3090单卡跑的3090 club项目，hermes很慢，可能是啥原因呢？

[c0aster]

使用的模型是：ik-llama/iq4ks-mtp ubergarm-iq4ks/mtp.yml 200K
飞书向hermes部署任务后会很慢，然后使用其他工具如opencode，首次回复也是等半天（提示等待大模型回复）

部分日志：
======== Prompt cache: cache size: 87920, n_keep: 0, n_discarded_prompt: 0, cache_ram_n_min: 0, f_keep: 1.00, cache_ram_similarity: 0.50

• looking for better prompt, base f_keep = 1.000, sim = 0.999, n_keep = 0, n_discarded_prompt = 0
• cache state: 2 prompts, 7057.293 MiB (limits: 8192.000 MiB, 0 tokens, 103215 est)
  • prompt 0x57fe3f66c340: 29210 tokens, 0 discarded, checkpoints: 9, 2043.364 MiB
  • prompt 0x57fe40af7ac0: 59709 tokens, 0 discarded, checkpoints: 25, 5013.929 MiB
    prompt cache load took 50.92 ms
    INFO [ launch_slot_with_task] slot is processing task | tid="132079807373312" timestamp=1781745060 id_slot=0 id_task=61295
    ======== Cache: cache_size = 87920, n_past0 = 86407, n_past1 = 86407, n_past_prompt1 = 86407, n_past2 = 87920, n_past_prompt2 = 87916
    INFO [ batch_pending_prompt] kv cache rm [p0, end) | tid="132079807373312" timestamp=1781745060 id_slot=0 id_task=61295 p0=87920
    slot create_check: id 0 | task 61295 | erasing old context checkpoint (pos_min = 64764, pos_max = 64764, n_tokens = 64765, size = 150.122 MiB)
    slot create_check: id 0 | task 61295 | created context checkpoint 32 of 32 (pos_min = 88043, pos_max = 88043, n_tokens = 88044, size = 150.299 MiB, took 280.99 ms)
    INFO [ batch_pending_prompt] kv cache rm [p0, end) | tid="132079807373312" timestamp=1781745061 id_slot=0 id_task=61295 p0=88044
    INFO [ log_server_request] request | tid="132067457601536" timestamp=1781745061 remote_addr="10.158.167.122" remote_port=52733 status=200 method="POST" path="/v1/chat/completions" params={}
    INFO [ process_single_task] slot data | tid="132079807373312" timestamp=1781745078 id_task=61443 n_idle_slots=0 n_processing_slots=1
    INFO [ log_server_request] request | tid="132067415638016" timestamp=1781745078 remote_addr="127.0.0.1" remote_port=44230 status=200 method="GET" path="/health" params={}
    slot print_timing: id 0 | task 61295 |
    prompt eval time = 732.55 ms / 129 tokens ( 5.68 ms per token, 176.10 tokens per second)
    eval time = 26252.92 ms / 631 tokens ( 41.61 ms per token, 24.04 tokens per second)
    total time = 26985.47 ms / 760 tokens
    draft acceptance rate = 0.91892 ( 408 accepted / 444 generated)
    statistics mtp: #calls(b,g,a) = 345 58543 58542, #gen drafts = 58543, #acc drafts = 53577, #gen tokens = 117086, #acc tokens = 101953, dur(b,g,a) = 0.611, 557016.191, 44.509 ms
    slot create_check: id 0 | task 61295 | erasing old context checkpoint (pos_min = 64838, pos_max = 64838, n_tokens = 64839, size = 150.122 MiB)
    slot create_check: id 0 | task 61295 | created context checkpoint 32 of 32 (pos_min = 88678, pos_max = 88678, n_tokens = 88679, size = 150.304 MiB, took 284.19 ms)
    INFO [ release_slots] slot released | tid="132079807373312" timestamp=1781745087 id_slot=0 id_task=61295 n_ctx=200192 n_past=88679 n_system_tokens=0 n_cache_tokens=88679 truncated=false
    INFO [ slots_idle] all slots are idle | tid="132079807373312" timestamp=1781745087
    ======== Prompt cache: cache size: 88679, n_keep: 0, n_discarded_prompt: 0, cache_ram_n_min: 0, f_keep: 1.00, cache_ram_similarity: 0.50
• looking for better prompt, base f_keep = 1.000, sim = 1.000, n_keep = 0, n_discarded_prompt = 0
• cache state: 2 prompts, 7057.293 MiB (limits: 8192.000 MiB, 0 tokens, 103215 est)
  • prompt 0x57fe3f66c340: 29210 tokens, 0 discarded, checkpoints: 9, 2043.364 MiB
  • prompt 0x57fe40af7ac0: 59709 tokens, 0 discarded, checkpoints: 25, 5013.929 MiB
    prompt cache load took 93.73 ms
    INFO [ launch_slot_with_task] slot is processing task | tid="132079807373312" timestamp=1781745090 id_slot=0 id_task=61521
    ======== Cache: cache_size = 88679, n_past0 = 86407, n_past1 = 86407, n_past_prompt1 = 86407, n_past2 = 88679, n_past_prompt2 = 88675
    INFO [ batch_pending_prompt] kv cache rm [p0, end) | tid="132079807373312" timestamp=1781745090 id_slot=0 id_task=61521 p0=88679
    slot create_check: id 0 | task 61521 | erasing old context checkpoint (pos_min = 68934, pos_max = 68934, n_tokens = 68935, size = 150.153 MiB)
    slot create_check: id 0 | task 61521 | created context checkpoint 32 of 32 (pos_min = 88743, pos_max = 88743, n_tokens = 88744, size = 150.305 MiB, took 384.42 ms)
    INFO [ batch_pending_prompt] kv cache rm [p0, end) | tid="132079807373312" timestamp=1781745090 id_slot=0 id_task=61521 p0=88744
    INFO [ log_server_request] request | tid="132078322556928" timestamp=1781745090 remote_addr="10.158.167.122" remote_port=44163 status=200 method="POST" path="/v1/chat/completions" params={}

