geohot/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests Actions Activity
14,527 Commits 4 Branches 15 Tags
112e50ec3a1a2f740e86ff2f5e70dc38abbcc8f0
Commit Graph
14527 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
chenyuandGitHub 112e50ec3a fix onnx.py type annotations (#17811) 2026-08-28 13:34:43 -04:00
chenyuandGitHub ae63f23557 fix onnx MaxUnpool (#17809) 
pads can be tuple
 2026-08-28 13:13:47 -04:00
chenyuandGitHub 287679a88a delete X86 not used codes [PR] (#17807) 2026-08-28 12:50:50 -04:00
Teddy TennantandGitHub 2eacd4fa68 fix Conv2d same padding with per axis dilation (#17805) 2026-08-28 12:44:43 -04:00
Joel LisenbyandGitHub 7fdc58b1cc add support for UD quants (#17806) 2026-08-28 09:38:34 -07:00
chenyuandGitHub cb857a5a40 remove X86 ymm thingy (#17803) 2026-08-28 10:18:10 -04:00
chenyuandGitHub 6083de5cff fix nan compares (#17801) 
* fix nan compares

* fix
 2026-08-28 09:17:18 -04:00
nimlgenandGitHub 55c8e18930 hcq2 spec2 (#17798) 2026-08-28 12:51:55 +03:00
nimlgenandGitHub 76dce1eb8d tiny hcq2 changes (#17797) 
* tiny hcq2 changes

* x

* x
 2026-08-28 11:29:00 +03:00
qazalandGitHub 6242b09066 cleaner mxfp4 gemm prelude (#17796) 
* cleaner mxfp4 prelude

* rename sgprs

* min diff
 2026-08-28 15:59:35 +09:00
wozeparrotandGitHub 92d92c70a9 gptoss: faster grad handling (#17795) 2026-08-27 22:06:33 -07:00
Teddy TennantandGitHub 26c4b6319e fix interpolate with align_corners and an output size of 1 (#17790) 2026-08-27 19:17:05 -07:00
qazalandGitHub 53cad325c7 copy movement op tests construct copies (#17792) 
* copy movement op tests construct copies

* hop 3 gpus

* assert the amount of bytes moved

* work

* arange 4, everyone copies 16B first
 2026-08-28 11:09:11 +09:00
sirhcmandGitHub 4456e62343 ci: skip apt steps if packages are already installed (#17794) 2026-08-27 22:06:08 -04:00
George HotzandGitHub 39d7fd0981 axis arg is dead (#17793) 2026-08-27 18:41:09 -07:00
geohot 4e6bdac412 hotfix: disable HCQ2 2026-08-27 18:23:52 -07:00
chenyuandGitHub bc485d271f fix CMPLT between float16 and const 0 (#17791) 
* fix CMPLT between float16 and const 0

* fix
 2026-08-27 21:14:34 -04:00
George HotzandGitHub 6b361e4de7 move shape into arg for param/buffer (#17783) 
* move shape into arg for param/buffer

* no param_from_shape

* drop gratuitous syntax changes

* image is a in-graph view, folded into the param arg at render; drop dead multi param sharding

* view_as helper, simpler resolve_function, spec update

* spec: param/buffer are flat storage, no shape input

* image dims live in the param arg from transform_to_image; tighten kernel graph spec

* kernel graph spec: only RESHAPE/SHRINK over storage values, not all movement

* kernel graph: call args are storage, not views (pm_no_view_args); assert in spec

* strip views at the kernel graph level (pm_no_views), move into rangeify

* touchups
 2026-08-27 18:13:24 -07:00
chenyuandGitHub 38e0fe103b fix where with float64 branches (#17789) 
* fix where with float64 branches

* fix
 2026-08-27 21:03:43 -04:00
sirhcmandGitHub e6ad2907a8 onnxruntime respects NUM_CPU_THREADS (#17788) 2026-08-27 20:11:27 -04:00
chenyuandGitHub fede358811 fix fancy indexing with uint8 index (#17787) 
* fix fancy indexing with uint8 index

* fix
 2026-08-27 19:59:27 -04:00
sirhcmandGitHub f06832bf6f test llm with --no_chat_template (#17785) 2026-08-27 19:01:55 -04:00
geohot ee3161e924 hotfix: decrease dims in test_attention to get below the 90s limit 2026-08-27 14:44:06 -07:00
geohot 38fa0643cf hotfix: update max test time from 300 -> 90 2026-08-27 14:27:22 -07:00
George HotzandGitHub f267a49639 UOp.param accepts single sint as shape (int,) (#17784) 
* UOp.param accepts single sint as shape (int,)

Change UOp.param signature from shape:tuple[sint, ...]|None to shape:tuple[sint, ...]|sint|None.
A single sint means 1D shape (sint,). Update all callers from (n,) to n syntax.
No param_from_shape — multi-dim shapes stay as tuples.

