Sync: `tinygrad/tinygrad:master` → `sunnypilot/tinygrad:master` - #8

Merged

sunnyhaibin merged 176 commits into

Jun 5, 2026

Commits on May 15, 2026

nv: move structs to vram (#16184)
nimlgen
authored
28b98e5
viz/cli: select all calls in graph by default (#16214)
Qazalin
authored
d54fa86
remove test_svd_nonfull_5_5 (#16217)
chenyuxyz
authored
c6cf9e8
remove noop requires_grad_ calls (#16213)
chenyuxyz
authored
07a172d
remove Tensor.eye override (#16219)
chenyuxyz
authored
d62c1d8
update spec for LOAD (#16221)
geohot
authored
2570bde
fix caformer onnx run (#16222)
geohot
authored
2549b14
cl: device extensions in arch (#16220)
sirhcm
authored
2c61f65

Commits on May 16, 2026

metal: arch is GPU family (#16223)
sirhcm
authored
79c0ae5
llama: fix running flat_llama (#16224)
wozeparrot
authored
1596943
remove more invalid (#16227)
wozeparrot
authored
2d48d7a
move debug prints (#16218)
nimlgen
authored
e575f77
fp8 gemm tests with scale args (#16231)
Qazalin
authored
ebcb7b7
fix nolocals and beam (#16232)
nimlgen
authored
bef5f71
do not gate backward on requires_grad (#16230)
chenyuxyz
authored
0ddc50d
fp8 gemm speedup (#16236)
Qazalin
authored
d95bf39
remove use of requires_grad in test/ (#16237)
chenyuxyz
authored
8631b6f
remove requires_grad use in extra/examples (#16238)
chenyuxyz
authored
dcee90a

Commits on May 17, 2026

remove Tensor.detach override [pr] (#16239)
chenyuxyz
authored
fcdd1af
remove requires_grad= in tinygrad/ (#16241)
chenyuxyz
authored
981c121

Commits on May 18, 2026

make slow tests faster (#16244)
chenyuxyz
authored
5ae4dbd
Revert "fp8 gemm speedup (#16236)" (#16245)
Qazalin
authored
bfb2d1f
deviceless const from UOp (#16243)
chenyuxyz
authored
db639eb
cl: check for cl_khr_fp64 (#16246)
sirhcm
authored
50481ec
to and shard is noop for deviceless uop (#16247)
chenyuxyz
authored
73e6b49
assign for deviceless const source (#16248)
chenyuxyz
authored
7543440
ci: actually use clang-20, enable bfloat16 (#16249)
sirhcm
authored
7515824
llama: use hipcc 7.1 version (#16250)
Qazalin
authored
98b8a2b
multi alu with deviceless const (#16251)
chenyuxyz
authored
d532b4f
llama: also use 7.1 comgr, update startup_walltime.sh (#16253)
Qazalin
authored
18b102f

Commits on May 19, 2026

llama: don't save weight (#16252)
wozeparrot
authored
a3d59fa
Tensor.device and UOp.device can be None (#16255)
chenyuxyz
authored
c1076ed
more deviceless const prerequisites [pr] (#16256)
chenyuxyz
authored
e694d7f
all tests pass with NOOPT=1 (#16257)
geohot
authored
3f2d401
tighten shape spec for broadcasting (#16206)
geohot
authored
a120709
set hipcc mac docker to 7.1 (#16261)
Qazalin
authored
e4350e7
hcq2: split into schedule/realize (#16216)
nimlgen
authored
b3dcf8f
PYTHONREMU: fix CDNA VOP3 conditional writes (#16258)
sirhcm
authored
a178301
fix: pass input device to ONNX helper internal tensors (#16242)
shettysach
authored
74567c1
25000 lines for x86 backend
geohot
committed
b2e8102
X86 with Ops.INS (#14873)

ttomsa
and
geohot
authored
aa1e59a
more prerequisuite test changed for deviceless const (#16264)
chenyuxyz
authored
890b731
fix mypy mockgpu and add tinygrad.renderer.isa to packages (#16265)
sirhcm
authored
bb2a51f
PYTHONREMU: V_CVT_PK_BF8_F32 saturation (#16268)
sirhcm
authored
7cdd9cb
Remove Ops.VCONST (#16267)
geohot
authored
5551574