[Xiaote]

@c0aster 3090 单卡跑 Hermes 慢，我分析几个可能的原因：

1. MTP 模型的开销

你用的是 ik-llama/iq4ks-mtp，这是带 MTP（Multi-Token Prediction）投机解码的量化版。MTP 在生成时需要同时跑 draft model 和验证，对单卡是有额外开销的。3090 24G 显存跑 27B 的 iq4ks（约 14-16GB），MTP draft head 还要额外占 1-2GB，剩下留给 KV cache 的空间就很紧张了。

推荐先试一下关掉 MTP，用普通模式跑同一个模型：

--no-mtp

如果速度明显提升，说明是 MTP 的 overhead 大于收益。

2. Context 长度设置

200K 上下文对 3090 来说太大了。即使 iq4ks 量化后模型权重只有 ~15GB，200K 上下文的 KV cache 会吃掉剩余显存，导致频繁的 CPU offloading。建议先设小 context 试速度：

--ctx-size 8192

如果变快了，再逐步增大到合适值。

3. 确保用 CUDA backend

llama.cpp 默认会用 CUDA，但确认一下启动参数里有：

-ngl 9999  # 把所有层都加载到 GPU

或者至少 -ngl 32（27B 模型约 32 层全放 GPU）。

4. 一个快速调试流程

# 最小化测试
llama-cli -m ik-llama/iq4ks-mtp.gguf -ngl 9999 --no-mtp -c 4096 -p "你好"
# 如果这个快，说明是 long context 或 MTP 导致的慢

# 再加 MTP 试试
llama-cli -m ik-llama/iq4ks-mtp.gguf -ngl 9999 -c 4096 -p "你好"

你现在的具体 tokens/s 是多少？跑 llama-cli --speed 的结果贴一下，更好定位问题。

[毅袁]

关掉thinking 立马起飞
--jinja --chat-template-kwargs "{"enable_thinking": false}"

[c0aster]

@毅袁 thinks，我试试，抓个网页信息都要10分钟，但用opencode写项目，启动后还是挺快的

[wwcd2016]

3090 带hermes。我测试了一个月，放弃。没有一个能满足要求的。一个月20元就足够养活hermes了。
现在模型主要用一下越狱版。

[stxpnet]

聊天不能用0.6,0.6是编程的，追求精准无错， 聊天是泛泛的发现一些问题，按官方推荐的。0.7试试

[c0aster]

@stxpnet 大佬具体是啥的0.7呢，我直接跑的3090 club的项目

[stxpnet]

