Sync: `tinygrad/tinygrad:master` → `sunnypilot/tinygrad:master` - #10

Merged

sunnyhaibin merged 307 commits into

Sep 2, 2026

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 15, 2026

fix cast to weak twice [pr] (#17548)
chenyuxyz
authored
5ca87f1
fix webgpu is_nan [pr] (#17551)
chenyuxyz
authored
417563c
device: fix remap (#17549)
nimlgen
authored
9702296

Commits on Aug 16, 2026

add max_shape/max_numel to mixins + pad_to (#17553)
geohot
authored
e688e07
iface in device (#17554)
nimlgen
authored
e7bf2a8

Commits on Aug 17, 2026

few torch_backend fix (#17558)
chenyuxyz
authored
c30bf11
fix emulated long cast to double (#17559)
chenyuxyz
authored
057a18a
no dtype in vconst_like [PR] (#17560)
chenyuxyz
authored
bfd4048
delete dead DType.scalar [PR] (#17561)
chenyuxyz
authored
138fb4a
more torch backend fixups (#17562)
chenyuxyz
authored
e25f867
add some dels to jit for OOM fixes (#17566)
geohot
authored
37a54dc
remove torch backend detach hack (#17565)
chenyuxyz
authored
821e80f
update a few is CONST check to check device None [pr] (#17563)
chenyuxyz
authored
42714e1
delete bad torch backend function override (#17568)
chenyuxyz
authored
7fe8e35
add back beautiful_mnist_torch in CI (#17569)
chenyuxyz
authored
954d4f7
hcq2: one submitter (#17556)
nimlgen
authored
58edff6
fix call arg indexing in shard scheduling (#17519)
b1tg
authored
2776c5b
hcq2: speed (#17570)
nimlgen
authored
00d6eed
llm: respect expert_gating_func (#17458)
b1tg
authored
b757437
some more torch backend cleanups (#17571)
chenyuxyz
authored
901d257
add back cast where rule [pr] (#17572)
chenyuxyz
authored
34c9b9d
clean up cstyle render inf and nan [PR] (#17573)
chenyuxyz
authored
d4a1f39

Commits on Aug 18, 2026

base_rewrite reorder [PR] (#17575)
chenyuxyz
authored
cb892e1
llm: refactor delta attention (#17564)
geohot
authored
8d2cc64
compile server for cuda on mac (#17574)
sirhcm
authored
a746861
viz: expanding srcs should not override history (#17577)
Qazalin
authored
7cbe8e0
Revert "disk: neable polling (#17538)" (#17578)
nimlgen
authored
0b757bb
alu(long, weakint) can do math in int too [pr] (#17579)
chenyuxyz
authored
a1366e2
hcq2: 64bit addresses (#17576)
nimlgen
authored
af2a43c
gptoss: zero-2 optim (#17581)
wozeparrot
authored
72280bb
start renderer casted const migration [pr] (#17582)
chenyuxyz
authored
49778d9
use KernelCountException when kernel count is being compared (#17584)
geohot
authored
c31038f
ci: cleanup deps (#17583)
sirhcm
authored
2cfb421

Commits on Aug 19, 2026

casted CONST migration for LLVM and PTX [pr] (#17585)
chenyuxyz
authored
e37b44d
ci: standardize all ubuntu runs-on to ubuntu-24.04 (#17586)
sirhcm
authored
df3b114
casted CONST migration for nir [pr] (#17588)
chenyuxyz
authored
ca86a42
fused_qkv_rope in UOp (#17591)
Qazalin
authored
8c2bf02
Revert "fused_qkv_rope in UOp (#17591)" (#17593)
Qazalin
authored
d539aaf
casted CONST migration for cstyle [pr] (#17587)
chenyuxyz
authored
68b4407
casted CONST migration for x86 [pr] (#17592)
chenyuxyz
authored
e8ba214
update NULL tests that depends on strong dtype CONST (#17594)
chenyuxyz
authored
b3e2f17
finish casted_consts migration [PR] (#17595)
chenyuxyz
authored
9550378
delete stale tests (#17596)
chenyuxyz
authored
bd6e70a
fold_bitcast should truncate its input [pr] (#17601)
chenyuxyz
authored
c218b48
fix TestDevCopySpeeds SIZE (#17602)
chenyuxyz
authored
a4fadcf
realize: fast stat (#17600)
nimlgen
authored
0c5307b
fix roll on zero-sized tensors (#17603)
EthanBnntt
authored
7064e76
no float in tensor shape [pr] (#17605)
chenyuxyz
authored
b8cc74e
fix TestDevCopySpeeds command (#17607)
chenyuxyz
authored
0a0b6cb
test updates for weak const change (#17606)
chenyuxyz
authored
fc214da
hcq2: speed (#17604)
nimlgen
authored
ab68c58
cpu: link with rt (#17608)
sirhcm
authored
2067133
gptoss: save more (#17613)
wozeparrot
authored
c89ae6c

