Commit Graph

Select branches

Hide Pull Requests

abstractions_4

add_local_buffer

after_qol

algebraic_upat

all_bv_no_idx

all_sink

amd_isel

amd_refac

amd_smu_reset_patch

amd_sqtt

amd_uop

amd_uop_cleanups

assign_copy

assign_try_3

autogen-comgr-2

before2

better_spec

better_syn

bitcast_spec

broadcast_shape_2

bt_ei

buf_heur

bump_nv_2404

cache_reshape

call_inline

call_is_kernels

callify

cast_vec_as_cast

cat_mop

changes_from_rangify

check_clsetarg

check_dims

ci-deps

cifar_30

cifar_30_clean

clean_load

clean_spec

clone_tg

codegen_try_2

contig_on_after

continue_mat

deepsek

default_late_allreduce

delete_kernel_py

delete_ones

delete_slow_rangeify

demote_op

disable_rules

dsp_search

dsp_search_merged

dtype_shape

earlier_gater

emu_cdna

emulate_ctx

end_all_ranges

endrange

er_prioity

execution_unit

explicit-dev

failing_assign

fancy_usb

fast_llm

fix_fa_fwd_clang_22

fix_fuse

fix_llvm_cast

fix_range_merging_sd

fix_rangeify_tests

fix_rc

fix_sched_copy

fix_slice_store

fix_src_spec

fixup_fix_kernel

full_spec

fun_w_egraphs

gate-on-load-store

gate_rewrite

gemma4_gpt

gen_pdf_fast

gh-pages

gitea-test-ci

gl_dims

gldims

glm_flash

grad_outerworld

grok_x86_novec

group_div_rules

heur_4096

heur_group

hk_gemm

http_threading

hwtest_fixes

id_store

if_range_cleanups

ignore_oob

image_idx_xy

image_in_mc

image_no_vec

image_refactors_stack

index_to_slice

invalid_try_3

jinja_template

kernel4

kernel_is_call

kernelless

kimi_slop

kitten_matmul

late_Locals

late_add_load

late_allreduce

late_codegen

late_decomp_is_late

late_lower_idx

late_view

late_weakint

lds_double_buffering

less_wmma_arg

lil_dg_change

lil_rangeify_td

llama_trainer

llm_262k_minimal

llm_flash_attn

llm_qwen

llm_serve_fixes

load_view_late

local_cache

locals_have_device

loop_end_op

loop_is_rng

lowerer_hacks

mac_pytest

manual_kernel_3

master

mega_lowerer

min_new_lin

min_outer_range

minor_views

mixins7

mlperf_training_v4.1

mlperf_training_v5.0

mlperf_training_v6.0

modern_scan

more_apps

more_early_comps

more_folding_rules

more_kernelcount

more_pcode

more_rangeify_pm_tests

more_ren_clean

more_uop_gc_test

more_uop_mm

more_vmap

move_devec

move_dtype_decomps

move_expander

move_gates_to_load_store

move_mop_cleanup

move_schedule

move_simp_views

moveleftright

mt_typing

multi_range

multioutput

new_ck

new_devec

new_invalid_and_image

new_linearizer

new_shape

new_test

no-chat-template

no_assign_2

no_buf_0

no_decimals

no_late_simplify_marg

no_merge_ends

no_merge_views

no_mstack_in_allreduce

no_one_ranges

no_sip_usbgpu

no_tsink_base

no_tuplize

no_user_ck_contig

no_vec_x86_4

noop_index

novecimg

num_batches_tracked_shape

olmo3_support

one_compiler

one_hcopt

only_assign_buffers

only_loop_backends

only_reg_emu

only_reg_emu2

onnx-cpu-count

outer_range

outerworld_work

pad_invalid

pad_invalid_work

pad_mixin

param_call_is_function

parse_sqtt_more

pci_iommu

port_nir

postrange_hcopts

precompile_backward

prefill_qwen_fast

print_special_ops

ps_clean

pyremu_try

python_prg

python_speed

q4k_support

qcom-openpilot-13ms

qcom_830

qcom_mmapeak

qcom_speed_codex

qol_string

qwen36_27b_amd_900

qwen36_27b_amd_minimal

qwen36_amd_bugfixes

r_cleanups

r_new_opt

ranged_call

rangeify_2

rangeify_const_folding

rdna3_vibes

rdna4_asm

rdna4_gemm

real_pyrender

recurse_sub

reg_coalese

relu_pattern

remove_assign

remove_cast_hack

remove_ck_contig

remove_define_var

remove_forced_re

remove_on_stack

remove_programspec

remove_pylint

remove_slow_tests

