This website requires JavaScript.
49951e1b1a
Sync custom model cards across nodes (#1768 )
ciaranbor
2026-03-24 12:51:36 +00:00
e06e70a835
Prefer higher model download % for placement (#1767 )
ciaranbor
2026-03-24 12:11:56 +00:00
d1f13f21a1
time to fiddle
Evan
2026-03-24 11:52:20 +00:00
a6cdc93d3c
disable
Evan
2026-03-24 11:22:09 +00:00
a9c7b1c68a
urgk
Evan
2026-03-23 19:14:28 +00:00
e9fdd8d4af
improve logging: add dates to verbose stderr, match file log level to verbosity (#1772 )
vskiwi
2026-03-23 19:06:26 +03:00
6355f3d8fb
man i gotta commit
Evan
2026-03-23 16:00:57 +00:00
b90538f078
Fix macmon fallback and install docs
Alex Cheema
2026-03-22 14:26:52 -07:00
973e4db085
Unnecessary further optimizations 3
Ryuichi Leo Takashige
2026-03-20 23:56:42 +00:00
8cd6191c70
gotta move the cache soon
Evan
2026-03-21 22:22:25 +00:00
016de1803b
Unnecessary further optimizations 2
Ryuichi Leo Takashige
2026-03-20 11:23:52 +00:00
532a8f0b07
we got it to BUILD
Evan
2026-03-20 15:38:16 +00:00
60a6ac1125
Unnecessary further optimizations
Ryuichi Leo Takashige
2026-03-19 21:56:47 +00:00
5422e831ce
Extra QOL that needs to be reworked
Ryuichi Leo Takashige
2026-03-19 19:15:50 +00:00
03ea3cf6cd
Performance optimizations
Ryuichi Leo Takashige
2026-03-19 11:51:57 +00:00
581a1fcd79
move chunks to core
Evan
2026-03-19 16:18:40 +00:00
ac7194dd90
mupdate
Evan
2026-03-19 16:01:35 +00:00
cfb04800c2
override macmon in flake
Evan
2026-03-17 11:34:41 +00:00
07598a3af1
teeny refactor (#1753 )
Evan Quiney
2026-03-19 15:57:50 +00:00
a2de281c67
Replace GDN projections with register-sharing batched kernel
dmcc73
2026-03-19 15:13:39 +00:00
f538b44211
shmovin
Evan
2026-03-19 12:54:14 +00:00
63f57fc193
Ciaran/dashboard download bug (#1755 )
ciaranbor
2026-03-19 14:47:45 +00:00
a3ce437fd4
first
Evan
2026-03-19 11:10:58 +00:00
6fa2cc1265
Non overlapping case
Ryuichi Leo Takashige
2026-03-18 23:19:24 +00:00
04197fe27b
Some QOL
Ryuichi Leo Takashige
2026-03-18 21:21:48 +00:00
d1490444a1
Implement prefill/decode really
Ryuichi Leo Takashige
2026-03-18 20:27:28 +00:00
ba472da84f
Implement prefill/decode
Ryuichi Leo Takashige
2026-03-18 11:31:55 +00:00
9394d04f5f
Add LCB TPS benchmark script
dmcc73
2026-03-18 17:37:36 +00:00
92c04b0aa5
Add batched fused Metal kernel patches for Qwen3.5 MoE decode
dmcc73
2026-03-18 17:36:28 +00:00
74b877dbcd
persist node ids in .cache
nid-persist
Evan
2026-02-24 11:49:42 +00:00
f208586092
Implement prefill/decode
Ryuichi Leo Takashige
2026-03-18 10:39:22 +00:00
be731d3a85
Merge main
leo/dgx-spark-integrations
Ryuichi Leo Takashige
2026-03-17 19:02:35 +00:00
655185cfe7
Address comments 4 - defer to the warmup into the exo batch generator and vllm batch engine and don't store model on the generators.
