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
12,069 commits 280 branches 15 tags 325 MiB
Commit graph

1,616 commits

Author SHA1 Message Date
qazal
cf73d7e2a7
 		hotfix: disable slower asm gemm shape from llama seqlen 8192 (#14582) 2026-02-06 15:05:19 +09:00
qazal
be77873974
 		llama: contig backward for wk / wv matmul backward (#14581) 2026-02-06 14:54:00 +09:00
wozeparrot
f73468d516
 		fa: block skipping for fa kv bwd (#14569) 2026-02-05 16:13:53 -08:00
chenyu
41a179f542
 		fix test_xlm_roberta_large (#14564) 
onnxruntime does not allow symlink that's outside model dir. update snapshot_download to use local_dir instead of cache_dir. some ad hoc migration step to copy the existing model too
 2026-02-05 14:56:06 -05:00
qazal
190042358f
 		llama: faster bf16 matmul / rope backward (#14558) 2026-02-05 23:57:25 +09:00
George Hotz
b398335f62
 		assembly/amd: fix saturation in python remu (#14557) 
* PYTHONREMU: failing test for V_SUB_NC_U32_E64 clamp

* fix saturation in PYTHON_REMU

* simpler

* more tests, less lines

---------

Co-authored-by: Christopher Milan <chrismilan@ucla.edu>
 2026-02-05 18:35:57 +08:00
wozeparrot
c1ea6687e5
 		fa: simpler is faster (#14548) 2026-02-05 01:13:17 -08:00
George Hotz
43e7eda4e7
 		grad_b uses custom gemm (#14550) 
* grad_b uses custom gemm

* fix multi backward, acc is in float32

* test_gemm_batched

* square gemm

---------

Co-authored-by: qazal <77887910+Qazalin@users.noreply.github.com>
Co-authored-by: qazal <qazal.software@gmail.com>
 2026-02-05 15:22:27 +09:00
qazal
f9cfb64cd9
 		test asm_gemm in CI (#14551) 
* test asm_gemm in CI

* default float16

* use a smaller shape for multi

* smaller size

* smaller for CI

* smaller for ci

* need half
 2026-02-05 13:32:22 +09:00
Christopher Milan
232848d086
 		PYTHONREMU: VOP3P integer operations with constants don't cast to fp16 (#14546) 
* PYTHONREMU: VOP3P integer operations with constants don't cast to fp16

* put that back

* cleaner

* do that once
 2026-02-04 20:10:59 -05:00
wozeparrot
2966619834
 		feat: llama uses enable_gqa during training (#14545) 2026-02-04 16:22:31 -08:00
Christopher Milan
5338ce6b74
 		test S_PACK in extra/assembly/amd/test/hw (#14537) 
* S_PACK_LL_B32_B16 in test/hw

* add rest of S_PACK instructions
 2026-02-04 14:17:16 -05:00
chenyu
9052db678f
 		remove allow_shape_mismatch in Tensor.replace (#14536) 
move all logic to torch_backend and not hacking Tensor method
 2026-02-04 12:38:18 -05:00
nimlgen
62786d488a
 		am: mi3xx perf (#14529) 2026-02-04 19:32:43 +03:00
chenyu
d57d24c7d4
 		Buffer.as_buffer -> Buffer.as_memoryview [pr] (#14535) 
it casts to memoryview. also inline the as_typed_buffer checks to Tensor._data
 2026-02-04 11:31:11 -05:00
chenyu
67f91e897b
 		UOp.is_contiguous -> UOp.has_buffer_identity [pr] (#14530) 
one more confusing buffer related method, but it's definitely not is_contiguous
 2026-02-04 09:21:26 -05:00
Christopher Milan
ecbce5269e
 		PYTHONREMU properly supports S_PACK_LL_B32_B16 (#14527) 
* PYTHONREMU properly supports S_PACK_LL_B32_B16