remove_vec_2

remove_vec_3

renderers_use_shapes_2

replace_if_with_range

reprocess_node_experiments

reshape_trait

resnet-ci

revert-11547-fix_mismatch_red

revert-11715-revert-11711-amd_llvm_tc_cast

revert-11809-no_ki_in_gpudims

revert-11811-better_div_nesting

revert-11876-memplan_vbig

revert-12104-simplify_valid_in_symbolic

revert-12308-bufferize_early

revert-12731-jitlink_init_check

revert-12813-after_cleanups

revert-13046-better_cleanup_arange_buffers

revert-13311-prof_link_sqtt

revert-13583-rocr_desc_2

revert-14734-tload

revert-17141-full_rewrite_to_sink

revert-17408-_broadcasted-weak

revert-9438-simple_matmul_np_rand

revert_wrong_ck_fix

rewrite_rangeify2

rtoposort

scan_assign

scan_op_8

sched_cache

sched_lin

sd_load_simple

self_type

set_testpath

shrink_in_render

simp_spec

simple_gidxs

simple_pool

simple_priority

simple_rangeify

simpler-sd-compile

simpler_end

simpler_fusion

simpler_image

simpler_postrange

simpler_reduce

slice_like_shrink

spec_3

spec_cleanup

spec_ren

speed_one_linear

split_buf_idx

split_decomp_pass

split_sub

stack_0

store_is_void

string_ranges

support_opts_in_contig

swizzle_in_kernel

sym_llm

sym_work

symbolic-chunk-split

szdiff_bot_test

tc_is_heur

tensor_graph_single_sink

test-lp

test_amd_crash

test_fa

test_from_postopt

test_hgemm

test_unit

tiny_scan

tmr

tool_call_support

topoprop

ttomsa-new_linearizer

tuplegtuple

two_stage_remove

typed_checks

typing_work

unused_permute_arg

uop_prg

upd_test4

update_benchmark

update_benchmark_search

update_mlperf

user_failure

vecless_x86_glm

verify_sink_dims_clean

view_after_const

view_in_codegen

viz_cbroadcast

warp_fun

wgpu-f16-packed

working_postopt

x86_moves

x86_numel

#1

#2

#3

#4

#5

v0.10.0

v0.10.1

v0.10.2

v0.10.3

v0.11.0

v0.12.0

v0.13.0

v0.4.0

v0.5.0

v0.6.0

v0.7.0

v0.8.0

v0.9.0

v0.9.1

v0.9.2

Mono Color

• a455e17539 llm: fully warm recurrent serving at startup geohot 2026-08-10 13:34:16 +00:00
• 778f8aee59 fix hevc (#17477) nimlgenandgeohot 2026-08-10 13:33:49 +03:00
• 9b347cc3e7 late loss.to("CPU") in llama (#17476) qazalandgeohot 2026-08-10 16:31:50 +08:00
• 32b9149040 llama: custom silu kernels (#17462) qazalandgeohot 2026-08-10 15:43:01 +08:00
• bbd4a77351 move platform tests to platform.yml (#17475) George Hotzandgeohot 2026-08-09 23:25:18 -07:00
• e103421a12 document Kimi K3 optimization handoff geohot 2026-08-10 06:14:30 -07:00
• 2b1b8c22a9 add fast exact-shape Kimi K3 benchmark geohot 2026-08-10 06:08:07 -07:00
• f53f0e7e79 document final Kimi K3 MI350 load profiling geohot 2026-08-10 05:34:56 -07:00
• 224bac0318 speed up Kimi K3 decode projections on MI350X geohot 2026-08-10 05:01:14 -07:00
• 1d86204718 optimize Kimi K3 serving on MI350X geohot 2026-08-10 03:59:15 -07:00
• c6ac4961d7 speed up Kimi K3 safetensor loading geohot 2026-08-10 02:23:34 -07:00
• 1b3732a6ed speed up Kimi K3 TP8 loading on MI350X geohot 2026-08-10 01:15:26 -07:00
• 8611fe22a7 fix hevc (#17477) nimlgenandGitHub 2026-08-10 13:33:49 +03:00
• 2821bd646f late loss.to("CPU") in llama (#17476) qazalandGitHub 2026-08-10 16:31:50 +08:00
• 44f1f45cd5 llama: custom silu kernels (#17462) qazalandGitHub 2026-08-10 15:43:01 +08:00
• 553bdf68e6 llm: defer recurrent server warmup geohot 2026-08-10 06:59:49 +00:00
• 4e1c0166f8 llm: stabilize recurrent serving across requests geohot 2026-08-10 06:43:27 +00:00
• 566f32fe9f move platform tests to platform.yml (#17475) George HotzandGitHub 2026-08-09 23:25:18 -07:00
• d28f5f261b llm: keep recurrent single-token shapes static geohot 2026-08-10 05:45:20 +00:00
