Add reproduction bundle: raw A/B results, harness, notes

README.md

@@ -0,0 +1,130 @@
+# Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)
+
+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.
+
+## 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.
+
+### 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.

ab_base.json

@@ -0,0 +1,101 @@
+{
+ "tag": "base",
+ "results": [
+  {
+   "name": "R1 reasoning",
+   "tps": 67.15504778583991,
+   "n": 128,
+   "wall": 2.0080041885375977,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "R2 reasoning",
+   "tps": 65.42088811688807,
+   "n": 128,
+   "wall": 2.139404773712158,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "R3 reasoning",
+   "tps": 66.57294035379063,
+   "n": 128,
+   "wall": 2.1015982627868652,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "C1 code",
+   "tps": 67.10459016532035,
+   "n": 128,
+   "wall": 2.090456962585449,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "C2 code",
+   "tps": 67.42175492988403,
+   "n": 128,
+   "wall": 2.083451509475708,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "C3 code",
+   "tps": 67.66148585688477,
+   "n": 128,
+   "wall": 2.074040412902832,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "M1 math",
+   "tps": 68.12873433977892,
+   "n": 128,
+   "wall": 2.061222791671753,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "M2 math",
+   "tps": 68.19275245279128,
+   "n": 128,
+   "wall": 2.051217794418335,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "F1 format",
+   "tps": 68.31218345102393,
+   "n": 128,
+   "wall": 2.0777087211608887,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "F2 format",
+   "tps": 68.30072108351834,
+   "n": 128,
+   "wall": 2.0450594425201416,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "Z1 chinese",
+   "tps": 68.48945693792805,
+   "n": 128,
+   "wall": 2.0459251403808594,
+   "dn": 0,
+   "da": 0
+  },
+  {
+   "name": "Z2 chinese",
+   "tps": 0.0,
+   "n": 1,
+   "wall": 0.21819567680358887,
+   "dn": 0,
+   "da": 0
+  }
+ ]
+}

ab_driver.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+set -u
+B=/home/zyl/bonsai2/mtp-runtime/source/llama/build/bin
+TAG=$1; shift
+systemctl --user stop llama-bonsai-mtp.service 2>/dev/null; systemctl --user stop llama-bonsai.service 2>/dev/null
+systemctl --user stop "mtpab-$TAG.service" 2>/dev/null; sleep 2
+systemd-run --user --collect --unit="mtpab-$TAG" /bin/bash -c "cd $B && LD_LIBRARY_PATH=$B HOME=/home/zyl \
+  ./llama-server -m /home/zyl/models/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 \
+  --host 127.0.0.1 --port 8080 -a bonsai-mtp $* > /home/zyl/bonsai2/logs/ab-$TAG.log 2>&1" >/dev/null 2>&1
+for i in $(seq 1 60); do sleep 3; curl -s --max-time 3 -o /dev/null http://127.0.0.1:8080/health 2>/dev/null && break; done
+python3 /home/zyl/bonsai2/mtp_ab.py 8080 "$TAG" 128
+systemctl --user stop "mtpab-$TAG.service"

ab_mtp.json

@@ -0,0 +1,101 @@
+{
+ "tag": "mtp",
+ "results": [
+  {
+   "name": "R1 reasoning",
+   "tps": 88.02853510813647,
+   "n": 128,
+   "wall": 1.6208324432373047,
+   "dn": 118,
+   "da": 68
+  },
+  {
+   "name": "R2 reasoning",
+   "tps": 82.02894394516592,
+   "n": 128,
+   "wall": 1.7505979537963867,
+   "dn": 130,
+   "da": 62
+  },
+  {
+   "name": "R3 reasoning",
+   "tps": 84.59436668127641,
+   "n": 128,
+   "wall": 1.700646162033081,
+   "dn": 118,
+   "da": 68
+  },
+  {
+   "name": "C1 code",
+   "tps": 99.84433717511283,
+   "n": 128,
+   "wall": 1.4969289302825928,
+   "dn": 100,
+   "da": 77
+  },
+  {
+   "name": "C2 code",
+   "tps": 114.81222325282013,
+   "n": 128,
+   "wall": 1.3252921104431152,
+   "dn": 88,
+   "da": 83
+  },
+  {
+   "name": "C3 code",
+   "tps": 89.18871278214709,
+   "n": 128,
+   "wall": 1.635740041732788,
+   "dn": 112,
+   "da": 70
+  },
+  {
+   "name": "M1 math",
+   "tps": 97.22071245175326,
+   "n": 128,
+   "wall": 1.5286304950714111,
+   "dn": 104,
+   "da": 75
+  },
+  {
+   "name": "M2 math",
+   "tps": 84.00415390619317,
+   "n": 128,
+   "wall": 1.715515375137329,
+   "dn": 121,
+   "da": 66
+  },
+  {
+   "name": "F1 format",
+   "tps": 111.53372929270928,
+   "n": 128,
+   "wall": 1.3411991596221924,
+   "dn": 90,
+   "da": 81
+  },
+  {
+   "name": "F2 format",
+   "tps": 112.33182289782492,
+   "n": 128,
+   "wall": 1.3317580223083496,
+   "dn": 89,
+   "da": 82
+  },
+  {
+   "name": "Z1 chinese",
+   "tps": 91.65472869839459,
+   "n": 128,
+   "wall": 1.5930490493774414,
+   "dn": 110,
+   "da": 71
+  },
+  {
+   "name": "Z2 chinese",
+   "tps": 0.0,
+   "n": 1,
+   "wall": 0.21099615097045898,
+   "dn": 0,
+   "da": 0
+  }
+ ]
+}

