nimlgen
afe14ccbfa
amd: aql default when several xccs ( #11832 )
2025-08-26 15:16:36 +03:00
qazal
3674c0754e
viz: small uop click changes ( #11846 )
...
* also highlight self
* can always unselect by clicking outside
* less layout
2025-08-26 14:56:13 +03:00
qazal
f2a3c27372
viz: g.edges() once ( #11845 )
2025-08-26 13:29:59 +03:00
qazal
b0df3e62a8
viz: light up srcs and paths on UOp click ( #11844 )
...
* viz: light up srcs and paths on UOp click
* safari doesn't have context-stroke
* safari also has a bug
* safari acceptance
2025-08-26 09:03:09 +03:00
qazal
6236749867
viz: move rect styles to classes ( #11842 )
...
* viz: move rect styles to classes
* add rect
2025-08-26 07:55:34 +03:00
qazal
81ffa07439
viz: pass through nodes without a link ( #11841 )
2025-08-26 07:00:43 +03:00
Sieds Lykles
265d287615
add decomp for !x&!y -> !(x|y) ( #11836 )
2025-08-26 05:21:06 +02:00
chenyu
337e979a59
call dtypes.as_const in Tensor(list) ( #11840 )
2025-08-25 22:08:26 -04:00
George Hotz
215818379b
new (post) group for reduce ( #11837 )
...
* new (post) group for reduce
* fixes
* leave if
* fix locals
* size
* no vectorized buf
* image fixes
* don't track that
* fix ptx
* name buffer with reduce range
* remove unused in lowerer
* yay DEFINE_REG refactor
2025-08-25 18:03:00 -07:00
chenyu
ac3449b0c8
truncate_fp16 cleanup ( #11838 )
...
native `@` is default
2025-08-25 19:03:41 -04:00
qazal
e146418f65
hotfix: profiler content-type is application/octet-stream ( #11831 )
2025-08-25 15:56:42 +03:00
qazal
a1f6823060
viz: memory layout in client side ( #11830 )
...
* viz: memory layout in client side
* update test_viz
2025-08-25 14:49:33 +03:00
George Hotz
a6dbb09058
changes for postrange ( #11828 )
2025-08-24 17:37:07 -07:00
George Hotz
27701ef823
add locals support to rangeify ( #11826 )
2025-08-24 14:03:12 -07:00
Sieds Lykles
a286a1a6f7
Fast idiv try removing factors of two before cast ( #11824 )
...
* try removing factors of two
* dont return if None
* add test
2025-08-24 20:04:25 +02:00
George Hotz
a03b930339
hotfix: green v2 in docs
2025-08-24 10:25:14 -07:00
George Hotz
6540bb32a6
move into codegen late [pr] ( #11823 )
2025-08-24 10:23:25 -07:00
nimlgen
bba088ef11
amd aql queue ( #11708 )
...
* amd aql queue
* xcc
* fiz
* aql better
* llvm
* no for aql
* wrap
* is_sql
* am support
* complete
* fix
* mypy
* minor
2025-08-24 19:53:00 +03:00
George Hotz
1fa09d9ede
BLOCK_REORDER is context var, heuristic cleanups [pr] ( #11819 )
...
* BLOCK_REORDER is context var, heuristic cleanups [pr]
* split get opt and do opt
* oops, should be on
2025-08-24 09:41:34 -07:00
qazal
8b18cc2a94
viz memory layout cleanup ( #11820 )
...
* rename to dtype_size
* cleanr memory shape creator
2025-08-24 19:37:31 +03:00
Sieds Lykles
dd69114573
Revert "Better div nesting ( #11811 )" ( #11818 )
...
This reverts commit 952f729b07
2025-08-24 18:11:24 +02:00
nimlgen
e19f901330
amd: rptr/wptr in create_queue ( #11817 )
2025-08-24 18:03:45 +03:00
nimlgen
d71444857e
amd: apply relocs for kernel_code_entry_byte_offset for AMD_LLVM ( #11816 )
...
* amd: apply relocs for kernel_code_entry_byte_offset for AMD_LLVM
* fix
2025-08-24 17:48:40 +03:00
George Hotz
44bc7dc73d
remove KernelInfo from GROUP_REDUCE ( #11814 )
2025-08-23 19:55:41 -07:00
George Hotz
229adfb7c3
Revert "remove KernelInfo from gpudims ( #11809 )" ( #11813 )
...
This reverts commit 846753f343
2025-08-23 19:37:10 -07:00
Sieds Lykles
952f729b07
Better div nesting ( #11811 )
...
* remove check
* use fold_divmod_congruence instead of simplify
* adjust tests
* shorten line
2025-08-24 04:17:40 +02:00
Sieds Lykles
e652062f92
tweak divmod_folding condition ( #11810 )
2025-08-24 02:59:02 +02:00
George Hotz
846753f343
remove KernelInfo from gpudims ( #11809 )
...
* remove KernelInfo from gpudims
* that's good in there
2025-08-23 16:32:45 -07:00
Sieds Lykles
07d4ed7e4c
one more symbolic add variation ( #11807 )
2025-08-24 01:15:04 +02:00
qazal
759ebea4eb
viz: reflect timeline API boundary in names ( #11808 )
...
