This website requires JavaScript.
b8d3c084c5
BACKLOG P: AD sign-in to Slurm user, tried end to end
Ilker Manap and Claude Opus 5.5
2026-11-14 23:00:28 +01:00
57f0fd2ed0
Slurm's jobs in containers: pyxis and enroot, installed by the agents
Ilker Manap and Claude Opus 5.5
2026-11-14 21:31:20 +01:00
c3a75dcf3f
A node being lent says when a replica with nowhere to go is stopped
Ilker Manap and Claude Opus 5.5
2026-11-14 20:45:57 +01:00
4b4306bb6a
Slurm users reach a node only through a job of theirs
Ilker Manap and Claude Opus 5.5
2026-11-14 20:30:49 +01:00
a00c92648e
Status says when the cluster's alerts go nowhere
Ilker Manap and Claude Opus 5.5
2026-11-12 22:49:37 +01:00
2c32a58a06
Suspending a person says their keys are refused, which they are
Ilker Manap and Claude Opus 5.5
2026-11-12 22:38:54 +01:00
d8ce573bdb
Slurm's shared files and users: one path everywhere, accounts as Linux users
Ilker Manap and Claude Opus 5.5
2026-11-12 21:32:56 +01:00
b9d20e7bba
Licence: FSL-1.1 with an Apache 2.0 future licence
Ilker Manap and Claude Fable 5.1
2026-11-12 19:55:24 +01:00
0d78f9bb34
A machine given by address is reached through its ssh config Host
Ilker Manap and Claude Opus 5.5
2026-11-11 21:31:22 +01:00
5ac6bef463
BACKLOG P: the fresh install tried, and what it left open
Ilker Manap and Claude Opus 5.5
2026-11-12 10:42:22 +01:00
557a5a1ef3
Found on a fresh three-machine install: what the screens said wrong
Ilker Manap and Claude Opus 5.5
2026-11-12 10:38:17 +01:00
21828fefce
BACKLOG P: slurmdbd done; P6, Slurm's jobs in llmgw's accounting
Ilker Manap and Claude Opus 5.5
2026-11-12 08:21:40 +01:00
748f777a56
Slurm's accounting: slurmdbd and a MariaDB on the controller
Ilker Manap and Claude Opus 5.5
2026-11-12 07:54:13 +01:00
02774769a7
Slurm: why priority/basic, said rightly
Ilker Manap and Claude Opus 5.5
2026-11-11 23:21:24 +01:00
bab0766ee1
Slurm: jobs start on submission, and a lent node stays open across restarts
Ilker Manap and Claude Opus 5.5
2026-11-11 23:18:22 +01:00
5db920bf2c
Slurm: a job that asks for no memory gets a share, not the whole node
Ilker Manap and Claude Opus 5.5
2026-11-11 23:11:24 +01:00
356f7e0ced
Slurm in the console: each node, and lend or take back from its row
Ilker Manap and Claude Opus 5.5
2026-11-11 22:53:04 +01:00
17287ee8b0
BACKLOG P: failover while lent tried, and fixed
Ilker Manap and Claude Opus 5.5
2026-11-11 22:45:05 +01:00
0774e901c0
Slurm: a new leader with nothing kept waits for the nodes a while
Ilker Manap and Claude Opus 5.5
2026-11-11 22:37:52 +01:00
141a1f46c5
Slurm: a new leader keeps the nodes it has not heard from yet
Ilker Manap and Claude Opus 5.5
2026-11-11 22:33:53 +01:00
3387f97f6a
Slurm's ports for the cluster only, and srun in a range of its own
Ilker Manap and Claude Opus 5.5
2026-11-11 22:10:25 +01:00
914985bf9f
Slurm in the first install: server init --slurm, and slurm.all_nodes
Ilker Manap and Claude Opus 5.5
2026-11-11 14:50:35 +01:00
b15944aa1f
A package install that is no engine's clears its progress; slurmd retried
Ilker Manap and Claude Opus 5.5
2026-11-11 13:23:53 +01:00
dcf4e256e6
BACKLOG P: tried on the PoC
Ilker Manap and Claude Opus 5.5
2026-11-11 13:20:32 +01:00
849760c67d
Slurm: no second restart; backfill every 10 s, MailProg set
Ilker Manap and Claude Opus 5.5
2026-11-11 13:18:35 +01:00
087552136f