mtp_ab.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+"""多 prompt A/B：12 条 × 5 类，temp0/topk1/seed7/n=128，读 timings 的 draft 计数。"""
+import json, sys, time, urllib.request, statistics
+
+PORT = int(sys.argv[1]); TAG = sys.argv[2]; NP = int(sys.argv[3]) if len(sys.argv) > 3 else 128
+URL = f"http://127.0.0.1:{PORT}/completion"
+H = {"Content-Type":"application/json","Authorization":"Bearer sk-lan-97f1c41bb1642ab23f810c65"}
+
+P = {
+ "R1 reasoning": "Quantum computing exploits superposition and entanglement to perform computations that are intractable for classical machines. The principal obstacle in practice is decoherence, which",
+ "R2 reasoning": "The reason ternary weights appeal to hardware designers is not only the storage density. A ternary multiply accumulator can be implemented as",
+ "R3 reasoning": "When a large language model is quantized below two bits per weight, the failure mode that appears first is usually not factual recall but rather",
+ "C1 code":      "def merge_intervals(intervals):\n    \"\"\"Merge overlapping intervals.\"\"\"\n    if not intervals:\n        return []\n    intervals = sorted(intervals)\n    merged = [intervals[0]]\n    for start, end in intervals[1:]:\n",
+ "C2 code":      "class RingBuffer:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.buf = [None] * capacity\n        self.head = 0\n        self.size = 0\n\n    def push(self, item):\n",
+ "C3 code":      "async def fetch_all(session, urls, concurrency=8):\n    sem = asyncio.Semaphore(concurrency)\n    async def one(u):\n        async with sem:\n            async with session.get(u) as r:\n",
+ "M1 math":      "A train leaves at 09:15 and travels 240 km at an average speed of 96 km/h. It then waits 20 minutes and returns at 80 km/h. Step by step, the arrival time back is",
+ "M2 math":      "We flip a fair coin until we see two heads in a row. Let E be the expected number of flips. Setting up the recursion gives",
+ "F1 format":    "Output the JSON object for a user record with fields name, age, city, tags (array), then output it again, then again, then again:",
+ "F2 format":    "Item 1: alpha\nItem 2: beta\nItem 3: gamma\nItem 4: delta\nItem 5:",
+ "Z1 chinese":   "把下面这段技术说明改写成更通俗的中文，保持信息不丢：折叠旋转基把正交变换折进权重，运行时只需对激活做一次同构变换即可。改写结果：",
+ "Z2 chinese":   "请用中文解释为什么三值权重（-1、0、+1）在存储上比 int4 更省，并给一个具体数字例子。回答：",
+}
+
+def run(name, prompt):
+    body = json.dumps({"prompt": prompt, "n_predict": NP, "temperature": 0.0,
+                       "top_k": 1, "seed": 7, "cache_prompt": False}).encode()
+    t0 = time.time()
+    try:
+        r = json.load(urllib.request.urlopen(urllib.request.Request(URL, data=body, headers=H), timeout=1800))
+    except Exception as e:
+        print(f"  {name:14} 失败: {e}"); return None
+    wall = time.time() - t0; t = r.get("timings", {})
+    dn, da = t.get("draft_n") or 0, t.get("draft_n_accepted") or 0
+    tps = t.get("predicted_per_second", 0.0)
+    print(f"  {name:14} {tps:6.1f} t/s | {t.get('predicted_n'):4} tok | {wall:5.1f}s wall | "
+          f"acc {da/dn*100:5.1f}% ({da}/{dn})" if dn else f"  {name:14} {tps:6.1f} t/s | {t.get('predicted_n'):4} tok | no draft")
+    return {"name": name, "tps": tps, "n": t.get("predicted_n"), "wall": wall, "dn": dn, "da": da}
+
+print(f"[{TAG}] n_predict={NP} prompts={len(P)}")
+res = [x for x in (run(k, v) for k, v in P.items()) if x]
+tps = [r["tps"] for r in res]
+tot_n = sum(r["n"] for r in res); tot_w = sum(r["wall"] for r in res)
+dn = sum(r["dn"] for r in res); da = sum(r["da"] for r in res)
+print(f"  ---- {TAG}: 均值 {statistics.mean(tps):.1f} t/s | 中位 {statistics.median(tps):.1f} | "
+      f"端到端 {tot_n/tot_w:.1f} t/s | 接受率 {da/dn*100:.1f}% ({da}/{dn})" if dn else
+      f"  ---- {TAG}: 均值 {statistics.mean(tps):.1f} t/s | 中位 {statistics.median(tps):.1f} | 端到端 {tot_n/tot_w:.1f} t/s")
+json.dump({"tag": TAG, "results": res}, open(f"/home/zyl/bonsai2/ab_{TAG}.json","w"), indent=1)

