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
12,524 commits 280 branches 15 tags 325 MiB
Commit graph

8,514 commits

Author SHA1 Message Date
George Hotz
37a40bf975 early lower cat 2026-03-07 11:39:20 +08:00
George Hotz
af1db22b25 simpler 2026-03-07 10:11:21 +08:00
George Hotz
be0f9d1055 min 2026-03-07 10:00:29 +08:00
George Hotz
5b9a6c5520 Add Ops.CAT movement op (ai slop) 2026-03-06 18:51:25 +08:00
George Hotz
6fd18ef875
 		rename CAT to VCAT (#15167) 2026-03-06 18:46:28 +08:00
Roelof van Dijk
059c6326c0
 		metal uint32 icb offset overflow (#15156) 
* metal uint32 icb offset overflow

fix: diff

supports_exec_item

GraphRunner.supports_exec_item

tests

fix: can't import on non-metal

stricter

* also test the non-metal buffer case

* imports on non-mac
 2026-03-06 00:54:39 +03:00
chenyu
da61088ca4
 		more divmod recombine (#15162) 2026-03-05 12:53:22 -05:00
chenyu
167a1d56a6
 		improve divmod folding (#15148) 
canonicalize to div than mod which enables more simplifcation
 2026-03-05 10:07:36 -05:00
Christopher Milan
b824579e4d
 		simplify image_conv2d pitch alignment hacks (#15158) 2026-03-05 07:17:34 -05:00
qazal
5bf542469d
 		viz: python traceback for USER device (#15160) 
* start

* ux

* unittests
 2026-03-05 20:22:09 +09:00
Roelof van Dijk
d65923bda5
 		tensor.py: add normalize function (#15159) 
* tensor.py: add normalize function

* p==0 should match torch
 2026-03-05 18:55:53 +08:00
Roelof van Dijk
fc0534910c
 		q5k is like q4k (#15155) 2026-03-05 17:02:49 +08:00
Ananta Ranganathan
8ef656324e
 		FIXED TEST Q5_K GGUF dequant (#15147) 
* q5_k gguf support as separate pr

* fix the problematic gemv test for q5_k

* add assert to make sure the gemv test cant fail with warning instead of error
 2026-03-05 16:32:36 +08:00
George Hotz
e97922a57c
 		LLM speedup with two jits, prefill/rollout (#15153) 
* START_TIME

* print cleanup

* fix tests
 2026-03-05 16:21:09 +08:00
wozeparrot
be23772d43
 		llama3 fixes part2 (#15150) 2026-03-04 23:43:50 -08:00
George Hotz
fb43b415f9
 		fix symbolic shape call + chunked prefill (#15149) 
* fix precompile for symbolic shape

* chunked prefill

* cleaner

* test that
 2026-03-05 14:02:26 +08:00
George Hotz
8a82b26522
 		llm: print the prefill cache size (#15146) 
* print the llm prefill cache size

* mock that too
 2026-03-05 12:13:28 +08:00
chenyu
b5370fd52d
 		use copy_multi in alu_multi [pr] (#15143) 
* use copy_multi in alu_multi [pr]

* copy to anything
 2026-03-04 22:53:00 -05:00
George Hotz
72a9ed6e23
 		fix render depth bug + add warmup to serve + no realize default (#15144) 
* fix render depth bug + add warmup to serve

* make realize not the default
 2026-03-05 11:21:16 +08:00
George Hotz
ac1847cbf7
 		fully symbolic llm (#15097) 
* work

* llm symbolic (almost)

* work

* revert that

* llm sym

* works

* cleanups

* cache tokens with the kv cache

* cleanups

* cleanups
 2026-03-05 10:22:11 +08:00
qazal
33a1970045
 		sqtt: simplify inst mapping, validate JUMP processing in CI (#15139) 
* jump cleanup

* assert there's a JUMP

* new example for JUMP

* regenerate examples

* rdna4 work

* new packets

* work

* less for branch handling

* less verbose

* fix err message
 2026-03-05 09:53:12 +09:00
chenyu
04da527a7a
 		minor div_and_mod_symbolic cleanups (#15138) 2026-03-04 19:05:44 -05:00
chenyu
106d18b792
 		use UOp methods in allreduce.py [pr] (#15137) 
except the one line with Ops.BUFFER and Ops.NOOP, not sure what that's for
 2026-03-04 17:15:33 -05:00
chenyu
34594bcaaf
 		Revert "bug in metal: offset is stored as uint32, overflow (#15129)" (#15136) 
This reverts commit 9c58db16fa.
 2026-03-04 16:54:42 -05:00
Roelof van Dijk
9c58db16fa
 		bug in metal: offset is stored as uint32, overflow (#15129) 
* metal uint32 icb offset overflow

