llama.cpp+qwen3.6-27b 初步测试
-
@老用户 说:
kv cache type 使用q4_0精度，有没有在长上下文的情况下测试过稳定性，智商、工具调用如何。我在使用中发现，上下文到7-8W的时候就开始飘了。所以有时候我认为256K上下文有没有必要，不如把kv cache type精度调高点
我今天正好做了个测试,
Hermes 长会话 KV
q8_0vsq4_0A/B（2026-05-26）结论
- 这次真实 Hermes 多轮长会话基准里，
KV q8_0和KV q4_0没有表现出可见的语义或结构优势差异。 - 两边都是：
12 / 12turn 全过- 最终精确召回通过
- 最终文件状态召回通过
chain_diff只有first_request + exact_message_append- 没有
message_prefix_drift - 没有
forcing full prompt re-processing - 没有 server 侧
class=prefix-drift
更直接地说：
在这条真实 Hermes replay/tool/file 多轮链路上，至少这一次跑下来，q8_0没有比q4_0明显更稳，q4_0也没有出现明显漂移退化。基准配置
- 日期：
2026-05-26 - benchmark 脚本：
~/custom-agent-stack/local-agent-setup/scripts/benchmark_hermes_long_session_kv.py - 结果目录：
~/.cache/local-agent-setup/benchmarks/hermes-long-kv-20260526-ab1 - Hermes 运行时：仓库版
~/custom-agent-stack/hermes - llama-server：
~/src/ik_llama.cpp/build-mmq/bin/llama-server - 模型：
~/models/Qwen3.6-27B-MTP-IQ4_KS.gguf - ctx：
128000 - block chars：
30000 - toolset：
file - compression：
off - ignore_rules：
on
任务形态
不是单问单答，也不是直接打 OpenAI-compatible
/chat/completions的伪 benchmark。
这次走的是真实 Hermes 多轮链路：HermesCLI.chat()连续多轮追加 history- 中间混合长 reference turn
- 中间混合
write_file / patch / read_file - 结尾做精确 JSON 召回
- 同时抓：
- Hermes request diagnostics
- llama-server console log
- 最终语义结果和文件状态
结果表
KV Passed Turns Final Recall Final File Recall Max Approx Tokens exact_message_appendmessage_prefix_driftforcing_fullprefix_driftacceptance avg q8_012/12yes yes 41221150000.73017q4_012/12yes yes 41221150000.75664解释
1. 结构稳定性
这部分两者完全一样：
first_request = 1exact_message_append = 15message_prefix_drift = 0session_changed = 0forcing full prompt re-processing = 0
这说明：
- Hermes replay 没有在这组任务里引入可见 prefix 漂移
- llama-server 的 checkpoint / prompt cache 路径工作正常
q4_0没有比q8_0更容易把 replay 链打崩
2. 语义结果
这部分两者也一样：
- 长 reference 块记忆没丢
- 文件工具链没错
- 结尾 JSON 精确召回通过
- 最终文件尾部状态召回通过
所以就“长会话漂移”这个问题看，这次没有证据表明
q8_0更稳。3. acceptance
这次单次 run 里：
q8_0:0.73017q4_0:0.75664
q4_0略高，但差距不大，而且这只是一次跑出来的平均值。
在没有重复样本之前，不能据此下结论说q4_0优于q8_0，更不能据此反推“q8_0在真实 Hermes 长会话里一定更聪明”。目前更可信的判断
至少在你这条链路里：
KV q8_0对patch/path这类错误并没有表现出明确额外优势- 真正影响 tool 成功率的主因，仍然更像是：
- 模型文件本身
- chat template / replay 一致性
- Hermes 侧 prefix 稳定化 hack
- llama-server 侧 checkpoint / single-slot 行为
而不是简单的：
- “把 KV 从
q4_0提到q8_0，模型就突然会分辨 patch/path”
@blackjack 相信你的测试个结果，但我实际跑hermes过程中，Q4_0确实拉胯，跑OpenClaw更是如此，就是经常会陷入死循环。
- 这次真实 Hermes 多轮长会话基准里，
-
@blackjack 相信你的测试个结果，但我实际跑hermes过程中，Q4_0确实拉胯，跑OpenClaw更是如此，就是经常会陷入死循环。
-
@blackjack 相信你的测试个结果，但我实际跑hermes过程中，Q4_0确实拉胯，跑OpenClaw更是如此，就是经常会陷入死循环。
qwen的工具调用极弱，让他专门做过patch工具测试，分不清工具名称patch和参数名称path。这个就是模型能力问题，再怎么提示也白扯，只能在hermes里把参数名称path改成路径等其他严重不让他花眼的文字，还有各种对他人性化的反馈。死循环基本就是掉入到各种工具调用的汪洋大海中了，你可以开个ai让他研究一下日志
@blackjack 没深入研究，我用Q8 kv就没这个问题了。
-
昨天碰到一个 oom, 好像是 host ram 的 oom, 没搞懂为啥它使用那么多 system ram...
