tinygrad/tinygrad
2
0
Fork 3
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-08-29 11:16:07 +00:00
Code Issues Packages Projects Releases Wiki Activity
14,184 Commits 340 Branches 16 Tags
f92ac85fbdd9d0c5c4716bb90348b59310deebd8
Commit Graph
14184 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
geohot f92ac85fbd llm: simplify AMD Qwen kernels 2026-08-02 14:17:43 +00:00
geohot fb53442c17 llm: use standard RoPE table allocation 2026-08-02 14:13:06 +00:00
geohot eba359128e llm: reuse master recurrent cache semantics 2026-08-02 14:09:36 +00:00
geohot 3804a87b60 llm: keep Qwen warmup changes scoped 2026-08-02 13:55:10 +00:00
geohot 9d02a3fe68 llm: remove unused Qwen AMD paths 2026-08-02 13:55:10 +00:00
geohot 35dbe53d5d llm: trim unused AMD quant paths 2026-08-02 13:55:10 +00:00
geohot ca29776f74 llm: fast packed Qwen3.6 inference on AMD 2026-08-02 13:55:10 +00:00
chenyuandGitHub 0258c7fefc minor argstr and alloc stack cleanup [PR] (#17361) 2026-08-01 22:37:06 -04:00
George HotzandGitHub fb607fb990 faster devectorizer with one line (#17360) 2026-08-01 13:06:35 -07:00
George HotzandGitHub 15c936db01 merge devectorize + indexing (#17354) 
* external benchmark schedule in 5 sec (codex slop)

* prune

* real?

* delete
 2026-08-01 11:40:15 -07:00
wozeparrotandGitHub 98b700bad1 gptoss optim fixes (#17356) 2026-08-01 09:10:06 -07:00
chenyuandGitHub 8e524ca467 CONST related cleanups [pr] (#17352) 
ConstFloat(nan) != nan should be False, and some Invalid bool cleanups
 2026-08-01 09:54:28 -04:00
nimlgenandGitHub 665822ab34 hcq2: faster replace (#17353) 
* hcq2: use runtime device for submit params

* hcq2: parameterize buffers in linear time
 2026-08-01 16:32:04 +03:00
chenyuandGitHub 6c0ec39279 UOp.is_invalid [PR] (#17351) 
helper to prep ConstArg
 2026-08-01 02:36:32 -04:00
chenyuandGitHub b502fc1367 more const arg -> val (#17350) 2026-08-01 02:18:10 -04:00
qazalandGitHub 161783d8f7 add _device_num back to ast.variables (kimi) (#17327) 2026-08-01 14:56:50 +09:00
George HotzandGitHub 5a1c641f79 more arg -> val (#17349) 
* more arg -> val

* kimi

* more
 2026-07-31 22:44:24 -07:00
Adeeb ShihadehandGitHub 20b8ecff50 support new USB vendor ID (#17348) 2026-07-31 21:45:06 -07:00
George HotzandGitHub 9082ecef5d use .val to access the value of Ops.CONST (#17347) 2026-07-31 19:51:42 -07:00
George HotzandGitHub 099d69ff7d ci: split macos unit test into metal and mock runners (#17346) 2026-07-31 19:41:18 -07:00
George HotzandGitHub a88f832f0c remove UOp.val (#17345) 2026-07-31 18:58:38 -07:00
chenyuandGitHub 850989115d __int__ and __float__ work for weak (#17342) 2026-07-31 19:56:37 -04:00
sirhcmandGitHub 15d515299e heuristics: try multiple TC axes (#17341) 2026-07-31 19:25:18 -04:00
sirhcmandGitHub 85ced44db6 tc: don't allow reduce over output dims (#17340) 2026-07-31 16:48:51 -04:00
chenyuandGitHub 277433259e fix sym_infer for CAST (#17338) 2026-07-31 14:38:38 -04:00
chenyuandGitHub 8dc225e28e dtype_from_uop(INS) is None [PR] (#17337) 2026-07-31 14:33:15 -04:00
chenyuandGitHub ad24750487 const(value, dtype) -> const(value).cast(dtype) in tests (#17335) 2026-07-31 13:24:31 -04:00
qazalandGitHub a11ee26bb8 viz: prep for faster cli DEBUG=3 (#17334) 
* move data

