Sync: `tinygrad/tinygrad:master` → `sunnypilot/tinygrad:master` - #9

Merged

Discountchubbs merged 795 commits into

Aug 10, 2026

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

Commits on Jul 23, 2026

AND OR XOR SHL SHR cannot have float operands [PR] (#17164)
chenyuxyz
authored
3e031d6
don't use promo_lattice in fast_idiv [PR] (#17165)
chenyuxyz
authored
0f9edb0

Commits on Jul 24, 2026

support loops in nir (kimi) (#17166)
geohot
authored
bf5989e
arange upcast to int64 with big N (#17167)
chenyuxyz
authored
f5d9c31
corner cases from weak const branch (#17173)
chenyuxyz
authored
a292dab
llm: fix generated tokens in usage accounting (#17174)
geohot
authored
08eceb0
don't support Tensor(list(np.array)) [pr] (#17175)
chenyuxyz
authored
2dd2e3d
hcq2 cleanups (#17162)
nimlgen
authored
709d7b2
hcq2 cleanup 2 (#17177)
nimlgen
authored
0ecef21
cpu prep for hcq2 (#17153)
nimlgen
authored
017d1ab
remove unneeded expand [PR] (#17176)
chenyuxyz
authored
40bfd84
gptoss: moe gemm kernels (#17178)
wozeparrot
authored
c9b60ca
add x86 loops (#17179)
GabrielNakamoto
authored
fcb1d3f
simpler _split_cumalu [PR] (#17180)
chenyuxyz
authored
e0b30b7
clean up unbroadcast [PR] (#17182)
chenyuxyz
authored
86b42d7
data_src concept in run_rangeify [pr] (#17183)
chenyuxyz
authored
38c04fa
convert COPY -> STORE early (#17172)
geohot
authored
dcad119
cleanups because copy is not allowed in rangeify (kimi) (#17186)
geohot
authored
f65001e
improve call for viz (#17187)
geohot
authored
a346e2e

Commits on Jul 25, 2026

update create_non_native_float_pats [PR] (#17191)
chenyuxyz
authored
de3508e
Invalid is bool, put weakint in promo lattice [pr] (#17188)
chenyuxyz
authored
0a6125e
standardize Program class (#17189)
sirhcm
authored
9fdaa4b
simpler lower_weakint_node [PR] (#17193)
chenyuxyz
authored
d923263
less wrong calculate_storage_offset (#17194)
chenyuxyz
authored
4c58b26
fix torch backend as_strided (#17195)
chenyuxyz
authored
8a10892
add one line viz mention (#17198)
Qazalin
authored
732e6bd
checked cast in torch backend unwrap (#17199)
chenyuxyz
authored
9f78504
fix torch backend batchnorm backward (#17201)
chenyuxyz
authored
983ad3b
refactor mlperf optim (#17200)
wozeparrot
authored
f0117e9
put weakint in dtypes.weaks [pr] (#17202)
chenyuxyz
authored
ee2ccb1
derive torch backend dispatch from the aten schema (#17203)
chenyuxyz
authored
f902513
promo (uint64, int) -> weakfloat like JAX [pr] (#17204)
chenyuxyz
authored
74c2121
delete _index_to_concrete_int [pr] (#17205)
chenyuxyz
authored
492dc6d
failing batch norm test (#17206)
chenyuxyz
authored
076b37e
hcq2, cpu is hcq2-ish (#17197)
nimlgen
authored
3946df7

Commits on Jul 26, 2026

fix torch backend out= into a view (#17210)
chenyuxyz
authored
a60b5f7
weak frompy prerequisite [PR] (#17213)
chenyuxyz
authored
eb88905
slight weak behavior tweak and cleanups [pr] (#17214)
chenyuxyz
authored
a96974e
minor normalize cleanup [pr] (#17215)
chenyuxyz
authored
acc2374
SHR/SHL are Broadcastable [pr] (#17216)
chenyuxyz
authored
5d1aa84
nv: set lower interleave level to reduce GPU hogging (ai slop) (#15518)

authored
ac12914
Revert "nv: set lower interleave level to reduce GPU hogging (ai slop) (#15518)" (#17218)
geohot
authored
960430a
fix nvrtc_check helper used for jitlink call (#16362)
IntendedConsequence
authored
d70a134
fix ValueError in UOp.axis for shard reshape crossing boundary (#16547)

mgisabsolute
and
geohot
authored
79c07a3
realize weak is no-op [pr] (#17219)
chenyuxyz
authored
a8d5109
clean up mixin cos and exp [PR] (#17221)
chenyuxyz
authored
94dad3d
fix _prepare_jit_inputs for weak [pr] (#17222)
chenyuxyz
authored
4b67605
hcq2: amd indirect (#17220)
nimlgen
authored
97a2265
speedy hcq2 (#17225)
nimlgen
authored
165f062
simpler minimum and copysign [PR] (#17226)
chenyuxyz
authored
afb25a6