[71766.725058] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 4.2025.05-1~bpo12+1 03/12/2026 [71766.725059] Call Trace: [71766.725073] <TASK> [71766.725076] dump_stack_lvl+0x5d/0x80 [71766.725082] dump_header+0x43/0x1aa [71766.725085] oom_kill_process.cold+0xa/0xb2 [71766.725088] out_of_memory+0x217/0x4b0 [71766.725091] __alloc_pages_slowpath.constprop.0+0xc3b/0xdd0 [71766.725098] __alloc_frozen_pages_noprof+0x2cd/0x320 [71766.725103] alloc_pages_mpol+0x7d/0x180 [71766.725107] folio_alloc_noprof+0x5d/0xe0 [71766.725110] __filemap_get_folio+0x1dd/0x330 [71766.725112] filemap_fault+0x10c/0x12f0 [71766.725116] __do_fault+0x30/0x180 [71766.725119] do_fault+0x310/0x540 [71766.725122] __handle_mm_fault+0x8ee/0xf20 [71766.725124] ? srso_alias_return_thunk+0x5/0xfbef5 [71766.725129] handle_mm_fault+0xec/0x2e0 [71766.725132] do_user_addr_fault+0x2c3/0x7f0 [71766.725135] exc_page_fault+0x74/0x180 [71766.725139] asm_exc_page_fault+0x26/0x30 [71766.725140] RIP: 0033:0x9bab0a [71766.725161] Code: Unable to access opcode bytes at 0x9baae0. [71766.725162] RSP: 002b:000000c00093cc60 EFLAGS: 00010216 [71766.725164] RAX: 0000000001b35410 RBX: 0000000001b77928 RCX: 0000000001b77928 [71766.725165] RDX: 0000000000a244e0 RSI: 00000000009baae0 RDI: 000000c000aaa6e0 [71766.725166] RBP: 000000c00093cca0 R08: 0000000000000040 R09: 0000000000000082 [71766.725167] R10: 00007f88af626fa8 R11: 00000000000000d0 R12: 0000000000000006 [71766.725168] R13: 000000c001548410 R14: 000000c000171880 R15: ffffffffffffffff [71766.725172] </TASK> [71766.725173] Mem-Info: [71766.725181] active_anon:3652772 inactive_anon:3225935 isolated_anon:0 active_file:74 inactive_file:592 isolated_file:0 unevictable:2004 dirty:0 writeback:0 slab_reclaimable:16967 slab_unreclaimable:49972 mapped:2076 shmem:4443 pagetables:27449 sec_pagetables:0 bounce:0 kernel_misc_reclaimable:0 free:57534 free_pcp:0 free_cma:0 [71766.725185] Node 0 active_anon:14611088kB inactive_anon:12903740kB active_file:296kB inactive_file:2004kB unevictable:8016kB isolated(anon):0kB isolated(file):0kB mapped:8304kB dirty:0kB writeback:0kB shmem:17772kB shmem_thp:0kB shmem_pmdmapped:0 kB anon_thp:17055744kB kernel_stack:12720kB pagetables:109796kB sec_pagetables:0kB all_unreclaimable? no Balloon:0kB [71766.725189] Node 0 DMA free:11264kB boost:0kB min:28kB low:40kB high:52kB reserved_highatomic:0KB free_highatomic:0KB active_anon:0kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15864kB managed:153 60kB mlocked:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB [71766.725193] lowmem_reserve[]: 0 1948 32063 32063 32063 [71766.725198] Node 0 DMA32 free:124216kB boost:0kB min:3760kB low:5576kB high:7392kB reserved_highatomic:0KB free_highatomic:0KB active_anon:1852244kB inactive_anon:14144kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present: 2061152kB managed:1995156kB mlocked:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB [71766.725202] lowmem_reserve[]: 0 0 30114 30114 30114 [71766.725207] Node 0 Normal free:94656kB boost:30720kB min:94508kB low:125332kB high:156156kB reserved_highatomic:0KB free_highatomic:0KB active_anon:12758844kB inactive_anon:12889596kB