Commits on Aug 20, 2026

benchmarks: split multigpu (#17615)
sirhcm
authored
c63d94e
test updates from weak const branch (#17618)
chenyuxyz
authored
e6324d1
fused_qkv_rope in UOp try 2 (#17619)
Qazalin
authored
a1263fa
bump amd firmware (#17621)
geohot
authored
57d1104
safe changes for mi350p (#17620)
geohot
authored
c117da9
hcq2: staging (#17622)
nimlgen
authored
8c3cb00
hcq2: faster beam (#17624)
nimlgen
authored
e68aa16
fix dtype_from_uop for invalid ALU [PR] (#17626)
chenyuxyz
authored
6732d05
clean up some index(dtype=) [PR] (#17634)
chenyuxyz
authored
f55c1a3
x64 imm uint64 (#17636)
chenyuxyz
authored
a8ecb73
benchmarks: cleanups (#17631)
sirhcm
authored
eba5b7e
remove kernel_cnt, names are no longer unique (#17633)
geohot
authored
1707dca
hotfix: disable HCQ2
geohot
committed
a57188e
x86 REX issue (#17580)
chenyuxyz
authored
0e7ab86
skip INT_MIN % -1 on X86 and LLVM (#17637)
sirhcm
authored
c773891
fa: remove swa recompute (#17635)
wozeparrot
authored
d80254c
fix float_to_bf16 on non-float32 inputs (#17638)
sirhcm
authored
3715006
update const selector where folding [pr] (#17640)
chenyuxyz
authored
592e3f8
ci: venv in /opt/venv (#17632)
sirhcm
authored
c655aaf
add extract support to fetch (#17641)
geohot
authored
0ccef54
add ansipad and PARALLEL contextvar (#17642)
geohot
authored
3fdbb82
compile kernels in parallel (#17629)
geohot
authored
9aa9e11

Commits on Aug 21, 2026

llm prefill failing test (#17630)
b1tg
authored
9216aa4
hotfix: qkv bw kernel requires hipcc and cdna4 (#17645)
Qazalin
authored
d716d0d
no-op weak.py reorder [PR] (#17646)
chenyuxyz
authored
1bcb6bd
fix llm vars regression (kimi) (#17644)
geohot
authored
1cb0600
profiler failing test for non unique kernel names (#17647)
Qazalin
authored
80bf60d
prereq viz cleanups for unique profile keys (#17649)
Qazalin
authored
dcc2d02
profiler tracing using unique keys (#17651)
Qazalin
authored
e8a8d99
remove setUp from test_sqtt_profiler (#17652)
Qazalin
authored
1c3c9e9
weak 1 for FDIV in get_late_rewrite_patterns [PR] (#17655)
chenyuxyz
authored
176377f
fix render_marg [pr] (#17656)
chenyuxyz
authored
554d078
cleanup the mxfp4 gemm (#17660)
Qazalin
authored
77f698e
don't match casted const in const_folding_pat [pr] (#17658)
chenyuxyz
authored
cc32aa1
usb amd: pipelined copyin over the 0xF2 engine (2.6x faster) (#17628)

geohot
and
tiny
authored
756e82e
ceildiv in wgsl _packed_size [pr] (#17662)
chenyuxyz
authored
3919ce8
hcq2: use shrink.bitcast (#17653)
nimlgen
authored
1cf8a2c
more wgsl pack cleanups [PR] (#17664)
chenyuxyz
authored
8f59041
few more self folding [pr] (#17657)
chenyuxyz
authored
8f9cbdf
l2i and sign_extend cleanups [pr] (#17668)
chenyuxyz
authored
402bea7
usb copyin: async arm and drain, 323 MB/s on comma (#17663)
geohot
authored
3082956
ci: remove setup-python (#17665)
sirhcm
authored
298748e
hcq2: parallel compile (#17667)
nimlgen
authored
12f889a
FLOORDIV to SHR for powers of 2 [pr] (#17669)
chenyuxyz
authored
daa154a
nv: hevc (#17661)
nimlgen
authored
4fd4eaf
keep IndexingContext scoped in indexing (#17670)
geohot
authored
f986829
compile3: log printed timings (#17671)
sirhcm
authored
07cce78
gptoss: fused ce (#17672)
wozeparrot
authored
52596db