mtp_bench.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+"""A/B 测 MTP 接受率与速度：同一 prompt 集、temp0/topk1/seed7，读 timings 的 draft_n/draft_n_accepted。"""
+import json, sys, time, urllib.request
+
+PORT = int(sys.argv[1]) if len(sys.argv) > 1 else 8080
+TAG  = sys.argv[2] if len(sys.argv) > 2 else "run"
+NP   = int(sys.argv[3]) if len(sys.argv) > 3 else 96
+URL  = f"http://127.0.0.1:{PORT}/completion"
+H    = {"Content-Type": "application/json",
+        "Authorization": "Bearer sk-lan-97f1c41bb1642ab23f810c65"}
+
+PROMPTS = {
+  "V(reasoning)": ("Quantum computing exploits superposition and entanglement to perform "
+                   "computations that are intractable for classical machines. The principal "
+                   "obstacle in practice is decoherence, which"),
+  "P(repetitive)": "The capital of France is",
+  "C(code)": ("def merge_intervals(intervals):\n    \"\"\"Merge overlapping intervals.\"\"\"\n"
+              "    if not intervals:\n        return []\n    intervals = sorted(intervals)\n    merged = [intervals[0]]\n    for start, end in intervals[1:]:\n"),
+}
+
+def run(name, prompt):
+    body = json.dumps({"prompt": prompt, "n_predict": NP, "temperature": 0.0,
+                       "top_k": 1, "seed": 7, "cache_prompt": False}).encode()
+    t0 = time.time()
+    try:
+        r = json.load(urllib.request.urlopen(
+            urllib.request.Request(URL, data=body, headers=H), timeout=1800))
+    except Exception as e:
+        print(f"  {name:14} 请求失败: {e}"); return None
+    wall = time.time() - t0
+    t = r.get("timings", {})
+    dn, da = t.get("draft_n"), t.get("draft_n_accepted")
+    acc = (da / dn) if (dn and da is not None) else None
+    print(f"  {name:14} gen {t.get('predicted_per_second',0):6.1f} t/s | {t.get('predicted_n'):4} tok | "
+          f"wall {wall:5.1f}s | draft {dn if dn is not None else '-'}"
+          f"{f'/{da}' if da is not None else ''}"
+          f"{f' = {acc*100:.1f}%' if acc is not None else ''}")
+    return {"name": name, "tps": t.get("predicted_per_second"), "n": t.get("predicted_n"),
+            "wall": wall, "draft_n": dn, "draft_accepted": da, "acc": acc,
+            "prompt_ms": t.get("prompt_ms"), "timings": t}
+
+print(f"[{TAG}] n_predict={NP}")
+res = [x for x in (run(k, v) for k, v in PROMPTS.items()) if x]
+json.dump({"tag": TAG, "results": res}, open(f"/home/zyl/bonsai2/mtp_{TAG}.json", "w"), indent=1)
+if res:
+    tot_n = sum(r["n"] for r in res); tot_w = sum(r["wall"] for r in res)
+    dn = sum(r["draft_n"] or 0 for r in res); da = sum(r["draft_accepted"] or 0 for r in res)
+    print(f"  合计: {tot_n} tok / {tot_w:.1f}s = {tot_n/tot_w:.1f} t/s | "
+          f"draft {dn}→{da}" + (f" ({da/dn*100:.1f}%)" if dn else ""))

