Commit Graph

5369 Commits

Author SHA1 Message Date
chenyu
fc5b94b902 fix UOp.where(const, const) (#15560)
* fix UOp.where(const, const)

* fix
2026-04-01 05:28:49 -04:00
Christopher Milan
acf239e4d2 specify renderer in DEV, <dev>_<ren>=1 is deprecated (#15551) 2026-03-31 18:35:14 -04:00
nimlgen
5181c8e23a llm: fix nan in kvcache (#15552) 2026-04-01 00:38:45 +03:00
qazal
a15345a53e viz/cli: improve --help message (#15546)
* viz/cli: improve --help message

* not the default

* more work

* -s

* respect colored
2026-03-31 22:31:33 +09:00
chenyu
4ac2552642 improve ReduceMixin.all (#15544)
use prod instead of min since `mul` lowered to `and` directly
2026-03-31 07:54:27 -04:00
chenyu
89ec22131a tests to show double negation in min is not cancelled (#15543) 2026-03-31 06:59:13 -04:00
qazal
8feb8edc68 gemm/asm: add fp8 support to cdna asm_gemm (#15542)
* work

* hmm, mixins

* rhs_transposed

* also fix the dtype

* check for hipcc

* Exception

* select dev

* default
2026-03-31 19:32:54 +09:00
qazal
467c0af8aa viz: skip flaky sever tests (#15538) 2026-03-31 17:20:30 +09:00
qazal
f88e255cea gemm/asm: split and parameterize dtype in llama gemm tests (#15408)
* gemm/asm: more tests for emulator, parameterize llama gemm tests

* bf16 atol
2026-03-31 17:12:44 +09:00
b1tg
a63392a565 llm: pairwise ranking topk for MoE expert selection (#15499) 2026-03-31 12:46:39 +08:00
qazal
bc866a93f0 viz: rename exec to sqtt (#15527)
* viz: rename exec to sqtt

* more
2026-03-31 08:06:51 +09:00
Christopher Milan
adbfd82d1d DEV is ContextVar, setting Device.DEFAULT is deprecated (#15508) 2026-03-30 17:10:49 -04:00
qazal
ad6347f6d8 sqtt: allow mapping sopk to IMMEDIATE packets (#15525)
* work

* with s_waitcnt

* with the sopp variants, increase threads

* remove that

* sdst=NULL produces IMMEDIATE, otherwise is SALU
2026-03-30 23:12:17 +09:00
chenyu
f0eaac4235 reduce mixin (#15523) 2026-03-30 05:23:58 -04:00
chenyu
f485d0b664 UOp.sum -> usum, prod -> uprod [pr] (#15522)
rename to prep reduce mixin
2026-03-29 04:51:55 -04:00
qazal
266fb07721 viz: show exec duration (#15484)
* duration

* handwritten tests

* rdna3 pickle

* rdna4 pickle

* asserts

* rm that

* wmma work

* r4

* this shows the overlap well

* ohh okay it goes back

* are ds_load and ds_store different queues on RDNA4?

* print msg, v_mul_lo_u32 is 4 cycles?

* discover

* wmma something

* wmma comment

* less

* less

* better comments

* work

* inst st

* delay column

* better cli

* emit_alt

* update test_handwritten

* work
2026-03-28 22:48:59 +09:00
chenyu
c0753ab62f XOR simplifcation rules (#15512)
x^-1 has good vmin/vmax, and x^y^y is x
2026-03-27 23:23:27 -04:00
qazal
ccaa6bfc19 viz/cli cleanups (#15511)
* one less function

* work

* layout

* better handling of rewrites

* mypy passes
2026-03-28 08:50:38 +09:00
qazal
dcc2a5d23b viz/cli: simplify to --source and --item flags (#15510)
* viz/cli: simplify to --source and --item flags

* update viz cli test
2026-03-28 04:46:39 +09:00
nimlgen
0d6fc0f571 jit: graphing in uops (#15489)
* jit: graphing as rewrite rule

* f

* +metal,cuda

* x

* cl

* x

* x

* simpler

* f

* m

* x

* revert?

