抡锤者

感谢，抄了作业，重新编译一下从原来～30TPS 提升到～40TPS,后面对coding微调了一下基本上确定大概eGPU +7900xtx能编程能测试了，等装上x99-cd3会来更新一下

# 7900 XTX (TB3 eGPU) + Qwen3.6-27B llama.cpp MTP — Bench Summary

Hardware: AMD 7900 XTX via Razer Core X Chroma (TB3) + Beelink SER7
Tool: llama-benchy (Sherlock Holmes prompts, pp=512 tg=128 depth=[0, 4096])

| # | Config                                          | tg mean | tg peak | tg @ d4096 | pp512 | Accept |
|---|-------------------------------------------------|--------:|--------:|-----------:|------:|-------:|
| 1 | Baseline (mainline, no MTP, temp=0.2)           |   30.26 |    31.5 |      29.79 |   459 |    n/a |
| 2 | + MTP enabled (old PR build 9117)               |   35.54 |    41.0 |      29.45 |   310 |    97% |
| 3 | + Rebuilt PR to latest (9173, GDN rollback fix) |   37.25 |    45.5 |      34.70 |   353 |    57% |
| 4 | + GPU power_dpm forced to `high`                |   45.00 |    54.8 |      37.94 |   351 |    57% |
| 5 | + Qwen "precise coding" sampling (current)      |   37.32 |    46.8 |      31.75 |   368 |    54% |

Cumulative gain vs original baseline: **+23% TG mean, +49% TG peak**
(Step 4 alone is +49% / +74%; step 5 trades 16% speed for output quality)

## Variant comparisons (PR 9173 + perf=high)

| Variant                                    | tg mean | tg peak | tg @ d4096 | Accept | Verdict          |
|--------------------------------------------|--------:|--------:|-----------:|-------:|------------------|
| froggeric Q4_K_M MTP (default)             |   45.00 |    54.8 |      37.94 |    67% | ✅ Best mean     |
| unsloth Q4_K_M MTP                         |   36.13 |    44.0 |      34.68 |    49% | ❌ -19% TG       |
| unsloth UD-Q4_K_XL MTP                     |   43.65 |    53.0 |      33.01 |    60% | ≈ Tied, worse @d |
| Chain: `ngram-mod,draft-mtp` (unsloth tip) |       — |       — |          — |      — | 🔴 CRASH (SSM)   |

## Sampling A/B (froggeric MTP, n=2, perf=high)

| Preset                  | temp / top_p / top_k / pp | tg mean | Accept@0 | Note          |
|-------------------------|---------------------------|--------:|---------:|---------------|
| Fast (temp=0.2)         | 0.2 / —   / 20  / —       |   45.00 |      67% | Fastest, repetitive |
| Precise coding (active) | 0.6 / 0.95 / 20 / 0.0     |   37.32 |      54% | ★ Current default   |
| Non-thinking general    | 0.7 / 0.8  / 20 / 1.5     |   36.26 |      57% | Best @ long ctx     |
| Thinking general        | 1.0 / 0.95 / 20 / 1.5     |   37.68 |      59% | Avoid (no MTP gain) |

## Other paths evaluated and rejected

| Option                       | Result on 7900 XTX                     |
|------------------------------|----------------------------------------|
| vLLM (ROCm)                  | ❌ -10–20%, no Qwen3.6 MTP, 4–8h install |
| TurboQuant (Vulkan port)     | ❌ Broken — 10 t/s, GPU util <30%      |
| DFlash / Hipfire             | ❌ Crashes >4k context, no MTP         |
| MLC-LLM (Vulkan)             | ⚠️ ~10 t/s slower, no MTP             |

## Hardware ceiling vs realistic upgrades

| Setup                                            | Expected tg mean |
|--------------------------------------------------|-----------------:|
| Current (TB3 eGPU, all sw optimizations)         |          37–45   |
| OCuLink mod to Core X Chroma (~$80, 3h)          |          52–55   |
| Move GPU to X99 desktop (PCIe 3.0 x16)           |          58–62   |
| Modern AM5 + PCIe 4.0 x16 (blog reference)       |              67  |