Slurm: slurmctld started again after a reconfigure; no accounting gathered
Ilker Manap and Claude Opus 5.5
2026-11-11 13:16:31 +01:00
d2c27e2c76
Slurm: slurmd started as its unit expects, and slurmd's own stepd not a job
Ilker Manap and Claude Opus 5.5
2026-11-11 13:11:41 +01:00
8a5a04ed77
Slurm: the version from the daemon, not scontrol
Ilker Manap and Claude Opus 5.5
2026-11-11 13:06:42 +01:00
410fe545ce
Slurm beside llmgw, P3 and P4: the leader hands nodes over; ctl slurm
Ilker Manap and Claude Opus 5.5
2026-11-11 12:33:14 +01:00
35de47444c
Slurm beside llmgw, P2: the agent keeps its node's part
Ilker Manap and Claude Opus 5.5
2026-11-11 12:25:41 +01:00
d439d507d7
Slurm beside llmgw, P1: its settings and the files a controller runs on
Ilker Manap and Claude Opus 5.5
2026-11-11 12:20:31 +01:00
89767c8be1
BACKLOG P: training beyond one card, in llmgw or Slurm beside it
Ilker Manap and Claude Opus 5.5
2026-11-11 12:12:07 +01:00
d864438272
A model that failed to start is placed elsewhere next
Ilker Manap and Claude Opus 5.5
2026-11-10 09:22:11 +01:00
6320dd5b13
A new stream is told the node's files again
Ilker Manap and Claude Opus 5.5
2026-11-08 22:14:52 +01:00
5732a5ebda
The serve forms start at the longest context that fits
Ilker Manap and Claude Opus 5.5
2026-11-08 14:38:41 +01:00
561a536477
The weight sweep names a set as its fetch did
Ilker Manap and Claude Opus 5.5
2026-11-08 14:01:27 +01:00
45fe7daaa0
The admin API document regenerated for CollectResult's left counts
Ilker Manap and Claude Opus 5.5
2026-11-08 11:38:08 +01:00
102ca9556f
An organization's admin, and forms only for those who may send them
Ilker Manap and Claude Opus 5.5
2026-11-08 11:38:06 +01:00
2cb1afda55
The introduction's comparison on one slide
Ilker Manap and Claude Opus 5.5
2026-11-04 17:06:49 +01:00
b6478e24b7
An introduction for decision makers, linked from the website
Ilker Manap and Claude Opus 5.5
2026-11-04 12:21:04 +01:00
cba4897987
Deleting unused model files says what went, and sets go whole
Ilker Manap and Claude Opus 5.5
2026-11-03 10:29:16 +01:00
3bb52742a4
v0.16.3: changelog and README
Ilker Manap and Claude Opus 5.5
2026-11-03 09:43:40 +01:00
1a3f385be4
Every setting of a served model from the mirror page, and its replicas per profile
Ilker Manap and Claude Opus 5.5
2026-11-03 09:40:06 +01:00
e23732fbbb
Replicas counted per profile: seven on cards and four on CPUs under one name
Ilker Manap and Claude Opus 5.5
2026-11-03 09:21:50 +01:00
6403e61f4b
How fast a model answers on the CPU, before it is served
Ilker Manap and Claude Opus 5.5
2026-11-03 08:52:15 +01:00
220110ed0d
v0.16.2: changelog and README
Ilker Manap and Claude Opus 5.5
2026-11-02 08:32:34 +01:00
a88624747e
A set wanted by several replicas on one node is fetched once
Ilker Manap and Claude Opus 5.5
2026-11-02 08:29:25 +01:00
540b518fde
Rate limits counted across the control planes with router.active_active
Ilker Manap and Claude Opus 5.5
2026-11-01 17:30:08 +01:00
95fb5eeb6e
BACKLOG: O2 tried on the PoC, two routers carry 1.5x the leader alone
Ilker Manap and Claude Opus 5.5
2026-11-01 16:51:54 +01:00
8b7eefef16
Standbys that answer the client API themselves, behind router.active_active
Ilker Manap and Claude Opus 5.5
2026-11-01 13:47:47 +01:00
f664592948
BACKLOG: N8 half done, ctl model context takes Strata and the longest answer alone
Ilker Manap and Claude Opus 5.5
2026-11-01 09:51:26 +01:00
