Comment by xlayn

Comment by xlayn

Hey, just in case.. I have been maintaining this fork of llama.cpp, I run also 7900XTX, 3 in my case... and I want to know if you see improved speed... notice the ngram mod where instead of loosing the ngrams they are stored in disk, I have a 31M slot and so far it's being working really nice for me

I'm seing around 2800t/s on pp at the very beginning of the conversation, and then going to around 800 by the 256k context size...

I've also update the value in the gguf to have instead of context size 256k be triple and have use it all the way to around 600k context size

use the cache eviction branch https://github.com/alainnothere/llama.cpp/tree/disk-cache-ev...

llama-server -m /path_to_model/Qwen3.8-27B-UD-IQ4XS_u.gguf -np 2 --host 0.0.0.0 --port 8089 --kv-unified --temp 0.2 --top-k 20 --min-p 0.0 --cache-disk-path /path_to_store_conversations_to_disk --cache-ram 100 --cpu-range 0-23 --reasoning-budget 4096 --reasoning-budget-message "... I am thinking for too long -- let me gather more info about the task." --no-cache-idle-slots --spec-draft-auto --spec-draft-sampling probabilistic --spec-draft-n-min 1 --spec-draft-p-min 0 --spec-ngram-auto --spec-ngram-mod-size 35544320 --spec-ngram-mod-cache /path_to_storage_of_ngram_mod_this_is_specific_per_model/ngram-mod-qwen3.8-27b.bin --fit on --device Vulkan3,Vulkan2,Vulkan1 --fit-target 1024,2560,2560 --ctx-checkpoints 25 --checkpoint-spill-max 25 --cache-type-k f16 --cache-type-v f16 -ctkd q8_0 -ctvd q8_0 -devd Vulkan3 --spec-type draft-mtp,ngram-mod --spec-draft-n-max 7 --spec-ngram-mod-n-min 7 --spec-ngram-draft-n-max 31 --spec-rs-depth 24 --fit-verify on

Use this if you have 3 cards, if you have less maybe just remove it -b 8192

In case you want to know how many tokens you have use, --token-ledger /path_to_where_you_want_to_store_stats/token_ledger.json

here is how logs look like I saw a 350t/s on tg but was because you hit ngram mod tables over things the model has already seen