Ryuichi Leo Takashige
2026-03-17 18:52:32 +00:00
1dd9c28842
Address comments 3, mainly refactors
Ryuichi Leo Takashige
2026-03-17 18:31:41 +00:00
cacd26e63c
Type error lol
Ryuichi Leo Takashige
2026-03-17 17:47:30 +00:00
6a3eb2f37d
close()
Ryuichi Leo Takashige
2026-03-17 17:31:15 +00:00
e1df77bc4c
No more future annotations
Ryuichi Leo Takashige
2026-03-17 17:07:32 +00:00
a6519ba006
Update mflux to 0.16.9 (#1751 )
ciaranbor
2026-03-17 16:58:23 +00:00
e78e53df6e
Address comments 2 including vllm capability in state
Ryuichi Leo Takashige
2026-03-17 15:50:42 +00:00
8dc3c8b816
flush old addrs
babbler
Evan
2026-03-17 14:28:18 +00:00
c70d9006e8
Address comments including task id interface
Ryuichi Leo Takashige
2026-03-17 14:45:23 +00:00
29eb21131f
just /64s i guess
Evan
2026-03-16 14:53:26 +00:00
2aa7aea538
remove iface flushing
Evan
2026-03-16 12:54:15 +00:00
0a3caa7498
correctness
Evan
2026-03-16 12:31:47 +00:00
458278a5a7
add leo's bandwidth test
Evan
2026-03-16 11:09:05 +00:00
36423936bd
NO BAD lo
Evan
2026-03-16 10:21:47 +00:00
bd290ab3a3
trying lo again
Evan
2026-03-13 14:18:03 +00:00
a2524577bc
ok didnt work lets try /112
Evan
2026-03-13 10:31:28 +00:00
2c7ec92bce
lets try just one thing per thing
Evan
2026-03-12 17:27:17 +00:00
6329896909
cleanup
Evan
2026-03-11 16:49:40 +00:00
c580d56d25
move private directories
Evan
2026-03-11 14:32:10 +00:00
b448b94558
error handling
Evan
2026-03-11 13:40:27 +00:00
7a71973d9c
uniform node ids
Evan
2026-03-10 18:26:06 +00:00
8371ec0c07
oops
Evan
2026-03-10 15:58:58 +00:00
7c09fe6479
handle receiver lagged better
Evan
2026-03-10 15:38:45 +00:00
bafdcd864e
more logs
Evan
2026-03-10 15:12:09 +00:00
d411d08d1d
clippy
Evan
2026-03-10 13:16:09 +00:00
ccc002bed6
some safety
Evan
2026-03-10 12:43:48 +00:00
8b0b03a7d8
babble babble
Evan
2026-03-10 10:54:07 +00:00
72cd8552ae
Merge branch 'main' into leo/dgx-spark-integrations
Ryuichi Leo Takashige
2026-03-17 13:58:49 +00:00
b713889f73
Fix exo bench again again (#1750 )
rltakashige
2026-03-17 13:47:24 +00:00
6ee673147d
Fix exo bench prefill and decode tps (#1749 )
rltakashige
2026-03-17 13:36:44 +00:00
ff4d20eed9
Fix image models through dashboard (#1746 )
ciaranbor
2026-03-17 11:22:01 +00:00
8cd1308336
Tidy pass 1
Ryuichi Leo Takashige
2026-03-16 23:31:44 +00:00
04dcdbd127
Merge main
Ryuichi Leo Takashige
2026-03-16 23:01:48 +00:00
ec5d62f935
Strip vllm generator
Ryuichi Leo Takashige
2026-03-16 21:31:57 +00:00
e96f084051
Distributed callbacks
Ryuichi Leo Takashige
2026-03-16 21:00:15 +00:00
dc68ddbac0
Prompt formatting
Ryuichi Leo Takashige
2026-03-16 20:37:38 +00:00
073f8c1690
add batching
Ryuichi Leo Takashige
2026-03-16 19:19:12 +00:00
01701ca609
try using different fork
leo/use-sm121-branch
Ryuichi Leo Takashige
2026-03-16 17:05:17 +00:00
3c29d0dd4c
test prefix caching
Ryuichi Leo Takashige
2026-03-13 17:46:21 +00:00
45c0bb0289
Bake scale into SDPA pass1 source, remove last scalar input