* use single int syntax in hcq2 copy_with_kernel
 2026-08-27 14:07:33 -07:00
sirhcmandGitHub 14cbecef8f autogen: use deb for comgr 2 (#17782) 2026-08-27 16:32:16 -04:00
chenyuandGitHub 54d69f439f fix test_assign for no subbuffer backends (#17757) 
* test_assign to test/backend

* fix
 2026-08-27 10:47:21 -04:00
nimlgenandGitHub 980c42d611 hcq2 tiny speedups (#17773) 2026-08-27 17:42:37 +03:00
nimlgenandGitHub 7eb763a3c2 bnxt to extra (#17772) 
* bnxt to extra

* x

* x

* les
 2026-08-27 17:29:17 +03:00
Teddy TennantandGitHub 5c3d044465 fix asinh gradient at zero (#17758) 2026-08-27 09:51:17 -04:00
nimlgenandGitHub f19d89e29a hcq2: a bit cleaner deps (#17771) 2026-08-27 14:24:59 +03:00
tonyfloatersuandGitHub 2d4e7afcbf mockgpu/amd: use generated KFD ioctl metadata (#17761) 
* init commit

* init commit

* minimize to single file scope
 2026-08-27 12:24:37 +03:00
qazalandGitHub 0d66815849 set BEAM_PADTO=0 in llama (#17769) 2026-08-27 16:45:33 +09:00
qazalandGitHub d16a2510ad remove some dead scheduler tests (#17768) 
* remove tests that use .uop

* that test does nothing
 2026-08-27 16:01:28 +09:00
qazalandGitHub 118e812b31 simplify more same device copies (#17766) 
* simplify more same device copies

* stack of contigs or copies

* rename
 2026-08-27 12:54:12 +09:00
sirhcmandGitHub 0820c7a424 benchmarks: openpilot matrix (#17714) 2026-08-26 18:53:55 -04:00
wozeparrotandGitHub 22ffff3258 gptoss: single jit (#17525) 2026-08-26 15:32:21 -07:00
George HotzandGitHub a0a901c8e4 faster qwen 3.8 (#17720) 
* faster qwen

* test fix

* dead code

* fix gguf issue

* pretty nt loads

* lil

* use warp
 2026-08-26 15:23:25 -07:00
sirhcmandGitHub c015351ac5 fix _get_cpu_count for docker --cpus=N in python 3.13+ (#17760) 2026-08-26 16:53:30 -04:00
nimlgenandGitHub 6074c002e1 hcq2: fix jit (#17747) 
* hcq2: fix reduce

* Dx

* inputs table

* emoty commit
 2026-08-26 22:36:36 +03:00
George HotzandGitHub 6042b87272 delete PCONTIG [PR] (#17756) 
* delete PCONTIG

* cleanups
 2026-08-26 12:07:28 -07:00
sirhcmandGitHub cc72b9f7be cleanup BENCHMARK_LOG (#17754) 2026-08-26 14:55:54 -04:00
chenyuandGitHub 6a3b297548 fix PTX NIR SPEC=2 for bool [pr] (#17753) 
* fix PTX NIR SPEC=2 for bool [pr]

storing bool with uint8 needs to pass SPEC

* the fix
 2026-08-26 14:07:29 -04:00
George HotzandGitHub ea6c82f3be small changes from new rangeify (#17752) 2026-08-26 10:33:31 -07:00
chenyuandGitHub 0abcf09b74 never bufferize_to_store weak input [PR] (#17751) 
github github
 2026-08-26 13:25:52 -04:00
chenyuandGitHub 4bdc865131 delete unused rewrite rules [PR] (#17748) 2026-08-26 10:50:42 -04:00
Teddy TennantandGitHub 4c20f1d357 fix asinh precision loss on negative inputs (#17749) 2026-08-26 10:08:04 -04:00
qazalandGitHub ecf79e260d better all2all schedule test (#17746) 
* better all2all schedule test

* deconstruct those numbers

* reorder
 2026-08-26 15:46:10 +09:00
b1tgandGitHub 9860e5d285 llm tokenizer: fix tekken, add gpt4o (#17733) 2026-08-25 23:30:35 -07:00
George HotzandGitHub 625c05df1e fix am_smi to respect dev (#17742) 2026-08-25 19:21:55 -07:00