sirhcm/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests 2 Actions Activity
11,106 Commits 336 Branches 15 Tags
python_speed
Commit Graph
11106 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
George HotzandGitHub 846753f343 remove KernelInfo from gpudims (#11809) 
* remove KernelInfo from gpudims

* that's good in there
 2025-08-23 16:32:45 -07:00
Sieds LyklesandGitHub 07d4ed7e4c one more symbolic add variation (#11807) 2025-08-24 01:15:04 +02:00
qazalandGitHub 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 HotzandGitHub 132f09fab7 global/locals from AxisType in range (#11806) 2025-08-23 15:49:17 -07:00
qazalandGitHub 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 HotzandGitHub 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
qazalandGitHub 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
qazalandGitHub 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 LyklesandGitHub 6a50ab6b87 adjust idiv min_max (#11802) 
* change div min_max

* add tests
 2025-08-23 22:25:51 +02:00
chenyuandGitHub 9d4cccd0f9 test_dtype_alu cleanups (#11799) 2025-08-23 15:11:17 -04:00
George HotzandGitHub 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
qazalandGitHub b975830424 add profile loader helper in test_viz (#11797) 2025-08-23 19:20:29 +03:00
chenyuandGitHub 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
qazalandGitHub aaea6b97ad viz memory: compute nbytes (#11795) 
* viz memory: compute nbytes

* local map
 2025-08-23 17:34:07 +03:00
qazalandGitHub 58653b5eae viz: store memory scale (#11794) 2025-08-23 16:19:44 +03:00
chenyuandGitHub fb8ee02424 Tensor.logaddexp (#11793) 2025-08-23 09:15:00 -04:00
Sieds LyklesandGitHub 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
chenyuandGitHub 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
chenyuandGitHub 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
nimlgenandGitHub b057a90d49 memory: rename is_huge_page -> is_page (#11786) 2025-08-22 20:08:58 +03:00
qazalandGitHub 38f0fa7bde viz: only send trace duration (#11789) 
* viz: only send trace duration

* can unwrap
 2025-08-22 20:00:48 +03:00
qazalandGitHub 1c81ec9248 viz: rename to start/end timestamp (#11788) 2025-08-22 19:47:49 +03:00
qazalandGitHub 9ff03680ba viz: store relative timestamps (#11787) 
* viz: store relative timestamps

* err

* update test
 2025-08-22 19:30:21 +03:00
nimlgenandGitHub 698392334f system: message for eaccess as well (#11785) 2025-08-22 18:21:32 +03:00
geohotstanandGitHub 1e679bd789 fix max_unpool2d inf (#11784) 
* start

* add regression test for maxunpool2d
 2025-08-22 08:31:24 -04:00
George HotzandGitHub 9832599c9e test_vmap + permute isn't a sint (#11783) 
* test_vmap + permute isn't a sint

* order
 2025-08-21 22:39:35 -07:00
George HotzandGitHub bb8de51e5f remove unused early cleanups + contig w range [pr] (#11780) 
* remove unused early cleanups [pr]

* contiguous with range

* woah, this works
 2025-08-21 20:04:45 -07:00
chenyuandGitHub 91a4de4ca7 fix getitem with inf in tensor (#11781) 2025-08-21 21:55:32 -04:00
George HotzandGitHub 66e9d54eed RANGEIFY=2 is partial contig (#11777) 2025-08-21 16:53:58 -07:00
Jordan ChalupkaandGitHub 8de6db15ac exclude .git from ruff (#11773) 2025-08-21 15:37:50 -07:00
George HotzandGitHub 5954a0975f fix some assigns on rangeify (#11774) 
* fix some assigns

* llvm test

* more tests

* upd test
 2025-08-21 15:15:54 -07:00
qazalandGitHub 2e0eb88549 viz: add metadata to UOp tracing (#11772) 
* viz: add metadata to UOp tracing

* place after tag

* optional field

* err, refcount of root must be 0
 2025-08-22 00:18:45 +03:00
George HotzandGitHub d6f9606e93 small cleanups to rangeify (#11769) 2025-08-21 11:15:09 -07:00
uuuvnGitHubwozeparrot
 		bd4a9473b0 Multihost exception handling (#11729) 
Co-authored-by: wozeparrot <[email protected]>
 2025-08-21 13:51:49 -04:00
George HotzandGitHub a2c7b807e0 don't bufferize 0s (#11766) 2025-08-21 10:10:56 -07:00
nimlgenandGitHub 9eff7cd1d8 am: support 64bit discovery (#11768) 2025-08-21 18:28:13 +03:00
b1tgGitHubb1tgchenyu
 		56cd47a159 fix amd llvm bf16 tc (#11713) 
* fix amd llvm bf16 tc