Commits on Jul 27, 2026

fix python_alu inf (#17227)
chenyuxyz
authored
456b5b5
skip test_float_to_fp8e4m3_extreme_values (#17228)
chenyuxyz
authored
0f98212
webgpu failing test for duplicate PARAM in CALL [pr] (#17230)
Qazalin
authored
95e3b00
test updates from weak flip (#17231)
chenyuxyz
authored
f45fc4c
validate json output of viz.cli in CI (#17232)
Qazalin
authored
19c4d73
gptoss: split no-wd params (#17233)
wozeparrot
authored
0569744
polish the viz readme (#17236)
Qazalin
authored
a3ca1e5
flip from_py to use weak dtypes [pr] (#17229)
chenyuxyz
authored
818a892
hcq2: do not cache beam (#17234)
nimlgen
authored
8eaeede
fix shard axis through symbolic reshape (#17238)
b1tg
authored
bdbb1d7
run_linear in external_test_gpu_crash (#17239)
nimlgen
authored
8b9ef15
don't match strong typed const in UPat [pr] (#17240)
chenyuxyz
authored
896afad
make Tensor(None) weakfloat (#17241)
chenyuxyz
authored
0bb36c9
clean up pm_lower_weak [pr] (#17243)
chenyuxyz
authored
550225f

Commits on Jul 28, 2026

update pm_long_decomp [pr] (#17245)
chenyuxyz
authored
654d475
don't match const dtype in UPat [pr] (#17247)
chenyuxyz
authored
f837ca3
64-bit UOp.variable support (#17246)
sirhcm
authored
ab8fb19
Revert "64-bit UOp.variable support (#17246)" (#17248)
chenyuxyz
authored
4b7022e
pm_long_decomp doesn't depend on operand dtype [pr] (#17249)
chenyuxyz
authored
1380d6c
test for extra copy in allreduce_cast (#16740)
Qazalin
authored
e3b3eea
arange stack regression test (#17250)
Qazalin
authored
05e4727
UOp.const(dtype=None) infers from from_py [pr] (#17253)
chenyuxyz
authored
37cf159
make the github actions runners generic for gitea (#17254)
geohot
authored
a9ad080
delete explicit casts [pr] (#17255)
chenyuxyz
authored
c9e1154
remove some dtype= when construct UOp [pr] (#17257)
chenyuxyz
authored
749e002
preserve typed ranges through reshape (#17259)
drukpa1455
authored
d1c3ae0
switch _device_num to AxisType.DEVICE range (kimi) (#17252)
geohot
authored
7d48926
delete dsp_pm_late (#17260)
geohot
authored
9ce65b7
CAPTURE_PROCESS_REPLAY=0 default chaging test [PR] (#17261)
chenyuxyz
authored
fde3a8f
remove unneeded default args in renderers (#17262)
geohot
authored
0cdddf3
rename CPU_COUNT to NUM_CPU_THREADS with cgroup awareness (#17263)

geohot
and
teeny-runner
authored
755dfb2
DEFAULT_FLOAT/DEFAULT_INT ContextVar [pr] (#17265)
chenyuxyz
authored
23e9e76
update dtypes.md (#17266)
chenyuxyz
authored
f11f884
add device range as src[1] to multi (kimi) (#17264)
geohot
authored
1757067
rename MULTI to UNSHARD (#17267)
geohot
authored
57ae1bc