active_file:660kB inactive_file:2004kB unevictable:8016kB writ epending:0kB present:31457280kB managed:30837388kB mlocked:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB [71766.725210] lowmem_reserve[]: 0 0 0 0 0 [71766.725215] Node 0 DMA: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 1*1024kB (U) 1*2048kB (M) 2*4096kB (M) = 11264kB [71766.725228] Node 0 DMA32: 0*4kB 1*8kB (M) 3*16kB (UM) 8*32kB (UM) 10*64kB (UM) 5*128kB (UM) 3*256kB (UM) 4*512kB (UM) 1*1024kB (U) 2*2048kB (M) 28*4096kB (M) = 124216kB [71766.725245] Node 0 Normal: 635*4kB (UME) 484*8kB (UME) 664*16kB (UME) 503*32kB (UME) 315*64kB (UME) 196*128kB (UME) 59*256kB (UME) 3*512kB (ME) 0*1024kB 0*2048kB 0*4096kB = 95020kB [71766.725261] Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=1048576kB [71766.725263] Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB [71766.725264] 12002 total pagecache pages [71766.725265] 6815 pages in swap cache [71766.725266] Free swap = 72kB [71766.725267] Total swap = 8496124kB [71766.725268] 8383574 pages RAM [71766.725269] 0 pages HighMem/MovableOnly [71766.725269] 171598 pages reserved [71766.725270] 0 pages hwpoisoned [71766.725271] Tasks state (memory values in pages): [71766.725272] [ pid ] uid tgid total_vm rss rss_anon rss_file rss_shmem pgtables_bytes swapents oom_score_adj name [71766.725279] [ 513] 0 513 16897 240 32 208 0 135168 256 -250 systemd-journal [71766.725283] [ 535] 990 535 23023 185 0 185 0 81920 256 0 systemd-timesyn [71766.725285] [ 541] 0 541 9456 707 544 163 0 98304 384 -1000 systemd-udevd [71766.725288] [ 819] 0 819 77254 161 64 97 0 106496 96 0 accounts-daemon [71766.725290] [ 821] 105 821 1564 239 0 239 0 53248 64 0 avahi-daemon [71766.725292] [ 824] 0 824 1716 185 64 121 0 49152 0 0 cron [71766.725294] [ 826] 989 826 2549 457 256 201 0 61440 256 -900 dbus-daemon [71766.725297] [ 839] 0 839 20199 275 96 179 0 61440 0 0 irqbalance [71766.725299] [ 845] 987 845 95842 717 569 148 0 114688 224 0 polkitd [71766.725301] [ 846] 0 846 54923 493 384 109 0 77824 0 0 rsyslogd [71766.725303] [ 848] 0 848 76962 254 128 126 0 102400 0 0 switcheroo-cont [71766.725305] [ 861] 0 861 4761 267 128 139 0 73728 192 0 systemd-logind [71766.725308] [ 862] 0 862 48059 131 0 131 0 126976 480 0 touchegg [71766.725310] [ 865] 0 865 117523 709 544 165 0 151552 0 0 udisksd [71766.725312] [ 881] 105 881 1518 151 37 114 0 53248 32 0 avahi-daemon [71766.725314] [ 964] 0 964 102609 882 704 178 0 167936 0 0 NetworkManager [71766.725317] [ 969] 0 969 4383 353 224 129 0 77824 0 0 wpa_supplicant [71766.725319] [ 1000] 0 1000 97731 195 64 131 0 126976 384 0 ModemManager [71766.725321] [ 1045] 0 1045 95236 179 0 179 0 110592 192 0 lightdm [71766.725323] [ 1048] 0 1048 2944 447 256 191 0 61440 32 -1000 sshd [71766.725325] [ 1082] 0 1082 654404 6572 5217 178 1177 782336 9472 0 Xorg [71766.725327] [ 1085] 0 1085 2042 147 32 115 0 57344 0 0 agetty [71766.725329] [ 1228] 0 1228 10995 185 40 145 0 86016 96 0 master [71766.725332] [ 1230] 116 1230 11129 208 32 176 0 77824 128 0 qmgr [71766.725334] [ 1246] 0 1246 43208 161 0 161 0 90112 256 0 lightdm [71766.725336] [ 1254] 1000 1254 5813 822 640 182 0 98304 32 