-
025ed9fd82
feat: add prefill progress bar for long prompts (#1181)
2026-02-18 19:18:25 -08:00
-
3e9eb93f82
exo theme
Sami Khan
2026-02-19 04:21:58 +05:00
-
19bc09550d
Add status=downloaded filter for model endpoint (#1539)
gh-screenshot-assets
rltakashigeandGitHub
2026-02-18 22:34:11 +00:00
-
7cadca4f27
Try multiple endpoints for internet connectivity check (#1516)
2026-02-18 14:10:07 -08:00
-
24e99ce197
Cleanup mistakes (#1537)
rltakashigeandGitHub
2026-02-18 22:05:26 +00:00
-
315992549b
fix: unblock MpReceiver.close() to prevent shutdown hang (#1511)
2026-02-18 13:59:02 -08:00
-
ce5a65d3b9
Add MiniMax M2.5 model cards (#1514)
2026-02-18 13:11:13 -08:00
-
c2f2111b88
Fix tool calling (#1529)
rltakashigeandGitHub
2026-02-18 20:29:18 +00:00
-
99b0db0634
chore: apply nix fmt formatting to svelte files
feat/dashboard-light-mode
Alex CheemaandClaude Sonnet 4.6
2026-02-18 11:46:00 -08:00
-
e3d91dc08a
feat: add light/dark mode toggle to dashboard
Alex CheemaandClaude Sonnet 4.6
2026-02-18 11:04:09 -08:00
-
-
-
6c322ebb72
feat: only show thinking toggle for models that support it (#1497)
2026-02-18 09:05:00 -08:00
-
523eff541e
aaa
aiohttp
Evan
2026-01-16 13:45:39 +00:00
-
-
-
2ebe6216b4
feat: add explicit --offline mode for air-gapped clusters (#1525)
vskiwiandGitHub
2026-02-18 19:18:09 +03:00
-
ea5fc8125f
Address comment
leo/add-glm-5
Ryuichi Leo Takashige
2026-02-18 16:06:30 +00:00
-
f54c80b121
Ciaran/image edit api (#1500)
ciaranborandGitHub
2026-02-18 16:05:39 +00:00
-
48b8f86395
Add support for GLM 5 (#1526)
rltakashigeandGitHub
2026-02-18 14:04:06 +00:00
-
14da9c3f59
Merge branch 'main' into leo/add-glm-5
rltakashigeandRyuichi Leo Takashige
2026-02-18 13:20:31 +00:00
-
-
-
-
5cbd6377a2
prioritize official model cards over custom model cards
Evan
2026-02-18 12:48:42 +00:00
-
242067e8eb
Update mlx to the nuclear solution
Ryuichi Leo Takashige
2026-02-18 13:16:30 +00:00
-
03267171ea
move mx_barrier to right before prefill
Ryuichi Leo Takashige
2026-02-18 12:39:40 +00:00
-
080995a409
try a longer join timeout to prevent crashes on instance deletion
Ryuichi Leo Takashige
2026-02-18 12:34:05 +00:00
-
0d92d001e6
wrong model card
Ryuichi Leo Takashige
2026-02-18 11:58:59 +00:00
-
0287ad71a1
format and lint
Ryuichi Leo Takashige
2026-02-18 11:51:30 +00:00
-
82ac16cc0f
format and lint
Ryuichi Leo Takashige
2026-02-18 11:48:45 +00:00
-
d78dd516f3
bad diffs 2
Ryuichi Leo Takashige
2026-02-18 11:47:55 +00:00
-
5c154d091d
bad diffs
Ryuichi Leo Takashige
2026-02-18 11:45:41 +00:00
-
e4e0517627
Merge branch 'main' into leo/add-glm-5
rltakashigeandGitHub
2026-02-18 11:44:43 +00:00
-
-
-
-
b28c1d9e92
Merge branch 'main' into leo/add-glm-5
Ryuichi Leo Takashige
2026-02-18 11:43:38 +00:00
-
-
8f01523ddb
remove dead code (#1496)
Evan QuineyandGitHub
2026-02-18 11:43:27 +00:00
-
-
-
5ddc2574ec
smore
splitting-rust-rewrite-3
Evan
2026-02-17 14:13:35 +00:00
-
d5e56999de
remove dead code
Evan
2026-02-17 11:54:07 +00:00
-
-
-
3addeadea8
Update mlx-lm to 0.30.7 (#1520)
