George Hotz
3cfd4c915f
warp_num
2025-10-29 19:08:10 +08:00
George Hotz
abb01ce6a0
last
2025-10-29 18:57:38 +08:00
George Hotz
6853da2a2f
works
2025-10-29 18:45:16 +08:00
George Hotz
154ddd98fd
it works, it's just slow...
2025-10-29 18:35:44 +08:00
George Hotz
e4ef94cf10
both is broken
2025-10-29 18:15:38 +08:00
George Hotz
0c274151ad
tensor core works
2025-10-29 18:06:05 +08:00
George Hotz
c4e32d4f63
demote op
2025-10-29 17:47:12 +08:00
George Hotz
a9d91ffcfc
DEMOTE op for putting globals in locals
2025-10-29 17:22:59 +08:00
George Hotz
819592ee67
hotfix: disable DoubleMatmul for PTX
2025-10-29 16:37:17 +08:00
George Hotz
30ca3f2af8
all double matmul ( #12993 )
...
* fix more double matmuls
* a few more
* all double matmul passes
* opts for flash attention
* fix spec
* comment
2025-10-29 16:25:27 +08:00
Sieds Lykles
9f39f6391c
shared_codegen_spec and fix index spec ( #12967 )
...
* split shared_codegen_spec and fix index
* add VCONST to program_spec and move index to shared_codegen_spec
* working ignore_oob=0
* cleanup
* fix spec
* undo that
* move barrier and special earlier
* fix more spec issues
* more updates
* remove special from program_spec
* cleanup and fixes
* move more to shared
* special is not in shared_spec
* some comments
* dont do bounds check there
2025-10-29 09:14:11 +01:00
George Hotz
1c362736aa
fix more double matmuls ( #12991 )
...
* fix more double matmuls
* a few more
2025-10-29 16:09:48 +08:00
George Hotz
e42b4edf8c
remove if stuff ( #12992 )
2025-10-29 15:29:35 +08:00
George Hotz
8c47cf4323
pcontig double matmul works ( #12899 )
...
* pcontig double matmul works
* tests
* contract
* closer
* works-ish
* add that broadcast
* 2 more work
* something
* disable broken ones
* llvm
* align 16
2025-10-29 13:06:43 +08:00
George Hotz
35b6f4148d
delete untested quantize ( #12990 )
2025-10-29 12:46:32 +08:00
Sieds Lykles
5ce8a1d2f2
Merge adjacent try all permutations for reduce ( #12972 )
2025-10-29 05:04:54 +01:00
George Hotz
b147e7e8e6
flatten bufferize ( #12984 )
...
* flatten bufferize
* simpler
* tests pass
* flat
* not flat
2025-10-29 11:23:43 +08:00
qazal
a7dac11aad
viz: keep rewrite step in back button history ( #12986 )
2025-10-29 11:09:43 +08:00
qazal
37967fa17b
viz: add integer query param helper and more typing ( #12985 )
...
* viz: query param helper
* json.dumps once
2025-10-29 10:44:01 +08:00
chenyu
fb53bdad5d
unused propagate_invalid rules [pr] ( #12983 )
...
named is not used, so you know it never matched
2025-10-28 22:16:50 -04:00
chenyu
ef16e6c68c
unwrap instead of cast [pr] ( #12982 )
2025-10-28 21:29:23 -04:00
chenyu
f55fcfecf9
ProgramSpec uops must end with SINK [pr] ( #12981 )
2025-10-28 17:12:22 -04:00
chenyu
9442442cb1
update variable names in search [pr] ( #12979 )
...
no lin nor linearize
2025-10-28 15:37:52 -04:00
wozeparrot
d66c997a39
feat: thunderkittens fa2 ( #12955 )
2025-10-28 11:27:45 -07:00
b1tg
bb307b9e81
fix fp8 vectorization ( #12977 )
...
* fix fp8 vectorization
* add fp8 tc to benchmark
2025-10-28 13:55:30 -04:00
nimlgen
c11dd56956
amd: cleanup import urls ( #12976 )
2025-10-29 00:43:02 +08:00
George Hotz
5e01cc299b
zero len ranges fail ( #12974 )
...
* zero len ranges fail
* fix Python backend
* fix llvm
* fix ptx
* yolo fix nir
* this works...
* always store...
* always store...
* Revert "always store..."