Commits on May 20, 2026

fix RUN_PICKLE and test it (#16272)
geohot
authored
da7414d
llama: match flat_llama with model_train (#16269)
wozeparrot
authored
361553c
clone can take device (#16271)
chenyuxyz
authored
188d7ec
relax policy ASSERT_MIN_STEP_TIME to 3.2 (#16273)
chenyuxyz
authored
7af7b67
clean up deviceless const _buffer (#16274)
chenyuxyz
authored
170b857
extent jit const error to deviceless inputs (#16276)
chenyuxyz
authored
e1715b3
fused ce llama kernel in UOps (#16263)
Qazalin
authored
1e0fffe
filter by base_class in pci_scan_bus on macOS (#16282)

sunkencity999
and
christopherbradford
authored
fe2d8d1
remove _force_unique from Tensor init (#16277)
chenyuxyz
authored
4dbe6a2
enable broadcasting in _shape (#16285)
geohot
authored
19535df
fix: qlinearconv quant params (#16234)
BraunPhilipp
authored
a01d591
hcq2: cleanups (#16278)
nimlgen
authored
a88feef
llama: apply_grad saves memory (#16275)
wozeparrot
authored
825f30b
remove DEVECTORIZE (#16290)
geohot
authored
58d58c1
fix x86 nondeterminism (#16293)
Qazalin
authored
a19fa29
UOp.clone [pr] (#16295)
chenyuxyz
authored
beea463
am: override smu 13_0_7 to 13_0_0 (#16292)
sirhcm
authored
da07b28
x86: don't use id (#16296)
Qazalin
authored
9e88b08
use clone instead of unique_const in allreduce [pr] (#16297)
chenyuxyz
authored
d548f8d

Commits on May 21, 2026

move is_dtype_supported to renderer (#16226)
sirhcm
authored
172f949
don't use dtype vec for image idx (#16298)
geohot
authored
ec54725
print dtype decomp warning in DEBUG=2 (#16300)
sirhcm
authored
dd279ee
set SPLIT_W13=0 for 8b DP by default (#16302)
Qazalin
authored
e33e058
hcq2: multi prereqs (#16304)
nimlgen
authored
a321700
llama: remove fused grad accum (#16301)
wozeparrot
authored
afc5bfa
dtype.vec shapes (#16287)
geohot
authored
6815f28
full(buffer=True) (#16311)
chenyuxyz
authored
73ea36f
llama: realize amax (#16308)
wozeparrot
authored
fb718a5
llama: update 405b scripts (#16309)
wozeparrot
authored
0c41317
remove many requires_grad= args (#16321)
chenyuxyz
authored
720a27b
only load tinymesa_cpu when LVP is explicitly requested (#16320)
sirhcm
authored
518e605
Tensor.requires_grad -> is_param (#16325)
chenyuxyz
authored
31424cd

Commits on May 22, 2026

start cleaning up dtype tests (#16324)
sirhcm
authored
150a82d
use more non-buffered const (#16327)
chenyuxyz
authored
9744d51
remove getenv(CI) from core tinygrad (#16326)
sirhcm
authored
c2d0657
more unique const removal prerequisite (#16328)
chenyuxyz
authored
3115952
quantize_fp8 kernels in uops (#16288)
Qazalin
authored
bbfe4f8
am: mec_pipe0_reset on gfx12 only (#16331)
nimlgen
authored
5d77a94
nv: fix boot mem (#16332)
nimlgen
authored
90dbb45
Fix no module named error (#16305)

googlefan256
and
chenyuxyz
authored
eeadf26
add missing init files, add chat.html to package-data (#16334)
chenyuxyz
authored
acc5197
feat: bump version to 0.13.0 (#16337)
wozeparrot
authored
2d48fe8
hcq2: move submit lowering to schedule (#16330)
nimlgen
authored
26b3b3f
start cleanup of the slowest tests (#16339)
sirhcm
authored
451f381
ci: cleanup some deps [pr] (#16340)
sirhcm
authored
35461d4