31f440ab6e
A served model's longest answer changed in place, and Strata's context
Ilker Manap and Claude Opus 5.5
2026-11-01 00:15:38 +01:00
f4b34dbf11
FreeToken, an engine for models larger than the card
Ilker Manap and Claude Opus 5.5
2026-10-31 22:10:05 +01:00
c866b3a8ee
How much of each GPU works, beside whether anything does
Ilker Manap and Claude Opus 5.5
2026-10-31 20:11:26 +01:00
7bfe70c8df
BACKLOG: the router profiled, the write per token removed, what is left
Ilker Manap and Claude Opus 5.5
2026-10-31 15:50:15 +01:00
5550fefaad
Stream events already in hand go to the client in one write, not one each
Ilker Manap and Claude Opus 5.5
2026-10-31 15:30:19 +01:00
808c4724d2
BACKLOG: O1 measured, the router's ceiling on the PoC
Ilker Manap and Claude Opus 5.5
2026-10-31 10:45:59 +01:00
9ff49cad8e
Requests missing the cache for the same key at once share one read and one allowance
Ilker Manap and Claude Opus 5.5
2026-10-31 09:30:18 +01:00
524f029603
External providers left out of admission's saturation; the scale-out section in the backlog
Ilker Manap and Claude Opus 5.5
2026-10-30 19:13:15 +01:00
be54f1617b
A mirror download cut off carries on from where it got to, and is tried again
Ilker Manap and Claude Opus 5.5
2026-10-30 17:47:32 +01:00
812eb8aac1
Regenerate the admin API document for the fetch phase
Ilker Manap and Claude Opus 5.5
2026-10-30 15:07:16 +01:00
76746715ec
A set's files checked four at once, the check said as it goes and not taken for a stuck download
Ilker Manap and Claude Opus 5.5
2026-10-30 14:58:44 +01:00
d5dcf78836
A set's download reported as the whole set's, not the part being fetched
Ilker Manap and Claude Opus 5.5
2026-10-30 14:48:45 +01:00
a5fefbdaaa
v0.16.1: changelog and README
Ilker Manap and Claude Opus 5.5
2026-10-30 13:14:25 +01:00
2eb96f4872
The plan says why Strata's DeepSeek does not run here
Ilker Manap and Claude Opus 5.5
2026-10-30 13:09:09 +01:00
f41c96d30f
Strata's DeepSeek not offered on cards its engine was not built for, and refused there by the agent
Ilker Manap and Claude Opus 5.5
2026-10-30 13:00:21 +01:00
104b60fbf8
A repository's files read from every directory, not only its top
Ilker Manap and Claude Opus 5.5
2026-10-30 12:16:11 +01:00
845de8eef4
The mirror page's Strata tab; Strata's ready-made packs served with Strata fetching them, strata-ple named
Ilker Manap and Claude Opus 5.5
2026-10-30 12:06:46 +01:00
4ee9ba54ef
v0.16.0: changelog and README
Ilker Manap and Claude Opus 5.5
2026-10-30 11:21:32 +01:00
12464ce0f6
Cluster upgrades: nodes told together with progress said, done when back on the version, the download kept past a broken stream
Ilker Manap and Claude Opus 5.5
2026-10-30 11:14:02 +01:00
d6f8f433f4
CHANGELOG: one list
Ilker Manap
2026-10-30 10:38:13 +01:00
7465c24830
Strata's families and sizes known: checked when a model is added, its memory reserved from them
Ilker Manap and Claude Opus 5.5
2026-10-30 10:17:42 +01:00
91f668fe7e
Strata named by hand is reported, its engine stopped with it, its memory required, Pascal allowed
Ilker Manap and Claude Opus 5.5
2026-10-30 09:49:16 +01:00
73091d7c97
A cached file checked as another is refused and left, not deleted for the next model
Ilker Manap and Claude Opus 5.5
2026-10-30 08:17:43 +01:00
b8ed368e31
Weights from Hugging Face pinned to a commit and checked against the sha256 it published