**Current `start_server start`:** llama.cpp PR 9173 + froggeric MTP Q4_K_M + `--spec-type draft-mtp --spec-draft-n-max 2` + KV q4_0 + FA on + Qwen precise coding sampling + GPU perf=high.

llama-benchy result:

cd /var/home/deck/tmp/llama-benchy
uv run llama-benchy \
  --base-url http://127.0.0.1:8081/v1 \
  --model froggeric/Qwen3.6-27B-MTP-GGUF \
  --served-model-name Qwen3.6-27B-Q4_K_M-mtp.gguf \
  --tokenizer Qwen/Qwen3.6-27B \
  --pp 2048 --tg 32 \
  --depth 0 8192 32768 \
  --runs 1 --no-cache --latency-mode generation --skip-coherence \
  --save-result results/qwen36-27b-mtp-8081-sample-20260513.json --format json

Results:

| context depth | pp t/s | tg t/s | peak tg t/s | TTFR | est PPT |
|---|---:|---:|---:|---:|---:|
| 0 | 457.92 | 29.75 | 30.0 | 4693 ms | 4477 ms |
| 8192 | 432.96 | 28.24 | 29.0 | 23870 ms | 23654 ms |
| 32768 | 329.57 | 25.24 | 27.0 | 105856 ms | 105640 ms |

.venv/bin/llama-benchy \
  --base-url http://127.0.0.1:8081/v1 \
  --model Qwen/Qwen3.6-27B \
  --served-model-name Qwen_Qwen3.6-27B-Q4_K_M.gguf \
  --tokenizer Qwen/Qwen3.6-27B \
  --pp 2048 \
  --tg 32 \
  --depth 0 8192 32768 \
  --runs 1 \
  --latency-mode generation \
  --save-result results/qwen36-27b-original-8081-20260513T235739Z.json \
  --format json


Results:
| depth | pp t/s | tg t/s | TTFR ms |
|---|---:|---:|---:|
| 0 | 685.49 | 30.63 | 3190.39 |
| 8192 | 640.61 | 30.00 | 16184.55 |
| 32768 | 486.52 | 28.16 | 71766.55 |

llama.cpp server config:

    #MODEL="/run/media/deck/ExternalSSD/.llama.cpp/models/froggeric_Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-Q4_K_M-mtp.gguf"
    MODEL="/var/run/media/deck/ExternalSSD/.llama.cpp/models/Qwen_Qwen3.6-27B-GGUF/Qwen_Qwen3.6-27B-Q4_K_M.gguf"

      # cd "/var/home/deck/tmp/llama-pr-22673-mtp-clean/build-vulkan-pr22673/bin"
      cd "/var/home/deck/code/llama.cpp/build-vulkan/bin"

      export VK_LOADER_LAYERS_DISABLE=VK_LAYER_LS_frame_generation

      exec ./llama-server \
        -m "$MODEL" \
        -ngl 99 \
        -dev Vulkan0 \
        -fa on \
        -c 200000 \
        -ctk q4_0 \
        -ctv q4_0 \
        -ub 256 \
	--temp 0.2 \
	--top-k 20 \
	--parallel 1 \
        -rea off \
        --reasoning-budget 0 \
        --host "$HOST" \
        --port "$PORT"

       # MTP flags:
       #       --spec-type mtp 
       #       --spec-draft-n-max 2

昨天测了一天感觉MTP打开没有变化（～30tok/s），用了几轮就会爆VRAM, 希望指正哪里出问题了。
我是用beelink ser7 + eGPU 7900xtx + bazzite + hermes agent + discrod
现在基本可以游戏/LLM随时切换, eGPU坑还是很多， 在等x99主板到装机

eGPU坑：
用all-ways-egpu可以点亮显卡+游戏
Kfd/ROCm没发使用,试了setup时不去设置iGPU kfd就能用了，但是bazzite不能进game mode了，还在找最后解决方案

确认是Hermes auxiliary.vision: auto 自动转到能支持的模型了