Commit graph

12,643 commits

Author SHA1 Message Date
George Hotz
34d2ef6447
Merge branch 'master' into precompile_backward 2026-03-17 15:17:41 +08:00
George Hotz
125d987886 no NOOPT 2026-03-17 15:14:25 +08:00
chenyu
6b6d1814ca
update no_vectorized_index [pr] (#15313)
combine no_vectorized_index and no_vectorized_index_broadcast
2026-03-17 03:05:23 -04:00
George Hotz
c83f2a7154 simpler 2026-03-17 15:01:46 +08:00
George Hotz
32a8b3aaa0 split v not split 2026-03-17 14:57:53 +08:00
George Hotz
39aae38b2a compact grad 2026-03-17 14:37:40 +08:00
George Hotz
f909d5c983 fix 2026-03-17 14:19:47 +08:00
George Hotz
f7512e9595 cleanups 2026-03-17 14:14:41 +08:00
b1tg
856a839efc
llm: fix qwen3 moe topk renormalization (#15201) 2026-03-17 12:57:33 +08:00
chenyu
1283b57b4e
update fix_store_after_hazard (#15309)
actual gate is just not CONTIGUOUS, also don't need to check against full backward_slice
2026-03-16 23:55:59 -04:00
George Hotz
9f591b42d1 add precompile backward support 2026-03-17 11:40:09 +08:00
Christopher Milan
575b40b93a
determine image shapes before index devectorization (#15304) 2026-03-16 23:16:33 -04:00
George Hotz
3ff03be413
call always has tuple (#15297)
* call always has tuple

* fix pre-commit and simplify

* update

* fix

* move that assert

* tuple

* fix multi

* cleanups

* fix merge
2026-03-17 10:58:46 +08:00
chenyu
1b8b151195
simpler Tensor.assign (#15302) 2026-03-16 22:37:25 -04:00
wozeparrot
674c760974
embedded bwd vocab shard (#15001)
* fix: remove more multi from call

* feat: embedding bwd vocab sharding

* clean: unused import

* clean: don't actually need this pattern
2026-03-16 19:37:16 -07:00
Christopher Milan
62bfd48d95
smarter padding in image_conv2d (#15289) 2026-03-16 22:17:48 -04:00
chenyu
e1fab4d2a9
UOp.store is always void [pr] (#15301) 2026-03-16 21:58:05 -04:00
chenyu
02afb45f29
remove UOp.assign [pr] (#15300)
* remove UOp.assign [pr]

it's all store and after, UOp is immutable

* fix test
2026-03-16 21:45:41 -04:00
qazal
33bd33e783
sqtt: add CDNA ops enum, show in viz (#15140) 2026-03-17 09:38:42 +09:00
chenyu
3e2b7803e6
view assign replaces at buffer identity (#15298)
matches what functions capture
2026-03-16 19:58:38 -04:00
qazal
346596cdce
viz: nanoseconds time axis in sqtt (#15299)
* ui

* secondaryTick is optional

* shader markers data

* instSt infra

* path forward

* details
2026-03-17 07:20:18 +09:00
nimlgen
1bc4cb254c
signed tinygpu as default (#15296)
* signed tinygpu as default

* f

* no sip
2026-03-16 19:29:41 +08:00
Christopher Milan
0de519c7c2
[pr] fewer simplify calls in image_fixup (#15283) 2026-03-16 06:57:52 -04:00
nimlgen
27e29127b5
system: remote prereqs (#15290)
* x

* new format for apl

* this

* typing

* rpc

* tuple

* linter+new tinygpu
2026-03-16 18:45:41 +08:00
chenyu
837b06c609
style cleanups in allocations.py [pr] (#15295) 2026-03-16 05:45:24 -04:00
George Hotz
476276f4b4
support grads on tuples (#15287)
* support grads on tuples

* simpler

* grad_fxn works

* cleanups

* unused
2026-03-16 17:39:34 +08:00
chenyu
20799df10b
remove Ops.ASSIGN [pr] (#15294)
goodbye
2026-03-16 05:22:21 -04:00
chenyu
b3378e7022
UOp.assign is store+after [pr] (#15292) 2026-03-16 04:51:50 -04:00
George Hotz
2e1c81c23f
allow_implicit to disable implicit params (#15291)
* allow_implicit to disable implicit params

