sirhcm/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests 2 Actions Activity
14,537 Commits 336 Branches 15 Tags
explicit-dev
Commit Graph
14537 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
George HotzandGitHub 477b573807 update llm kv cache to be half (#17690) 
* update llm kv cache to be half / chunk_size to always be 32

* just dtype
 2026-08-23 14:02:54 -07:00
geohot bb0e99acbf hotfix: skip that nan test on mac 2026-08-23 08:40:46 -07:00
nimlgenandGitHub 93865e2c66 hcq2: sunday housekeeping (#17686) 
* hcq2: sunday housekeeping

* x
 2026-08-23 17:20:40 +03:00
George HotzandGitHub 5b60a09ab0 some fixes for the AMD emulator (#17684) 
* some fixes for the AMD emulator

* simpler

* revert

* min
 2026-08-22 22:48:07 -07:00
chenyuandGitHub b0a1285330 lil decomp cleanup [PR] (#17682) 2026-08-22 18:11:14 -04:00
nimlgenandGitHub a2e64e16aa hcq2: early usb (#17683) 
* hcq2: usb interface and submit

* x

* x

* x

* x

* r

* x
 2026-08-23 00:22:39 +03:00
chenyuandGitHub a9069c177a make decomp pass SPEC=2 [PR] (#17681) 2026-08-22 08:01:15 -04:00
chenyuandGitHub 8950942e75 remove explicit dtype for NOOP and decomp [PR] (#17678) 2026-08-21 22:27:36 -04:00
chenyuandGitHub 356f665377 test update for weak const (#17675) 2026-08-21 21:50:29 -04:00
chenyuandGitHub 7204d46786 delete dtype_from_uop INDEX exempt (#17674) 2026-08-21 21:08:06 -04:00
George HotzandGitHub af242819d8 refactor the AMD emulator slop (kimi) (#17673) 
* refactor the AMD emulator slop (kimi)

* mypy
 2026-08-21 18:00:58 -07:00
wozeparrotandGitHub 52596dbf38 gptoss: fused ce (#17672) 2026-08-21 16:19:28 -07:00
sirhcmandGitHub 07cce78cec compile3: log printed timings (#17671) 2026-08-21 19:18:32 -04:00
George HotzandGitHub f986829461 keep IndexingContext scoped in indexing (#17670) 2026-08-21 14:12:39 -07:00
nimlgenandGitHub 4fd4eafb23 nv: hevc (#17661) 
* nv: hevc

* x

* nv: zero the nvdec scratch buffers
 2026-08-21 23:37:09 +03:00
chenyuandGitHub daa154aa22 FLOORDIV to SHR for powers of 2 [pr] (#17669) 2026-08-21 16:28:04 -04:00
nimlgenandGitHub 12f889aaad hcq2: parallel compile (#17667) 
* hcq2: parallel compile

* Dx
 2026-08-21 23:27:21 +03:00
sirhcmandGitHub 298748ebd3 ci: remove setup-python (#17665) 2026-08-21 16:19:47 -04:00
George HotzandGitHub 3082956a17 usb copyin: async arm and drain, 323 MB/s on comma (#17663) 
* usb copyin: can safely arm before drain

* perf counter

* 294 MB/s for comma

* free speed with async transfers
 2026-08-21 13:12:03 -07:00
chenyuandGitHub 402bea7ddd l2i and sign_extend cleanups [pr] (#17668) 
towards good threefry decomp
 2026-08-21 15:55:45 -04:00
chenyuandGitHub 8f9cbdf0cc few more self folding [pr] (#17657) 2026-08-21 15:21:28 -04:00
chenyuandGitHub 8f59041ee5 more wgsl pack cleanups [PR] (#17664) 2026-08-21 14:34:25 -04:00
nimlgenandGitHub 1cf8a2c7fe hcq2: use shrink.bitcast (#17653) 
* hcq2: shrink.bitcast

* x

* x

* x

* s

* x

* Dx

* Revert "hotfix: disable HCQ2"

This reverts commit a57188ea6d.

* x
 2026-08-21 21:24:34 +03:00
chenyuandGitHub 3919ce8427 ceildiv in wgsl _packed_size [pr] (#17662) 2026-08-21 14:02:20 -04:00
George HotzGitHubtiny
 		756e82e055 usb amd: pipelined copyin over the 0xF2 engine (2.6x faster) (#17628) 
Stream 240KB chunks into two alternating 256KB SRAM bounce windows; each chunk
ends in a unique 512B sentinel that a prebuilt SDMA ring polls before copying
the chunk to VRAM, followed by an in-order drain fence that the host waits on
before re-arming a window. No timing assumptions in either direction: the
sentinel is in-stream proof of data landing, the fence proves a full drain.