457.04.985.808 I srv print_line: token ledger: requests = 27077, prompt = 3928887769 (cached = 3874707808, processed = 47151296), generated = 22126308 (reasoning = 15129037), t> 457.04.985.913 I slot rq_write_fin: id 0 | task -1 | request timeline: task 77431, ttft 1580.0 ms (http 20.4, tokenize 65.1, queue 247.1, select 0.0, restore 2.7, prompt 1244.3),> 457.05.202.640 I srv handle_compl: task created: conversation_id=64d89c6d7ff76923 (tokens 0..18489, system+first_user) 457.05.202.757 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.994 (> 0.100 thold), f_keep = 1.000 457.05.206.552 I slot launch_slot_: id 0 | task 77458 | processing task, is_child = 0 457.05.207.188 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81032..81036 of 81552 457.05.246.405 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81036..81548 of 81552 457.05.310.149 I srv lazify_check: checkpoint: converted pos_min=80943 to disk-backed ('cp_64d89c6d7ff76923_80943.bin', 149.6 MiB released) 457.05.872.722 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81548..81552 of 81552 457.05.943.921 I srv lazify_check: checkpoint: converted pos_min=81035 to disk-backed ('cp_64d89c6d7ff76923_81035.bin', 149.6 MiB released) 457.09.000.990 I slot print_timing: id 0 | task 77458 | n_gen = 340, tg = 112.88 t/s, tg_3s = 113.20 t/s 457.11.161.223 I slot print_timing: id 0 | task 77458 | prompt eval time = 790.96 ms / 520 tokens ( 1.52 ms per token, 657.43 tokens per second) 457.11.161.226 I slot print_timing: id 0 | task 77458 | eval time = 5163.27 ms / 642 tokens ( 8.06 ms per token, 124.15 tokens per second) 457.11.161.226 I slot print_timing: id 0 | task 77458 | total time = 5954.23 ms / 1162 tokens 457.11.161.227 I slot print_timing: id 0 | task 77458 | graphs reused = 72319 (pp skipped 525) 457.11.161.228 I slot print_timing: id 0 | task 77458 | graphs prebuilt = 7295 (hit 6975, miss 320) 457.11.161.229 I slot print_timing: id 0 | task 77458 | early prebuilt = 28686 (fit 44246, hit 26411, miss 2275, aborted 0, 96584.2 ms host) 457.11.161.238 I slot print_timing: id 0 | task 77458 | draft acceptance = 0.83204 ( 535 accepted / 643 generated), mean len = 6.00 457.11.161.243 I slot print_timing: id 0 | task 77458 | draft auto = n=6 | 1:55.4 t/s [2.00 tok / 36.1 ms] (4) 2:67.7 t/s [2.50 tok / 37.0 ms] (4) 3:101.0 t/s [4.00 tok / 39.6 ms> 457.11.161.251 I slot print_timing: id 0 | task 77458 | draft by impl: draft-mtp | steps 101 | built 5.5 | offered 5.5 | accepted 4.9 (0.90) | full 80 | step 45.2 ms | 131.5 t/s > 457.11.161.257 I slot print_timing: id 0 | task 77458 | draft by impl: ngram-mod | steps 6 | built 64.0 | offered 14.3 | accepted 6.0 (0.42) | full 1 | step 90.8 ms | 77.1 t/s | > 457.11.161.258 I slot print_timing: id 0 | task 77458 | ngram policy: auto on | cap now 12 | cap min 7 median 12 max 31 over 6 chains | cap bound 6 | whole ema 0.17 | resume 0 | > 457.11.161.259 I slot print_timing: id 0 | task 77458 | spec rollback: prebatch restores 1 | host checkpoints 0 | checkpoint restores 0 457.11.161.260 I slot print_timing: id 0 | task 77458 | chain truth: ngram | chains 6 | censored 0 | would land mean 7.0 | full 0 | hist 0:0 1-4:1 5-8:3 9-16:2 17-32:0 33-48:0 49> 457.11.161.261 I slot print_timing: id 0 | task 77458 | ngram gate: held 38 | agree 5 (would land 0:0 1-4:1 5-16:4 17+:0, pending 0) | disagree 33 (would land 0:33 1-4:0 5-16:0 1> 457.11.161.261 I slot print_timing: id 0 | task 77458 | ext shadow: mtp+ngram | attempts 101 | hits 85 | tested 68 | landed mean 0.4 (over tested) | hist 0:65 1-4:1 5-8:0 9-16:2 > 457.11.161.262 I slot print_timing: id 0 | task 77458 | alt table: n=48 | attempts 101 | hits 45 | would land mean 0.0 | hist 0:45 1-4:0 5-8:0 9-16:0 17-32:0 33-48:0 49-64:0 65+:> 457.11.161.263 I slot print_timing: id 0 | task 77458 | mtp ahead: tried 53 (0 after a p_min stop) | verify full 49 | bonus matched 48 (0 after p_min) | too short 0 | served 48 s> 457.11.161.265 I slot print_timing: id 0 | task 77458 | mtp confidence: stop cap:101(acc 4.9) pmin:0(acc 0.0) other:0 | p bins landed/tested: <0.5 1/3 0.5 0/1 0.6 1/1 0.7 0/0 0.8> 457.11.163.234 I slot release: id 0 | task 77458 | stop processing: n_tokens = 82195, truncated = 0 457.11.163.426 I srv print_line: token ledger: requests = 27078, prompt = 3928969321 (cached = 3874788840, processed = 47151816), generated = 22126950 (reasoning = 15129071), t> 457.11.163.526 I slot rq_write_fin: id 0 | task -1 | request timeline: task 77458, ttft 884.3 ms (http 21.2, tokenize 68.3, queue 0.1, select 0.0, restore 3.3, prompt 791.0), gen> 457.11.396.231 I srv handle_compl: task created: conversation_id=64d89c6d7ff76923 (tokens 0..18489, system+first_user)

Replies xlayn · 15h
Open on HN
Loading the discussion…

Domain filters

Stories from these domains are hidden from every list. Subdomains match too: blocking substack.com also hides danluu.substack.com. Type a domain, or paste a link to block its site, and press Enter.

