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
11,106 commits 280 branches 15 tags 325 MiB
Commit graph

11,106 commits

This branch
This branch
All branches
Author SHA1 Message Date
Ignacio Sica
eddafb84e5
 		Bugfix for TC=3 (#9464) 
* wrong but uses less shared

* for size 8 tc1 with devectorize in 0 loads into local before wmma and works

* improvements over tc1 devectorize

* fix tc=3

* works for handcoded tc opts

* clean bugfix tc=3

* fix

* revert changes
 2025-03-21 16:43:42 -07:00
chenyu
6da78164f9
 		assert Kernel ast.op to be Ops.SINK [pr] (#9539) 
rest of the code assumes self.ast is defined anyway
 2025-03-21 18:09:44 -04:00
chenyu
c33679c47b
 		increase size in test_multinomial_counterexample (#9540) 
should be less flaky
 2025-03-21 17:46:52 -04:00
Francis Lata
1a1087e3a0
 		cleanups on losses and dataset tests (#9538) 2025-03-21 17:03:18 -04:00
Francis Lata
8cbe4009fc
 		RetinaNet losses (#9536) 
* add sigmoid_focal_loss and l1_loss

* update ref implementation comment
 2025-03-21 15:52:54 -04:00
Francis Lata
e6389184c5
 		update comment for retinanet dataloader implementations (#9534) 
Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-03-21 15:07:45 -04:00
chenyu
ee3d313b34
 		Revert "update ruff to 0.11.2 (#9531)" (#9535) 
This reverts commit d8d65e2747.
 2025-03-21 14:52:25 -04:00
chenyu
b46b8ee15e
 		add a flag to log when beam surpassed max limit [pr] (#9533) 2025-03-21 13:37:02 -04:00
Francis Lata
eb95825eea
 		RetinaNet dataloader (#9442) 
* retinanet dataloader

* remove batch_size from generate_anchors

* refactor kits19 dataset tests

* add tests for dataloader

* fix testing setup and cleanups

* remove unused import
 2025-03-21 13:36:41 -04:00
b1tg
58206fa8a9
 		add amd llvm compiler (#9519) 
Co-authored-by: b1tg <b1tg@users.noreply.github.com>
Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-03-21 23:13:27 +08:00
chenyu
d8d65e2747
 		update ruff to 0.11.2 (#9531) 
0.11.2 fixed the false alert from 0.11.1. also pinned the version in setup for now to prevent broken CI from ruff upgrade
 2025-03-21 10:32:59 -04:00
qazal
ee3ed73ed1
 		add reorder_view matcher to scheduler [pr] (#9528) 2025-03-21 17:46:20 +08:00
George Hotz
8e555c586c
 		switch quantization to unsigned/unsigned + add Ops.REDUCE (#9527) 
* switch quantization to unsigned/unsigned + add Ops.REDUCE

* tests

* nhwc + replay pkl
 2025-03-21 17:02:37 +08:00
nimlgen
a35b0a88bf
 		am: just rename and reorder ip init funcs (#9504) 2025-03-21 15:57:32 +08:00
nimlgen
8a131ab271
 		am: allow allocations as small as a page (#9523) 
* am: fix allocs

* bettermsg

* comment

* next time
 2025-03-21 15:53:32 +08:00
Sieds Lykles
3ad3ac4d1e
 		Change dtypes.int to dtypes.ints (#9517) 2025-03-20 17:24:26 -04:00
chenyu
b9fab9b914
 		pin ruff to 0.11.0 in CI (#9520) 
0.11.1 had a bug https://github.com/astral-sh/ruff/issues/16874 that breaks ci
 2025-03-20 13:12:50 -04:00
George Hotz
3c5161b4cb
 		add validation of the bounds of Ops.INDEX (#9503) 
* add validation of the bounds of Ops.INDEX

* do mask properly

* more validation

* correct

* fix gated

* add CAST support to vmin/vmax

* fix ptx and image

* ptx no diff

* upat.index also stays

---------

Co-authored-by: qazal <qazal.software@gmail.com>
 2025-03-20 12:15:55 +08:00
qazal
0b20f91ce7
 		remove move_mask from the devectorizer (#9511) 
* remove move_mask from the devectorizer

