qazal and GitHub
e684fcc689
pm_reduce_collapse fix for re enabling stack for cat of same shape (codex) ( #17251 )
...
* change
* stack cat
* remove nop
* change
2026-07-29 14:47:28 +09:00
George Hotz and GitHub
57ae1bc7a7
rename MULTI to UNSHARD ( #17267 )
...
* rename MULTI to UNSHARD
* comment updates (glm)
* rename method to unshard
2026-07-28 16:51:41 -07:00
chenyu and GitHub
23e9e76e8c
DEFAULT_FLOAT/DEFAULT_INT ContextVar [pr] ( #17265 )
2026-07-28 19:08:18 -04:00
chenyu and GitHub
fde3a8faff
CAPTURE_PROCESS_REPLAY=0 default chaging test [PR] ( #17261 )
...
* CAPTURE_PROCESS_REPLAY=0 default chaging test [PR]
quick patch before making default dtype context vars, which is quite annoying
* few more
2026-07-28 17:32:35 -04:00
qazal and GitHub
05e4727feb
arange stack regression test ( #17250 )
...
* simple failing test
* commend out stack
* a little smaller
* work
2026-07-28 20:00:22 +09:00
b1tg and GitHub
bdbb1d702f
fix shard axis through symbolic reshape ( #17238 )
...
* fix shard axis through symbolic reshape
* bind
2026-07-27 11:32:15 -04:00
chenyu and GitHub
818a892ebc
flip from_py to use weak dtypes [pr] ( #17229 )
2026-07-27 10:07:00 -04:00
chenyu and GitHub
456b5b5060
fix python_alu inf ( #17227 )
2026-07-26 20:04:12 -04:00
chenyu and GitHub
74c2121d99
promo (uint64, int) -> weakfloat like JAX [pr] ( #17204 )
2026-07-25 15:20:42 -04:00
chenyu and GitHub
ee2ccb1f24
put weakint in dtypes.weaks [pr] ( #17202 )
...
* put weakint in dtypes.weaks [pr]
* custom_add_var
2026-07-25 14:00:34 -04:00
chenyu and GitHub
0a6125e571
Invalid is bool, put weakint in promo lattice [pr] ( #17188 )
2026-07-25 00:06:53 -04:00
George Hotz and GitHub
08eceb0ad9
llm: fix generated tokens in usage accounting ( #17174 )
2026-07-23 21:38:07 -07:00
chenyu and GitHub
1b616669d2
few more symbolic rules [pr] ( #17161 )
...
bitcast twice can skip middle one; bool != False is self; two more where combine
2026-07-23 18:05:32 -04:00
sirhcm and GitHub
6ee16eb1fa
split out LLVM compilers ( #17160 )
2026-07-23 17:45:23 -04:00
chenyu and GitHub
250de4b142
use stack for same shape cat [pr] ( #17156 )
...
same device multi cat benefits too
2026-07-23 13:50:28 -04:00
chenyu and GitHub
9267fca91a
use STACK in wino ( #17155 )
...
STACK is great
2026-07-23 13:02:42 -04:00
chenyu and GitHub
e9a86c99ed
more symmetrical x//d<c -> x<c*d [pr] ( #17146 )
...
* more symmetrical x//d<c -> x<c*d [pr]
* parse_valid too
2026-07-23 00:19:53 -04:00
chenyu and GitHub
6a9c23b1c1
more symmetrical div_and_mod_symbolic [pr] ( #17145 )
2026-07-22 23:34:03 -04:00
chenyu and GitHub
855175123b
fix z3 AND in validate ( #17139 )
2026-07-22 16:09:28 -04:00
chenyu and GitHub
a7c693d2fd
rework pm_lower_weakint [pr] ( #17136 )
...
deleted pm_no_index and many "remove hanging casts"
2026-07-22 15:53:11 -04:00
chenyu and GitHub
557e674861
delete pm_index_invalid [pr] ( #17130 )
2026-07-22 11:50:08 -04:00
chenyu and GitHub
17557d7fdf
bring back fold_where_closure [pr] ( #17134 )
...
* bring back fold_where_closure [pr]
generic enough to delete delete pm_index_invalid
* cannot cache self
2026-07-22 11:08:14 -04:00
chenyu and GitHub
b1060ca708
don't promote dtype in _pad_constant [pr] ( #17125 )
2026-07-21 18:13:50 -04:00
chenyu and GitHub
46b82d4755
don't auto cast cond for WHERE ( #17115 )
...
no or_casted all WHEREs with single mixin, matched torch
2026-07-21 13:00:07 -04:00
chenyu and GitHub
f64f96ec59
broadcast_axes [PR] ( #17112 )
...
prerequisite to simplify broadcasting logic and make it implicit
2026-07-21 11:53:12 -04:00
sirhcm and GitHub
f3a5337825
correct spelling of coalesce ( #17103 )
2026-07-20 21:44:04 -04:00
chenyu and GitHub
13ca9bd8a6
remove dtypes.index again ( #17104 )
...
also reverted some dtype change, the split made things needlessly complicated
2026-07-20 20:30:04 -04:00
sirhcm and GitHub
980748ccfc
add multiple_of to ParamArg ( #17101 )
2026-07-20 20:11:54 -04:00
George Hotz and GitHub
f7ce7f330d
llm: minor fixes + tests ( #17099 )
...