mtp_n3.json

@@ -0,0 +1,74 @@
+{
+ "tag": "n3",
+ "results": [
+  {
+   "name": "V(reasoning)",
+   "tps": 88.1621441033446,
+   "n": 96,
+   "wall": 1.2502977848052979,
+   "draft_n": 117,
+   "draft_accepted": 56,
+   "acc": 0.47863247863247865,
+   "prompt_ms": 169.653,
+   "timings": {
+    "cache_n": 0,
+    "prompt_n": 32,
+    "prompt_ms": 169.653,
+    "prompt_per_token_ms": 5.30165625,
+    "prompt_per_second": 188.62030143881924,
+    "predicted_n": 96,
+    "predicted_ms": 1077.56,
+    "predicted_per_token_ms": 11.342736842105262,
+    "predicted_per_second": 88.1621441033446,
+    "draft_n": 117,
+    "draft_n_accepted": 56
+   }
+  },
+  {
+   "name": "P(repetitive)",
+   "tps": 103.1431518375767,
+   "n": 96,
+   "wall": 1.1303412914276123,
+   "draft_n": 99,
+   "draft_accepted": 61,
+   "acc": 0.6161616161616161,
+   "prompt_ms": 100.431,
+   "timings": {
+    "cache_n": 0,
+    "prompt_n": 5,
+    "prompt_ms": 100.431,
+    "prompt_per_token_ms": 20.086199999999998,
+    "prompt_per_second": 49.785424819029984,
+    "predicted_n": 96,
+    "predicted_ms": 921.05,
+    "predicted_per_token_ms": 9.695263157894736,
+    "predicted_per_second": 103.1431518375767,
+    "draft_n": 99,
+    "draft_n_accepted": 61
+   }
+  },
+  {
+   "name": "C(code)",
+   "tps": 111.25203034955388,
+   "n": 96,
+   "wall": 1.087364912033081,
+   "draft_n": 91,
+   "draft_accepted": 64,
+   "acc": 0.7032967032967034,
+   "prompt_ms": 131.773,
+   "timings": {
+    "cache_n": 0,
+    "prompt_n": 54,
+    "prompt_ms": 131.773,
+    "prompt_per_token_ms": 2.4402407407407405,
+    "prompt_per_second": 409.7956333998619,
+    "predicted_n": 96,
+    "predicted_ms": 853.917,
+    "predicted_per_token_ms": 8.9886,
+    "predicted_per_second": 111.25203034955388,
+    "draft_n": 91,
+    "draft_n_accepted": 64
+   }
+  }
+ ]
+}

run_mtp_ab2.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+# 用法: run_mtp_ab.sh <tag> <额外 llama-server 参数...>
+set -u
+TAG=$1; shift
+MODEL=/home/zyl/models/Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf
+UNIT=bonsai-mtp-$TAG
+LOG=/home/zyl/bonsai2/logs/mtp-$TAG.log
+systemctl --user stop llama-bonsai.service 2>/dev/null
+systemctl --user stop "$UNIT.service" 2>/dev/null
+sleep 2
+systemd-run --user --collect --unit="$UNIT" /bin/bash -c \
+  "cd /home/zyl/bonsai2/mtp-runtime/source/llama/build/bin && LD_LIBRARY_PATH=/home/zyl/bonsai2/mtp-runtime/source/llama/build/bin HOME=/home/zyl \
+   ./llama-server -m $MODEL -ngl 99 -fa on -c 131072 -ctk q4_0 -ctv q4_0 -np 1 \
+   -t 16 --temp 0 --host 127.0.0.1 --port 8080 -a bonsai-mtp $* > $LOG 2>&1" 2>&1 | tail -1
+for i in $(seq 1 60); do
+  sleep 3
+  curl -s --max-time 3 -o /dev/null http://127.0.0.1:8080/health 2>/dev/null && break
+  systemctl --user is-active -q "$UNIT" || { echo "!! 单元退出，日志尾部："; tail -5 "$LOG"; exit 1; }
+done
+echo "== 服务就绪 (${TAG}) =="
+grep -aiE "nextn|mtp|spec|draft" "$LOG" | head -5
+python3 /home/zyl/bonsai2/mtp_bench.py 8080 "$TAG" 96
+echo "== 服务端日志里的接受率行 =="
+grep -a "draft acceptance" "$LOG" | tail -3
+systemctl --user stop "$UNIT.service"