mirrors/tinygrad
2
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-24 02:14:17 +00:00
Code Issues Projects Releases Packages Wiki Activity
13,360 commits 280 branches 15 tags 325 MiB
Commit graph

13,360 commits

This branch
This branch
All branches
Author SHA1 Message Date
chenyu
beea4633fc
 		UOp.clone [pr] (#16295) 
generates the store after structure
 2026-05-20 17:47:49 -04:00
qazal
a19fa2908f
 		fix x86 nondeterminism (#16293) 2026-05-21 05:48:05 +09:00
George Hotz
58d58c1659
 		remove DEVECTORIZE (#16290) 
* remove DEVECTORIZE

* fully remove DEVECTORIZE
 2026-05-20 13:25:49 -07:00
wozeparrot
825f30bf18
 		llama: apply_grad saves memory (#16275) 2026-05-20 13:14:06 -07:00
nimlgen
a88feef40f
 		hcq2: cleanups (#16278) 
* s

* simpler

* simler
 2026-05-20 21:48:50 +03:00
Philipp Braun
a01d5918af
 		fix: qlinearconv quant params (#16234) 
* fix: qlinearconv quant params

* fix: simplify reshape

---------

Co-authored-by: Philipp Braun <braunphilipp@users.noreply.github.com>
 2026-05-20 11:31:41 -07:00
George Hotz
19535df53c
 		enable broadcasting in _shape (#16285) 2026-05-20 11:21:51 -07:00
chenyu
4dbe6a2ee7
 		remove _force_unique from Tensor init (#16277) 2026-05-20 14:13:05 -04:00
Christopher Bradford
fe2d8d1ecf
 		filter by base_class in pci_scan_bus on macOS (#16282) 
The Linux path of pci_scan_bus reads /sys/bus/pci/devices/.../class and
skips devices whose base class doesn't match. The macOS (IOKit) path
appended every IOPCIDevice unconditionally, so callers that supplied
base_class to narrow down to e.g. display devices would also get the
audio companion function of a multifunction GPU.

Concretely, an NVIDIA RTX Pro 6000 Blackwell exposes:
  10de:2bb1  class 0x030000 (display)
  10de:22e8  class 0x040300 (multimedia audio)

A PROBE for base_class=3 returned both. With the sorted() at the end of
pci_scan_bus, 22e8 (audio) came first, so the NV runtime picked the
audio function as device 0 and stalled on RESIZE_BAR.

This mirrors the Linux filter on line 70 using the existing read_prop
helper.

Co-authored-by: Christopher Bradford <christopher.bradford@joby.aero>
 2026-05-20 20:09:35 +03:00
qazal
1e0fffe256
 		fused ce llama kernel in UOps (#16263) 
* work

* using uops

* delete things

* work

* work

* higher level uops

* cleanups
 2026-05-20 19:45:28 +09:00
chenyu
e1715b3b92
 		extent jit const error to deviceless inputs (#16276) 2026-05-20 02:02:45 -04:00
chenyu
170b857da9
 		clean up deviceless const _buffer (#16274) 
process on CPU similar to multi
 2026-05-19 22:47:45 -04:00
chenyu
7af7b6703a
 		relax policy ASSERT_MIN_STEP_TIME to 3.2 (#16273) 2026-05-19 22:29:09 -04:00
chenyu
188d7ec15e
 		clone can take device (#16271) 
useful to materialize const on a specific device
 2026-05-19 21:29:27 -04:00
wozeparrot
361553c0a8
 		llama: match flat_llama with model_train (#16269) 2026-05-19 17:25:56 -07:00
George Hotz
da7414d6dc
 		fix RUN_PICKLE and test it (#16272) 
* add test for openpilot RUN_PICKLE

* fix RUN_PICKLE and test it
 2026-05-19 17:00:25 -07:00
George Hotz
55515747b7
 		Remove Ops.VCONST (#16267) 
* start removing vconst

