github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-07 22:23:55 -05:00
Code Issues Packages Projects Releases Wiki Activity
11,449 Commits 204 Branches 13 Tags
8dcba2e2cc74c81f57c9ea81c0bad881f0d5144e
Commit Graph

11449 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
George Hotz
8dcba2e2cc no full_rewrite [pr] (#13809) 
* no full_rewrite [pr]

* fix

* fix docs
 2025-12-22 23:20:01 -05:00
George Hotz
edce2303f4 rewrite to program (#13808) 2025-12-22 20:03:33 -05:00
George Hotz
2af2b4da5d Revert "rewrites for renderer and compiler (#13646)" (#13806) 
This reverts commit 339dadf056.
 2025-12-22 19:21:33 -05:00
George Hotz
339dadf056 rewrites for renderer and compiler (#13646) 
* rewrites for renderer and compiler

* full_rewrite_to_program

* fix pre-commit

* compiler passed into get_program

* no pkl compiler

* lib on program spec

* fix spec

* fix test

* no device

* compiler_device

* nm

* fix nir

* fix

* simplest

* fix tests

* revert
 2025-12-22 18:58:43 -05:00
Daniel Xu
4edaaf19e5 Handle tied embeddings for llama 3.2 1B (#13796) 
Previously the output.weight layer would not be loaded, and would only
contain randomly initialized values. This led to junk when doing a
forward pass.

Signed-off-by: Daniel Xu <daniel@thinkingmachines.ai>
 2025-12-22 16:31:40 -05:00
chenyu
7f1d41c9f9 delete files that import ShapeTracker (#13805) 2025-12-22 15:54:18 -05:00
qazal
b31373ca70 remove llvm-mca stuff from viz (#13802) 2025-12-23 01:41:51 +08:00
chenyu
27d899ce97 TRAIN=0 to only eval llama (#13804) 2025-12-22 11:55:46 -05:00
chenyu
39d962106f update llama logging (#13803) 
```
REWRITE_STACK_LIMIT=1000000 SMALL=1 BASEDIR=/raid/datasets/c4-8b SAMPLES=1000 BS=8 DP=8 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=8B SEQLEN=1024 PYTHONPATH=. MODEL=llama3 python3 examples/mlperf/model_train.py

    1 93.44 s run, 11.8750 loss, 0.000000000001 LR, 642.43 GB used,  19644.30 GFLOPS
    2 101.78 s run, 11.8750 loss, 0.000000000001 LR, 1454.57 GB used,  17039.35 GFLOPS
    3 7.34 s run, 11.8750 loss, 0.000000000002 LR, 1454.57 GB used, 236258.78 GFLOPS
    4 4.32 s run, 11.8750 loss, 0.000000000002 LR, 1454.57 GB used, 401488.40 GFLOPS
    5 4.36 s run, 11.9375 loss, 0.000000000003 LR, 1454.57 GB used, 398116.13 GFLOPS
    6 4.32 s run, 11.8750 loss, 0.000000000003 LR, 1454.57 GB used, 401878.60 GFLOPS
    7 4.34 s run, 11.8750 loss, 0.000000000004 LR, 1454.57 GB used, 399822.57 GFLOPS
    8 4.35 s run, 11.8750 loss, 0.000000000004 LR, 1454.57 GB used, 398512.24 GFLOPS
    9 4.36 s run, 11.8750 loss, 0.000000000005 LR, 1454.57 GB used, 397832.61 GFLOPS
   10 4.40 s run, 11.8750 loss, 0.000000000005 LR, 1454.57 GB used, 394520.83 GFLOPS
```
 2025-12-22 11:28:29 -05:00
qazal
389f01c7f4 viz: amdgpu assembly basic block graph (#13755) 2025-12-22 23:17:16 +08:00
George Hotz
df0f9d6860 add olmoe support to llm (#13792) 
* add olmoe support to llm

* cleanups

* simpler

* clean

* fix mypy

* lil

* remove dumb assert
 2025-12-22 10:41:35 -04:00
qazal
81d9053013 roc: cast to nullptr instead of changing header (#13801) 2025-12-22 22:34:06 +08:00
nimlgen
d299d30f2c am_smi: fix with new autogen (#13800) 2025-12-22 16:53:26 +03:00
nimlgen
f6bda6ae4e am: continue from saved state (#13799) 
* am: gfx queue cont

* f

* reset

* f

* l
 2025-12-22 15:55:07 +03:00
qazal
6237bd86f6 sqtt/pmc viz improvements (#13797) 2025-12-22 18:16:35 +09:00
Sitananda Prasad
3000b8d762 symbolic: add x ^ x -> 0 folding pattern (#13794) 2025-12-21 21:47:28 -04:00
chenyu
5cb827f7bf clean up can_lossless_cast and add missing pairs [p] (#13793) 2025-12-21 12:18:33 -05:00
George Hotz
75a6a03664 add qwen3 moe support to tinygrad.apps.llm (#13775) 
* qwen moe works

* simple moe

* one test

* integration
 2025-12-21 12:36:02 -04:00
chenyu
29ef0809bb can_safe_cast -> can_lossless_cast (#13789) 
safe cast in numpy only means the result won't overflow, so lossless is more precise
 2025-12-21 11:29:19 -05:00
chenyu
ed1fd7023b use getattr in dtype.truncate [pr] (#13788) 2025-12-21 11:05:43 -05:00
qazal
9839838fdd viz UOp layout cleanup (#13787) 
* use the same names in server and client

* first layout args, then renderer args
 2025-12-21 22:11:40 +08:00
nimlgen
e523971028 am: make mqd contig (#13786) 2025-12-21 17:00:33 +03:00
qazal
09e060eab5 simplify viz node labels (#13784) 2025-12-21 16:45:06 +08:00
qazal
dc660c9fc0 remove stale / untested viz related files (#13785) 2025-12-21 16:42:48 +08:00
George Hotz
59c02dd87f does this fix the dtype test? (#13779) 
* does this fix the dtype test?

