3 | 111.5 | 1.63x | 90.0% | 81/90 |
90
- | F2 | list continuation | 68.3 | 112.3 | 1.64x | 92.1% | 82/89 |
91
- | Z1 | Chinese rewrite | 68.5 | 91.7 | 1.34x | 64.5% | 71/110 |
92
- | **median** | | **67.5** | **90.4** | **1.338x** | **68.1%** | **803/1180** |
93
-
94
- By category: code 1.32–1.70x (acceptance 62–94%), format/repetitive 1.63–1.64x (90–92%), math 1.23–1.43x,
95
- reasoning prose 1.25–1.31x (48–58%). The pattern is the usual one: speculation pays where text is predictable.
96
-
97
- > A 12th prompt (a second Chinese prompt) stopped after 1 token with `stop_type=eos` and empty content in **both**
98
- > arms — an artifact of raw `/completion` without a chat template, not a model defect. It is recorded in
99
- > `ab_base.json` / `ab_mtp.json` and excluded from the statistics above.
100
-
101
- ### Draft length
102
-
103
- `--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%).
104
- Roughly a wash; lower draft length favors repetitive text, higher favors code.
105
-
106
- ### Prefix reuse (same config, no speculation)
107
-
108
- Re-sending an **identical** 12,485-token prompt evaluates only `prompt_n = 4` tokens in 0.28 s (0.4 s wall). So a
109
- continuing conversation does not repay prefill; switching conversations does (single KV slot, `-np 1`).
110
-
111
- ### 16 GB VRAM: KV precision is the trap, context size is not
112
-
113
- | config (`-c 262144`) | prefill (12.5k prompt) | decode | VRAM |
114
- |---|---:|---:|---:|
115
- | PQ2_0 + MTP, `-ctk q8_0 -ctv q8_0` | **101 → 35 t/s** (10,240 tokens took 293 s) | — | 15.7 GiB |
116
- | PQ2_0 + MTP, `-ctk q4_0 -ctv q4_0` | **1726 t/s** | 90.4 t/s (median) | 15.7 GiB |
117
- | PTQ1_0 (no MTP), `-ctk q4_0 -ctv q4_0` | 1073 t/s | ~76 t/s | 12.4 GiB |
118
-
119
- The `common_fit_params: failed to fit params to free device memory` warning appears in the q4_0/262144 state too,
120
- but performance is unaffected — the prefill collapse tracks **KV cache size**, not the warning. Also note PQ2_0
121
- prefill here is 1.6x faster than PTQ1_0 at the same context, consistent with the model card.
122
-
123
- ## Files
124
-
125
- | file | what |
126
- |---|---|
127
- | `mtp_ab.py` | the 12-prompt A/B harness (reads `draft_n`/`draft_n_accepted` from timings) |
128
- | `ab_driver.sh` | launches one arm as a transient unit, runs the harness, tears down |
129
- | `mtp_bench.py` | smaller 3-prompt quick bench |
130
- | `run_mtp_ab2.sh` | single-arm driver |
131
- | `ab_base.json`, `ab_mtp.json` | raw per-prompt results and timings for both arms |
132
- | `mtp_n3.json` | `--spec-draft-n-max 3` arm |
133
-
134
- ## What was NOT measured
135
-
136
- The official 14-benchmark suite; multi-run means per prompt (each prompt was run once); contexts beyond 262,144;
137
- CPU-only runs; DFlash2 head comparison. These are single-machine measurements and should be read as such.
138
-
139
- ## Credits
140
-
141
- Model: `prism-ml/Ternary-Bonsai-2-27B-gguf` (Apache 2.0). MTP head, patch and patched runtime source:
142
- `ProCreations/Ternary-Bonsai-2-27B-MTP`. This bundle only records an independent reproduction on Ada/SM89 plus
143
- the measurement protocol.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ pretty_name: Ternary-Bonsai-2-27B + in-file MTP reproduction (Ada/SM89)
4
+ tags:
5
+ - ternary
6
+ - speculative-decoding
7
+ - mtp
8
+ - llama.cpp
9
+ - bonsai
10
+ ---
11
+
12
+ # Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)
13
+
14
+ > **This repository holds the raw data. The method (build script, harness, launch units, write-up) lives on GitHub:**
15
+ > **<https://github.com/zhaoyilun/bonsai2-27b-mtp-repro>**
16
+ >
17
+ > Both are the same piece of work: the GitHub repo has the code and the how-to, this dataset has the
18
+ > measurements it produced. Cross-linked in both directions.
19
+
20
+ Raw measurements, scripts and notes for the two discussions:
21
+
22
+ - official model repo: https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/discussions/23
23
+ - drafter repo: https://huggingface.co/ProCreations/Ternary-Bonsai-2-27B-MTP/discussions/2
24
+
25
+ **Headline**: with the MTP block packaged *inside* the target GGUF (so the draft context is created against the
26
+ target and shares its vocabulary), MTP speculative decoding gives a **median 1.338x decode speedup**
27
+ (range 1.23x–1.70x, aggregate draft acceptance **68.1%**, 803/1180) on a single 16 GB Ada card — after
28
+ working around a hard guard in the official fork that refuses the MTP graph on Hadamard-folded weights.