Commits on Jul 29, 2026

add UNSHARD to spec (#17269)
geohot
authored
a17387d
qwen3.6 for 27b and 35b-a3b (#17268)
chenyuxyz
authored
291ee43
apply shrink bugfix for 3.11 (#17271)
geohot
authored
dd16d5a
remove no-op explicit dtype= or cast [PR] (#17276)
chenyuxyz
authored
2f8f2d2
make .barrier implicit (kimi) (#17275)
geohot
authored
451120c
pm_reduce_collapse fix for re enabling stack for cat of same shape (codex) (#17251)
Qazalin
authored
e684fcc
fix smu reset for kernel >= 7 (#17277)
geohot
authored
527e573
test permuted input in custom Ops.PROGRAM test (#17279)
Qazalin
authored
6ea7d36
fix add_raw_barrier [pr] (#17278)
chenyuxyz
authored
3803f15
Patch half precision ops for older gpu archs (#17274)
NoahSchiro
authored
3df1b07
enable alloc_fragment support with UNSHARD (kimi) (#17272)
geohot
authored
bd296a7
weak const cleanups [PR] (#17282)
chenyuxyz
authored
6c2b9fa
rename LOOP -> WEAK and STRONGLOOP -> LOOP (#17283)
geohot
authored
52c9e5a
test case for weakfloat cast to weakint INDEX (#17287)
chenyuxyz
authored
dd86a30
64-bit UOp.variable support, try 2 (#17256)
sirhcm
authored
aab51fb
support 2d on UNSHARD (kimi) (#17285)
geohot
authored
b30c7e0
hcq2: use stack (#17286)
nimlgen
authored
d4ba8b6
fix c0*x<c1 symbolic [pr] (#17290)
chenyuxyz
authored
027907a
improve fragment example + index unshard (kimi) (#17288)
geohot
authored
138676a
generic c0*x<c1 [pr] (#17291)
chenyuxyz
authored
fd912b3
qcom: match cl for SP_CS_INSTR_SIZE (#17289)
sirhcm
authored
060f447

Commits on Jul 30, 2026

nicer TinyJit decorator (#17293)
sirhcm
authored
bfc9fc6
fix do_devectorize dtype (#17295)
chenyuxyz
authored
d52ef30
update rangeify comment for mop after AFTER (#17298)
Qazalin
authored
aba5ba4
spec tests for movement ops before custom_kernel (#17299)
Qazalin
authored
7b7b9c1
hcq2: less graph_rewrites (#17300)
nimlgen
authored
3f6f0a1
gptoss moe routing (#17284)
wozeparrot
authored
b5a2a56
kimi delta attention (#17281)
b1tg
authored
e25bf77
rewrite writes new dtype with dtype_from_uop [pr] (#17302)
chenyuxyz
authored
417245a
Revert "rewrite writes new dtype with dtype_from_uop [pr] (#17302)" (#17303)
chenyuxyz
authored
d05a3e6
rewrite writes new dtype with dtype_from_uop try 2 [pr] (#17305)
chenyuxyz
authored
341c4ed
update dtype_from_uop for SHR/SHL (#17307)
chenyuxyz
authored
b488cc7
update openpilot benchmarks (#17304)
sirhcm
authored
da15c43
ftdi reset chestnut before running comma benchmark (#17306)
sirhcm
authored
b290372
broadcast doesn't cast CONST [pr] (#17308)
chenyuxyz
authored
ce500c1
fix weak for image gate fusion [pr] (#17310)
chenyuxyz
authored
fe8ece7

Commits on Jul 31, 2026

Add softmin (#17292)
NoahSchiro
authored
f2c2f44
adjust lower weak order [pr] (#17314)
chenyuxyz
authored
6608b9d
cleanup gemm fragment + add store unshard (#17313)
geohot
authored
d65ea46
fp4 asm gemm 6+ pflops (#17315)
Qazalin
authored
a8c1e89
benchmark comma big model (#17312)
sirhcm
authored
13452b3
add mxfp4 quantize and layout kernels (#17320)
Qazalin
authored
0a3325f
llama with MXFP4 (#17321)
Qazalin
authored
f7964ac
gptoss: grouped moe (#17322)
wozeparrot
authored
93b74c7
failing test for unbound _device_num err in BEAM (#17326)
Qazalin
authored
6d2700f
coalesce ints (#17323)
nimlgen
authored
0c4bfae
gptoss ci (#17325)
wozeparrot
authored
4f5cadd
hcq2 faster schedule (#17324)
nimlgen
authored
155b84e
const(dtype, b) -> const(b, dtype) [PR] (#17328)
chenyuxyz
authored
8e2f175
remove shape= from UOp.const [PR] (#17331)
chenyuxyz
authored
7f4dbb8
hcq2: fix ib reuse (#17330)
nimlgen
authored
1095bbe
don't reset chestnut in benchmark (#17333)
sirhcm
authored
b95bd5b
viz: prep for faster cli DEBUG=3 (#17334)
Qazalin
authored
a11ee26
const(value, dtype) -> const(value).cast(dtype) in tests (#17335)
chenyuxyz
authored
ad24750
dtype_from_uop(INS) is None [PR] (#17337)
chenyuxyz
authored
8dc225e
fix sym_infer for CAST (#17338)
chenyuxyz
authored
2774332
tc: don't allow reduce over output dims (#17340)
sirhcm
authored
85ced44
heuristics: try multiple TC axes (#17341)
sirhcm
authored
15d5152
__int__ and __float__ work for weak (#17342)
chenyuxyz
authored
8509891