* simpler
 2025-12-20 17:31:46 -04:00
George Hotz
5228f7bd06 hotfix: opencode should not reformat files 2025-12-20 15:55:29 -04:00
chenyu
733ef0452c update test_uop_resolve (#13777) 
plain @unittest.expectedFailure is too broad
 2025-12-20 12:40:59 -05:00
nimlgen
3db2104fb8 am: timeout sos start (#13776) 2025-12-20 17:41:33 +03:00
qazal
94f97f6988 generic viz cleanups from the basic blocks branch (#13774) 
* simpler codeblock highlight

* simpler append

* status enum
 2025-12-20 18:18:03 +08:00
George Hotz
a987a8ed44 add neg VIZ support to not start server (#13772) 2025-12-20 00:36:38 -04:00
qazal
b7c2f0dd1b remove stale extra/sched directory (#13770) 2025-12-20 11:57:30 +08:00
George Hotz
86cd1e9e81 remove UPatAny for typing fix [pr] (#13766) 
* remove UPatAny for typing fix [pr]

* fix dtype
 2025-12-19 17:41:18 -04:00
George Hotz
4702da41d5 hotfix: mkdir for extra/disassemblers 2025-12-19 17:18:37 -04:00
George Hotz
45c459848d remove more stale stuff (#13765) 
* remove more stale stuff

* remove disassemblers/adreno

* stale
 2025-12-19 17:14:56 -04:00
George Hotz
744af193f0 remove ScheduleItem and merge it with ExecItem (#13759) 
* remove ExecItem and merge it with ScheduleItem

* less diff

* fix issues

* min diff

* don't change bufs in _lower

* min diff

* update

* revert

* fixes

* diff
 2025-12-19 17:04:24 -04:00
George Hotz
df6cde8a00 cleanup stale examples/extra (#13764) 
* cleanup stale files

* examples

* move those back

* old

* delete more
 2025-12-19 16:27:37 -04:00
chenyu
80b84f5267 ruff lint tinykitten (#13762) 
deleted used import and double spaces. a few ignore to not change the real code
 2025-12-19 14:31:00 -05:00
Christopher Milan
97103831c5 Revert "remove image from BufferSpec (#13636)" (#13761) 
This reverts commit 2571a1eb47.
 2025-12-19 13:54:36 -05:00
Christopher Milan
2571a1eb47 remove image from BufferSpec (#13636) 
* remove image from BufferSpec

* cl tiny_gemm (64) works

* mypy

* padding

* openpilot CL

* reshape properly

* remove extra qcom checks

* pad output

* mypy

* update compile test

* move undo

* TestImageCopy valid images

* TestImageRealization valid images

* TestImageDType valid images

* cleanups

* test_renderer_failures

* ruff

* mypy

* simplify ops_qcom

* bump step time
 2025-12-19 13:41:20 -05:00
chenyu
185a000882 gradient of COPY (#13760) 2025-12-19 13:33:59 -05:00
nimlgen
57fe4d0a59 am: no_update_ptr for master (#13757) 2025-12-19 19:37:37 +03:00
chenyu
7fcd3cf991 hotfix SPEC for AFTER(CONTIGUOUS) (#13752) 
fixed spec error in `PYTHONPATH="." REWRITE_STACK_LIMIT=5000000 NULL=1 DEFAULT_FLOAT="HALF" BERT_LAYERS=2 BENCHMARK=10  BS=128 GPUS=1 MODEL=bert python3 examples/mlperf/model_train.py`
 2025-12-19 10:05:45 -04:00
qazal
81b5815a66 viz: minimal data to render a graph (#13754) 2025-12-19 16:19:28 +08:00
Christopher Milan
849e46da21 DLL: _PATH variables can be parent dir (#13753) 2025-12-19 00:28:02 -05:00
qazal
159c0e92fa viz: infrastructure for basic block graphs (#13751) 2025-12-19 13:08:19 +08:00
George Hotz
fa40df972f fix tests for NV (#13744) 
* small fix

* min diff

* bfloat16 out
 2025-12-18 13:20:21 -04:00
nimlgen
77191fb744 hive_reset for mi350 (#13746) 2025-12-18 12:02:28 +03:00
nimlgen
ceff388f3d am: extend va space (#13745) 2025-12-18 11:20:43 +03:00
wozeparrot
99e667bdcd tk fa bwd (#13480) 2025-12-17 23:56:37 -08:00
George Hotz
aeb7516c8a tests passing on tinybox h3 (#13742) 2025-12-17 19:04:34 -04:00