Commit graph

13,633 commits

Author SHA1 Message Date
George Hotz
009b484dc0 bump 2026-06-17 16:22:13 -07:00
George Hotz
0b5796e3c8 fix gemm group + END shape 2026-06-17 16:16:29 -07:00
George Hotz
b2f4f6f6c4 spec for stack 2026-06-17 16:07:50 -07:00
George Hotz
a417b6c144 STACK 0 is dtype void 2026-06-17 16:04:04 -07:00
George Hotz
aef85ddc4d
addrspace special/range (#16647)
* addrspace special/range

* just include indexing

* define var is alu

* bring old ignore indexing back

* mults to fix

* fixes

* ALU

* fixes
2026-06-17 15:57:37 -07:00
chenyu
1e08c0a07c
remove NOOP from AFTER with multiple srcs (#16646) 2026-06-17 14:35:02 -04:00
chenyu
1acc40600d
indexing an after with all fully invalid stores is invalid (#16643)
* indexing an after with all fully invalid stores is invalid

* typing cast
2026-06-17 11:06:36 -04:00
nimlgen
0f0c622086
hcq2: multi folders (#16642) 2026-06-17 15:20:25 +03:00
George Hotz
be9b570cb2
late numbering of var params (#16640)
* do_number_param

* fix sort order in x86

* we don't want this
2026-06-17 00:36:08 -07:00
qazal
c7055d658f
viz: only store kernel info (#16641) 2026-06-17 16:21:57 +09:00
George Hotz
d631716858
remove const without STACK (#16639)
* remove const without STACK

* fix GEP rewrite

* fix null tests

* fix openpilot regression

* it's 10 in CI
2026-06-16 21:25:42 -07:00
wozeparrot
36f6d1b064
gemm: fix bf16 atb for mp sharding (#16637) 2026-06-16 15:58:47 -07:00
qazal
1cb6b88d37
viz: show contents of vconst (#16636)
* failing test

* render vconst

* simpler test

* reorder
2026-06-17 02:31:03 +09:00
nimlgen
5644605d92
hcq2: pack bufs (#16635)
* hcq2: pack bufs

* x
2026-06-16 18:58:16 +03:00
chenyu
d5d59a2be6
remove dead rangeify rules [PR] (#16634) 2026-06-16 10:03:08 -04:00
chenyu
f0998e9bba
Revert "invalid clone is anonymous buffer" (#16613) (#16633) 2026-06-16 08:27:48 -04:00
qazal
7d2b0b697d
simple failing test for invalid extra E kernel (#16632)
* simple failing test for invalid extra E kernel

* 6 kernels
2026-06-16 17:57:44 +09:00
wozeparrot
70cac72781
llama: realize weight init (#16623) 2026-06-15 23:00:19 -07:00
Christopher Milan
443f976305
fix buffer overrun in dcache_flush (#16630) 2026-06-15 23:26:32 -04:00
chenyu
aa2bef24a8
no_vectorized_alu in cstyle does nothing now [PR] (#16631) 2026-06-15 23:07:20 -04:00
chenyu
efd03d7153
invalid clone is anonymous buffer [PR] (#16613) 2026-06-15 20:14:26 -04:00
nimlgen
4a0488ae97
hcq2: optims (#16624)
* hcq2: optims

* x
2026-06-15 23:58:28 +03:00
George Hotz
41aa2fe119
test_gemm needs .clone() on eye (#16629) 2026-06-15 12:48:27 -07:00
qazal
10bdb9c9d0
viz: check node exists before anchoring zoom (#16627) 2026-06-15 21:03:24 +09:00
qazal
f998b9930a
fp8 gemm inv_scale in epilogue (#16625)
* fuse scale

* remove python inv_scale

* more inv_scale removal

* more cleanups

* cleaner

* diff polish

* work

* rename

* simpler

* simpler

* compute

* c

* Revert "c"

This reverts commit 8941fec7ca.

* Revert "compute"

This reverts commit 9db573a6d3.

* Revert "simpler"

This reverts commit 910ad33f87.

* Revert "simpler"

This reverts commit bf75d235a1.