* revert2

* back

* back

* t

* x

* m

* x

* c

* x

* l

* x

* comment

* smaller

* rv

* x

* x
2026-03-27 19:09:02 +03:00
chenyu
30ebbe7f17 few more fold valid tests (#15509)
from remove CORRECT_DIVMOD_FOLDING attempt
2026-03-27 10:38:42 -04:00
chenyu
323fcefd7d Revert "DEV is a ContextVar (#15505)" (#15506)
This reverts commit fdb30cba96.
2026-03-27 02:22:40 -04:00
Christopher Milan
fdb30cba96 DEV is a ContextVar (#15505) 2026-03-27 00:57:09 -04:00
Christopher Milan
67a50fb738 move where on load with casts (#15492) 2026-03-26 22:11:27 -04:00
qazal
586c49642f viz/cli: test in CI (#15501)
* viz cli work

* baseline test

* make cli test work without subprocess

* more checks

* check itrace

* s/return/return None

* change

* minimal

* colored
2026-03-27 06:47:15 +09:00
qazal
3f9f0fa846 viz: yield sqtt alt events (#15500)
* yield other

* less

* work

* less
2026-03-27 04:43:41 +09:00
nimlgen
7193f90746 test view input in jit (#15497)
* will anything fail?

* add test
2026-03-26 16:59:47 +03:00
Christopher Milan
bc180a963c deprecate <dev>=1 in favor of DEV=<dev> (#15467)
* start work on target

* add test

* update actions to use DEV

* update docs

* update readmes

* tests need that too

* update example

* update tests (comments)

* fix that test

* ruff

* mypy

* oops

* remove getenvs

* don't add Target yet

* and the test

* lint

* and docs

* more stuff

* assert

* few more fixes

* test assert
2026-03-26 03:48:03 -04:00
chenyu
7c8f992894 move EXPAND dtype cast back to gradient.py (#15481)
only a concern for gradient, not mixin
2026-03-25 19:25:26 -04:00
qazal
60bd546593 sqtt: add cycle count to rdna3 enums (#15473)
* update rdna3 sqtt enums to include cycle_count

* dispatch_to_exec
2026-03-25 23:19:54 +09:00
chenyu
713b322e70 add weakint to promo_lattice (#15463)
sits between bool and smallest int
2026-03-25 00:27:34 -04:00
George Hotz
fe2690399b llm: support assistant prefill + refactor to TransformerConfig (#15457)
* llm: support assistant prefill

* refactor to ModelConfig

* TransformerConfig

* more
2026-03-25 10:50:48 +08:00
qazal
1b3d00d6ac viz/cli: remove --offset and --limit flags (#15439)
* work

* also no more no-color

* reorder

* update llama

* sqtt readme

* itertools

* rm that

* signals back
2026-03-25 09:52:27 +09:00
qazal
652bab8aad viz: support nested track_rewrites (#15454)
* simple test

* stack active groups
2026-03-25 05:01:30 +09:00
chenyu
b7960841af support shape broadcast in UOp.alu (#15442)
i think it can integrate tighter, but now Tensor also does ufix from UOp and implicit dtype upcast
2026-03-24 10:14:57 -04:00
George Hotz
a33ac869aa llm server: temperature + test client (#15444)
* improvements to the llm server

* eval script

* eval llm

* better eval gets 58.71

* cleanups

* add temperature, but multinomial is absurdly slow

* claude is so smart

* lint

* remove slop

* no more stop
2026-03-24 21:07:15 +08:00
George Hotz
85dee83f5d amd flash attention cleanups + emulator fixes (#15431)
* amd flash attention cleanups

* simpler

* params

* fix emulator bugs

* fix idiv bug

* remove that test

* more emu fixes
2026-03-24 10:10:46 +08:00
qazal
a590eded87 sqtt: rdna4 decoder work (#15434)
* sqtt: rdna4 decoder work