100 systemd [71766.725338] [ 1256] 1000 1256 6387 531 421 110 0 81920 128 100 (sd-pam) [71766.725340] [ 1278] 1000 1278 2343 644 480 164 0 53248 0 200 dbus-daemon [71766.725342] [ 1279] 1000 1279 25582 1141 960 181 0 98304 0 200 pipewire [71766.725345] [ 1280] 1000 1280 21187 311 128 183 0 77824 0 200 pipewire [71766.725347] [ 1281] 1000 1281 119867 1068 896 172 0 155648 0 200 wireplumber [71766.725349] [ 1282] 1000 1282 41570 791 561 230 0 90112 0 200 pipewire-pulse [71766.725351] [ 1283] 1000 1283 119267 162 0 162 0 151552 544 0 cinnamon-sessio [71766.725353] [ 1297] 1000 1297 1810 183 32 151 0 53248 0 200 mpris-proxy [71766.725355] [ 1300] 114 1300 5369 193 32 161 0 61440 0 0 rtkit-daemon [71766.725358] [ 1361] 1000 1361 2637 153 71 82 0 49152 192 0 ssh-agent [71766.725360] [ 1372] 1000 1372 61760 178 21 157 0 225280 2816 0 fcitx [71766.725362] [ 1378] 1000 1378 2054 139 32 107 0 57344 64 0 dbus-daemon [71766.725364] [ 1382] 1000 1382 1279 117 3 114 0 49152 32 0 fcitx-dbus-watc [71766.725367] [ 1395] 1000 1395 45784 376 224 152 0 102400 0 200 gnome-keyring-d [71766.725369] [ 1404] 1000 1404 191033 682 517 165 0 462848 3232 0 csd-media-keys [71766.725371] [ 1407] 1000 1407 76111 218 32 186 0 98304 96 0 csd-screensaver [71766.725373] [ 1408] 1000 1408 79234 246 64 182 0 114688 256 0 csd-print-notif [71766.725375] [ 1411] 1000 1411 154236 268 0 268 0 438272 3648 0 csd-automount [71766.725377] [ 1416] 1000 1416 136010 270 1 269 0 425984 3872 0 csd-wacom [71766.725379] [ 1419] 1000 1419 176584 726 530 196 0 458752 3104 0 csd-color [71766.725381] [ 1421] 1000 1421 119372 1360 1084 276 0 421888 2496 0 csd-housekeepin [71766.725384] [ 1422] 1000 1422 119717 724 544 180 0 421888 3168 0 csd-xsettings [71766.725387] [ 1425] 1000 1425 127752 1320 1062 226 32 446464 2624 0 csd-background [71766.725391] [ 1427] 1000 1427 100813 282 97 185 0 413696 3520 0 csd-clipboard [71766.725394] [ 1428] 1000 1428 175014 1286 1073 213 0 454656 2560 0 csd-power [71766.725396] [ 1430] 1000 1430 60105 130 32 98 0 86016 128 0 csd-a11y-settin [71766.725398] [ 1431] 1000 1431 59835 114 0 114 0 81920 160 0 csd-settings-re [71766.725400] [ 1432] 1000 1432 154174 248 0 248 0 434176 3584 0 csd-keyboard [71766.725402] [ 1435] 1000 1435 95328 128 0 128 0 102400 128 0 at-spi-bus-laun [71766.725404] [ 1446] 1000 1446 41342 209 128 81 0 69632 0 200 dconf-service [71766.725406] [ 1447] 1000 1447 2120 172 0 172 0 53248 128 0 dbus-daemon [71766.725408] [ 1466] 1000 1466 78186 373 160 213 0 110592 0 200 gvfsd [71766.725410] [ 1487] 1000 1487 118027 287 160 127 0 122880 0 200 gvfsd-fuse [71766.725413] [ 1510] 1000 1510 42226 171 0 171 0 86016 192 0 at-spi2-registr [71766.725415] [ 1514] 1000 1514 134662 585 320 265 0 147456 0 200 gvfs-udisks2-vo [71766.725417] [ 1520] 115 1520 79008 173 0 173 0 122880 1120 0 colord [71766.725419] [ 1527] 0 1527 79740 509 320 189 0 122880 192 0 upowerd [71766.725421] [ 1539] 1000 1539 76984 252 32 220 0 98304 128 200 gvfs-mtp-volume [71766.725423] [ 1552] 1000 1552 97513 145 0 145 0 126976 288 200 gvfs-afc-volume [71766.725425] [ 1558] 1000 1558 77224 388 128 260 0 102400 32 200 gvfs-gphoto2-vo [71766.725428] [ 1563] 