* s_g

* update types

* less diff noise

* remove
2026-06-15 18:44:41 +09:00
nimlgen
4dc51aff6e
hcq2: jit (#16621)
* hcq2: jit

* x

* x

* minor
2026-06-15 06:35:35 +07:00
chenyu
2adedf5ccb
clean up fold_divmod_general [pr] (#16622)
genralized fold_binary_numerator in fold_divmod_congruence
2026-06-14 17:15:52 -04:00
George Hotz
a6d7fb9d4d
only SHRINK for non scalar access (#16619) 2026-06-14 10:08:37 -07:00
George Hotz
b1fb39502d delete that test 2026-06-14 09:42:58 -07:00
chenyu
2e181f4259
simpler cancel_divmod [PR] (#16616) 2026-06-14 11:41:31 -04:00
chenyu
5d5ead78da
inline unique_const in invalids [PR] (#16612) 2026-06-13 10:14:32 -04:00
Sieds Lykles
b00dd754a9
Remove if-condition from nested div rule [pr] (#16611)
* add rules and test

* trigger [pr]
2026-06-13 15:47:21 +02:00
nimlgen
5a9227b30a
hcq2: rebind var params (#16610) 2026-06-13 14:55:52 +03:00
nimlgen
8efc8d064f
unique based on opaque in from_buffer (#16609) 2026-06-13 14:31:58 +03:00
nimlgen
c43091a464
fix missing cast in cstyle (#16608)
* fix missing cast in cstyle

* x

* x
2026-06-13 10:04:06 +03:00
qazal
2e77bd01db
fp8 gemm cleanup (#16607) 2026-06-13 13:17:32 +09:00
Christopher Milan
bcdb988df0
split comma benchmark, dsp on c4 [PR] (#16598) 2026-06-12 23:26:05 -04:00
George Hotz
6b8fdfe4ca
alu addrspace is where the math happens (#16606)
* alu addrspace

* fix cstyle/llvm

* on ptx, reg+alu are the same thing
2026-06-12 20:01:28 -07:00
wozeparrot
67a4f129c2
llama: fix bf16 gemm oob (#16603) 2026-06-12 19:43:05 -07:00
Christopher Milan
8862c7549c
new-style dcache_flush (#16602) 2026-06-12 22:25:08 -04:00
chenyu
9e72a6b376
more indexing cleanup [PR] (#16600) 2026-06-12 21:33:47 -04:00
chenyu
aa32d309db
fix rangeify indexing for pad/reduce (#16599) 2026-06-12 20:26:15 -04:00
George Hotz
96b86aad7b
move new style transform up more (#16593)
* move new style transform up more

* pm_move_gates_from_index works on new style
2026-06-12 17:20:12 -07:00
chenyu
a35964493e
UPat method cleanups [PR] (#16596) 2026-06-12 17:22:54 -04:00
chenyu
3036b15ed9
remove Tensor.ufix [PR] (#16594)
* remove Tensor.ufix [PR]

* inline _ufix_keep_dtype
2026-06-12 14:40:28 -04:00
qazal
b2e95b2db3
rangeify: no copies for write+read of same slice (#16585)
* failing test

* cleaner failing tests

* assign and read of same slice shouldn't create copies

* err in the changes

* shrink with no overlapping regions in dest is fine
2026-06-13 02:19:47 +09:00
George Hotz
833cb37574
move up new style transform (#16592)
* simpler names

* move up new style transform

* fix that rule
2026-06-12 10:13:37 -07:00
George Hotz
51100d2c5c
new style cleanups (#16584)
* spec tighten

* revert

* lin fix

* lin fix

* needed for x86

* revert
2026-06-12 08:10:38 -07:00
Philip Sinitsin
76c10cd635
jit: don't memplan buffers reachable from live tensors (#16588)
The memory planner was suballocating BUFFERs created during JIT capture that are still referenced by external lazy tensor graphs, like the .grad tensors assigned by backward(). The replay then only writes the arena slices, so realizing such a tensor after the call reads freshly allocated memory and silently returns zeros. Hold every BUFFER reachable from a live Tensor instead of only the parameters of the return value; true internals are still planned. Fixes #16571.
2026-06-12 17:51:54 +03:00
nimlgen
2bfdf85f87
hcq2: move pre bufferize (#16589)
* hcq2: move pre bufferize

* x
2026-06-12 16:11:59 +03:00