* diff cleanup

* more diff

* test

* work

* works

* TS_DELTA_SHORT
2026-03-24 03:49:32 +09:00
qazal
109472c37e sqtt: new s_barrier pickles, handle rdna4 barriers in emulator (#15437) 2026-03-24 03:25:28 +09:00
nimlgen
fa4cdb422e memplan on linears (#15422)
* memplan

* test

* x

* arenas

* correct

* set any size

* ugh

* make hevc happy

* x

* x

* held

* rm old

* del

* x

* fu

* f

* cl

* cl

* ok
2026-03-23 19:50:16 +08:00
nimlgen
2da008ae3b jit: rm replan (#15433) 2026-03-23 19:31:51 +08:00
qazal
c4c53418f8 sqtt: comment out flaky rocprof timestamp assert (#15432)
* comment out rocprof assert, add new assert

* better than > 0 assert

* string
2026-03-23 19:24:04 +09:00
George Hotz
c62dea6881 ai slop flash attention (it works) (#15401)
* ai slop flash attention (it works)

* speed up, 2 TFLOPS + 7 GB/s

* simpler

* simpler

* optimize

* faster

* warp shuffle

* sqtt: link dispatch to exec (#15396)

* sqtt packet linking infra

python

* javascript

* ~doubly linked list

* ui works

* work

* exec can also highlight the pc, coloring work

* more work

* rm sqtt/model.py, doesn't need to be upstreamed

* viz: no context enters in cli, update llama profile (#15404)

* removed unused named arg in rules [pr] (#15414)

* viz: sqtt printer in viz/cli.py (#15411)

* work

* sqtt timeline in CLI

* format all printers nicely

* s/Showed/Printed

* ansistrip

* sys.exit

* keep colors in list

* work from amd_copy_matmul

* has_more always gets returned

* linter

* don't print colors

* more colors

* wow this is so deep

* work

* minor details

* selected

* improve progress bar

* remove it

* 22, global_load_vaddr is so long

* remove *0 hack in sign, gradient materializes zeros for unconnected nodes (#15416)

Amp-Thread-ID: https://ampcode.com/threads/T-019d1612-6322-706b-a94d-a812400a55cb

Co-authored-by: Amp <amp@ampcode.com>

* works

* cnt=20

* revert that

* uop slice tests

* simpler

---------

Co-authored-by: qazal <77887910+Qazalin@users.noreply.github.com>
Co-authored-by: chenyu <chenyu@fastmail.com>
Co-authored-by: gg <ggordbegli@gmail.com>
Co-authored-by: Amp <amp@ampcode.com>
2026-03-23 16:15:10 +08:00
nimlgen
c74fa9bbe1 fix jitbeam not triggered (#15424)
* um

* beam

* x

* f
2026-03-23 15:34:59 +08:00
qazal
c7b18e6108 viz: sqtt printer in viz/cli.py (#15411)
* work

* sqtt timeline in CLI

* format all printers nicely

* s/Showed/Printed

* ansistrip

* sys.exit

* keep colors in list

* work from amd_copy_matmul

* has_more always gets returned

* linter

* don't print colors

* more colors

* wow this is so deep

* work

* minor details

* selected

* improve progress bar

* remove it

* 22, global_load_vaddr is so long
2026-03-23 00:17:05 +09:00
nimlgen
9656d97d97 jit: captures linears, not execitems (#15399)
* jit: captures linears, not execitems

* x

* um

* etsts

* mockcuda
2026-03-21 16:32:12 +08:00
Christopher Milan
a12d3951de fix test_export_model imports (#15389) 2026-03-20 07:27:01 -04:00
Christopher Milan
1560b534a5 remove IMAGE=2 (#15312) 2026-03-20 06:26:52 -04:00
chenyu
c491345766 pass device into Tensor._frompy (#15385)
* pass device into Tensor._frompy

with this, canonicalize_device is the only usage of Device in tensor.py

* export_model.py
2026-03-20 05:09:01 -04:00
George Hotz
3b75d8a7a2 fix double after bug in rangeify (#15381) 2026-03-20 14:53:46 +08:00