1000 1563 76958 244 96 148 0 102400 0 200 gvfs-goa-volume [71766.725430] [ 1568] 1000 1568 129472 1092 928 164 0 233472 0 200 goa-daemon [71766.725432] [ 1576] 1000 1576 103338 155 32 123 0 167936 448 0 csd-printer [71766.725435] [ 1582] 1000 1582 112521 2526 2302 224 0 229376 2176 0 cinnamon-launch [71766.725437] [ 1593] 1000 1593 96929 398 192 206 0 114688 0 200 goa-identity-se [71766.725439] [ 1609] 1000 1609 1585603 21400 19495 175 1730 1376256 9984 0 cinnamon [71766.725441] [ 1666] 1000 1666 96417 143 0 143 0 118784 960 0 ibus-daemon [71766.725443] [ 1673] 1000 1673 42299 171 32 139 0 86016 96 0 ibus-memconf [71766.725445] [ 1674] 1000 1674 69794 271 27 244 0 167936 3040 0 ibus-extension- [71766.725447] [ 1676] 1000 1676 44591 237 32 205 0 98304 320 0 ibus-x11 [71766.725449] [ 1681] 1000 1681 77136 234 128 106 0 102400 0 200 ibus-portal [71766.725452] [ 1693] 1000 1693 102902 252 64 188 0 163840 1280 0 xapp-sn-watcher [71766.725454] [ 1715] 1000 1715 42298 196 32 164 0 86016 96 0 ibus-engine-sim [71766.725456] [ 1721] 1000 1721 153619 785 526 227 32 299008 5696 0 nemo-desktop [71766.725458] [ 1724] 1000 1724 132589 3439 3227 212 0 270336 2816 0 blueman-applet [71766.725460] [ 1726] 1000 1726 75200 278 64 214 0 204800 3904 0 cinnamon-killer [71766.725462] [ 1729] 1000 1729 263369 806 525 281 0 450560 4480 0 evolution-alarm [71766.725464] [ 1770] 1000 1770 375036 2744 2536 208 0 446464 1472 200 evolution-sourc [71766.725466] [ 1775] 1000 1775 113530 795 608 187 0 180224 32 200 obexd [71766.725469] [ 1788] 1000 1788 206201 1077 928 149 0 274432 0 200 evolution-addre [71766.725498] [ 1791] 1000 1791 225026 209 0 209 0 241664 864 200 evolution-calen [71766.725501] [ 1823] 0 1823 320552 10808 10778 30 0 282624 797 0 netbird [71766.725504] [ 1824] 0 1824 77293 260 128 132 0 98304 0 0 power-profiles- [71766.725506] [ 1910] 1000 1910 133521 405 192 213 0 126976 0 200 gvfsd-trash [71766.725509] [ 2079] 1000 2079 263021 2578 2306 272 0 413696 9664 0 mintUpdate [71766.725511] [ 2137] 1000 2137 16271 287 115 172 0 172032 4544 0 applet.py [71766.725513] [ 2145] 0 2145 5552 717 512 205 0 81920 32 0 sshd-session [71766.725516] [ 2155] 1000 2155 4453 169 0 169 0 81920 480 0 ssh [71766.725518] [ 2158] 1000 2158 4323 265 64 201 0 69632 352 0 ssh [71766.725520] [ 2160] 1000 2160 114180 159 0 159 0 102400 352 0 sshfs [71766.725523] [ 2161] 1000 2161 77314 35 35 0 0 86016 32 0 sshfs [71766.725526] [ 2169] 1000 2169 145847 5382 5163 219 0 315392 4672 0 mintreport-tray [71766.725528] [ 2184] 1000 2184 134921 2068 1716 224 128 225280 192 200 gnome-terminal- [71766.725531] [ 2189] 1000 2189 157663 1011 832 179 0 184320 0 200 xdg-desktop-por [71766.725533] [ 2194] 1000 2194 77191 257 96 161 0 102400 0 200 xdg-permission- [71766.725535] [ 2199] 1000 2199 134930 273 128 145 0 139264 0 200 xdg-document-po [71766.725538] [ 2205] 1000 2205 646 83 0 83 0 49152 0 200 fusermount3 [71766.725540] [ 2210] 1000 2210 102736 1353 1120 233 0 167936 192 200 xdg-desktop-por [71766.725543] [ 2228] 1000 2228 103011 1398 1248 150 0 172032 64 200 xdg-desktop-por [71766.725546] [ 2239] 1000 2239 2256 491 384 107 0 61440 160 200 bash [71766.725548] [ 