* default
 2026-02-03 23:45:33 -05:00
wozeparrot
720c9597a9
 		feat: llama uses is_causal on sdpa during training (#14528) 2026-02-03 20:24:30 -08:00
qazal
d1bfbe9ce3
 		isolate slow llama gemm (#14525) 2026-02-04 12:20:10 +09:00
George Hotz
d59e6e7a37
 		move more tests to test/null, split some existing ones (#14512) 
* move more tests to test/null, split some existing ones

* null work

* null work

* move more

* fixes

* move PIL

* PIL in CLIP

* don't move that
 2026-02-03 20:20:20 +08:00
qazal
a98c53769a
 		ASM_GEMM=1 runs the UOp gemm on non cdna (#14516) 
* ASM_GEMM=1 runs the UOp gemm on non cdna

tests run on mac in 3 seconds

* min diff
 2026-02-03 20:42:02 +09:00
qazal
5c1d21349e
 		viz: profiler command line tool (#14515) 2026-02-03 19:51:25 +09:00
George Hotz
dd2de4f838
 		rename all DEFINE_GLOBAL to PARAM (#14511) 2026-02-03 15:09:38 +08:00
wozeparrot
bbcd3d67a3
 		fa: faster (#14453) 2026-02-02 21:34:17 -08:00
chenyu
66d2b02f11
 		delete files that depends on extra.optimization.helpers (#14499) 2026-02-02 13:33:33 -05:00
George Hotz
6e958dbfd4
 		assembly/amd: add RDNA4 support to emulator (#14341) 
* start new rdna4

* work

* plus works

* more pass

* rdna4

* assembly/amd: fix RDNA4 emulator for float16 and VOP3 clamp

* stale

* rev

* rr

* rdna4 emu tests

* cleanup

* cleanup

* simp

* works

* better factorizaion

* hacks

* fix mockgpu

* guard both

* cleaner

* gate

* bug fix and a few tests

* all test_tiny
 2026-02-02 21:35:59 +08:00
qazal
965940dd00
 		sqtt: update examples after event field change (#14493) 
* regen sqtt examples

* cdna

* rdna4

* packet counts for rdna3

* sqttmap work
 2026-02-02 21:39:48 +09:00
George Hotz
965149a46d
 		assembly/amd: add ds perm instructions (#14486) 
* assembly/amd: add ds perm instructions

* NO SKIP

* fix preexisting RDNA3 issues

* pcode

* assert

* asserts

* unify

* simp

* good fix
 2026-02-02 16:02:00 +08:00
Robbe Derks
d75a1b0d5a
 		usbgpu: use BOT interface for patch.py (#13644) 
* BOT usage

* cleanup

* fix lint

* fix ruff

* fix -7?
 2026-02-02 11:54:46 +08:00
qazal
616e9c1483
 		CDNA assembly gemm in tensor.py with flag (#14310) 
* work

* work

* the assembly

* remove the old one

* remove ws bufs, assert splitk

* notes cleanup

* work

* gemm args

* gemm in mixins would be nice

* add gemm gradient

* print counters

* the realize is for DEBUG=2 aesthetics

* dedup

* rewrite to python dsl, no list copies

* leave that

* add B, M, N, K to gemm name

* it's M0 not NULL

* fp16 support

* test cleanup + more gemms

* work from viz

* more work

* gemm batch_size

* xccg path work

* tiny comments on the label naming

* s_waitcnt
 2026-01-31 22:34:14 +09:00
qazal
d69bc5aa1a
 		make DEV=NULL EMULATE=AMD amd_asm_matmul run (#14460) 2026-01-31 20:45:24 +09:00
George Hotz
b705c9143c
 		assembly/amd: test more instructions (#14365) 
* assembly/amd: test more instructions

* more

* passing

* revert

* no const fold

* remove junk

* cleaner
 2026-01-31 12:40:22 +08:00
Christopher Milan
e575dd8275
 		prevent UB in long decomp and more emulated tests (#14447) 2026-01-30 19:38:41 -05:00
nimlgen
1998e0bb28
 		nv: add prof props to dev (#14437) 2026-01-30 12:51:43 +03:00
wozeparrot
c2fb8b208f
 		fa: 32 block size (#14416) 2026-01-29 13:59:13 -08:00
nimlgen
dc977a03b0
 		nv_pma: bw decoder (#14424) 
* nv_pma: bw decoder