Adds a small pooled async bulk-OUT layer to USB3 so staging the next chunk
overlaps the wire, and a slot_start parameter to scsi_write for the second
window. 107 -> 276 MB/s copyin on tinyc8 (Kryo-3XX host).

Co-authored-by: tiny <tiny@local>
 2026-08-21 10:38:50 -07:00
chenyuandGitHub cc32aa18db don't match casted const in const_folding_pat [pr] (#17658) 2026-08-21 12:16:07 -04:00
qazalandGitHub 77f698e55b cleanup the mxfp4 gemm (#17660) 2026-08-22 00:59:04 +09:00
chenyuandGitHub 554d078ac4 fix render_marg [pr] (#17656) 2026-08-21 10:03:22 -04:00
chenyuandGitHub 176377ff6e weak 1 for FDIV in get_late_rewrite_patterns [PR] (#17655) 2026-08-21 09:33:43 -04:00
qazalandGitHub 1c3c9e96f6 remove setUp from test_sqtt_profiler (#17652) 
* remove setUp from test_sqtt_profiler

* remove that

* cleaner

* do not need that
 2026-08-21 17:29:11 +09:00
qazalandGitHub e8a8d99b99 profiler tracing using unique keys (#17651) 
* profiler tracing using unique keys

* cleaner + comment
 2026-08-21 17:06:01 +09:00
qazalandGitHub dcc2d021e7 prereq viz cleanups for unique profile keys (#17649) 
* cleaner

* just use VIZ=-2

* better
 2026-08-21 16:18:25 +09:00
qazalandGitHub 80bf60d782 profiler failing test for non unique kernel names (#17647) 
* err

* expected
 2026-08-21 13:52:35 +09:00
George HotzandGitHub 1cb0600086 fix llm vars regression (kimi) (#17644) 
* fix llm regression (kimi)

* unneeded

* more principled
 2026-08-20 21:31:43 -07:00
chenyuandGitHub 1bcb6bdc62 no-op weak.py reorder [PR] (#17646) 2026-08-20 23:56:06 -04:00
qazalandGitHub d716d0d927 hotfix: qkv bw kernel requires hipcc and cdna4 (#17645) 2026-08-21 12:38:45 +09:00
b1tgandGitHub 9216aa494c llm prefill failing test (#17630) 2026-08-20 18:58:34 -07:00
George HotzandGitHub 9aa9e11301 compile kernels in parallel (#17629) 
* compile kernels in parallel (slop)

* cleanups

* fixes

* hand cleanups

* no PARALLEL with VIZ

* ugh, we need to remove the n from names

* no hcq2 in that test

* main for abstractions3

* fix compile server to be fork safe

* fix num cpu threads in context

* just use a lock

* fix __main__ on spawn

* compileonce

* python3 is double

* xdist sizing

* parallel compile

* fetch supports extract

* fix issues

* revert compiler server to master, drop compileonce

* normal workflow

* PARALLEL=0 for QCOMCL compiletest
 2026-08-20 16:41:45 -07:00
George HotzandGitHub 3fdbb82bfe add ansipad and PARALLEL contextvar (#17642) 2026-08-20 16:26:02 -07:00
George HotzandGitHub 0ccef542e0 add extract support to fetch (#17641) 2026-08-20 16:20:16 -07:00
sirhcmandGitHub c655aaf3a2 ci: venv in /opt/venv (#17632) 
default python is now 3.14, not 3.12
 2026-08-20 18:54:54 -04:00
chenyuandGitHub 592e3f8363 update const selector where folding [pr] (#17640) 
folding a strong dtype WHERE to a weak const branch keeps the strong dtype
 2026-08-20 18:49:34 -04:00
sirhcmandGitHub 3715006a21 fix float_to_bf16 on non-float32 inputs (#17638) 2026-08-20 18:14:30 -04:00
wozeparrotandGitHub d80254c1d9 fa: remove swa recompute (#17635) 2026-08-20 14:26:44 -07:00
sirhcmandGitHub c773891e3f skip INT_MIN % -1 on X86 and LLVM (#17637) 2026-08-20 17:26:26 -04:00
chenyuandGitHub 0e7ab863a0 x86 REX issue (#17580) 
* failing test

* fix
 2026-08-20 16:49:04 -04:00
geohot a57188ea6d hotfix: disable HCQ2 2026-08-20 13:39:48 -07:00
George HotzandGitHub 1707dca3b4 remove kernel_cnt, names are no longer unique (#17633) 
* remove kernel_cnt, names are no longer unique

* check uops, not names

* fix hcq2
 2026-08-20 13:28:25 -07:00
sirhcmandGitHub eba5b7e750 benchmarks: cleanups (#17631) 2026-08-20 16:12:14 -04:00
chenyuandGitHub a8ecb73363 x64 imm uint64 (#17636) 
* x64 imm uint64

* fix
 2026-08-20 15:30:17 -04:00