Misc. bug: 10% slower prefill start speed than baseline? (using Q8 KV)

未关闭
#157 4 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
22/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp, linux
领域
performance

调研方向

Reproduce with the same model and launch script on both builds; the key signal is the PP prefill lines from llama-server and the final prompt eval time in the logs. Profile the prefill path around Q8 KV cache handling and compare the fork against upstream commit 1dfe94e04, then narrow the roughly 10% gap to a kernel or scheduling difference. Done means the slowdown is explained and either fixed or shown to be configuration-related.

由索引模型根据 Issue 内容生成。

描述

Name and Version

Beellama:
0.4.6 main

LLama.cpp:
version: 0.4.0-dev (build 10910, commit 1dfe94e04)
built with GNU 15.2.0 for Linux x86_64

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line
#!/bin/sh
# Launch script exported with Grun for profile: Qwen-27B-MTP-TP
CUDA_VISIBLE_DEVICES=0,1 ./llama-server \
  --api-key sk-ik-llama \
  --batch-size 2048 \
  --cache-ram 24480 \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --chat-template-file /mnt/AI/models/qwen-chat-template-froggeric-22.1.jinja \
  --checkpoint-min-step 1024 \
  --ctx-checkpoints 32 \
  --ctx-size 192000 \
  --fit off \
  --gpu-layers all \
  --image-min-tokens 1024 \
  --load-mode none \
  --main-gpu 1 \
  --min-p 0.0 \
  --mmproj /mnt/AI/models/Qwen3.8-27B-mmproj-BF16.gguf \
  --model /mnt/AI/models/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-MID-HIGH.gguf \
  --parallel 1 \
  --presence-penalty 0.0 \
  --repeat-penalty 1.0 \
  --spec-draft-n-max 4 \
  --spec-draft-n-min 0 \
  --spec-draft-ngl all \
  --spec-draft-p-min 0.3 \
  --spec-draft-type-k q8_0 \
  --spec-draft-type-v q8_0 \
  --spec-type draft-mtp,ngram-map-k4v \
  --split-mode tensor \
  --temp 0.9 \
  --tensor-split 50,50 \
  --top-k 20 \
  --top-p 0.95 \
  --n-gpu-layers-draft all \
  --no-prefill-assistant \
  --reasoning-preserve
Problem description & steps to reproduce

With the same launch arguments, I noticed that prefill speed is always starting at ~2700 tps on BeeLlama and ~29xx tps on Llama.cpp. Tested with first prompt after new llama-server launch. About 30k identical system prompt + initial instruction.