Ilker Manap and Claude Opus 5.5
2026-10-30 07:33:15 +01:00
29d7f54cfd
A path escaped as it came reaches the leader through a standby; F5 in the backlog
Ilker Manap and Claude Opus 5.5
2026-10-29 22:54:27 +01:00
cacaa35c3f
The gateway embeds the documents llmgw-kb sends as text
Ilker Manap and Claude Opus 5.5
2026-10-29 21:52:33 +01:00
01ad8c9207
v0.15.3: changelog and README
Ilker Manap and Claude Opus 5.5
2026-10-29 20:54:53 +01:00
59187cdae5
A model can be stopped and kept, or removed with its files from the mirror and the nodes
Ilker Manap and Claude Opus 5.5
2026-10-29 20:45:48 +01:00
5726ceb84e
BACKLOG: J4 done, tried on the PoC with the first mirror stopped
Ilker Manap and Claude Opus 5.5
2026-10-29 20:01:47 +01:00
435ffe1171
A file removed outside the running mirror is no longer reported, and its removal reaches the peers
Ilker Manap and Claude Opus 5.5
2026-10-29 19:25:00 +01:00
785e06ecbf
A mirror with peers records the checksums of files that had none, so they are copied too
Ilker Manap and Claude Opus 5.5
2026-10-29 18:36:59 +01:00
0033a530c7
mirror add-peer on a mirror machine of its own says what to run on a control plane
Ilker Manap and Claude Opus 5.5
2026-10-29 18:30:30 +01:00
8cd376615b
v0.15.2: changelog and README
Ilker Manap and Claude Opus 5.5
2026-10-29 16:46:15 +01:00
e1cff0d4e3
A served model can be moved to more cards, or fewer
Ilker Manap and Claude Opus 5.5
2026-10-29 16:40:49 +01:00
51aeadae3e
v0.15.1: changelog and README
Ilker Manap and Claude Opus 5.5
2026-10-29 15:59:12 +01:00
986981ac81
A served model's context can be changed, engine and memory with it
Ilker Manap and Claude Opus 5.5
2026-10-29 15:29:03 +01:00
1908bfb4b2
Strata on cuda01 end to end: its engine is the shard's, its metrics read
Ilker Manap and Claude Opus 5.5
2026-10-29 15:48:15 +01:00
c0d5194ea0
Strata runs as an engine: a 125B MoE on one card
Ilker Manap and Claude Opus 5.5
2026-10-29 15:09:13 +01:00
eb9c448d68
BACKLOG: J4 tried on two machines
Ilker Manap and Claude Opus 5.5
2026-10-29 15:04:11 +01:00
1c16f2337d
A second mirror: llmgw mirror add-peer, and on a store machine at setup
Ilker Manap and Claude Opus 5.5
2026-10-29 13:43:27 +01:00
8a76a53afb
The control plane uses whichever mirror answers
Ilker Manap and Claude Opus 5.5
2026-10-23 14:45:42 +02:00
5bc71b3d0a
Mirrors kept in step: llmgw mirror serve --peer
Ilker Manap and Claude Opus 5.5
2026-10-23 09:59:59 +02:00
c09d5ad4ee
Outside services read answers redacted and whole; a card's own lanes
Ilker Manap and Claude Opus 5.5
2026-10-29 13:22:39 +01:00
1948a787ae
Fixes from the second review: grants, flags, updates, setup
Ilker Manap and Claude Opus 5.5
2026-10-29 12:31:07 +01:00
ecf5545a7a
VS Code's sign-in keeps its place; recordings checked for secrets
Ilker Manap and Claude Opus 5.5
2026-10-29 11:40:19 +01:00
258591b3f2
What an engine downloads by a Hugging Face id is asked about too
Ilker Manap and Claude Opus 5.5
2026-10-29 11:06:11 +01:00
5eac62d556
Guardrail services merged the stricter way; the request log keeps no restored value
Ilker Manap and Claude Opus 5.5
2026-10-29 11:03:57 +01:00
f94d902cc3
MCP servers: fields that can be emptied, addresses without passwords
Ilker Manap and Claude Opus 5.5
2026-10-29 10:03:27 +01:00
721705b29d
A card in the wrong PCIe slot is said so
Ilker Manap and Claude Opus 5.5
2026-10-29 08:49:16 +01:00