29
+
30
+ ![A/B results](https://huggingface.co/datasets/zhaokeqi/bonsai2-27b-mtp-repro/resolve/main/chart_en.png)
31
+
32
+ ## Environment
33
+
34
+ | item | value |
35
+ |---|---|
36
+ | GPU | NVIDIA RTX 4080 SUPER, 16376 MiB (Ada, SM89) |
37
+ | Driver / toolkit | CUDA UMD 13.3, CUDA toolkit 13.3.73 |
38
+ | Host | WSL2 (kernel 5.15), 32-core CPU, 46 GB RAM |
39
+ | 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`) |
40
+ | Runtime | built from `runtime/prism-dflash2-source.tar.gz`, sha256 `8c0f589673b25574f27f013bb3278824384eb35eb984034a2445af3c437b9d05`, base `d8f26ee`, `-DCMAKE_CUDA_ARCHITECTURES=89` |
41
+
42
+ The published prebuilt runtime archive is CUDA 13.3 **SM120 (Blackwell) only**, so on Ada it has to be built from
43
+ the source they ship. Build on 32 cores: 484 targets, ~9 minutes.
44
+
45
+ One gotcha worth knowing: with `nvcc` not on `PATH`, CMake fails with `CMAKE_CUDA_COMPILER-NOTFOUND` unless the
46
+ compiler is passed explicitly:
47
+
48
+ ```
49
+ cmake -S llama -B llama/build -G Ninja -DGGML_CUDA=ON \
50
+ -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
51
+ -DCMAKE_CUDA_ARCHITECTURES=89 -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=OFF
52
+ cmake --build llama/build -j 28 --target llama-server llama-bench
53
+ ```
54
+
55
+ ## The blocker on the official fork, and the fix
56
+
57
+ `prism-b10683-d8f26ee` refuses to create the MTP context for this file:
58
+
59
+ ```
60
+ E llama_init_from_model: failed to initialize the context: Hadamard-latent table 'token_embd.weight' is read without the inverse transform
61
+ E common_speculative_init_result: failed to create MTP context
62
+ ```
63
+
64
+ The guard is `llama_verify_hadamard_graph` (both strings are visible in `libllama.so`); no flag bypasses it. The MTP
65
+ graph reads `token_embd` with a plain `ggml_get_rows`, but the table is stored in the rotated (Hadamard-latent) basis.
66
+ The fix is to apply the inverse transform to that lookup — `runtime/bonsai-mtp-embedding.patch` in the drafter repo
67
+ does exactly that in `src/models/qwen35.cpp::graph_mtp`, and the source archive already contains it.
68
+
69
+ ## Results
70
+
71
+ ### A/B, 12 prompts x 5 categories, n_predict = 128
72
+
73
+ Both arms identical except `--spec-type`:
74
+
75
+ ```
76
+ llama-server -m Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf \
77
+ -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -np 1 -t 16 --temp 0 \
78
+ --spec-type none # arm A
79
+ --spec-type draft-mtp --spec-draft-n-max 2 # arm B
80
+ ```
81
+
82
+ Sampling: `POST /completion`, `temperature=0, top_k=1, seed=7, cache_prompt=false`. Acceptance is read from the
83
+ response `timings` object (`draft_n`, `draft_n_accepted`) — `llama-cli` does not print acceptance, only the server does.
84
+
85
+ | # | prompt (abbrev.) | baseline t/s | draft-mtp t/s | speedup | acceptance | accepted/drafted |
86
+ |---|---|---:|---:|---:|---:|---:|
87
+ | R1 | reasoning prose | 67.2 | 88.0 | 1.31x | 57.6% | 68/118 |
88
+ | R2 | reasoning prose | 65.4 | 82.0 | 1.25x | 47.7% | 62/130 |
89
+ | R3 | reasoning prose | 66.6 | 84.6 | 1.27x | 57.6% | 68/118 |
90
+ | C1 | Python continuation | 67.1 | 99.8 | 1.49x | 77.0% | 77/100 |
91
+ | C2 | Python continuation | 67.4 | 114.8 | 1.70x | 94.3% | 83/88 |
92
+ | C3 | async Python | 67.7 | 89.2 | 1.32x | 62.5% | 70/112 |
93
+ | M1 | step-by-step math | 68.1 | 97.2 | 1.43x | 72.1% | 75/104 |
94
+ | M2 | probability recursion | 68.2 | 84.0 | 1.23x | 54.5% | 66/121 |
95
+ | F1 | JSON repetition | 68.3 | 111.5 | 1.63x | 90.0% | 81/90 |
96
+ | F2 | list continuation | 68.3 | 112.3 | 1.64x | 92.1% | 82/89 |
97
+ | Z1 | Chinese rewrite | 68.5 | 91.7 | 1.34x | 64.5% | 71/110 |
98
+ | **median** | | **67.5** | **90.4** | **1.338x** | **68.1%** | **803/1180** |
99
+
100
+ By category: code 1.32–1.70x (acceptance 62–94%), format/repetitive 1.63–1.64x (90–92%), math 1.23–1.43x,
101
+ reasoning prose 1.25–1.31x (48–58%). The pattern is the usual one: speculation pays where text is predictable.