* add (wrong) ptx

* reason

* enable index addition in PTX, we won't have the INDEX anyways

* space
 2025-03-20 11:53:12 +08:00
qazal
9302738263
 		hotfix: more consistent wgsl.py spacing + cleanups [pr] (#9515) 
* hotfix: more consistent wgsl.py spacing + cleanups [pr]

* free things up
 2025-03-20 11:07:15 +08:00
George Hotz
68053d0510
 		dsp stuff / sniff ioctls from snpe (#9490) 
* sniff ioctls from snpe

* dump input buffers

* snpe logs from dsp

* NHWC support

* knum 3

* this run?

* revert those

---------

Co-authored-by: Comma Device <device@comma.ai>
 2025-03-20 10:38:23 +08:00
qazal
2223b93338
 		add UPat.or_casted [pr] (#9513) 2025-03-20 10:08:32 +08:00
qazal
1839e8c9b3
 		place masks in INDEX for TestGatedStoreRewrite [pr] (#9512) 2025-03-20 09:46:53 +08:00
b1tg
bd731a8624
 		AMDCompiler refactor (no_comgr prereq) (#9497) 
* add amdgpu_disassemble to helpers

* refactor hip compiler

---------

Co-authored-by: b1tg <b1tg@users.noreply.github.com>
 2025-03-20 09:44:07 +08:00
geohotstan
8c0d0a122c
 		Add return_indices to max_pool (#9506) 
* wow argmax is so good

* 1 less line

* clean up and better variable names

* is this torch thing right...?

* add more tests

* slap a TODO on it

* clean ups

* prettier looking code and fix ceil mode test

* add return types and some docs

* ok that was a bad example since indices == value, just no example
 2025-03-19 15:25:37 -04:00
chenyu
189f62d44f
 		add rounding to tqdm unit scale (#9507) 
fixed `AssertionError: ' 1.00/10.0  1000it/s]' != ' 1.00/10.0  1.00kit/s]'`
 2025-03-19 12:08:46 -04:00
nimlgen
a5c971ff3a
 		am: prereqs for rdna4 1/n (#9495) 
* am: ip_ver rename for acc

* am: refactor this

* fix version

* ugh
 2025-03-19 17:14:57 +08:00
Francis Lam
1e5d9ad8f7
 		extra/gemm/max_matmul: start of custom kernels for GEMM (#6926) 
* extra/gemm/max_matmul: start of custom kernels for GEMM

* add an unoptimized FP16/FP16 MMA example

* add slow 3-stage fp16 acc example

* add correct 3-stage pipeline with unswizzled/flat smem input (slow)

* add acc fp16 example with 3 stages and swizzle (no bank conflicts)

* add max version of NV fp16_fp16_fp16

* fix up comments and removed unused code in max variations

* add start of no_xor example

* fix to account for UOps to Ops
 2025-03-19 15:04:57 +08:00
George Hotz
865f23dd7b olmoe memory usage cleanups 2025-03-19 12:28:18 +08:00
b1tg
2c87a22cf2
 		fix prg size calculation when there are adjacent mapped ranges (#9498) 
Co-authored-by: b1tg <b1tg@users.noreply.github.com>
 2025-03-19 11:55:03 +08:00
b1tg
1d71436e6a
 		use libllvm19 in ci (#9494) 
Co-authored-by: b1tg <b1tg@users.noreply.github.com>
 2025-03-19 11:53:32 +08:00
b1tg
a95b489a55
 		nanoGPT train works with tiny torch backend (#9283) 
* train_shakespeare_char.py works

* move aten.where.self_out to tiny_backend_out

* fix memory leak

* corealize in the backward_hook

* Update backend.py

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
 2025-03-19 11:51:02 +08:00
chenyu
f8976dd2eb
 		enable more webgpu tests (#9502) 
OSX has larger buffer number limit, and it supports fp16 now
 2025-03-18 23:03:54 -04:00
qazal
ae688e4103
 		simple failing test for scheduling parallel reduce [pr] (#9501) 
* simple failing test for scheduling parallel reduce [pr]