2026-02-18 03:14:23 -08:00
-
ab622f79c3
EXO iOS app
Sami Khan
2026-02-18 06:40:07 +05:00
-
44845be117
fix instance type mismatch in /instance/previews endpoint
fix/instance-preview-type-mismatch
Alex CheemaandClaude Opus 4.6
2026-02-17 10:09:17 -08:00
-
-
-
ce45af58e3
fix: --no-downloads no longer blocks model loading (#1510)
fix/no-downloads-model-loading
Alex CheemaandClaude Opus 4.6
2026-02-17 14:33:47 -08:00
-
-
-
f2be929211
Leo/address rdma gpu locks 2 (#1515)
2026-02-17 22:00:52 +00:00
-
f8748eade1
Add MLX compute device backend selection (cpu/gpu/auto)
support-mlx-device-backends
Alex CheemaandClaude Opus 4.6
2026-02-17 10:52:11 -08:00
-
-
-
d923f77629
chore: retrigger CI
alexcheema/rdma-health-detection
Alex CheemaandClaude Opus 4.6
2026-02-17 10:44:43 -08:00
-
c4d24a24e4
Handle MessageTooLarge error in router to prevent node crash
handle-message-too-large
Alex CheemaandClaude Opus 4.6
2026-02-17 10:34:22 -08:00
-
-
-
9c4ceef575
fix: pass tasks arg to get_transition_events in reconciler
Alex CheemaandClaude Opus 4.6
2026-02-17 10:26:13 -08:00
-
0ad9968eb7
chore: retrigger DMG build
Alex Cheema
2026-02-16 07:15:38 -08:00
-
7d2245d197
chore: retrigger build
Alex Cheema
2026-02-16 06:58:33 -08:00
-
08c94bc283
fix: apply nix fmt to node_timeout.py
Alex CheemaandClaude Opus 4.6
2026-02-16 06:30:53 -08:00
-
7913e1a03f
fix: only show RDMA cable warning for directly-connected Thunderbolt pairs
Alex CheemaandClaude Opus 4.6
2026-02-16 05:43:37 -08:00
-
539be54edb
fix: evict stale peer when node restarts with new ID (merge blocker)
Alex CheemaandClaude Opus 4.6
2026-02-15 19:20:09 -08:00
-
483c52f3b1
fix: clean up orphaned runners when node times out
Alex CheemaandClaude Opus 4.6
2026-02-15 19:14:00 -08:00
-
fbf9007591
fix: clean up stale runners from state when instance is deleted
Alex CheemaandClaude Opus 4.6
2026-02-15 18:39:52 -08:00
-
1b68eb5be5
fix: shard redistribution blocked when pinned node dies (BUG-B)
Alex CheemaandClaude Opus 4.6
2026-02-15 18:32:04 -08:00
-
afaf778ba6
fix: RDMA warning uses all edges to find peers, not just RDMA-tagged ones
Alex CheemaandClaude Opus 4.6
2026-02-15 17:21:03 -08:00
-
59e44df71b
fix: matchesSelectedRuntime tautology and unifiedDisplayItems derived pattern
Alex CheemaandClaude Opus 4.6
2026-02-15 16:47:57 -08:00
-
cd7799cd9a
fix: RDMA dashboard warning shows 'between A and B' for all unhealthy pairs
Alex CheemaandClaude Opus 4.6
2026-02-15 16:44:33 -08:00
-
d9f82034b5
Add proactive RDMA device health detection with dashboard warning
Alex CheemaandClaude Opus 4.6
2026-02-12 16:15:57 -08:00
-
5ea3778674
fix: use force=True for multiprocessing set_start_method
Alex CheemaandClaude Opus 4.6
2026-02-14 07:02:32 -08:00
-
a9b835d481
fix: eliminate command/reconciler interleaving race in meta-instance
Alex CheemaandClaude Opus 4.6
2026-02-13 12:50:57 -08:00
-
7feb5045e2
test: add 25 edge-case tests for MetaInstance lifecycle
Alex CheemaandClaude Opus 4.6
2026-02-13 12:17:41 -08:00
-
64beeb58d8