* fix: diff

* supports_exec_item

* GraphRunner.supports_exec_item

* tests

* fix: can't import on non-metal
 2026-03-04 22:52:12 +03:00
chenyu
1f96cc2b51
 		update non-contiguous buffer error message [pr] (#15131) 
* update non-contiguous buffer error message [pr]

also cleaned up the tests

* order
 2026-03-04 11:13:26 -05:00
George Hotz
47faa2d7b4 hotfix: llm kv cache uses clone instead of realize to avoid many realize 2026-03-04 19:07:03 +08:00
George Hotz
8ebd24637b
 		fix fa forward building with clang 22 (#15124) 
* fix fa forward building with clang 22

* fix: override rocm path

---------

Co-authored-by: Woze Parrot <wozeparrot@gmail.com>
 2026-03-04 02:32:25 -08:00
Christopher Milan
592f9bf6c6
 		set OPENPILOT_HACKS=1 to enable replace assign (#15123) 2026-03-04 05:26:04 -05:00
Christopher Milan
5623cea7b1
 		move openpilot contiguous hacks to schedule (#15120) 2026-03-04 03:04:06 -05:00
George Hotz
5ecfe549e7
 		allreduce is a function with LATE_ALLREDUCE=1 (#15119) 
* allreduce as a function

* allreduce function

* support allreduce function

* LATE_ALLREDUCE
 2026-03-04 15:17:58 +08:00
Christopher Milan
e7e70a3c95
 		simplify idx before counting backward_slice (#15117) 2026-03-03 23:53:50 -05:00
George Hotz
2d72a4a90c
 		fix copying padded const (#15116) 
* fix const padding cpu

* remove comment
 2026-03-04 10:39:45 +08:00
chenyu
b5ebb4d06d
 		contiguous_view_offset returns only offset [pr] (#15113) 
size is always input.size
 2026-03-03 15:23:39 -05:00
nimlgen
abd830b260
 		am: setup_rinf returns only doorbell (#15112) 2026-03-03 19:27:41 +03:00
nimlgen
4b42bb54aa
 		am: reset sdma to start from 0 (#15109) 2026-03-03 18:14:46 +03:00
George Hotz
01ddb4c267
 		add precompile to call (#15099) 
* add precompile to call

* put get back

* something

* after structure

* alt

* keep it call

* resolve call

* resolve linear call

* precompile works with llm

* revert rangeify

* color for debugging

* getenv PRECOMPILE

* clean up deco pattern

* fully recursive sink scheduling

* revert llama

* fix SPEC=2
 2026-03-03 22:32:42 +08:00
qazal
c7f908b788
 		sqtt: fix rdna4 structs (#15111) 
* work

* DEBUG=2
 2026-03-03 23:32:14 +09:00
qazal
8dd691761d
 		sqtt: remove old files (#15108) 2026-03-03 22:43:24 +09:00
Christopher Milan
5f6b610da1
 		FLOAT16 logic for IMAGE==1 goes back to image_conv2d (#15105) 2026-03-03 05:37:57 -05:00
George Hotz
7d025089e3
 		no after removal (#15102) 
* no after removal

* we are using walk

* null schedule test

* pytest deps

* Revert "pytest deps"

This reverts commit 5e1c5304ec.

* Revert "null schedule test"

This reverts commit 02da66053e.

* clean null tests
 2026-03-03 17:50:31 +08:00
wozeparrot
92c16810ac
 		feat: per device mem_used (#15100) 2026-03-03 01:31:28 -08:00
qazal
e3a0598d0b
 		viz: the whole pc should be in view (#15101) 2026-03-03 17:17:53 +09:00
Christopher Milan
c70e8af068
 		move IMAGE FLOAT16 logic to allocations (#15095) 
* FLOAT16 logic in allocations

* cleanup

* separate that

* only apply when IMAGE == 1

* test passing now

* create image buffers earlier
 2026-03-02 22:00:05 -05:00
George Hotz
d483e4153a
 		buffer view is like buffer (#15082) 
* buffer view is like buffer

* fix

* swap_reshape_shrink

* contiguous on gguf, fix overlap

* revert that

* _device_supports_view

* this

* fix that test

* 0 buffers

* that test was wrong

* this

* check correct size

* contig BUFFER_VIEW

* this

* fix tests

* buffer view tests

* om

* fix torch

* no MOCKGPU

* skip
 2026-03-03 09:52:33 +08:00
qazal
848f5cea96
 		viz: sqtt instruction packet trace (#15065) 2026-03-03 07:55:04 +09:00
chenyu
f80b1033c5
 		simpler Tensor.all (#15089) 
same generated kernel
 2026-03-02 11:08:55 -05:00
chenyu
4008f7d4e8
 		move Tensor.one_hot +1 to python (#15088) 2026-03-02 10:56:41 -05:00
nimlgen
dafbe9733a
 		am: cleanup (#15086) 2026-03-02 17:06:21 +03:00
George Hotz
5ff278446c
 		add contiguous_view_offset (#15084) 
* add contiguous_view_offset

* no int
 2026-03-02 18:05:04 +08:00