Commits on May 23, 2026

deviceless const cleanups (#16341)
chenyuxyz
authored
149a87d
update test_stack (#16345)
chenyuxyz
authored
926d125

Commits on May 24, 2026

hcq2: multi (#16347)
nimlgen
authored
a891727
remove Tensor.from_uop (#16344)
chenyuxyz
authored
5d5e028

Commits on May 25, 2026

use get_buf in exec_kernel (#16356)
nimlgen
authored
64095cf
fix mailbox comment, add types (#16360)
Qazalin
authored
eecd470
llama: offload master weights (#16355)
wozeparrot
authored
68d2102
use windows-specific alias nvcuda when loading cuda on windows (#16260)
IntendedConsequence
authored
2ab90f3
viz/cli: add --interval (#16363)
Qazalin
authored
b73d2d1
invalids is empty [pr] (#16353)
chenyuxyz
authored
cdeb861
viz: sidebar toggles always recenter (#16366)
Qazalin
authored
3506eb0
viz: stable node position for +- toggle (#16367)
Qazalin
authored
4bcc53e
webgpu: shader-f16 support in arch (#16370)
sirhcm
authored
d8f86be

Commits on May 26, 2026

move buffer view offset to src (#16364)
geohot
authored
689ab6a
hotfix: bump Mac pytest timeout to 4 minutes
geohot
committed
695a006
remove getenv(CI) (#16365)
sirhcm
authored
8ddd132
Revert "hotfix: bump Mac pytest timeout to 4 minutes"
geohot
committed
942cb42
gather to single device (#16354)
wozeparrot
authored
76fc39c
script to generate testsig for DSP (#16371)
geohot
authored
41ee7da
hotfix: bump Mac pytest timeout to 4 minutes (try 2)
geohot
committed
322693d
hcq2: simpler (#16361)
nimlgen
authored
032905d
viz: add +- toggle for hidden UOps (#16368)
Qazalin
authored
846a809
bufferview offset is units of input dtype (#16378)
geohot
authored
7f1b028
remove unique_const (#16372)
chenyuxyz
authored
09019d6
Revert "remove unique_const (#16372)" (#16380)
chenyuxyz
authored
9b00def
fix emu bug in gfx950 (#16381)
geohot
authored
bac82d4
remove unique_const (#16382)
chenyuxyz
authored
d861c50
remove CONST(UNIQUE) (#16383)
chenyuxyz
authored
0b88827
remove allow_any_len from store (#16385)
geohot
authored
fd96303
cleanup some old test skips (#16384)
sirhcm
authored
d23659d
disable flaky cl test (#16388)
sirhcm
authored
3adf7f5

Commits on May 27, 2026

rename BUFFER_VIEW to SLICE (#16391)
geohot
authored
156a443
llama: fakedata uses real vocab size (#16389)
wozeparrot
authored
3a7a6da
test update for arange and eye (#16394)
chenyuxyz
authored
fa14cde
ci: use ubuntu 24.04 (#16393)
sirhcm
authored
aacc8ad
revert const unique changes (#16395)
chenyuxyz
authored
945ed4f
skip test_setitem_fancy_on_unrealized_view (#16400)
chenyuxyz
authored
3e80f37
test_custom_kernel_precompile_multidevice (#16401)
chenyuxyz
authored
6da7855
webgpu: fix enums (#16402)
sirhcm
authored
bacabf0
bring back test and torch backend change for unique const (#16403)
chenyuxyz
authored
c33b767
hcq2 rewrite (#16375)
nimlgen
authored
0c385e3
llama: don't allocate grad_xw13 in bf16 (#16359)
Qazalin
authored
452c7d4
skip modulo by zero in test_dtype_alu (#16404)
sirhcm
authored
171401e
shrink/pad use (new_shape, offset) (#16405)
geohot
authored
8ee3a37
llama: delayed scaling in optim (#16407)
wozeparrot
authored
dac3743
marg line cleanup (#16408)
geohot
authored
b21afb4

