github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-26 07:18:40 -05:00
Code Issues Packages Projects Releases Wiki Activity
10,417 Commits 206 Branches 14 Tags
reprocess_node_experiments
Commit Graph

10417 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
chenyu
b190d85ad7 benchmark script bert softmax (#9759) 2025-04-07 00:31:18 -04:00
Ignacio Sica
58785181a8 AMD bf16xf32 TC (#9717) 
* dont test bf16 for emulated amd tc

* skip bf16 tc test in ci

* skip bf16 for AMD in test_tensor_cores_codegen

* add simple bf16 gemm test to benchmark
 2025-04-07 11:41:04 +08:00
chenyu
43e4565148 weighted linear in external_benchmark_bert_matmuls (#9757) 
include the linear to get qkv, and permute so that stride matches with the real run
 2025-04-06 23:35:42 -04:00
George Hotz
28e06d2d44 minor cleanups from patternmatcher [pr] (#9756) 2025-04-07 11:28:14 +08:00
qazal
1ce4912770 viz profiler ui (#9664) 
* localhost:8000/prof

* selector + table

* add pid

* on null selection reset filters

* table sort

* charset=utf-8

* clear the rest

* sort by duration

* render table

* format

* nothing in copy thread

* keep starts

* sort back

* less javascript

* diff

* works on firefox
 2025-04-07 00:30:17 +08:00
chenyu
8a585dc5c1 benchmark script for matmuls in bert (#9752) 
2 main matmuls in the bert layers. getting these to be fast makes bert fast
 2025-04-06 19:34:25 +08:00
qazal
139999c6d7 map viz files + query params cleanup [pr] (#9754) 
* map viz files + query params cleanup [pr]

* .width + fix
 2025-04-06 16:20:00 +08:00
Francis Lata
71b8890dd6 use validation dataloader inside retinanet eval (#9747) 2025-04-05 16:46:55 -04:00
nimlgen
5f7c79676f jit: prune independent copies (#9749) 
* jit: prune independent copies

* linter

* check kernel cnt
 2025-04-05 20:50:28 +03:00
nimlgen
c2573b247c jit: rename optimize_weights -> replan_buffers_memory_layout (#9751) 2025-04-05 20:35:15 +03:00
uuuvn
493fb315b1 fix RDNA2 support (#9700) 
linux amdgpu_discovery.c:amdgpu_discovery_set_ip_blocks is a ton of
switch cases with sometimes weird choices like replacing nbio 3.X with
2.3 while nbio 2.5 is somehow nbio 7.0. `import_module` currently just
tries to replace revision and minor with zeroes if there is no exact
match, but that's not enough to cover all that weirdness
 2025-04-05 18:42:47 +03:00
chenyu
5a04f4d4ba revert bert hparams for green and red (#9744) 
did more runs and it's not really better and not worth the change. only useful for BS=1024
 2025-04-05 07:38:01 -04:00
chenyu
407ca54382 symbolic fold double where (#9436) 
* symbolic fold double where

a.where(b.where(c, d), d) -> (a & b).where(c, d). a pattern in optimizer

* test case
 2025-04-05 05:12:17 -04:00
Sieds Lykles
9c2fc695b5 cond.logical_not().where(a,b) -> cond.where(b,a) (#9741) 
* Add rule for negation in where, simplifies arange patterns

* 0 becomes 0.0 again

* Only if cond is bool

* ne is never None

* Add a test

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-04-04 19:13:32 -04:00
Sieds Lykles
e9a3ac02a5 Remove ne from arange pattern (#9743) 2025-04-04 18:31:13 -04:00
nimlgen
86c55414d7 ops_amd: simplify gfx version (#9742) 
* ops_amd: simplify gfx version

* fix

* all vesrsion compact style

* mypy

* revert this

* rename back to target
 2025-04-04 22:18:11 +03:00
qazal
16d6aa15f1 record unittest name in process replay (#9731) 
* record unittest name in process replay

* getitem

* filename + (optional) name

* del

* get_test_method

* not solved

* try with linecache

* test: print_loc

* format

* without linecache

* checkout master
 2025-04-05 01:39:48 +08:00
qazal
354db961c6 viz refactor to prep for profiler [pr] (#9739) 2025-04-04 17:13:14 +08:00
chenyu
fe998798fb linearizer failure test for OUT OF BOUNDS ACCESS (#9738) 2025-04-04 03:48:43 -04:00
George Hotz
8b5a523743 fix minimum length in pattern matcher (#9736) 2025-04-04 14:57:01 +08:00
chenyu
640ff681c3 rename bert script to 8xMI300X (#9734) 
and adds a script for single MI300X
 2025-04-03 23:36:24 -04:00
George Hotz
b719aa1fb0 only check once for divisible fold lengths (#9732) 2025-04-04 11:27:34 +08:00
George Hotz
926b0bcc57 cache folded upcast [pr] (#9733) 2025-04-04 11:23:19 +08:00
George Hotz
8206c7281e move const multiply after REDUCE (#9730) 2025-04-04 11:07:46 +08:00
chenyu
6b3480ec70 update mi300x bert haparams (#9716) 
* update mi300x bert haparams

borrowed from previous submission that also did BS=1024

* update
 2025-04-03 22:30:00 -04:00
George Hotz
cac8bcf8b5 use Ops.REDUCE (#9721) 
* decrease bert python time [pr]

* order copies

* Revert "order copies"

This reverts commit 3f62c8693b.