Fix JACCL SideChannel bytes serialization for JSON round-trip
Alex CheemaandClaude Opus 4.6
2026-02-12 14:23:44 -08:00
-
3238157834
Use named pipes (FIFOs) for JACCL SideChannel relay
Alex CheemaandClaude Opus 4.6
2026-02-12 12:06:37 -08:00
-
9859ebd67a
Add pipe-based JACCL SideChannel relay via exo control plane
Alex CheemaandClaude Opus 4.6
2026-02-12 08:55:40 -08:00
-
9298bab09e
Preserve last_failure_error across instance recreation, fix RDMA banner wording
Alex CheemaandClaude Opus 4.6
2026-02-11 16:48:30 -08:00
-
f4b7f376b0
chore: remove temporary screenshot files
Alex CheemaandClaude Opus 4.6
2026-02-11 16:40:16 -08:00
-
6c7c1079e5
temp: add jaccl warning screenshots for PR comment
Alex CheemaandClaude Opus 4.6
2026-02-11 16:38:53 -08:00
-
c2a356014c
dashboard: show warning banner for [jaccl] RDMA driver errors
Alex CheemaandClaude Opus 4.6
2026-02-11 16:38:42 -08:00
-
4f538e553a
Retry runners within the same Instance instead of recreating
Alex CheemaandClaude Opus 4.6
2026-02-11 14:21:11 -08:00
-
1e5e626725
Remove timestamp-based retry cooldown
Alex CheemaandClaude Opus 4.6
2026-02-11 12:59:39 -08:00
-
a171c96c41
Consolidate failure state onto MetaInstance, add 5s retry cooldown
Alex CheemaandClaude Opus 4.6
2026-02-11 12:55:47 -08:00
-
229fb29514
Show retry attempt count with error message, e.g. (2/3)
Alex CheemaandClaude Opus 4.6
2026-02-11 12:43:20 -08:00
-
89f0a4a69d
Include node friendly names in runner error messages
Alex CheemaandClaude Opus 4.6
2026-02-11 12:41:10 -08:00
-
a896ecca84
Remove permanent retry blocking, allow continuous retry batches
Alex CheemaandClaude Opus 4.6
2026-02-11 12:35:03 -08:00
-
49ada3821f
Show retry count in exceeded retry limit message (3/3)
Alex CheemaandClaude Opus 4.6
2026-02-11 12:28:17 -08:00
-
a6f7131fe0
Collect all runner error messages instead of just the last one
Alex CheemaandClaude Opus 4.6
2026-02-11 12:27:49 -08:00
-
617d94ebc0
Stop infinite retries after 3 failures, show errors persistently in dashboard
Alex CheemaandClaude Opus 4.6
2026-02-11 12:21:11 -08:00
-
e20af70176
Add instance retry logic with max 3 retries and failure tracking
Alex CheemaandClaude Opus 4.6
2026-02-11 12:09:42 -08:00
-
46219fcd07
Fix MetaInstance.node_ids frozenset failing JSON deserialization
Alex CheemaandClaude Opus 4.6
2026-02-11 10:54:56 -08:00
-
37c5e878b9
Send node_ids from placement preview when launching instances
Alex CheemaandClaude Opus 4.6
2026-02-11 10:49:37 -08:00
-
4d8ef4f462
Enforce min_nodes >= 2 for RDMA (MlxJaccl) instances
Alex CheemaandClaude Opus 4.6
2026-02-11 10:42:21 -08:00
-
77afead0ef
Ensure min_nodes >= node filter size when launching
Alex CheemaandClaude Opus 4.6
2026-02-11 10:36:51 -08:00
-
959865f713
Send node_ids from dashboard, error on RDMA when unavailable
Alex CheemaandClaude Opus 4.6
2026-02-11 10:26:29 -08:00
-
4317cf2bfb
Fix use_default validator silently ignoring sharding/instance_meta
Alex CheemaandClaude Opus 4.6
2026-02-11 10:05:00 -08:00
-
848b5139f5