Commits on Aug 1, 2026

remove UOp.val (#17345)
geohot
authored
a88f832
ci: split macos unit test into metal and mock runners (#17346)
geohot
authored
099d69f
use .val to access the value of Ops.CONST (#17347)
geohot
authored
9082ece
support new USB vendor ID (#17348)
adeebshihadeh
authored
20b8ecf
more arg -> val (#17349)
geohot
authored
5a1c641
add _device_num back to ast.variables (kimi) (#17327)
Qazalin
authored
161783d
more const arg -> val (#17350)
chenyuxyz
authored
b502fc1
UOp.is_invalid [PR] (#17351)
chenyuxyz
authored
6c0ec39
hcq2: faster replace (#17353)
nimlgen
authored
665822a
CONST related cleanups [pr] (#17352)
chenyuxyz
authored
8e524ca
gptoss optim fixes (#17356)
wozeparrot
authored
98b700b
merge devectorize + indexing (#17354)
geohot
authored
15c936d
faster devectorizer with one line (#17360)
geohot
authored
fb607fb

Commits on Aug 2, 2026

Commits on Aug 3, 2026

no hard coded dtype int for rangeify debuf [PR] (#17373)
chenyuxyz
authored
23c7813
deflake test_hcq with MOCKGPU (#17370)
chenyuxyz
authored
314df72
hcq2: inputs table (#17374)
nimlgen
authored
e22935c
llama: refactor amax stuff and skip in fp4 (#17375)
Qazalin
authored
be5f62d
hcq2: epoch (#17376)
nimlgen
authored
7c1ce50
gptoss: fix moe routing (#17377)
wozeparrot
authored
3331944
MAX_LINE_COUNT=26000 (#17378)
chenyuxyz
authored
a2385ae
usb: wait for PCIe link after power on (#17380)
YassineYousfi
authored
104ee90
improve threefry codegen [pr] (#17379)
chenyuxyz
authored
33755a3
scalar ALU index fix + llm: preserve_thinking (#17381)
geohot
authored
c2625c7
hotfix: revert test_scalar_alu_index, violates spec
geohot
committed
87289a7
llm: update test_llm_server tests (#17382)
geohot
authored
3cb786f

Commits on Aug 4, 2026

bitcast in python for _bits_to_rand [PR] (#17383)
chenyuxyz
authored
67dc02d
llm: bugfixes + warmup (#17384)
geohot
authored
c21a552
use python bitcast in fold_bitcast [PR] (#17385)
chenyuxyz
authored
0170a30
const_like cleanups [PR] (#17386)
chenyuxyz
authored
13fff4f
more custom kernel contig input edge case tests (#17387)
Qazalin
authored
c9cd44b
llama: accurate mxfp4 mfu (#17388)
Qazalin
authored
f993228
Revert "usb: wait for PCIe link after power on (#17380)"
geohot
committed
568bfb6
support symbolic shapes in allreduce (#17364)
b1tg
authored
0796853
dtype_from_uop for CUSTOM, CUSTOMI, PYLITERAL [PR] (#17390)
chenyuxyz
authored
6eedca5
update dtype_from_uop for IMAGE INDEX [PR] (#17391)
chenyuxyz
authored
0db63e1
keep weak consts weak in symbolic [PR] (#17392)
chenyuxyz
authored
85e9440
fa: fix dq hazard with D=64 (#17393)
wozeparrot
authored
80d2073
add test_eye + color INDEX (#17394)
geohot
authored
f489f4b
more weak const without cast in const_like [PR] (#17395)
chenyuxyz
authored
7b6d2dd
fix _min_max for CAST from float to int [pr] (#17396)
chenyuxyz
authored
c1a10e0
fix pow on extreme inputs (#17397)
chenyuxyz
authored
d79772f
use weak 0 in convert_pad_to_where_to_keep_behavior_local [pr] (#17398)
chenyuxyz
authored
f295f9f