* remove a lot of vconst

* const folding + strict ordering

* update tests

* spec from minigen

* move that
 2026-05-19 16:35:24 -07:00
Christopher Milan
7cdd9cbdeb
 		PYTHONREMU: V_CVT_PK_BF8_F32 saturation (#16268) 2026-05-19 19:29:59 -04:00
Christopher Milan
bb2a51f1ea
 		fix mypy mockgpu and add tinygrad.renderer.isa to packages (#16265) 2026-05-19 16:45:03 -04:00
chenyu
890b731b1e
 		more prerequisuite test changed for deviceless const (#16264) 2026-05-19 15:43:45 -04:00
ttomsa
aa1e59ab97
 		X86 with Ops.INS (#14873) 
* draft

* cleanup test_encodings

* cleanup test_isel

* model flag state and support rematerialization

* woops

* add vbroadcastss instruction

* don't fuse load if used multiple times in src

* add movabs instruction and fix idiv

* fixes

* add x86 backend to tests

* float16 fix

* rm TwoAddress2nd

* add BARRIER

* test windows ci

* yup isel fixes the mask stuff too and its beautiful

* add cmoves to the spec

* support storing imms

* no TUPLE_ORDER, breaks tests

* fix remaining seg faults

* add float max

* always fuse index

* minor

* fix DEFINE_VAR/SPECIAL and enable multithreading

* linter

* more linter

* more

* more

* more

* let's try this

* perhaps

* start new scheduler

* more scheduling info

* cleaner shuffle functions

* fixup isel tests

* skip bounds check when NOOPs exist

* skip inf rewrite tests

* fix const tag hack and add x86ops to _shape

* fix

* skip a few tests

* func arg order independent from op value

* x86 goes in own linearize

* switch to PARAM

* more

* add min x86op and neg in decomps

* do mulacc in isel

* use def_reg in test_encodings

* enable emulated int64 tests

* how much does this fix

* Ops becomes OpType

* fix

* rm noqa

* rm machine scheduler stuff

* and this

* allow for extending enums and move X86Ops out of uop

* fix imports

* rm X86GroupOp from ops.py

* spacing

* tell mypy to shut up

* more linter

* add x86op test

* allow set[X86Ops] in upat

* move NOOPs to pre_isel_matcher and rm NOOP from spec

* more asserts

* also this

* cleanup encode

* simplify live range

* fix idiv

* add Ops.INS to x86

* more changes

* more changes

* more changes

* fix

* fix

* fix

* fix

* print formatted assembly

* fix 8bit idiv?

* oops

* enable float16  and unaligned vector load/store

* actually no

* move x86 tests

* no more bool cast

* fix

* linter

* linter

* move X86Ops to x86.py

* fix vpbroadcast

* cleanups

* linter

* print correct reg names

* canonical max

* move max/min and add test

* support float16 vector load/store

* rm bad rewrite

* vpsrldq can't access memory

* regalloc takes renderer

* enable vector load/store on all dtypes

* more isel tests

* rm this for now

* a lot better

* fix

* fix

* fix

* deal with flags correctly

* fix

* enable gep noop rule

* fix

* fix

* fix

* add callee saved registers

* use Ops.CONST instead of X86Ops.IMM

* fix

* enable TUPLE_ORDER

* fix

* rm x86 code in linearizer

* fix

* fix

* fix

* move isa rewrites to codegen

* fix

* fix

* skip test_linearizer.py

* skip more tests

* fix

* fix for idiv/mod changes

* fix

* don't use fmadd if it duplicates fused op

* hacky

* fix

* cleanups

* cleanups

* fix

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
 2026-05-19 12:42:54 -07:00
George Hotz
b2e8102209 25000 lines for x86 backend 2026-05-19 11:27:41 -07:00
Sachith Shetty
74567c1958
 		fix: pass input device to ONNX helper internal tensors (#16242) 
* fix: pass input device to onnx methods internal tensors

* test: onnx helper internal tensors use input device
 2026-05-19 11:16:33 -07:00
Christopher Milan
a178301dbe
 		PYTHONREMU: fix CDNA VOP3 conditional writes (#16258) 2026-05-19 13:31:31 -04:00
nimlgen
b3dcf8f452
 		hcq2: split into schedule/realize (#16216) 
* hcq2: split into schedule/realize