* get both Tensor and UOp

* no implicits in llm
2026-03-16 16:40:14 +08:00
chenyu
a0d1444790
Tensor.assign is store+after [pr] (#15288)
* Tensor.assign is store+after [pr]

* put that back
2026-03-16 04:04:55 -04:00
George Hotz
08662bc4ab
add TUPLE/GETTUPLE, simple tests pass (#15286)
* simple tuple stuff passes

* resolved
2026-03-16 15:06:02 +08:00
nimlgen
e7705fe311
system: pcidev doesn't care about bars (#15284) 2026-03-16 14:45:43 +08:00
nimlgen
ff0bcc8de0
system: iface p1 changes (#15278) 2026-03-16 10:48:25 +08:00
qazal
4445f50356
viz: variable duration rdna barriers (#15277)
* viz: variable length rdna barriers

* work

* tiny changes

* simple wave simd test

* small wave sync test

* good multi barrier bug find

* simple fix

* wave_sync asserts

* rdna4 work

* more rdna4

* find more bugs in my model

* it's so much simpler

* wave_sync tests duration

* r4

* should just call this rdna4
2026-03-16 06:06:19 +09:00
qazal
5cd1daa3bc
cdna asm_gemm in one file, remove old rdna3 asm (#15281) 2026-03-16 04:32:30 +09:00
chenyu
cd14e8e64b
allocations contiguous is store+after (#15280) 2026-03-15 11:58:40 -04:00
qazal
7b6211fdd7
sqtt: remove discover_ops script (#15279) 2026-03-15 22:17:06 +09:00
wozeparrot
473e5e4368
feat: make USE_ATOMICS embedding bwd faster (#15151) 2026-03-14 21:21:10 -07:00
qazal
3858bfc83d
sqtt: CDNA inst decodes (#15274)
* sqtt: CDNA inst decodes

* JUMP packets other way

* cdna insts

* r3

* r4

* lds from simd1 and simd2
2026-03-14 21:03:46 +09:00
Christopher Milan
d753c5d7e5
IMAGE=1 image_conv2d pads for bank conflicts (#15252) 2026-03-14 07:59:16 -04:00
Christopher Milan
9047249a7c
m.where(x.pad_to(m.shape), Invalid) ranges shrink (#15275) 2026-03-14 07:26:36 -04:00
nimlgen
f392c53c66
system: merge remote into pciiface (#15273)
* system: merge remote into pciiface

* clenaer

* move

* mypy

* fix
2026-03-14 18:44:20 +08:00
chenyu
13eec8fbe8
remove unused assign rules [pr] (#15268) 2026-03-14 05:37:49 -04:00
Christopher Milan
dabdc986df
shrink guarded ranges, try 2 (#15272) 2026-03-14 04:24:05 -04:00
Christopher Milan
7cf4b16c91
Revert "shrink guarded ranges" (#15271) 2026-03-14 03:44:38 -04:00
Christopher Milan
d9951e2f8e
shrink guarded ranges (#15263) 2026-03-14 03:38:48 -04:00
qazal
43ffd66fda
viz: oneline inst list (#15269)
* viz: oneline inst list

* save 5 chars

* gradual padding
2026-03-14 15:37:18 +09:00
George Hotz
86f17468ed
store in spec + USB BOT fix (#15265)
* move spec to store

* usb bot flag

* Revert "usb bot flag"

This reverts commit 7b8b7824f0.

* fix assert
2026-03-14 13:25:05 +08:00
George Hotz
06d7cddb33
amd_copy_matmul is cleaner (#15248)
* amd_copy_matmul is cleaner

* it runs

* replicated stuff

* add tid there

* it runs

* cleanup

* x.src[1]

* flatten

* move that

* keep that assert
2026-03-14 12:56:09 +08:00
chenyu
b3600e4774
don't emit assign in transform_precompiled_call [pr] (#15262) 2026-03-13 22:42:35 -04:00