* define shapes once
* depth isn't an event property
* update server naming
2025-08-24 02:12:12 +03:00
George Hotz
132f09fab7
global/locals from AxisType in range ( #11806 )
2025-08-23 15:49:17 -07:00
qazal
0d86288bd7
viz: calculate timeline fixed points in client side ( #11805 )
...
* viz: calculate timeline fixed points in client side
* 26 bytes / event
* math
2025-08-24 01:44:40 +03:00
George Hotz
a75da49951
use AxisType for UPCAST/UNROLL ( #11800 )
...
* use AxisType for UPCAST/UNROLL
* fixes
* fix the bug
* fix hack
* bad test
* flaky test
2025-08-23 14:44:48 -07:00
qazal
2407fecdae
viz bytepack format ( #11792 )
...
* viz bytepack format
Training a 1B llama yields ~20M profiler events.
With JSON serialization, the browser tries to load 6GB to memory. This OOMs since each tab is limited to <3-4GB memory usage. Using a packed format, we only need ~600MB.
**Design decisions:**
- Timestamps are in microseconds relative to start time. They're stored in u32, which can express up to ~1 hr of trace events.
- Strings (kernel names, metadata, etc) are deduped.
- Buffer sizes are in u64 nbytes.
More optimization possible:
- The string lookup is a JSON dumped array, we can compress this.
- Can store less for memory by moving the layout to client.
**Results**
| | Events | JSON | bytepack |
|----------------|---------|-------------|-------------|
| DP=8 llama 1B train (`command: [1]`) | 24M | 5.8GB | 640MB |
| examples/beautiful_mnist.py | 16K | 3.7MB | 745KB |
| examples/gpt2.py | 55K | 12.54MB | 1.40MB |
`[1]`: `VIZ=1 FAKEDATA=1 OFFLOAD_OPTIM=1 DP=8 BS=8 GRADIENT_ACC_STEPS=2 BLOCK_REORDER=0 LR=3e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=8192 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py`
* python reference decoder
* 27 bytes / event, 1hr hard limit
2025-08-23 23:50:21 +03:00
qazal
b12d1d866c
count bytes per kernel in test_viz ( #11801 )
...
Currently at ~100 bytes/kernel with JSON.
2025-08-23 23:35:27 +03:00
Sieds Lykles
6a50ab6b87
adjust idiv min_max ( #11802 )
...
* change div min_max
* add tests
2025-08-23 22:25:51 +02:00
chenyu
9d4cccd0f9
test_dtype_alu cleanups ( #11799 )
2025-08-23 15:11:17 -04:00
George Hotz
aefabaf774
add AxisType to range ( #11798 )
...
* add AxisType to range
* missed them
* fix that test
* fix that test
2025-08-23 11:15:00 -07:00
qazal
b975830424
add profile loader helper in test_viz ( #11797 )
2025-08-23 19:20:29 +03:00
chenyu
7123df3928
Use Tensor.logaddexp to implement Tensor.softplus ( #11796 )
...
instead of piecewise linear, numerical is handled by logaddexp. jax does this and i think it's more elegant than torch's approach
2025-08-23 11:52:29 -04:00
qazal
aaea6b97ad
viz memory: compute nbytes ( #11795 )
...
* viz memory: compute nbytes
* local map
2025-08-23 17:34:07 +03:00
qazal
58653b5eae
viz: store memory scale ( #11794 )
2025-08-23 16:19:44 +03:00
chenyu
fb8ee02424
Tensor.logaddexp ( #11793 )
2025-08-23 09:15:00 -04:00
Sieds Lykles
5a6817d5f8
Fix z3 rendering of floats in indexing ( #11740 )
...
* Fix floating point comparison in indexing
* wrap in noop
* update tests
* improve rules for loading and comparing floats
* add test cast to bool
2025-08-23 05:56:19 +02:00
chenyu
4267c45db3
non-supported dtype in transcendental ( #11754 )
...
* non-supported dtype in transcendental
`CPU=1 python3 test/test_dtype_alu.py TestDTypeALU.test_bfloat16_unary` works
* test
* works on real mac
2025-08-22 23:13:45 -04:00
chenyu
e39b25cd36
upcast float exp to at least float32 ( #11758 )
...
* upcast float exp to at least float32
* unlucky seed
2025-08-22 20:16:34 -04:00
nimlgen
b057a90d49
memory: rename is_huge_page -> is_page ( #11786 )
2025-08-22 20:08:58 +03:00
qazal
38f0fa7bde
viz: only send trace duration ( #11789 )
...
* viz: only send trace duration
* can unwrap
2025-08-22 20:00:48 +03:00
qazal
1c81ec9248
viz: rename to start/end timestamp ( #11788 )
2025-08-22 19:47:49 +03:00
qazal
9ff03680ba
viz: store relative timestamps ( #11787 )
...
* viz: store relative timestamps
* err
* update test
2025-08-22 19:30:21 +03:00