把 你的完整启动命令贴出来。
温度 要0.7才可以聊天，比如下面这套配置（注意最后一行），用来聊天就能撑起多轮对话，适合一步一步的问答，探究出一个小的结论，不适合编写程序： killall llama-server 2>/dev/null; sleep 3
#6-19pm18测试 给hermes用的。
#--cash-ram 2560预期多用2.5G显存，如果显存不够慎用
#--ctx-checkpoints 64 将检查点从32增加到64，可能会增加显存，注意观察
#用了之后的好处,不会再出现 forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory,
#erased invalidated context checkpoint (pos_min = 87014, pos_max = 87014
killall llama3-server 2>/dev/null; sleep 3
killall llama-server 2>/dev/null; sleep 3
export LD_LIBRARY_PATH=/data/models/beellma616-kv.cpp/build/bin:$LD_LIBRARY_PATH
/data/model2/beellma616-kv.cpp/build/bin/llama-server
--host 0.0.0.0 --port 8025
--api-key "sk-my-tnt-secret-key-1234567890"
-m /data/model3/Qwen3.6-27B-Omnimerge-v4-IQ4_XS-mtp-522.gguf
--spec-type draft-mtp
--spec-draft-n-max 3
-ngl all -n 8192
--ctx-size 160000
--ctx-checkpoints 64
-b 4096 -ub 1024 -np 1
--cache-type-k kvarn4
--cache-type-v kvarn4 --kv-unified
--no-mmap --mlock
--cache-ram 3072 --cache-reuse 384
--no-host --jinja
--chat-template-kwargs '{"preserve_thinking": true}'
--chat-template-file /data/model2/chat_template-fixed-v20.jinja
--no-warmup --reasoning on --reasoning-budget 768 -fa on
--temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0

[c0aster]

@stxpnet

services:
  ik-llama-qwen36-27b-iq4ks-mtp:
    image: ${IK_LLAMA_IMAGE:-ghcr.io/ikawrakow/ik-llama-cpp@sha256:5f914f1ccade922417af58c94bd1cbb558052c8852d86678ead3fe693eec0143}
    container_name: "${ESTATE_CONTAINER:-ik-llama-qwen36-27b}"
    restart: unless-stopped
    ports:
      - "${ESTATE_PORT:-${PORT:-8020}}:8080"
    volumes:
      - "${MODEL_DIR:-../../../../../../models-cache}:/models:ro"
    # server target ENTRYPOINT is /app/llama-server — args only below.
    # ⚠ -np 1 is intentional on a single 24 GB card — do NOT raise it to
    #   "parallelize." One GPU is compute-bound: extra slots divide its
    #   throughput, they don't multiply it. At -np 4 each slot fell to
    #   ~14 tok/s here — slow enough to trip agentic clients' per-request
    #   timeouts (aider ran 1/30) — and -np>1 also auto-disables MTP and
    #   can OOM the spec-context buffer. On a higher-throughput card (e.g.
    #   5090) or multi-GPU the trade may flip — re-validate before raising.
    command: >-
      --host 0.0.0.0
      --port 8080
      --model /models/${GGUF_FILE:-qwen3.6-27b-gguf/ubergarm-mtp-iq4ks/Qwen3.6-27B-MTP-IQ4_KS.gguf}
      -ngl 99
      --ctx-size ${CTX_SIZE:-200000}
      -b ${BATCH_SIZE:-4096}
      -ub ${UBATCH_SIZE:-1024}
      -np ${NP:-1}
      -ctk ${KV_TYPE:-q4_0}
      -ctv ${KV_TYPE:-q4_0}
      -khad
      -vhad
      -ngld 99
      --spec-type mtp:n_max=${MTP_DRAFT_N_MAX:-2},p_min=${DRAFT_P_MIN:-0.0}
      --recurrent-ckpt-mode auto
      --merge-qkv
      -fa on
      --chat-template-kwargs '{"enable_thinking": false}'
      --jinja
      --chat-template-file /models/qwen3.6-27b-gguf/ubergarm-mtp-iq4ks/chat_template.jinja
      --parallel-tool-calls
      --reasoning ${REASONING:-off}
      --reasoning-format ${REASONING_FORMAT:-deepseek}
      --temp ${TEMP:-${TEMPERATURE:-0.6}}
      --top-p ${TOP_P:-0.95}
      --top-k ${TOP_K:-20}
      --min-p ${MIN_P:-0.0}
      --repeat-penalty ${REPEAT_PENALTY:-1.0}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${ESTATE_GPUS:-${CUDA_VISIBLE_DEVICES:-0}}"]
              capabilities: [gpu]

[wwcd2016]

其实，楼主的问题，就是一个目前没有彻底解决方案的问题。
很期待有大拿能搞定。然后我们 3090 单卡用户抄作业。
个人意见懵逼了。已经发给放弃单卡 玩 hermes。
当然，本质还是 deepseek 足够便宜。

[c0aster]

@wwcd2016 是的，看来3090还是显存不太够