* atol
 2025-03-19 10:52:13 +08:00
leopf
e4dad99145
 		nn.state docs cleanup (#8332) 
* doc cleanup

* extension cleanup

* manual definition

* bring back accept_filename for gguf_load

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-03-18 17:16:40 -04:00
chenyu
1ea4876dfa
 		olmoe touchups (#9499) 
GlobalCounters.reset() and only validate if temperature is 0
 2025-03-18 15:25:45 -04:00
geohotstan
f7506c6c25
 		JIT OLMoE (#9396) 
* jit the forward

* might timeout, idk just send it

* this is dumb

* naive bitonic lol

* idk if this is correct, but that squeeze before is definitly not

* vectorized bitonic sort, but still slow

* yay 1 layer is correct

* alright its pretty good

* good enough

* rerun CI

* nit improve comment
 2025-03-18 14:49:02 -04:00
Ignacio Sica
5c56cac0a0
 		MI300 mfma support (#9417) 
* add f16/f32 mfma support for MI300

- add 16x16 mfma shape support for f16 with f32 acc
- add ops_python mfma emulation
- add arch to AMDRenderer

* minor cleanup

* minor cleanup

* add mfma emulation task to ci

* add back todo

* hotfix: comment

* add tc=3 job to ci
 2025-03-18 14:33:30 -03:00
hooved
5500887eed
 		improve reproducibility of WebGPU CI puppeteer test (#9496) 
* try to make CI test fail with slow JS import

* prevent race between model import and reference

* revert artificial delay in JS module import
 2025-03-18 09:27:38 -04:00
qazal
cde4fd3be3
 		do not view_left assign + elementwise sources always have a shape [pr] (#9491) 2025-03-18 17:42:51 +08:00
George Hotz
117b7a16ef
 		VALIDATE_WITH_CPU [pr] (#9488) 
* VALIDATE_WITH_CPU [pr]

* fix test
 2025-03-18 15:15:04 +08:00
qazal
935cd01f56
 		simple failing test for graph_rewrite children [pr] (#9489) 
* simple failing test for graph_rewrite children [pr]

* lint

* update too
 2025-03-18 13:07:21 +08:00
George Hotz
d20494e6d7
 		move buffer logic to Buffer [pr] (#9487) 
* move buffer logic to Buffer [pr]

* pass shape into as_typed_buffer

* pass shape into as_typed_buffer

* work

* cleaner

* fix tests
 2025-03-18 11:21:21 +08:00
qazal
3be228182f
 		unbind Tensor variables last [pr] (#9486) 
* reorder do_realize [pr]

* move merge_views

* unbind all variables at the end [pr]
 2025-03-18 09:52:01 +08:00
qazal
b44f9c409a
 		reorder do_realize [pr] (#9485) 
* reorder do_realize [pr]

* move merge_views
 2025-03-18 09:30:10 +08:00
nimlgen
a82c9332d3
 		am: rename soc21 to soc (#9482) 2025-03-18 08:54:26 +08:00
qazal
b100fc0b20
 		split the rule that uses context in scheduler simplifier [pr] (#9484) 
* split the rule that uses context in scheduler simplifier [pr]

* add
 2025-03-18 08:12:26 +08:00
Anish Umale
5e58f4b65b
 		Tiny backend test_ops fix part 3 (#9483) 
* extract straightforward things from https://github.com/tinygrad/tinygrad/pull/9302

* pass dtype and device for ones_like
 2025-03-17 18:01:51 -04:00
TJ
9fcef4d009
 		add masked_select to tensor.py (#9468) 
* add masked_select to tensor.py

* fix tests

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
 2025-03-17 16:05:36 -04:00
chenyu
4f8eac59ea
 		failed test case for threefry (#9469) 
* failed test case for threefry

not sure if it's always like this, but increment before _threefry_random_bits is incorrect. the counts should start with random numbers generated so far.

use jax to generate 20 + 20 + 10 random numbers, the first 20 + 20 matches and the last 10 are different. just moving increment after _threefry_random_bits matches the number but jit test failes

* workaround

* why is this different?

* revert those

* and that
 2025-03-17 14:52:10 -04:00