102
+
103
+ > A 12th prompt (a second Chinese prompt) stopped after 1 token with `stop_type=eos` and empty content in **both**
104
+ > arms — an artifact of raw `/completion` without a chat template, not a model defect. It is recorded in
105
+ > `ab_base.json` / `ab_mtp.json` and excluded from the statistics above.
106
+
107
+ ### Draft length
108
+
109
+ `--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%).
110
+ Roughly a wash; lower draft length favors repetitive text, higher favors code.
111
+
112
+ ### Long context (chat path, natural text, exact depths via /tokenize)
113
+
114
+ | depth (tokens) | prefill t/s (MTP) | decode t/s (MTP) | acceptance | decode t/s (no spec) | MTP speedup |
115
+ |---:|---:|---:|---:|---:|---:|
116
+ | 8,020 | 1855 | 86.9 | 65.8% | 63.0 | 1.38x |
117
+ | 31,939 | 1674 | 62.4 | 52.2% | 53.7 | 1.16x |
118
+ | 64,084 | 1332 | 55.4 | 67.9% | 43.4 | 1.28x |
119
+ | 127,870 | 974 | 46.1 | 82.3% | 33.3 | 1.38x |
120
+ | 191,099 | 722 | 35.0 | 84.1% | 26.5 | 1.32x |
121
+
122
+ Decode falls hard with depth (87 -> 35 t/s with MTP; 63 -> 26.5 without) — that is KV cache traffic, and it
123
+ dominates long-context interactivity far more than the weight packing does. MTP's edge does **not** decay with
124
+ depth (1.16-1.38x); acceptance actually rises (66% -> 84%) because a long natural-text context constrains the
125
+ continuation while the draft's own cost stays flat.
126
+
127
+ Gotcha for anyone measuring this: `timings.prompt_n` reports only the **newly evaluated** tokens. A
128
+ 191k-token prompt can come back as `prompt_n = 63745` because llama-server reuses the cached prefix; the server
129
+ log shows `n_tokens = 191139, truncated = 0`, so nothing is being dropped. Interleaving a short "cache buster"
130
+ request does not evict the checkpoints — use `POST /tokenize` for the true length, or sum the incremental
131
+ `prompt_ms` across a cumulative ladder.
132
+
133
+ ### Prefix reuse (same config, no speculation)
134
+
135
+ Re-sending an **identical** 12,485-token prompt evaluates only `prompt_n = 4` tokens in 0.28 s (0.4 s wall). So a
136
+ continuing conversation does not repay prefill; switching conversations does (single KV slot, `-np 1`).
137
+
138
+ ### 16 GB VRAM: KV precision is the trap, context size is not
139
+
140
+ | config (`-c 262144`) | prefill (12.5k prompt) | decode | VRAM |
141
+ |---|---:|---:|---:|
142
+ | PQ2_0 + MTP, `-ctk q8_0 -ctv q8_0` | **101 → 35 t/s** (10,240 tokens took 293 s) | — | 15.7 GiB |
143
+ | PQ2_0 + MTP, `-ctk q4_0 -ctv q4_0` | **1726 t/s** | 90.4 t/s (median) | 15.7 GiB |
144
+ | PTQ1_0 (no MTP), `-ctk q4_0 -ctv q4_0` | 1073 t/s | ~76 t/s | 12.4 GiB |
145
+
146
+ The `common_fit_params: failed to fit params to free device memory` warning appears in the q4_0/262144 state too,
147
+ but performance is unaffected — the prefill collapse tracks **KV cache size**, not the warning. Also note PQ2_0
148
+ prefill here is 1.6x faster than PTQ1_0 at the same context, consistent with the model card.
149
+
150
+ ## Files
151
+
152
+ | file | what |
153
+ |---|---|
154
+ | `mtp_ab.py` | the 12-prompt A/B harness (reads `draft_n`/`draft_n_accepted` from timings) |
155
+ | `ab_driver.sh` | launches one arm as a transient unit, runs the harness, tears down |
156
+ | `mtp_bench.py` | smaller 3-prompt quick bench |
157
+ | `run_mtp_ab2.sh` | single-arm driver |
158
+ | `ab_base.json`, `ab_mtp.json` | raw per-prompt results and timings for both arms |
159
+ | `mtp_n3.json` | `--spec-draft-n-max 3` arm |
160
+
161
+ ## What was NOT measured
162
+
163
+ The official 14-benchmark suite; multi-run means per prompt (each prompt was run once); contexts beyond 262,144;
164
+ CPU-only runs; DFlash2 head comparison. These are single-machine measurements and should be read as such.
165
+
166
+ ## Credits
167
+
168
+ Model: `prism-ml/Ternary-Bonsai-2-27B-gguf` (Apache 2.0). MTP head, patch and patched runtime source:
169
+ `ProCreations/Ternary-Bonsai-2-27B-MTP`. This bundle only records an independent reproduction on Ada/SM89 plus
170
+ the measurement protocol.