This reverts commit 0816cf344d
2025-10-28 22:49:55 +08:00
George Hotz
e936aa7974
cleanups from if range branch ( #12973 )
2025-10-28 20:58:47 +08:00
qazal
901d27b3ba
viz: optional text dims try 2 ( #12971 )
2025-10-28 18:54:28 +08:00
George Hotz
f5a3b33d33
add fun with nhwc convs
2025-10-28 17:12:22 +08:00
George Hotz
907499b02c
clean up GROUP/SINK ( #12969 )
...
* clean up GROUP/SINK
* fix end
* range_str color
2025-10-28 16:08:10 +08:00
Sieds Lykles
e22c5e7e73
process_replay uses opts argument for KernelInfo.opts_to_apply ( #12946 )
...
* opts_to_apply is opts
* skip beamed kernels
* simpler change
* fix the tensor cores tests for process replay
* use opts
2025-10-28 09:00:28 +01:00
George Hotz
6c9560a846
more syntactic sugar for pyrender ( #12968 )
2025-10-28 15:24:33 +08:00
George Hotz
b0da173f2f
add unique to const, fix longstanding bug ( #12965 )
...
* add unique to const, fix longstanding bug
* _force_unique=True
* fix tests
* fix more tests
2025-10-28 15:11:37 +08:00
Sieds Lykles
e110f4632a
split cat (on cpu) ( #12864 )
...
* split ranges but only on cpu
* except KernelOptError for threads
* use GROUP and END
* no more flatten_range needed
* remove noop end
* always process replay for openpilot
* update test
* skip test
* fix in outs calculation
With the new linearizer the toposort is a problem, this matches the spec
now
* undo that
2025-10-28 07:55:19 +01:00
qazal
3b82dee625
viz: match DEBUG=2 for exec item metadata ( #12966 )
...
* viz: match DEBUG=2 for exec item metadata
* remove repr from kernel
2025-10-28 14:53:57 +08:00
qazal
99589dea81
move viz edge tagging to UOp graph ( #12964 )
2025-10-28 12:46:23 +08:00
George Hotz
bbe0bebbf3
no range tags in kernels ( #12962 )
2025-10-28 12:33:48 +08:00
George Hotz
39c2117dea
cleanup pyrender ( #12961 )
2025-10-28 10:47:39 +08:00
George Hotz
2832954bcb
test with IGNORE_OOB=0 ( #12960 )
2025-10-28 10:32:19 +08:00
George Hotz
7784cec48e
pytest-split on spec ( #12959 )
2025-10-28 10:09:01 +08:00
George Hotz
4d817a289e
simplify spec ( #12958 )
...
* simplify spec
* more
2025-10-28 09:52:32 +08:00
George Hotz
62e62d8760
move verify to spec / cleanup ( #12956 )
...
* move verify to spec / cleanup
* lil
* more explicit
2025-10-28 08:58:10 +08:00
wozeparrot
24884c6768
fix: don't use KITTENS_HOPPER for 4090 ( #12954 )
2025-10-27 17:19:53 -07:00
nimlgen
372d9e5753
hcq: helper for visible devices ( #12950 )
...
* hcq: helper for visible devices
* fix
* f
2025-10-28 02:27:56 +08:00
Justin Erenkrantz
f2ffe9c8cf
Apply an override for nbio 7.3.0 to 7.2.0. ( #12949 )
2025-10-27 11:10:10 -07:00
qazal
63484d837e
Revert "viz graph drawing cleanups ( #12933 )" ( #12947 )
...
This reverts commit 189582db5e
2025-10-28 00:39:37 +08:00
chenyu
a79832b01f
control_flow.py -> linearizer.py [pr] ( #12948 )
2025-10-27 12:38:13 -04:00
b1tg
45e2f916a3
add quantize fp8 in llama3 ( #12893 )
...
* add quantize fp8 in llama3
* don't truncate fp8 alu result
* cast to float32 before matmul
* --model weights/LLaMA-3/8B-SF-DPO/
---------
Co-authored-by: chenyu <chenyu@fastmail.com >
2025-10-27 10:22:57 -04:00
George Hotz
25c2da1579
check SPEC=2 in CI ( #12945 )
...
* check SPEC=2 in CI
* split SPEC=2
* fast enough
2025-10-27 21:53:57 +08:00