Commits on Aug 5, 2026

use check_schedule in tests where possible (#17400)
geohot
authored
6122b3c
update minimum to not create strong type const [PR] (#17401)
chenyuxyz
authored
3eab809
add new schedule tests + format better (#17402)
geohot
authored
e1f4268
remove invalid special case in cast [PR] (#17405)
chenyuxyz
authored
9b508df
kill nvidia pids at benchmarks start (#17406)
sirhcm
authored
de57be1
more KernelCountException (#17407)
geohot
authored
46f0003
don't cast weak in _broadcasted [pr] (#17408)
chenyuxyz
authored
b45058b
fix AMD WMMA emulation and test in CI (#17184)

GabrielNakamoto
and
geohot
authored
77e124e
Revert "don't cast weak in _broadcasted [pr] (#17408)" (#17409)
chenyuxyz
authored
3bf9e70
hcq2 benchmark (#17235)
nimlgen
authored
874d331
viz/cli: faster and more complete rewrites print (#17411)
Qazalin
authored
ad32bd2
remove debug from test (#17410)
nimlgen
authored
5b0b68e
regression test for bert nan with weak (#17412)
chenyuxyz
authored
6cb419b
hcq2: cleaner (#17413)
nimlgen
authored
9b27ea8
chat: display reasoning_content from streamed responses (#17414)
geohot
authored
2cce85a
move callify into tensor (#17416)
geohot
authored
757a727
fix weak cast to strong dtype [pr] (#17418)
chenyuxyz
authored
c2f1e5a
few weak and decomp tweaks [PR] (#17419)
chenyuxyz
authored
07ac911
merge track_rewrites and profile_matches into rewrite_group [PR] (#17420)
geohot
authored
581bfdd
add benchmark_llm script
geohot
committed
a8a8030
fix slice + non contig kernels (#17423)
geohot
authored
470c032
split pm_fold_cast_const [PR] (#17425)
chenyuxyz
authored
d726e5f

Commits on Aug 6, 2026

scope variable names inside CALLs (#17424)
sirhcm
authored
be25207
remove some pm_fold_cast_const [pr] (#17426)
chenyuxyz
authored
d51e55a
revert wrong custom kernel fix (#17427)
geohot
authored
b4372df
move weak function and pm to uop/weak [PR] (#17429)
chenyuxyz
authored
7a9cd8e
one less strong dtype const in symbolic [pr] (#17431)
chenyuxyz
authored
969df86
fix postopt symbolic [pr] (#17433)
chenyuxyz
authored
28e6ef6
llama: custom quantize_mxfp4+transpose kernel (codex) (#17434)
Qazalin
authored
f258708
test MXFP4 llama without hipcc (#17435)
Qazalin
authored
9636dd1
hcq2: fence inputs (#17436)
nimlgen
authored
46230e9
fa: swa support (#17367)
wozeparrot
authored
1fd6b10

Commits on Aug 7, 2026

update linear interpolate to use int math for indices (#17441)
chenyuxyz
authored
d8cbc11
truncate float in DType.const [pr] (#17439)
chenyuxyz
authored
9020a88
fix f2f from fp8e5m2fnuz to half (#17442)
chenyuxyz
authored
28195d5
remove contiguous from custom_kernel (#17149)

Qazalin
and
geohot
authored
f253c44
viz: collapse PROGRAM nodes like CALL (codex) (#17438)

geohot
and
Qazalin
authored
1858f1f
fix var of large half input (#17444)
chenyuxyz
authored
baa6148
fix prod gradients at zero (#17404)
Robertboy18
authored
0c96cdc
clean up reduce MUL gradient (#17447)
chenyuxyz
authored
fca695a
c0+x<c1 -> x < c1-c0 is ints only [pr] (#17448)
chenyuxyz
authored
73e670c
cpu: fix eintr (#17449)
nimlgen
authored
b6189db
gptoss: fix sharded invalids (#17450)
wozeparrot
authored
1827ec5
fix cast to float _min_max [pr] (#17451)
chenyuxyz
authored
f76422b
move pm_fold_cast_const [pr] (#17453)
chenyuxyz
authored
59b88ea
better _drop_valid_stmts [pr] (#17454)
chenyuxyz
authored
4a3b8f6
fix ci emu (gpt) (#17437)
nimlgen
authored
4c206a5
nolocals supports variables (#17457)
sirhcm
authored
c0d2f9a

Commits on Aug 8, 2026

Commits on Aug 9, 2026

hcq2: fix beam (#17467)
nimlgen
authored
8c8b43d

Commits on Aug 10, 2026