* llm: minor fixes + tests
* error
2026-07-20 14:31:20 -07:00
Pol Puigdemont Plana and GitHub
ef77963cfd
derivative of logsumexp is independent of max ( #17088 )
...
same as #7009 but for logsumexp and logcumsumexp.
fwd+bwd kernel count 5 -> 3 for both. gradients unchanged
(ties, -inf masks, torch-compared at grad_atol=1e-7).
2026-07-20 06:52:16 -07:00
chenyu and GitHub
0146a30125
improve cast to unsign min_max [pr] ( #17078 )
2026-07-18 21:58:41 -04:00
George Hotz and GitHub
b53cd35cff
llm: make tokenizer fast (kimi) ( #17077 )
...
* llm: make tokenizer fast
* simpler
* re.escape + qcom mypy fix
2026-07-18 17:31:59 -07:00
chenyu and GitHub
f315df29a0
no weak Tensor from and to real buffer ( #17067 )
...
* no weak Tensor from and to real buffer
creation, assign, safe_save
* is_numpy_ndarray to tensor
* one more
2026-07-17 16:09:10 -04:00
George Hotz and GitHub
86a6ad8ed2
llm: split cli.py into serve.py with the HTTP server ( #17065 )
...
* llm: split cli.py into serve.py with the HTTP server
* min edit
2026-07-17 10:45:37 -07:00
George Hotz and GitHub
3ee2baf71d
llm: add tool calling support (kimi) ( #17061 )
...
* llm: add tool calling support
* simpler
* cls
* gpt cleanup
* more gpt cleanups
* tests for tools calling
2026-07-17 10:20:01 -07:00
George Hotz and GitHub
46172bb7c7
llm: add optional jinja template support (kimi) ( #17058 )
...
* add jinja template support (kimi)
* fix tests
* lil
* more crap to fallback
2026-07-16 19:02:58 -07:00
George Hotz and GitHub
61e104bdfb
use UOp.wmma everywhere ( #17045 )
...
* use UOp.wmma everywhere
* fix
2026-07-16 10:40:48 -07:00
chenyu and GitHub
5a4156c5d1
bitcast and element_size raise for weak dtypes ( #17046 )
2026-07-16 13:07:45 -04:00
chenyu and GitHub
be075b200a
weak dtypes in dtype_from_uop [PR] ( #17032 )
...
* weak dtypes in dtype_from_uop [PR]
* no weak in spec_program
* weak const fold tests
2026-07-15 16:54:31 -04:00
nimlgen and GitHub
d6fddb066f
usb: keep only custom ( #17029 )
...
* usb: keep only custom
* mockgpu by gpt
* gpt said sorry
* revert
* reset
* fix
* flash
2026-07-15 22:40:52 +03:00
chenyu and GitHub
0d19970edc
least_upper_dtype in dtype_from_uop [PR] ( #17028 )
2026-07-15 09:27:53 -04:00
chenyu and GitHub
47ddf94f17
remove InvalidType lt and gt ( #17023 )
...
not really used
2026-07-14 21:59:15 -04:00
sirhcm and GitHub
c9baa2ef79
use pattern matcher in contiguous_view_offset [PR] ( #17022 )
2026-07-14 19:37:36 -04:00
chenyu and GitHub
0410c9325d
make test/null follow the SPEC ( #17010 )
2026-07-13 14:01:41 -04:00
George Hotz and GitHub
03ecad9486
full removal of dtype.vec ( #16996 )
...
* full removal of dtype.vec
* fix typo
2026-07-12 09:25:50 -07:00
chenyu and GitHub
5a0751076d
simpler vconst_like ( #16990 )
...
also always use const_like in symbolic, fixed a crash
2026-07-12 08:21:57 -04:00
chenyu and GitHub
047a467bf9
delete UOp._stack and UOp.vectorize [PR] ( #16988 )
2026-07-11 15:26:55 -04:00
chenyu and GitHub
9d47014fd8
first class STACK [PR] ( #16986 )
2026-07-11 13:55:32 -04:00
chenyu and GitHub
2fda6b3888
fix shape broadcast for symbolic ( #16970 )
...
* fix shape broadcast for symbolic
0 or smax was wrong for the unresolved case
* test with null
2026-07-10 15:13:01 -04:00
chenyu and GitHub
8d546f55e9
dtypes.weakfloat ( #16956 )
2026-07-09 19:32:21 -04:00