Commits on Aug 22, 2026

refactor the AMD emulator slop (kimi) (#17673)
geohot
authored
af24281
delete dtype_from_uop INDEX exempt (#17674)
chenyuxyz
authored
7204d46
test update for weak const (#17675)
chenyuxyz
authored
356f665
remove explicit dtype for NOOP and decomp [PR] (#17678)
chenyuxyz
authored
8950942
make decomp pass SPEC=2 [PR] (#17681)
chenyuxyz
authored
a9069c1
hcq2: early usb (#17683)
nimlgen
authored
a2e64e1
lil decomp cleanup [PR] (#17682)
chenyuxyz
authored
b0a1285

Commits on Aug 23, 2026

Commits on Aug 24, 2026

don't hardcode dtype.int const in decode_hevc_frame (#17696)
chenyuxyz
authored
a31aca9
disk cache: thread-local db conn (#17694)
YassineYousfi
authored
11edcc1
hotfix: switch benchmark to qwen3.8
geohot
committed
9d0cd0e
llama rope freqs in fp32 (#17698)
Qazalin
authored
4ee114b
Revert "disk cache: thread-local db conn (#17694)"
geohot
committed
df52849
usb: don't fail async transfers on signal interruption (#17692)
YassineYousfi
authored
a5ea95d
mergable fast RDNA3 Qwen 3.8 (#17512)
geohot
authored
7e561fc
llama script changes from the speedups branch (#17701)
Qazalin
authored
9b0c656
dsp: cleanup temp linker script (#17704)
sirhcm
authored
9b481a4
PYLITERAL has dtypes.void (#17705)
chenyuxyz
authored
7b21ffa
am: use gc regs for xgmi (#17707)
nimlgen
authored
7a887e8
fix Muon weight decay being a no-op (#17709)
geohot
authored
5aabbb1
AGENTS: no amend commits
geohot
committed
08fbd25
llm: add DEBUG=2 for weights loading (#17710)
geohot
authored
76bf6b7
delete GETTUPLE on UNSHARD rule [PR] (#17711)
chenyuxyz
authored
f4cc288
mi350p: aqua vanjaram bringup for the raw PCI driver (#17693)
geohot
authored
ed11099
remove redundant UOp dtype [PR] (#17713)
chenyuxyz
authored
8df3dac
benchmarks: cleanup dsp benchmark (#17706)
sirhcm
authored
2e7790f
benchmarks: more aggressive timeouts (#17715)
sirhcm
authored
021c015
feat: bump version to 0.14.0 (#17716)
wozeparrot
authored
6f87158

Commits on Aug 25, 2026

use native shifts in Payne Hayek instead of pow2 mul/divs (#17717)
GabrielNakamoto
authored
d9004cf
const are weak 3 [pr] (#17695)
chenyuxyz
authored
a7df1a1
const cleanups [PR] (#17719)
chenyuxyz
authored
65ca685
remove more explicitly set dtype [PR] (#17721)
chenyuxyz
authored
1d694dd
split rangeify to prepare.py (#17722)
geohot
authored
a567831
bring back sqtt examples tests (#17725)
Qazalin
authored
ab79879
delete more explicitly set dtype to UOp [PR] (#17726)
chenyuxyz
authored
9607787
dtype_from_uop for THREEFRY and FDIV (#17727)
chenyuxyz
authored
9f01775
fix viz for hcq2 (#17729)
nimlgen
authored
023bfdb
wgsl cast before load for packed [pr] (#17731)
chenyuxyz
authored
7dc8b66
hcq2: fix copy/call in usb (#17732)
nimlgen
authored
d851aca
Reapply "disk cache: thread-local db conn (#17694)"
geohot
committed
b831ca6
benchmarks: remove usage of sudo from linux runners (#17735)
sirhcm
authored
7fef98c
usb amd: yield between signal polls (#17712)
YassineYousfi
authored
2824504
pin onnxruntime==1.24.1 (#17738)
sirhcm
authored
5503251
fix: external_test_gpu_crash on python 3.14 (#17739)
wozeparrot
authored
07268b7