First Bad Commit

No response

Relevant log output
Logs (Llama.cpp)
./run_Qwen-27B-MTP-TP.sh
00:00.26 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
00:00.26 I srv    load_model: loading model '/mnt/AI/models/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-MID-HIGH.gguf'
00:13.67 I cmn          init: llama threadpool init, n_threads = 6
00:13.83 I common_speculative_init_result: creating MTP draft context against the target model '/mnt/AI/models/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-MID-HIGH.gguf'
00:14.27 I srv    load_model: loaded multimodal model, '/mnt/AI/models/Qwen3.8-27B-mmproj-BF16.gguf'
00:14.32 I srv    load_model: initializing, n_slots = 1, n_ctx_slot = 192000, kv_unified = 'false'
00:14.32 W set_sampler: backend sampling not supported with SPLIT_MODE_TENSOR; using CPU
00:14.32 W spec common_specu: backend offload failed for seq_id=0; using CPU sampler
00:14.34 I srv  llama_server: model loaded
00:14.34 I srv  llama_server: listening on http://127.0.0.1:8080
00:14.34 W srv  llama_server: NOTICE: server default port will be changed to :9931 in a future release
00:14.34 W srv  llama_server:         ref: https://github.com/ggml-org/llama.cpp/pull/26508
00:16.30 I slot- 0 get_availa | task -1 | selected slot by LRU, t_last = -1
00:16.31 I slot- 0 launch_slo | task 0 | processing task, is_child = 0
00:19.96 I slot- 0 print_timi | task 0 | PP, n_tokens =  10240, progress = 0.30, t =   3.5 s / 2955.5 t/s
00:20.68 I slot- 0 print_timi | task 0 | PP, n_tokens =  12288, progress = 0.36, t =   4.2 s / 2941.9 t/s
00:21.41 I slot- 0 print_timi | task 0 | PP, n_tokens =  14336, progress = 0.42, t =   4.9 s / 2923.3 t/s
00:22.16 I slot- 0 print_timi | task 0 | PP, n_tokens =  16384, progress = 0.48, t =   5.6 s / 2903.2 t/s
00:22.91 I slot- 0 print_timi | task 0 | PP, n_tokens =  18432, progress = 0.54, t =   6.4 s / 2881.9 t/s
00:23.68 I slot- 0 print_timi | task 0 | PP, n_tokens =  20480, progress = 0.60, t =   7.2 s / 2859.1 t/s
00:24.47 I slot- 0 print_timi | task 0 | PP, n_tokens =  22528, progress = 0.66, t =   7.9 s / 2834.7 t/s
00:25.28 I slot- 0 print_timi | task 0 | PP, n_tokens =  24576, progress = 0.72, t =   8.7 s / 2811.0 t/s
00:26.09 I slot- 0 print_timi | task 0 | PP, n_tokens =  26624, progress = 0.78, t =   9.6 s / 2787.1 t/s
00:26.38 I slot- 0 print_timi | task 0 | PP, n_tokens =  27245, progress = 0.80, t =  10.0 s / 2715.8 t/s
00:27.38 I slot- 0 print_timi | task 0 | PP, n_tokens =  29293, progress = 0.86, t =  10.9 s / 2698.9 t/s
00:28.23 I slot- 0 print_timi | task 0 | PP, n_tokens =  31341, progress = 0.92, t =  11.7 s / 2683.4 t/s
00:29.09 I slot- 0 print_timi | task 0 | PP, n_tokens =  33389, progress = 0.98, t =  12.5 s / 2663.1 t/s
00:29.20 I slot- 0 print_timi | task 0 | PP, n_tokens =  33555, progress = 0.98, t =  12.9 s / 2607.9 t/s
00:29.61 I slot- 0 print_timi | task 0 | PP, n_tokens =  34067, progress = 1.00, t =  13.1 s / 2597.7 t/s
00:32.87 I slot- 0 print_timi | task 0 | n_gen =    251, tg =  82.49 t/s, tg_3s =  82.81 t/s
00:33.28 I slot- 0 print_timi | task 0 | prompt eval time =   13.5 s / 34071 tok (  0.4 ms/tok, 2516.8 t/s)
00:33.28 I slot- 0 print_timi | task 0 |        eval time =    3.4 s /  293 tok ( 11.8 ms/tok,  85.0 t/s)
00:33.28 I slot- 0 print_timi | task 0 |       total time =   17.0 s / 34364 tok
00:33.28 I slot- 0 print_timi | task 0 |    graphs reused =         78
00:33.28 I slot- 0 print_timi | task 0 | draft acceptance =  53.3% ( 204 /  383 accepted), mean len =  3.4
00:33.28 I slot- 0    release | task 0 | stop processing: n_tokens = 34363, truncated = 0
00:33.52 I slot- 0 get_availa | task -1 | selected slot by LCP similarity, f_sim_best = 0.850 (> 0.100 thold), f_keep = 0.998

