qazal
a80fb4e641
viz: better ordering of device engines in profiler ( #14590 )
2026-02-06 23:08:09 +09:00
qazal
b7e3fbe07e
llama: add VIZ=-1 to dev_run ( #14583 )
...
* llama: add VIZ=-1 to dev_run
* readme
* cleaner
* add profile.sh script
* better grouping of options
* add other row
* readme edits
* work
2026-02-06 22:59:22 +09:00
nimlgen
fbeb978170
diff devices for sdma ( #14589 )
...
* start
* x
* fix
* sdma
* c
* clean
* x
* hm
* cleaer
2026-02-06 16:39:12 +03:00
George Hotz
7cb996e153
bottom up earliest rewrites ( #14587 )
...
* better
* bottom up earliest rewrites
* fix
2026-02-06 18:13:07 +08:00
George Hotz
03af2404e2
small changes and test fixes from kernel is call ( #14586 )
2026-02-06 17:08:33 +08:00
George Hotz
3c26ce29b2
make disk tensor tests process safe ( #14584 )
2026-02-06 15:39:55 +08:00
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
chenyu
15d3344d9e
use int inputs in test_assign ( #14580 )
...
int is less flaky
2026-02-06 00:07:31 -05:00
qazal
50a166a5fa
viz: cleanup amdgpu target mapping ( #14579 )
...
* viz: cleanup amdgpu target mapping
* linter
* unwraps
2026-02-06 13:51:51 +09:00
chenyu
b09dc646f5
revert some late_buffer_view change ( #14578 )
...
revert #14478 which breaks tinyfs
2026-02-05 22:51:40 -05:00
chenyu
d41836f135
remove KERNEL special case in realize_assign [pr] ( #14573 )
2026-02-05 21:55:44 -05:00
George Hotz
6cbcf98627
KernelInfo is required on get_program ( #14571 )
...
* rangeify always adds KernelInfo
* fix tests
* skip flaky test
2026-02-06 10:49:27 +08:00
George Hotz
28c56a783c
add CallInfo and viz call toggle ( #14570 )
2026-02-06 09:30:58 +08:00
wozeparrot
f73468d516
fa: block skipping for fa kv bwd ( #14569 )
2026-02-05 16:13:53 -08:00
chenyu
b7ef775677
more cleanup in create_schedule [pr] ( #14566 )
...
fixed wrong comments and simplified queue building
2026-02-05 16:12:17 -05:00
Garret Castro
cee7ef7ab2
disable threads ( #14555 )
2026-02-05 16:11:32 -05:00
chenyu
79b7799dba
clean up linearize schedule [pr] ( #14565 )
...
* clean up linearize schedule [pr]
don't mix ScheduleItem and UOp in schedule queue
* ok
2026-02-05 15:24:09 -05:00
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
Christopher Milan
aa9dc50577
dtype decomps don't require bitshifts ( #14542 )
...
* dtype decomps don't require bitshifts
* simplify shr/shl
* ruff
2026-02-05 14:42:30 -05:00
Christopher Milan
b47397ab17
list ml_dtypes as dependency for DSP ( #14562 )
...
* pin onnxruntime to 1.23.2 for DSP
* list ml_dtypes instead
This reverts commit 84bb2cc0fc
2026-02-05 14:27:50 -05:00
chenyu
2b47a9a1b5
skip test_xlm_roberta_large ( #14563 )
...
symlink model not allowed in latest onnxruntime
2026-02-05 14:00:24 -05:00
chenyu
42c18da88a
add Ops asserts in toposort sched_sink [pr] ( #14561 )
...
more explicit
2026-02-05 12:40:02 -05:00
nimlgen
483bba4f05
nv: use prof_exec_counter ( #14559 )
2026-02-05 19:00:14 +03:00
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
chenyu
c0ca7f9c51
use more UOp.sum and UOp.prod [pr] ( #14549 )
2026-02-04 22:05:20 -05:00
chenyu
e8dace41b6
clean up UOp.vars [pr] ( #14547 )
2026-02-04 20:52:25 -05:00
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
chenyu
664f1bf76d
minor ops/jit cleanups [pr] ( #14543 )
2026-02-04 17:21:34 -05:00
chenyu
03d0fa9c3f
merge as_buf into buf_uop [pr] ( #14541 )
2026-02-04 16:32:23 -05:00
chenyu
43ef24a8af
remove buf_target [pr] ( #14540 )
...
not really needed
2026-02-04 15:03:47 -05:00
chenyu
8b7343b950
clean up is_realized [pr] ( #14538 )
...
base cannot be Ops.MULTI since MULTI is a view now
2026-02-04 14:24:10 -05:00
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
ec2b6bbda8
hcq: update signal logic ( #14531 )
2026-02-04 19:32:56 +03:00
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
024f57ecf5
jit input_buffers cleanup [pr] ( #14532 )
2026-02-04 10:14:38 -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
George Hotz
fb9df1e031
pretty print binary ( #14520 )
2026-02-04 18:04:35 +08:00
Christopher Milan
8c3c026d86
decomp float16 to float32 ( #14417 )
...
* decomp float16 to float32
* denormals arent zero
* add test
* denormals are zero
* fix
* oops
* bitcast works
* fix LOADs
* test_dtype passing
* cleanup
* mypy
* debug print
* only emulate if EMULATED
* very ugly, but passes spec
* add test_dtype_alu tests
* Revert "very ugly, but passes spec"
This reverts commit fdc3999b65
2026-02-04 01:37:47 -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
chenyu
9c2fc118ef
relax setitem target check ( #14526 )
...
old check was too conservative
2026-02-03 22:32:49 -05:00
qazal
d1bfbe9ce3
isolate slow llama gemm ( #14525 )
2026-02-04 12:20:10 +09:00