* is_cdna

---------

Co-authored-by: b1tg <[email protected]>
Co-authored-by: chenyu <[email protected]>
 2025-08-21 09:33:28 -04:00
George HotzandGitHub a044648111 rangeify load cleanups + multi support (#11765) 
* use the old buf_uop + cleanups

* simpler handling of load

* everything needed for multi too
 2025-08-20 20:55:49 -07:00
George HotzandGitHub 9f94c25a25 fix symbolic usage. use shrink, not reshape (#11762) 
* fix test_var

* revert those things

* fix the ones in test tiny

* use better syntax

* it's the same, but that's clearer

* fix pad
 2025-08-20 18:35:42 -07:00
chenyuandGitHub 5276fbc9c5 fix gather with inf values (#11760) 
(mask * x) is wrong because 0*inf is nan. i feel we have a lot of those still...
 2025-08-20 20:35:40 -04:00
wozeparrotandGitHub b979162c5d llama3 eval train (#11706) 2025-08-20 19:56:35 -04:00
chenyuandGitHub dbd3b67657 clamp GRAD_CLIP_NORM in llama (#11761) 2025-08-20 19:55:50 -04:00
George HotzandGitHub 9635592141 ** rangeify, try 3 (#11683) 
* ** rangeify, try 3

* bring that over

* bufferize, don't use contig tag

* work

* ish

* fix rangeify

* flash attention is back

* fix rangeify tests

* stuff passes

* fix test_log_softmax

* more stuff passes

* progress children

* new endrange solution

* progress

* progress counter

* basic assign

* contigs only

* symbolic in schedule

* unbind_kernel

* late children

* ops fixed

* beautiful mnist is close

* that seems to work

* mnist works

* improve names

* fix bmnist

* no pcontig

* testing backward

* work

* clone movement ops

* new_range helper

* MBLOCK/MERGE

* ops tests pass

* revert mblock stuff

* cleanups...but it breaks ops

* remove reindex

* hack for relu

* disable the hacks

* more hacks

* upd

* mostly works with cleanups disabled

* ndr

* ops tests pass

* terrible hacks for indexing to work

* context mismatch

* pcontig

* split pcontig v contig

* z3 trunc

* null

* no fuse in rangeify

* ops test passes

* lnorm

* fix assign

* nd rangeify

* both should work

* tests for rangeify

* cleanups

* stores pass the pointer through

* disable pcontig for now

* PARTIAL_CONTIG is a flag
 2025-08-20 14:22:44 -07:00
chenyuandGitHub d7553721d1 clean up test_dtype_alu (#11757) 
remove the check that looks into schedule, only test if output matches
 2025-08-20 14:36:18 -04:00
chenyuandGitHub 5f08a3e928 hotfix: cast half to float in Tensor.tolist (#11755) 
workaround for python < 3.12
 2025-08-20 12:18:35 -04:00
qazalandGitHub de4cb722a4 viz: add metadata and var_vals tracing (#11753) 
* viz: add metadata and var_vals tracing

* add test_trace_metadata

* set TRACEMETA=1
 2025-08-20 18:39:51 +03:00
nimlgenandGitHub 6589c9e643 hcq: better errors for ifaces (#11751) 
* hcq: better errors for ifaces

* fix linter

* typo

* space
 2025-08-20 17:50:51 +03:00
chenyuandGitHub be7b0b6970 TRANSCENDENTAL_SUPPORTED_DTYPES->TRANSCENDENTAL_DTYPES (#11752) 2025-08-20 10:29:36 -04:00
ttomsaandGitHub 220a2a88d7 a*(1/b) -> a/b on LLVM, CPU (#11743) 
* add fdiv rewrite

* :)

* use float_lop

* use reciprocal()

* revert

* move to decompositions
 2025-08-20 09:35:10 -04:00
George HotzandGitHub 12ab3f8b06 correct row_count in process replay (#11748) 2025-08-19 22:21:07 -07:00