Commits on May 28, 2026

viz: click on +- toggles sources (#16409)
Qazalin
authored
88e88d6
fix slice store to remove the index (#16410)
geohot
authored
5ee6b6b
viz: make call toggle easier to click on (#16411)
Qazalin
authored
202adc6
refactor webgpu (#16406)
sirhcm
authored
0ae957b
.const() should not ignore dtype (#16412)
chenyuxyz
authored
d72d8ee
flip offset and shape in pad and shrink (#16414)
geohot
authored
edca5df
hcq2: new multi (#16413)
nimlgen
authored
b0e49af
disable process replay by default (#16419)
sirhcm
authored
c8af163
addrspace property (#16418)
geohot
authored
c9adde7
use namespace runners (#16387)

geohot
and
sirhcm
authored
c87f343
llama: use old scale for dequant in optim (#16417)
wozeparrot
authored
36c8ff7

Commits on May 29, 2026

ci: dev matrix (#16420)
sirhcm
authored
7d38edf
add ParamArgs [pr] (#16421)
geohot
authored
1e7f1dc
move llvmspeed to benchmarks (#16422)
sirhcm
authored
a909acb
rename clangjit to clang (#16423)
sirhcm
authored
69aa205
ci: autocancel outdated pr jobs (#16424)
sirhcm
authored
6e0d526
llama: optim amax margin (#16425)
wozeparrot
authored
f86966a
viz: set label offset for asm (#16426)
Qazalin
authored
814d414
viz: addrspace little colored box (#16427)
Qazalin
authored
54cfb79
do not require clearing method_cache in viz tests (#16428)
Qazalin
authored
81bc00c
remove dead branch in UOp [PR] (#16429)
chenyuxyz
authored
fa666ce
remove vec= from UPat.cvar [PR] (#16430)
chenyuxyz
authored
2d7e5ba
llama: fix mp (#16434)
wozeparrot
authored
6787de9
full derivation of addrspace (#16433)
geohot
authored
e2a0434
hcq2: rework pm_bufferize (#16431)
nimlgen
authored
d69aca4
fix viz addrspace (#16437)
geohot
authored
9506b78
limit CONST(UNIQUE) to invalids only (#16432)
chenyuxyz
authored
b728070
ci: no fetch in backend tests (#16438)
sirhcm
authored
434cfa9
ci: macos dev matrix (#16436)
sirhcm
authored
ef50a49
fix AffineGrid fusion (#16439)
chenyuxyz
authored
8ac62b2
ci: remove duplicate op compile test (#16441)
sirhcm
authored
d943493

Commits on May 30, 2026

llama: minimize peak init mem (#16440)
wozeparrot
authored
c23652e
ci: run dsp on tinygrad[testing] (#16442)
sirhcm
authored
c377d01
python prg is pkl uops (#16443)
geohot
authored
cf55aaf
llama: zero grad this way (#16445)
wozeparrot
authored
6795c2d
llama: update local amax implementation after ParamArgs change (#16446)
Qazalin
authored
29b47a0
fix word wrap (#16450)
nimlgen
authored
2c2cb33

Commits on May 31, 2026

support negative pads in circular pad mode (#16448)
javierdejesusda
authored
989f713
minor updates from vec removal (#16456)
geohot
authored
556defa

Commits on Jun 5, 2026

Merge branch 'upstream/master' into sync-20260605
Test User
committed
03dbaa0