* rewrite count

* Ops.REDUCE

* acc first in the add chain

* Fix tensor core acc

* arange patterns look good

* fix multireduce gate

* reduce rewrite rule

* bump that to 15 minutes

* multiwmma isn't fusing

* gep through wmma is gep pushing

* bump that timeout too, it's all env setup

* add failing test
 2025-04-04 10:14:34 +08:00
nimlgen
949459fdd6 jit: fix deallocate on unallocated buffers in free_intermediates (#9699) 2025-04-03 18:32:51 +03:00
qazal
52a8ecb15e record unittest location in process replay [pr] (#9727) 2025-04-03 20:50:09 +08:00
geohotstan
ac713e04db ONNX add output shape validation (#9720) 
* add output shape validation and remove support for sequence_type

* nit better err msg

* add sequence_type back

* improve err msg

* Revert "improve err msg"

This reverts commit dc9eaea4bb.

* Revert "add sequence_type back"

This reverts commit 288170b2d9.

* do explicit shape equality

* small nit
 2025-04-03 05:44:53 -04:00
chenyu
7dadbf3697 insert float() in bert acc (#9726) 
sum of bool by default uses default_float for acc. So without float, it might overflow with a large BS and default_float=HALF.

fixed clsf_accuracy to not be inf in mi300x bert
 2025-04-03 05:44:09 -04:00
chenyu
79145e3d40 cleanup truncate_bf16 [pr] (#9725) 
use torch bfloat16 for groundtruth in test. also a TODO for discrepancy
 2025-04-03 05:43:49 -04:00
Ignacio Sica
bc2d86195e increase test tolerance (#9719) 2025-04-03 15:24:09 +08:00
chenyu
1d25844d44 Revert "disable CI red llama 3 4 gpu beam (#9690)" (#9709) 
This reverts commit 6a5eacba8b.
 2025-04-03 02:34:39 -04:00
George Hotz
49dafe6d43 add gc tests [pr] (#9718) 
* add gc tests [pr]

* del

* more gc tests

* add NullGraph
 2025-04-03 14:08:32 +08:00
Ignacio Sica
bc91fffc5d fix gated store with index in python backend (#9703) 
* add default gate in index

* assert store

* add TestRendererFailures

- move test_gated_store_with_alu to new TestRenderFailures class for
tests that fail on multiple renderers
- add test_renderer_failures.py run on python CI

* add test for gated index in 2d

* test TestRenderFailures
 2025-04-03 12:48:28 +08:00
qazal
f2bd65ccfc delete Ops.EMPTY and Tensor._metaop (#9715) 
* delete Ops.EMPTY and Tensor._metaop [pr]

* test_creation

* arg=

* abstractions2
 2025-04-03 12:29:02 +08:00
George Hotz
5c7b549eab use functools.cache instead of lru_cache(None) [pr] (#9714) 
* use functools.cache instead of lru_cache(None) [pr]

* more cache
 2025-04-03 11:47:13 +08:00
qazal
bbd13191f4 cleanup tensor BIND + remove outdated comments in tensor.py [pr] (#9712) 
* cleanup tensor BIND + remove outdated comments in tensor.py [pr]

* from_blob whitespace

* assert
 2025-04-03 11:21:53 +08:00
geohotstan
e1d7e47cca fix ONNX IsInf unintended dtype promotion (#9711) 
* add IsInf

* add corresponding test

* that float16 is kinda silly
 2025-04-02 22:46:15 -04:00
qazal
11ae254dc5 construct BUFFER UOps directly when device in known [pr] (#9710) 
* construct BUFFER UOps directly when device in known [pr]

* diff
 2025-04-03 10:41:44 +08:00
George Hotz
1714fc3ba4 start work on speed [pr] (#9707) 
* fix get_location

* fix get_location try 2

* clean up split_load_store [pr]

* SHR fixup [pr]
 2025-04-03 10:39:01 +08:00
George Hotz
0f1ffc2050 hotfix: cat tests 2048 instead of 256 2025-04-03 10:37:56 +08:00
uuuvn
5bd485c027 Fix double SDMA_OP_FENCE (#9705) 
Introduced in #9585, probably when i incorrectly resolved merge conflict
while rebasing an old, mi300x-only branch. Seems to be the source of
multi gpu beam llama hangs
 2025-04-03 09:43:37 +08:00
chenyu
a6fec2f5ae dev_run for bert on mi300x (#9706) 2025-04-02 21:12:55 -04:00
nimlgen
d96b4983ac amd: support rdna4 in runtime again (#9702) 2025-04-03 01:19:23 +07:00
Ignacio Sica
2d6d8b7355 add bf16 mfma support (#9695) 
* add bf16 mfma support

* skip tc if emulated_amd and dtypes is bf16

* hotfix
 2025-04-02 21:44:49 +08:00
nimlgen
a6733f519f dsp: make relro sections contiguous (#9701) 2025-04-02 18:02:16 +07:00
George Hotz
ea5caefef0 gep should look at count, not vcount (#9698) 
* gep should look at count, not vcount

* gep in order is a rule

* min change

* gep on void
 2025-04-02 18:10:57 +08:00
George Hotz
f72a87fd0e add proper support for Ops.IGNORE to remove store masks (#9692) 
* add proper support for Ops.IGNORE to remove store masks

* remove useless NHWC

* revert that
 2025-04-02 16:38:01 +08:00
chenyu
3b8d923692 remove skip LLVM in test_div_int (#9686) 2025-04-02 04:15:00 -04:00