Add placement error feedback and per-node loading status
Alex CheemaandClaude Opus 4.6
2026-02-11 10:01:07 -08:00
-
72d9574fe8
Show MetaInstance sharding/type while PLACING, fix MlxIbv references
Alex CheemaandClaude Opus 4.6
2026-02-11 09:23:44 -08:00
-
14ff11eec2
Extract reconciler into ProcessManager protocol, fix RDMA instance type
Alex CheemaandClaude Opus 4.6
2026-02-11 09:19:13 -08:00
-
4e9cbd09b0
Extract reconciler into ProcessManager protocol
Alex CheemaandClaude Opus 4.6
2026-02-11 09:06:41 -08:00
-
46bc69d65d
Simplify MetaInstance binding: put meta_instance_id on Instance
Alex CheemaandClaude Opus 4.6
2026-02-10 16:15:29 -08:00
-
9c49ea3a89
Add explicit MetaInstance binding, slim MetaInstance to use ModelId
Alex CheemaandClaude Opus 4.6
2026-02-10 15:53:07 -08:00
-
21b594b176
add GLM-5 model support
add-glm5-support
Alex CheemaandClaude Opus 4.6
2026-02-17 10:20:34 -08:00
-
0c6e943cea
replace model ID string matching with architecture-based checks in tokenizer loading
architecture-based-tokenizer-checks
Alex CheemaandClaude Opus 4.6
2026-02-17 10:02:05 -08:00
-
-
-
26368f9837
Add MetaInstance declarative layer with connection health checking
Alex CheemaandClaude Opus 4.6
2026-02-10 13:43:50 -08:00
-
83af8c63fa
Revert "Use custom fork that resolves GPU locks" (#1502)
rltakashigeandGitHub
2026-02-17 18:18:54 +00:00
-
-
-
eccc6298d1
Revert "Add MetaInstance declarative layer (#1447)"
Evan Quiney
2026-02-17 18:02:52 +00:00
-
-
-
de5c1212ee
fix: add missing tasks arg to get_transition_events()
alexcheema/mlx-distributed-transfer
Alex CheemaandClaude Opus 4.6
2026-02-17 09:39:46 -08:00
-
000058aece
fix: call weight_loader in Step35ShardingStrategy for distributed transfer
Alex CheemaandClaude Opus 4.6
2026-02-15 14:19:53 -08:00
-
b5c4df2700
fix: match load_model's selective quantization for large model transfer
Alex CheemaandClaude Opus 4.6
2026-02-13 11:04:26 -08:00
-
89e3159871
Fix Step35ShardingStrategy override and pytest collection
Alex CheemaandClaude Opus 4.6
2026-02-13 10:17:30 -08:00
-
2765f95da1
fix: always transfer metadata files to all nodes before weight loading
Alex CheemaandClaude Opus 4.6
2026-02-12 15:01:00 -08:00
-
48cc931ba4
feat: replace has_weight_files with state-based has_local_model + fix pipeline peak memory
Alex CheemaandClaude Opus 4.6
2026-02-12 14:26:23 -08:00
-
3409f19496
feat: stream weight broadcast layer-by-layer to reduce peak memory
Alex CheemaandClaude Opus 4.6
2026-02-12 13:38:32 -08:00
-
aab8aa38a9
fix: free broadcast weights before sharding to halve peak memory
Alex CheemaandClaude Opus 4.6
2026-02-12 13:14:12 -08:00
-
b594c2f58b
refactor: invert metadata detection to exclude .safetensors
Alex CheemaandClaude Opus 4.6
2026-02-12 13:00:53 -08:00
-
d0c0085f57
fix: include .jinja files in metadata transfer
Alex CheemaandClaude Opus 4.6
2026-02-12 12:55:33 -08:00
-
60b0ba2841
fix: prevent OOM on receiver by using lazy=True with broadcast weights
Alex CheemaandClaude Opus 4.6
2026-02-12 12:50:14 -08:00