david/new-kernels-v2
dmcc73
2026-03-15 13:20:21 +00:00
327dcb1740
Bake constants into GDN projections kernel too
dmcc73
2026-03-15 13:15:51 +00:00
9dc52092b0
Bake constants into Metal source, use params arrays for varying values
dmcc73
2026-03-15 13:05:40 +00:00
11b86396cc
Ensure cache_offset is Python int for 0-d scalar
dmcc73
2026-03-15 12:51:46 +00:00
f94e7e5b95
Revert "Precompute RoPE cos/sin in Python, avoid scalar kernel input"
dmcc73
2026-03-15 02:35:39 +00:00
058ccfba24
Precompute RoPE cos/sin in Python, avoid scalar kernel input
dmcc73
2026-03-15 02:30:31 +00:00
3332a2fbe5
Revert RoPE kernel to original (needs MLX >= 0.31.0)
dmcc73
2026-03-15 02:20:45 +00:00
7b303a4fbb
Fix RoPE kernel: dereference float position pointer
dmcc73
2026-03-15 01:56:43 +00:00
632726ac49
Fix RoPE kernel: pass position as float32 to avoid Metal cast issue
dmcc73
2026-03-15 01:54:55 +00:00
cbc610d321
Fix RoPE kernel: implicit int-to-float conversion for position
dmcc73
2026-03-15 01:51:23 +00:00
5d45aba21f
Fix RoPE kernel: dereference position pointer
dmcc73
2026-03-15 01:47:18 +00:00
aa67606ddb
Fix RoPE kernel: use float() constructor instead of C-style cast
dmcc73
2026-03-15 01:43:09 +00:00
b45801f1d4
Adding more kernels to qwen3.5 8bit models
dmcc73
2026-03-15 01:14:15 +00:00
adee2f2111
test adding new kernels in exo
davidmcc73
2026-03-09 19:57:57 +00:00
3a1fc000f9
Bake scale into SDPA pass1 source, remove last scalar input
david/new-kernels
dmcc73
2026-03-15 13:20:21 +00:00
d527d8bc46
Bake constants into GDN projections kernel too
dmcc73
2026-03-15 13:15:51 +00:00
ed54eb73e9
Bake constants into Metal source, use params arrays for varying values
dmcc73
2026-03-15 13:05:40 +00:00
3e953daf5f
Ensure cache_offset is Python int for 0-d scalar
dmcc73
2026-03-15 12:51:46 +00:00
2be3615994
Revert "Precompute RoPE cos/sin in Python, avoid scalar kernel input"
dmcc73
2026-03-15 02:35:39 +00:00
97d7f03032
Precompute RoPE cos/sin in Python, avoid scalar kernel input
dmcc73
2026-03-15 02:30:31 +00:00
ce101091cd
Revert RoPE kernel to original (needs MLX >= 0.31.0)
dmcc73
2026-03-15 02:20:45 +00:00
16b22ba71d
Fix RoPE kernel: dereference float position pointer
dmcc73
2026-03-15 01:56:43 +00:00
5434fd00e6
Fix RoPE kernel: pass position as float32 to avoid Metal cast issue
dmcc73
2026-03-15 01:54:55 +00:00
035c2d1bbd
Fix RoPE kernel: implicit int-to-float conversion for position
dmcc73
2026-03-15 01:51:23 +00:00
de3fa5f060
Fix RoPE kernel: dereference position pointer
dmcc73
2026-03-15 01:47:18 +00:00
1ddf14d6d6
Fix RoPE kernel: use float() constructor instead of C-style cast
dmcc73
2026-03-15 01:43:09 +00:00
cc8308d919
Rebuild: fix Metal compile error in RoPE kernel
dmcc73
2026-03-15 01:37:25 +00:00
d3507d2cd6
Adding more kernels to qwen3.5 8bit models
dmcc73
2026-03-15 01:14:15 +00:00
594ed99734
new uv lock for fastsafetensors
rltakashige
2026-03-13 12:59:21 +00:00