* split
 2026-08-01 02:16:30 +09:00
sirhcmandGitHub b95bd5b2a5 don't reset chestnut in benchmark (#17333) 2026-07-31 12:39:04 -04:00
nimlgenandGitHub 1095bbe409 hcq2: fix ib reuse (#17330) 
* hcq2: initialize IB reuse counters at link

* x
 2026-07-31 18:43:23 +03:00
chenyuandGitHub 7f4dbb8090 remove shape= from UOp.const [PR] (#17331) 
inlined to const_like
 2026-07-31 11:18:59 -04:00
chenyuandGitHub 8e2f175542 const(dtype, b) -> const(b, dtype) [PR] (#17328) 
prep for dtype removal
 2026-07-31 09:46:37 -04:00
nimlgenandGitHub 155b84ee80 hcq2 faster schedule (#17324) 
* avoid quadratic STACK dtype promotion

* build HCQ patch stacks directly

* pack HCQ command buffers linearly

* remove HCQ command buffer simplification
 2026-07-31 16:08:18 +03:00
wozeparrotandGitHub 4f5cadd15d gptoss ci (#17325) 2026-07-31 05:51:58 -07:00
nimlgenandGitHub 0c4bfaeb48 coalesce ints (#17323) 
* merge ints

* fix z3 validation of coalesced loads

* fix uint vector names in CUDA and Metal
 2026-07-31 15:49:56 +03:00
qazalandGitHub 6d2700f0b7 failing test for unbound _device_num err in BEAM (#17326) 
* min failing test

* switch to cpu

* err
 2026-07-31 20:24:44 +09:00
wozeparrotandGitHub 93b74c75fc gptoss: grouped moe (#17322) 2026-07-31 03:25:32 -07:00
qazalandGitHub f7964acb64 llama with MXFP4 (#17321) 
* mxfp4 in llama

* less

* name
 2026-07-31 18:25:25 +09:00
qazalandGitHub 0a3325f9c2 add mxfp4 quantize and layout kernels (#17320) 2026-07-31 14:50:18 +09:00
sirhcmandGitHub 13452b3775 benchmark comma big model (#17312) 2026-07-31 01:37:34 -04:00
qazalandGitHub a8c1e89500 fp4 asm gemm 6+ pflops (#17315) 
* fp4 gemm

* better kernargs structure

* move to .s files

* work

* work

* p2

* style

* use .py

* move to dsl

* cleanup

* add MFMA_SCALE_X2_ENCODING

* cleanup mfma

* fma docs

* gemm_mxfp4

* more cleanup

* move

* move to cdna_asm_gemm

* change

* rm

* change

* mx
 2026-07-31 14:25:23 +09:00
George HotzandGitHub d65ea465ed cleanup gemm fragment + add store unshard (#17313) 
* cleanup gemm fragment + add store unshard

* multi

* fix
 2026-07-30 20:53:28 -07:00
chenyuandGitHub 6608b9d8e2 adjust lower weak order [pr] (#17314) 
* adjust lower weak order [pr]

* no symbolic_simple
 2026-07-30 23:36:12 -04:00
Noah SchiroandGitHub f2c2f4456b Add softmin (#17292) 
* Add softmin

* Remove extra tests
 2026-07-30 20:05:11 -04:00
chenyuandGitHub fe8ece7efa fix weak for image gate fusion [pr] (#17310) 2026-07-30 19:47:47 -04:00
chenyuandGitHub ce500c1946 broadcast doesn't cast CONST [pr] (#17308) 
* broadcast doesn't cast CONST [pr]

it might change arg from python int to python float, but not cast the CONST, so we don't need or_casted in symbolic. we lower the weak dtype right before decomp

* fix
 2026-07-30 18:51:59 -04:00
sirhcmandGitHub b290372121 ftdi reset chestnut before running comma benchmark (#17306) 2026-07-30 14:28:45 -04:00
sirhcmandGitHub da15c43e51 update openpilot benchmarks (#17304) 2026-07-30 14:22:20 -04:00
chenyuandGitHub b488cc7df2 update dtype_from_uop for SHR/SHL (#17307) 
certain backend cast the distance, and we should not re-broadcast it
 2026-07-30 13:57:58 -04:00
chenyuandGitHub 341c4ed4f5 rewrite writes new dtype with dtype_from_uop try 2 [pr] (#17305) 
with NIR fix which is INDEX related

some Ops currently depends on explicitly set dtype and not dtype_from_uop
 2026-07-30 13:47:21 -04:00