New collection

Delete this collection?

About YAVCHN

YAVCHN is a reader for Hacker News and Lobsters, with articles and discussions in separate windows or Classic pages.

Created by Paul Parks and built with PUDL.

YAVCHN source code on GitHub

Privacy policy · Terms of use

Help

Keyboard

j / k
Move down and up the story list. The arrow keys scroll whatever has focus.
Enter
Read the marked story in the article reader.
[ / ]
Read the story above or below in the list, in the same article reader. Back returns to the story you came from.
p
Pin or unpin the marked story, which keeps it in Pinned.
n / N
Move to the next or previous top-level comment in the window in front.
c
Collapse or expand that comment.
f
Hide or show the story list.
Esc
Close a menu or this help.
Access key m
Go to the menu bar. Most browsers take it with Alt on Windows and Linux, and Safari with Control and Option.
?
Show this help.

Windows

Each story opens in a window holding its article above its discussion; drag the bar between them to share the room differently. A window can be moved by its title bar, resized from any edge, snapped to a half or a corner by dragging it there, maximised, or minimised to the bar at the foot of the page. Use Window > New reader window to open an empty reader, or Story > Open in new reader window to open another reader for the current article. Docked readers keep their articles when you select another story from the sidebar. Minimized readers can be restored and reused for their site. A window's Previous and Next buttons read up and down the list in the same window.

A link in a comment or an article to another Hacker News or Lobsters thread opens that thread in a window too. A link to a single HN comment opens the comment above its replies.

While a story's window is in front, the Story and Discussion menus in the menu bar hold its commands: pinning, Previous and Next story, sorting, collapsing every thread, jumping to the first new comment. Each window also remembers where you were in its article and discussion, so a reload, or Back to a story that Next took you past, finds your place again. Closing a window forgets it.

The whole arrangement lives in the address, so a bookmark or a shared link brings it back, and Back undoes the last change. Moving between Hacker News, Lobsters, their lists, Pinned and Find changes only the list, and leaves the windows open.

The list

The pin at the start of a row keeps the story in Pinned, and the cross at its end hides it. Pinned can be narrowed by words in the title, site or author, by source, and to the stories you haven't opened yet, and ordered by when you pinned them, by points or by comments; the filters are part of the address, so a filtered view can be bookmarked. Scroll past the end of the list to load more. Domain filters, in the View menu, hide every story from a site.

Collections are named lists of stories. Story > Add to collection files the story in front into one or more of them, and the Collections feed shows them all or one at a time; the menu that chooses collections also creates, renames, and deletes them. A note is your own text on a story. Choose Add note in a story's toolbar to write one; it saves as you type. Rows with a note carry the note mark, and the Notes feed lists every noted story and searches the text of your notes.

Browsing view

View > Windowed and View > Classic select the browsing view and save your default in this browser. Window view reuses a reader for each feed. Classic view opens stories and applets as pages. Open as a page is a one-off action that does not change your saved default. Use the Windowed selector to return an article to a window. Direct page links always open as pages.

Applets

The Applets menu in the menu bar holds three tools, each a window of its own. Replies to me takes your Hacker News user name and lists the replies to your last thirty comments and stories, checking again every three minutes while it is open, and marking what is new since you last marked them read. Look up a user opens a profile on Hacker News or Lobsters, with their submissions and recent comments, as a commenter's name in any discussion does; the bar at the top of a profile looks up someone else in the same window, and Back returns to the one before. Who is hiring? filters the posts of HN's monthly hiring threads by the words you type.

They read only what the sites publish to everyone, so none of them asks for a login, and your user name stays in this browser anonymously and is included in retained applet state when signed in.

Find

Find takes any link and lists every time it was submitted to Hacker News and Lobsters, so you can read each discussion of it.

About

YAVCHN never sees your Hacker News or Lobsters login. The discussion is fetched from each site's public API; to vote or reply, follow the link above the discussion, or the arrow beside a comment, to the source's own site. Without a YAVCHN account, your data stays in this browser. When signed in, pins, collections, notes, blocked domains, and retained reading state are stored with your account and synchronized across devices. Hidden stories and layout stay in this browser. The privacy policy has the details.

Open source: github.com/paulmooreparks/yavchn. Built with PUDL.