* missing

* x

* f

* clean

* cleaner

* x

* x

* x

* x

* x
 2026-05-19 16:40:17 +03:00
qazal
e4350e7de9
 		set hipcc mac docker to 7.1 (#16261) 
* set hipcc mac docker to 7.1

* pull from amd
 2026-05-19 21:30:39 +09:00
George Hotz
a120709671
 		tighten shape spec for broadcasting (#16206) 
* tighten shape spec for broadcasting

* use IndexError, not ValueError

* needs size
 2026-05-18 22:12:04 -07:00
George Hotz
3f2d401464
 		all tests pass with NOOPT=1 (#16257) 
* all tests pass with NOOPT=1

* fix a few more

* noopt 100% pass

* noopt 100% pass
 2026-05-18 20:39:51 -07:00
chenyu
e694d7f222
 		more deviceless const prerequisites [pr] (#16256) 
* more deviceless const prerequisites [pr]

* remove that

* arange.contiguous -> arange.clone in tests

arange will become deviceless const soon, update tests where it needs to be a buffer
 2026-05-18 23:14:12 -04:00
chenyu
c1076ed56c
 		Tensor.device and UOp.device can be None (#16255) 2026-05-18 22:08:10 -04:00
wozeparrot
a3d59faef6
 		llama: don't save weight (#16252) 2026-05-18 17:05:45 -07:00
qazal
18b102f355
 		llama: also use 7.1 comgr, update startup_walltime.sh (#16253) 2026-05-19 08:59:02 +09:00
chenyu
d532b4f533
 		multi alu with deviceless const (#16251) 2026-05-18 19:31:53 -04:00
qazal
98b8a2b407
 		llama: use hipcc 7.1 version (#16250) 2026-05-19 08:09:57 +09:00
Christopher Milan
7515824a6d
 		ci: actually use clang-20, enable bfloat16 (#16249) 2026-05-18 19:06:43 -04:00
chenyu
754344087a
 		assign for deviceless const source (#16248) 2026-05-18 17:39:53 -04:00
chenyu
73e6b4963b
 		to and shard is noop for deviceless uop (#16247) 2026-05-18 16:11:10 -04:00
Christopher Milan
50481ec9b4
 		cl: check for cl_khr_fp64 (#16246) 2026-05-18 14:42:43 -04:00
chenyu
db639ebe3e
 		deviceless const from UOp (#16243) 2026-05-18 14:14:12 -04:00
qazal
bfb2d1f89a
 		Revert "fp8 gemm speedup (#16236)" (#16245) 
This reverts commit d95bf394e1.
 2026-05-19 02:01:44 +09:00
chenyu
5ae4dbd599
 		make slow tests faster (#16244) 2026-05-18 11:42:02 -04:00
chenyu
981c12182f
 		remove requires_grad= in tinygrad/ (#16241) 2026-05-17 16:55:37 -04:00
chenyu
fcdd1af880
 		remove Tensor.detach override [pr] (#16239) 2026-05-16 23:58:12 -04:00
chenyu
dcee90aa3f
 		remove requires_grad use in extra/examples (#16238) 
except the ones fed into optimizer
 2026-05-16 18:40:26 -04:00
chenyu
8631b6f17d
 		remove use of requires_grad in test/ (#16237) 2026-05-16 17:21:07 -04:00
qazal
d95bf394e1
 		fp8 gemm speedup (#16236) 
* add asm_gemm option

* milestone

* work

* edit

* only the fast kernel

* diff
 2026-05-17 04:58:28 +09:00
chenyu
0ddc50d050
 		do not gate backward on requires_grad (#16230) 
DETACH is filtered in _deepwalk. instead of None, it gets 0 grad now
 2026-05-16 12:29:49 -04:00
nimlgen
bef5f717bc
 		fix nolocals and beam (#16232) 2026-05-16 18:09:19 +03:00
qazal
ebcb7b7cc0
 		fp8 gemm tests with scale args (#16231) 
* update atol

* update fp8 path

* more work

* update profile.sh
 2026-05-16 20:47:58 +09:00
nimlgen
e575f778f9
 		move debug prints (#16218) 
* move debug prints

* x
 2026-05-16 13:57:34 +03:00