* decoder fix

* better
 2026-01-30 00:12:39 +03:00
nimlgen
2d5c24879f
 		nv: pma for 5090 (#14420) 
* nv: pma for 5090

* hm

* 4090
 2026-01-29 20:06:01 +03:00
nimlgen
acb2fc36ba
 		nv_pma: add decoder (#14404) 
* nv_pma: add decoder

* cl
 2026-01-28 20:44:02 +03:00
nimlgen
544928766d
 		hcq_smi: kill mac pids (#14398) 2026-01-28 15:00:28 +03:00
George Hotz
202b74b369
 		assembly/amd: continue refactors (#14386) 
* simpler

* merge

* flat

* no ctx

* use the correct apis

* dup code

* write clean code

* remove bad helpers

* bits junk remove

* junk remove

* smem test

* fix tests

* correct fix + tests

* Fmt matters it seems

* wmma refactor

* a lil more

* kimi cleanups

* line
 2026-01-28 17:33:03 +08:00
qazal
0294014108
 		fix bufferize cost function for multi, improve VIZ=-1 cli (#14394) 
* improve cli

* remove_bufferize change
 2026-01-28 15:53:18 +09:00
George Hotz
88bc5ee212
 		assembly/amd: rename to better names (#14384) 
* assembly/amd: rename to better names

* might help fuzzing segfault

* emu2 -> emu
 2026-01-28 10:00:54 +08:00
chenyu
db010a31be
 		IGNORE_OOB -> CHECK_OOB [pr] (#14374) 
flip the meaning
 2026-01-27 12:20:59 -05:00
George Hotz
e5df7e640b
 		fix branches in amd_asm_matmul (#14369) 2026-01-27 20:48:42 +08:00
George Hotz
bfc88bcfb8
 		assembly/amd: emu refactors + enable PYTHON_REMU by default (#14361) 
* assembly/amd: start refactors

* cleanups

* those are global

* methods on ctx

* const cleanup

* range helper

* types and imports

* cleanups

* cleanups

* remove stale name

* fix emu2 types

* more typing

* more mypy

* cleanups

* fxns

* scc cleanup

* cleanups

* cleanups

* simpler parse_pcode

* laneid

* no defaults for pcode

* pcode is not optional

* cleanups

* functions cleanup

* splat

* expr_parser functions

* single tok

* invert global loops

* try_eat

* minor

* run parser on all

* no silent 0

* tests
 2026-01-27 17:42:24 +08:00
qazal
f866b2a513
 		mfma loop in asm dsl (#14349) 
* mfma loop in asm dsl

* work
 2026-01-27 11:11:37 +09:00
George Hotz
204f51e739
 		assembly/amd: bug fixes for PYTHON_REMU (#14347) 
* default PYTHON_REMU to 1

* mockgpu

* less size

* normal compile path

* uniqie

* more

* fix clamp

* Change PYTHON_REMU default to 0 in _try_dlopen_remu
 2026-01-27 00:48:22 +08:00
George Hotz
3b43d26f10
 		assembly/amd: emu speed (#14344) 
* assembly/amd: emu speed

* fix spec

* go

* don't do this

* simpler

* no stupid consts

* hack

* simpler

* no index

* no where

* faster linearizer

* fix spec

* no index dtype
 2026-01-26 22:21:34 +08:00
George Hotz
774a454bb5
 		assembly/amd: fix scratch SVE (#14340) 
* assembly/amd: default python REMU

* mem_used

* no lane

* sve

* remove that

* needs s_code_end in tests
 2026-01-26 21:03:51 +08:00
qazal
2d91fe6310
 		use amdgpu dsl in mmapeak (#14342) 
* use amdgpu dsl in mmapeak

* don't rely on llvm for vgpr counting

* llvm roundtrip assert

* rm it, add ci

* vgpr_count

* move emulated test to amd, it needs comgr

* env

* arch

* inst._fields -> inst.operands

* vgpr offset
 2026-01-26 22:03:43 +09:00