2285] 1000 2285 5666 788 557 231 0 81920 128 0 sshd-session [71766.725550] [ 2286] 0 2286 5554 711 512 199 0 90112 0 0 sshd-session [71766.725552] [ 2288] 1000 2288 2282 165 32 133 0 57344 512 0 bash [71766.725554] [ 2319] 1000 2319 5595 770 590 180 0 90112 0 0 sshd-session [71766.725557] [ 2331] 1000 2331 675 117 0 117 0 45056 0 0 sftp-server [71766.725559] [ 4065] 1000 4065 4282 188 80 108 0 73728 576 0 tmux: client [71766.725561] [ 4067] 1000 4067 8892 235 72 163 0 102400 1664 0 tmux: server [71766.725563] [ 5508] 1000 5508 2323 258 128 130 0 61440 480 0 bash [71766.725566] [ 38132] 1000 38132 2290 187 64 123 0 57344 512 0 bash [71766.725569] [ 38178] 1000 38178 5185 1453 1275 178 0 73728 0 0 nvtop [71766.725571] [ 50374] 0 50374 127230 1314 1115 199 0 266240 5024 0 fwupd [71766.725574] [ 50951] 1000 50951 1771 206 64 142 0 57344 0 0 bash [71766.725576] [ 137326] 0 137326 5246 368 224 144 0 81920 224 0 cupsd [71766.725579] [ 137327] 0 137327 48293 395 224 171 0 131072 544 0 cups-browsed [71766.725581] [ 416850] 0 416850 5553 758 544 214 0 90112 0 0 sshd-session [71766.725584] [ 416920] 1000 416920 5594 714 558 156 0 90112 32 0 sshd-session [71766.725586] [ 416921] 1000 416921 2282 562 480 82 0 65536 96 0 bash [71766.725588] [ 416960] 1000 416960 4282 280 112 168 0 65536 64 0 tmux: client [71766.725591] [ 416961] 1000 416961 2323 454 384 70 0 57344 224 0 bash [71766.725593] [ 438401] 0 438401 5552 722 576 146 0 94208 0 0 sshd-session [71766.725596] [ 438453] 1000 438453 5633 850 621 229 0 94208 32 0 sshd-session [71766.725598] [ 438468] 0 438468 5552 745 512 233 0 81920 0 0 sshd-session [71766.725601] [ 438470] 1000 438470 2282 591 480 111 0 53248 96 0 bash [71766.725603] [ 438525] 1000 438525 5593 792 621 171 0 81920 0 0 sshd-session [71766.725605] [ 438547] 1000 438547 675 81 0 81 0 49152 0 0 sftp-server [71766.725607] [ 438608] 1000 438608 4282 1338 1136 202 0 73728 0 0 tmux: client [71766.725610] [ 793687] 1000 793687 1771 170 32 138 0 57344 32 0 run-qwen3-vulka [71766.725612] [ 793688] 1000 793688 17163172 6781252 6781035 217 0 88559616 1995769 0 llama-server [71766.725615] [ 945076] 116 945076 11117 188 32 156 0 77824 128 0 pickup [71766.725619] [1047815] 1000 1047815 1395 103 0 103 0 53248 0 0 sleep [71766.725621] oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0,global_oom,task_memcg=/user.slice/user-1000.slice/[email protected]/tmux-spawn-ff073762-1455-4c6b-8b1c-b737ee738d0c.scope,task=llama-server,pid=793688,uid=1000 [71766.725781] Out of memory: Killed process 793688 (llama-server) total-vm:68652688kB, anon-rss:27124140kB, file-rss:868kB, shmem-rss:0kB, UID:1000 pgtables:86484kB oom_score_adj:0 -
@rock-shi 那就对了，24G跑128K上下文+MTP资源不够
我27 q4量化，kv均q8_0量化，上下文128k，MTP, 5090laptop 24GRAM，开thinking，50+tps，快的起飞啊
@blackjack 我在论坛大神的指点下，也起飞了
llama-server的启动参数
--reasoning off ^
--n-gpu-layers -1 ^
--ctx-size 131072 ^
--batch-size 2048 ^
--ubatch-size 1024 ^
--flash-attn on ^
--cache-type-k q4_0 ^
--cache-type-v q4_0 ^
--spec-type draft-mtp,ngram-mod ^
--spec-draft-n-max 3 ^
--spec-ngram-mod-n-max 5 ^
--spec-ngram-mod-n-min 3 ^
--temp 0.7 ^
--parallel 1 -
7900xtx+Ubuntu性能这么好？