Commits on Aug 26, 2026

CUSTOM arg is (str, dtype) [PR] (#17737)
chenyuxyz
authored
6ece327
lil fixes from new_rangeify (#17741)
geohot
authored
dc04c78
benchmarks: don't use sudo on mac (#17740)
sirhcm
authored
b49c03f
fix am_smi to respect dev (#17742)
geohot
authored
625c05d
llm tokenizer: fix tekken, add gpt4o (#17733)
b1tg
authored
9860e5d
better all2all schedule test (#17746)
Qazalin
authored
ecf79e2
fix asinh precision loss on negative inputs (#17749)
teddytennant
authored
4c20f1d
delete unused rewrite rules [PR] (#17748)
chenyuxyz
authored
4bdc865
never bufferize_to_store weak input [PR] (#17751)
chenyuxyz
authored
0abcf09
small changes from new rangeify (#17752)
geohot
authored
ea6c82f
fix PTX NIR SPEC=2 for bool [pr] (#17753)
chenyuxyz
authored
6a3b297
cleanup BENCHMARK_LOG (#17754)
sirhcm
authored
cc72b9f
delete PCONTIG [PR] (#17756)
geohot
authored
6042b87
hcq2: fix jit (#17747)
nimlgen
authored
6074c00
fix _get_cpu_count for docker --cpus=N in python 3.13+ (#17760)
sirhcm
authored
c015351
faster qwen 3.8 (#17720)
geohot
authored
a0a901c
gptoss: single jit (#17525)
wozeparrot
authored
22ffff3
benchmarks: openpilot matrix (#17714)
sirhcm
authored
0820c7a

Commits on Aug 27, 2026

simplify more same device copies (#17766)
Qazalin
authored
118e812
remove some dead scheduler tests (#17768)
Qazalin
authored
d16a251
set BEAM_PADTO=0 in llama (#17769)
Qazalin
authored
0d66815
mockgpu/amd: use generated KFD ioctl metadata (#17761)
tonyfloatersu
authored
2d4e7af
hcq2: a bit cleaner deps (#17771)
nimlgen
authored
f19d89e
fix asinh gradient at zero (#17758)
teddytennant
authored
5c3d044
bnxt to extra (#17772)
nimlgen
authored
7eb763a
hcq2 tiny speedups (#17773)
nimlgen
authored
980c42d
fix test_assign for no subbuffer backends (#17757)
chenyuxyz
authored
54d69f4
autogen: use deb for comgr 2 (#17782)
sirhcm
authored
14cbece
UOp.param accepts single sint as shape (int,) (#17784)
geohot
authored
f267a49
hotfix: update max test time from 300 -> 90
geohot
committed
38fa064
hotfix: decrease dims in test_attention to get below the 90s limit
geohot
committed
ee3161e
test llm with --no_chat_template (#17785)
sirhcm
authored
f06832b
fix fancy indexing with uint8 index (#17787)
chenyuxyz
authored
fede358

Commits on Aug 28, 2026

onnxruntime respects NUM_CPU_THREADS (#17788)
sirhcm
authored
e6ad290
fix where with float64 branches (#17789)
chenyuxyz
authored
38e0fe1
move shape into arg for param/buffer (#17783)
geohot
authored
6b361e4
fix CMPLT between float16 and const 0 (#17791)
chenyuxyz
authored
bc485d2
hotfix: disable HCQ2
geohot
committed
4e6bdac
axis arg is dead (#17793)
geohot
authored
39d7fd0
ci: skip apt steps if packages are already installed (#17794)
sirhcm
authored
4456e62
copy movement op tests construct copies (#17792)
Qazalin
authored
53cad32
fix interpolate with align_corners and an output size of 1 (#17790)
teddytennant
authored
26c4b63
gptoss: faster grad handling (#17795)
wozeparrot
authored
92d92c7
cleaner mxfp4 gemm prelude (#17796)
Qazalin
authored
6242b09
tiny hcq2 changes (#17797)
nimlgen
authored
76dce1e
hcq2 spec2 (#17798)
nimlgen
authored
55c8e18
fix nan compares (#17801)
chenyuxyz
authored
6083de5
remove X86 ymm thingy (#17803)
chenyuxyz
authored
cb857a5
add support for UD quants (#17806)
JoelLisenby
authored
7fdc58b
fix Conv2d same padding with per axis dilation (#17805)
teddytennant
authored
2eacd4f
delete X86 not used codes [PR] (#17807)
chenyuxyz
authored
287679a
fix onnx MaxUnpool (#17809)
chenyuxyz
authored
ae63f23
fix onnx.py type annotations (#17811)
chenyuxyz
authored
112e50e
fix onnx MeanVarianceNormalization arg (#17812)
chenyuxyz
authored
f196af2
am: vf (#17802)
nimlgen
authored
9599581
ci: use fetch to download libraries (#17814)
sirhcm
authored
49627ae
update X86 cast related stuff (#17816)
chenyuxyz
authored
b3dad4f
put dtype to Ops.INS arg (#17817)
chenyuxyz
authored
bd2711b
don't set UOp dtype in x86 (#17818)
chenyuxyz
authored
793c166
dtype_from_uop is never None (#17819)
chenyuxyz
authored
cb0db2d
dtype is not a UOp field anymore (#17821)
chenyuxyz
authored
086f45c
remove spec dtype check that's trivially true [PR] (#17823)
chenyuxyz
authored
0640cfc
delete Ops.FUNCTION (#17824)
geohot
authored
812d5ce
make UOp.dtype recursive_property [PR] (#17827)
chenyuxyz
authored
a9830e2