Logs (BeeLlama)
./run_Qwen27B-MTP-TP.sh 
00:00.41 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
00:00.42 I srv    load_model: loading model '/mnt/AI/models/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-MID-HIGH.gguf'
00:17.36 I cmn          init: llama threadpool init, n_threads = 6
00:17.55 I common_speculative_init_result: creating MTP draft context against the target model '/mnt/AI/models/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-MID-HIGH.gguf'
00:18.28 I srv    load_model: loaded multimodal model, '/mnt/AI/models/Qwen3.8-27B-mmproj-BF16.gguf'
00:18.28 I srv    load_model: initializing, n_slots = 1, n_ctx_slot = 192000, kv_unified = 'false'
00:18.29 W set_sampler: backend sampling not supported with SPLIT_MODE_TENSOR; using CPU
00:18.29 W spec common_specu: backend offload failed for seq_id=0; using CPU sampler
00:18.29 I srv  llama_server: model loaded
00:18.29 I srv  llama_server: listening on http://127.0.0.1:8080
00:18.29 W srv  llama_server: NOTICE: server default port will be changed to :9931 in a future release
00:18.29 W srv  llama_server:         ref: https://github.com/ggml-org/llama.cpp/pull/26508
00:45.89 I slot- 0 get_availa | task -1 | selected slot by LRU, t_last = -1
00:45.89 I slot- 0 launch_slo | task 0 | processing task, is_child = 0
00:48.98 I slot- 0 print_timi | task 0 | PP, n_tokens =   8192, progress = 0.27, t =   3.0 s / 2713.6 t/s
00:49.71 I slot- 0 print_timi | task 0 | PP, n_tokens =  10240, progress = 0.34, t =   3.7 s / 2736.4 t/s
00:50.45 I slot- 0 print_timi | task 0 | PP, n_tokens =  12288, progress = 0.41, t =   4.5 s / 2739.9 t/s
00:51.20 I slot- 0 print_timi | task 0 | PP, n_tokens =  14336, progress = 0.48, t =   5.2 s / 2738.3 t/s
00:51.97 I slot- 0 print_timi | task 0 | PP, n_tokens =  16384, progress = 0.55, t =   6.0 s / 2728.8 t/s
00:52.76 I slot- 0 print_timi | task 0 | PP, n_tokens =  18432, progress = 0.61, t =   6.8 s / 2716.8 t/s
00:53.56 I slot- 0 print_timi | task 0 | PP, n_tokens =  20480, progress = 0.68, t =   7.6 s / 2701.3 t/s
00:54.37 I slot- 0 print_timi | task 0 | PP, n_tokens =  22528, progress = 0.75, t =   8.4 s / 2683.5 t/s
00:55.20 I slot- 0 print_timi | task 0 | PP, n_tokens =  24576, progress = 0.82, t =   9.2 s / 2665.5 t/s
00:56.04 I slot- 0 print_timi | task 0 | PP, n_tokens =  26624, progress = 0.89, t =  10.1 s / 2647.2 t/s
00:56.35 I slot- 0 print_timi | task 0 | PP, n_tokens =  27252, progress = 0.91, t =  10.4 s / 2613.6 t/s
00:57.37 I slot- 0 print_timi | task 0 | PP, n_tokens =  29300, progress = 0.98, t =  11.4 s / 2574.6 t/s
00:57.50 I slot- 0 print_timi | task 0 | PP, n_tokens =  29509, progress = 0.98, t =  11.6 s / 2543.9 t/s
00:57.89 I slot- 0 print_timi | task 0 | PP, n_tokens =  30021, progress = 1.00, t =  12.0 s / 2506.7 t/s
01:01.13 I slot- 0 print_timi | task 0 | n_gen = 259, tg =   85.4 t/s, tg_3s =   85.8 t/s
01:04.17 I slot- 0 print_timi | task 0 | n_gen = 399, tg =   65.8 t/s, tg_3s =   46.2 t/s
01:05.81 I slot- 0 print_timi | task 0 | prompt eval time =   12.2 s / 30025 tok (  0.4 ms/tok, 2455.5 t/s)
01:05.81 I slot- 0 print_timi | task 0 |        eval time =    7.7 s /  500 tok ( 15.4 ms/tok,  64.9 t/s)
01:05.81 I slot- 0 print_timi | task 0 |       total time =   19.9 s / 30525 tok
01:05.81 I slot- 0 print_timi | task 0 |    graphs reused =        148
主要语言
C++
星标
1.2k
派生
80
平均合并
4 天 20 小时
30 天内合并 PR
2

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

Anbeeld/beellama.cpp 的其他 Issue

查看 Anbeeld/beellama.cpp 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。