我Windows10+RTX3090Ti，--n-gpu-layers 999 ^
--ctx-size 131072 ^
--batch-size 2048 ^
--ubatch-size 1024 ^
--flash-attn on ^
--cache-type-k q4_0 ^
--cache-type-v q4_0 ^
--cache-type-k-draft q4_0 ^
--cache-type-v-draft q4_0 ^不开MTP跑Qwen3.6 27B只能跑到30tokens/s;
开MTP变得更慢特别是在长上下文时，例如：我让模型分析一个大约128K的md文件，然后就爆了
你可以查一下编译llama-server的时候，用的mmq还是cuBLAS，或者有没有fallback到cuBLAS。亲测，两者性能差距巨大。
@blackjack 我用的是llama.cpp官方release：llama-b9329-bin-win-cuda-12.4-x64
之前我自己编译，不知道是什么参数不对，build后的llama-server怎么调参数都只有10tokens/s
-
@rock-shi 那就对了，24G跑128K上下文+MTP资源不够
我27 q4量化，kv均q8_0量化，上下文128k，MTP, 5090laptop 24GRAM，开thinking，50+tps，快的起飞啊
@rock-shi 那就对了，24G跑128K上下文+MTP资源不够
我27 q4量化，kv均q8_0量化，上下文128k，MTP, 5090laptop 24GRAM，开thinking，50+tps，快的起飞啊
厉害！一样的卡，大哥能给个作业抄吗？14900k，32g内存，llama.cpp，感谢！
-
@rock-shi 那就对了，24G跑128K上下文+MTP资源不够
我27 q4量化，kv均q8_0量化，上下文128k，MTP, 5090laptop 24GRAM，开thinking，50+tps，快的起飞啊
厉害！一样的卡，大哥能给个作业抄吗？14900k，32g内存，llama.cpp，感谢！
-
关于 hermes 接入 llama-server, 这两天有两个观察:
- hermes 新会话的第一条 prompt 大概是不到 20k token, 7900xtx大概需要30~40秒pp, 然后tg. 观察 llama-server 的log, 发现第一条prompt之后紧接着会更一个小 prompt, 这个 prompt 会把前面的 checkpoint (大概每8k个token一个 checkpoint)都冲掉, 这样下一次再接着聊, 前面的~20K prompt 还得重新 pp. 让 hermes 自己检查了一下, 第二个小 prompt 是它发的 title generation request. 为了避免这种情况, 可以禁止 title generation, 或者设一个辅助 aux model来生成 title (比如我让在线的 deepseek-v4-flash 干所有的 aux 工作);
- hermes stream mode 下有一个环境变量
HERMES_STREAM_READ_TIMEOUT, 它 控制收到第一个回复token的 timeout, 缺省为120s. 而在7900xtx 下pp花的时间大概是这样的(q5_1/q5_1 cache quant):
- 20k: 40s - 40k: 100s - 60k: 170s - 80k: 260s - 100k: 360s - 120k: 460s - 140k: 580s - 160k: 710s如果hermes有~50k的prompt, 赶上 llama-server cache checkpoint 刚好都是清空的情况下, pp没有完成之前, hermes就超时了.超时以后 hermes会中断当前请求,再发第二次(共三次). 如果第二次又刚好checkpoint被清空(我碰到过,具体原因还没搞明白),那么三次必然都会失败. 碰到这种情况, 可以把这个环境变量增大一下.
今天更新llama.cpp到最新, 还碰到了 llama-server RSS 超大给 oom-kill的情况, 以及 llama-server/hermes 进入死循环的情况. 具体还没有时间搞清楚. 目前我回到了 b9305.
目前的启动脚本如下:
#!/bin/bash LLAMA_SERVER=/home/bruin/github/llama.cpp/build-vulkan/bin/llama-server TOKEN_PER_CKPT=8192 # token per checkpoint, seems llama.cpp hardcoded NUM_CKPT=32 CTX_SIZE=$((TOKEN_PER_CKPT * NUM_CKPT)) ARGS=( --model /home/bruin/Qwen3.6-27B-Q4_K_M.gguf --mmproj /opt/gguf-models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf #--chat-template-file /opt/gguf-models/froggeric/Qwen-Fixed-Chat-Templates/chat_template.jinja --spec-type draft-mtp --spec-draft-n-max 2 # Max draft tokens --ctx-checkpoints ${NUM_CKPT} # 8k token per ckpt --ctx-size ${CTX_SIZE} # 262144 for 256k context #--swa-full # qwen3.6-27b does not support it --parallel 1 # Single slot --flash-attn on # Enable FlashAttention --n-gpu-layers 999 # All layers to GPU --cache-type-k q5_1 # Quantize KV cache keys --cache-type-v q5_1 # Quantize KV cache values #--fit off # --threads 16 # CPU threads helping tg --threads-batch 16 # CPU threads helping pg --batch-size 2048 # Batch size --ubatch-size 1024 # Micro‑batch size --cache-ram 0 # seems not working --reasoning auto # Auto reasoning --reasoning-format deepseek # Reasoning format --reasoning-budget 1024 # Reasoning budget --log-verbosity 4 # Log verbosity --host 0.0.0.0 --port 8000 # Listen on all interfaces, port 8000 --cont-batching # Continuous batching --no-warmup # Skip warmup --no-mmap # Don’t memory‑map model --mlock # Lock model in RAM --jinja # Jinja chat template --metrics # View metrics by accessing http://<ip:port>/metrics ) # print the cmdline echo "${LLAMA_SERVER}" for ((i=0; i<${#ARGS[@]}; i+=2)); do echo "${ARGS[i]} ${ARGS[i+1]}" done # run the cmd ${LLAMA_SERVER} "${ARGS[@]}" -