Commits on Aug 29, 2026

some tri cleanups [PR] (#17829)
chenyuxyz
authored
12f34d8
parallel beam and dataset cache for digitalocean_mi350x (#17830)
Qazalin
authored
02d67a8
use am in digitalocean_mi350x writeup (#17832)
Qazalin
authored
9aa3c48
fix bufferize_to_store for weak input (#17833)
chenyuxyz
authored
661440e
weak.py cleanups [PR] (#17834)
chenyuxyz
authored
8c7906d
fix logsumexp on fully masked slices (#17822)

teddytennant
and
chenyuxyz
authored
0a1bfc9
UPat(RandMixin) [PR] (#17835)
chenyuxyz
authored
2f5572a
fix weak const that overflows int32 [pr] (#17837)
chenyuxyz
authored
d7913e2
fix weak const overflow with movements [pr] (#17840)
chenyuxyz
authored
0bd7253
cannot decomp long Variable (#17842)
chenyuxyz
authored
0b58ff6

Commits on Aug 30, 2026

weak dtype overflow touchups [pr] (#17843)
chenyuxyz
authored
1a5a364
remove dead symbolic rules [pr] (#17844)
chenyuxyz
authored
61f7fd3
clean up unneeded dtype check in rules [PR] (#17845)
chenyuxyz
authored
7cc973a
fix promote weak in _broadcasted [pr] (#17849)
chenyuxyz
authored
9c49e35
Revert "fix promote weak in _broadcasted [pr] (#17849)" (#17851)
chenyuxyz
authored
902a1ef
fix weak const promote dropping the movement ops (#17846)

teddytennant
and
chenyuxyz
authored
f5b00de
fix elu/celu/selu gradient being nan on large inputs (#17852)
teddytennant
authored
1c183e6
more const cast cleanups [PR] (#17854)
chenyuxyz
authored
85841ce
fix logcumsumexp on fully masked prefixes (#17853)
teddytennant
authored
fdb4109
fix strided setitem being a no-op on a realized tensor (#17855)
teddytennant
authored
593aa1e

Commits on Aug 31, 2026

mi350p: fix pmc (#17858)
Qazalin
authored
a6c810e
viz: add cdna waveend (#17859)
Qazalin
authored
b48cb9f
fix erf gradient being zero at x=0 (#17861)

authored
3123f94
fix logaddexp and asin at 0 (#17862)
chenyuxyz
authored
e3c4b3d
llm: fix q6k fastpath (#17847)
b1tg
authored
af079e8
hcq2 rewrite patches (#17838)
nimlgen
authored
d025993
add Ops.RETURNED, remove TUPLE/GETTUPLE (#17804)
geohot
authored
60cc1d7
Revert "ci: use fetch to download libraries" (#17863)
sirhcm
authored
32bbf82
openpilot: no NOLOCALS=1 (kimi) (#17762)
sirhcm
authored
8f4bd50
delete CPU threading support (#17866)
geohot
authored
4197f74
remove NOLOCALS [pr] (#17868)
sirhcm
authored
79250d7

Commits on Sep 1, 2026

run llvmspeed Speed Test (BEAM=2) without cache (#17869)
chenyuxyz
authored
c12eefe
onnx: add tinygrad contiguous custom op (#17871)

YassineYousfi
and
sirhcm
authored
37738a7
merge OptOps.UNROLL and OptOps.UPCAST (#17867)
chenyuxyz
authored
a3bde51
delete dead OptOps tests (#17872)
chenyuxyz
authored
b87159c
OptOps.GROUP is OptOps.LOCAL (#17873)
chenyuxyz
authored
cf6fff1
Merge branch 'cf6fff13aa4dbf61d930c5c6da607ab2954edede' into sync-20260901
sunnyhaibin
committed
966a8f5