关于 hermes 接入 llama-server, 这两天有两个观察:
- hermes 新会话的第一条 prompt 大概是不到 20k token, 7900xtx大概需要30~40秒pp, 然后tg. 观察 llama-server 的log, 发现第一条prompt之后紧接着会更一个小 prompt, 这个 prompt 会把前面的 checkpoint (大概每8k个token一个 checkpoint)都冲掉, 这样下一次再接着聊, 前面的~20K prompt 还得重新 pp. 让 hermes 自己检查了一下, 第二个小 prompt 是它发的 title generation request. 为了避免这种情况, 可以禁止 title generation, 或者设一个辅助 aux model来生成 title (比如我让在线的 deepseek-v4-flash 干所有的 aux 工作);
- hermes stream mode 下有一个环境变量
HERMES_STREAM_READ_TIMEOUT, 它 控制收到第一个回复token的 timeout, 缺省为120s. 而在7900xtx 下pp花的时间大概是这样的(q5_1/q5_1 cache quant):
- 20k: 40s - 40k: 100s - 60k: 170s - 80k: 260s - 100k: 360s - 120k: 460s - 140k: 580s - 160k: 710s如果hermes有~50k的prompt, 赶上 llama-server cache checkpoint 刚好都是清空的情况下, pp没有完成之前, hermes就超时了.超时以后 hermes会中断当前请求,再发第二次(共三次). 如果第二次又刚好checkpoint被清空(我碰到过,具体原因还没搞明白),那么三次必然都会失败. 碰到这种情况, 可以把这个环境变量增大一下.
今天更新llama.cpp到最新, 还碰到了 llama-server RSS 超大给 oom-kill的情况, 以及 llama-server/hermes 进入死循环的情况. 具体还没有时间搞清楚. 目前我回到了 b9305.
目前的启动脚本如下:
#!/bin/bash LLAMA_SERVER=/home/bruin/github/llama.cpp/build-vulkan/bin/llama-server TOKEN_PER_CKPT=8192 # token per checkpoint, seems llama.cpp hardcoded NUM_CKPT=32 CTX_SIZE=$((TOKEN_PER_CKPT * NUM_CKPT)) ARGS=( --model /home/bruin/Qwen3.6-27B-Q4_K_M.gguf --mmproj /opt/gguf-models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf #--chat-template-file /opt/gguf-models/froggeric/Qwen-Fixed-Chat-Templates/chat_template.jinja --spec-type draft-mtp --spec-draft-n-max 2 # Max draft tokens --ctx-checkpoints ${NUM_CKPT} # 8k token per ckpt --ctx-size ${CTX_SIZE} # 262144 for 256k context #--swa-full # qwen3.6-27b does not support it --parallel 1 # Single slot --flash-attn on # Enable FlashAttention --n-gpu-layers 999 # All layers to GPU --cache-type-k q5_1 # Quantize KV cache keys --cache-type-v q5_1 # Quantize KV cache values #--fit off # --threads 16 # CPU threads helping tg --threads-batch 16 # CPU threads helping pg --batch-size 2048 # Batch size --ubatch-size 1024 # Micro‑batch size --cache-ram 0 # seems not working --reasoning auto # Auto reasoning --reasoning-format deepseek # Reasoning format --reasoning-budget 1024 # Reasoning budget --log-verbosity 4 # Log verbosity --host 0.0.0.0 --port 8000 # Listen on all interfaces, port 8000 --cont-batching # Continuous batching --no-warmup # Skip warmup --no-mmap # Don’t memory‑map model --mlock # Lock model in RAM --jinja # Jinja chat template --metrics # View metrics by accessing http://<ip:port>/metrics ) # print the cmdline echo "${LLAMA_SERVER}" for ((i=0; i<${#ARGS[@]}; i+=2)); do echo "${ARGS[i]} ${ARGS[i+1]}" done # run the cmd ${LLAMA_SERVER} "${ARGS[@]}"@laobenxiong 你这套参数接近目前的极限了,还有一个SMITHY可以试试,--cache-ram多轮对话好像有用。
-
@laobenxiong 你这套参数接近目前的极限了,还有一个SMITHY可以试试,--cache-ram多轮对话好像有用。
@John-Ato smithy是什么? 没搜到
-
最近在用这个llama.cp +qwen 35 a3b。 m5 max 蛮好的 百来token