• 14595b9ae8 llm: prepare direct Kimi K3 serving on MI350X geohot 2026-08-10 05:28:18 +00:00
• eaf7822239 llm: accelerate Kimi serving on gfx11 geohot 2026-08-10 03:58:44 +00:00
• 77e5be99bc llm: prepare Kimi K3 and accelerate recurrent prefill geohot 2026-08-10 02:03:14 +00:00
• 6edb5f9698 get kimi-linear running on 4x7900XTX (codex slop) geohot 2026-08-09 23:21:09 +00:00
• 8c8b43de62 hcq2: fix beam (#17467) nimlgenandGitHub 2026-08-09 16:53:47 +03:00
• e17c21e102 hcq2: timings (#17464) nimlgenandGitHub 2026-08-08 22:00:32 +03:00
• d4d537c8ae add SPEC checking for the kernel graph (#17432) George HotzandGitHub 2026-08-08 10:00:06 -07:00
• abe2256299 fix symbolic sharded reshape (#17463) b1tgandGitHub 2026-08-09 00:18:02 +08:00
• 8c49a7a34b support symbolic shapes in copy (#17461) b1tgandGitHub 2026-08-09 00:16:59 +08:00
• 9dd3b8402e default llama 8b to MXFP4=1 (#17465) qazalandGitHub 2026-08-09 00:13:32 +08:00
• c0d2f9ac0c nolocals supports variables (#17457) sirhcmandGitHub 2026-08-07 14:54:38 -07:00
• 4c206a52b1 fix ci emu (gpt) (#17437) nimlgenandGitHub 2026-08-07 22:36:59 +03:00
• 4a3b8f6501 better _drop_valid_stmts [pr] (#17454) chenyuandGitHub 2026-08-07 15:18:35 -04:00
• 59b88ea5e2 move pm_fold_cast_const [pr] (#17453) chenyuandGitHub 2026-08-07 13:31:56 -04:00
• f76422b8af fix cast to float _min_max [pr] (#17451) chenyuandGitHub 2026-08-07 11:56:19 -04:00
• 1827ec57f7 gptoss: fix sharded invalids (#17450) wozeparrotandGitHub 2026-08-07 23:42:02 +08:00
• b6189db8e9 cpu: fix eintr (#17449) nimlgenandGitHub 2026-08-07 17:46:57 +03:00
• 73e670c10f c0+x<c1 -> x < c1-c0 is ints only [pr] (#17448) chenyuandGitHub 2026-08-07 10:44:01 -04:00
• fca695a36f clean up reduce MUL gradient (#17447) chenyuandGitHub 2026-08-07 10:08:04 -04:00
• 0c96cdc300 fix prod gradients at zero (#17404) Robert JosephandGitHub 2026-08-07 09:56:52 -04:00
• baa6148066 fix var of large half input (#17444) chenyuandGitHub 2026-08-06 23:14:32 -04:00
• 1858f1fd9a viz: collapse PROGRAM nodes like CALL (codex) (#17438) 3 people 2026-08-06 19:44:34 -07:00
• f253c4469d remove contiguous from custom_kernel (#17149) 3 people 2026-08-07 10:20:59 +08:00
• 28195d51fb fix f2f from fp8e5m2fnuz to half (#17442) chenyuandGitHub 2026-08-06 21:19:46 -04:00
• 9020a88f03 truncate float in DType.const [pr] (#17439) chenyuandGitHub 2026-08-06 20:19:08 -04:00
• d8cbc11105 update linear interpolate to use int math for indices (#17441) chenyuandGitHub 2026-08-06 20:18:39 -04:00
• 38e131e796 pci: use VFIO_DEVICE_RESET for device resets when the device is bound to vfio geohot 2026-08-06 11:29:52 -07:00
• c8ac28ac79 test: IOMMU containment of wild device DMA geohot 2026-08-06 10:47:14 -07:00
• 997685493e pci: route DMA through vfio type1v2 when the device is behind an IOMMU geohot 2026-08-06 10:41:56 -07:00
• 1fd6b1035f fa: swa support (#17367) wozeparrotandGitHub 2026-08-06 23:07:30 +08:00
• 46230e9f17 hcq2: fence inputs (#17436) nimlgenandGitHub 2026-08-06 16:27:12 +03:00
• 9636dd1a25 test MXFP4 llama without hipcc (#17435) qazalandGitHub 2026-08-06 16:31:40 +08:00
• f258708d7d llama: custom quantize_mxfp4+transpose kernel (codex) (#17434) qazalandGitHub 2026-08-06 15:13:28 +08:00
• 28e6ef6937 fix postopt symbolic [pr] (#17433) chenyuandGitHub 2026-08-06 00:12:30 -04:00
• 969df866a3 one less strong dtype const in symbolic [pr] (#17431) chenyuandGitHub 2026-08-05 23:35:06 -04:00
• 7a9cd8e329 move weak function and pm to uop/weak [PR] (#17429) chenyuandGitHub 2026-08-05 22:27:40 -04:00
• c355c9baf5 Merge branch 'master' into qwen36_27b_amd_900 George HotzandGitHub 2026-08-05 19:14:24 -07:00
• 13afd14ec2 non removable no_user_ck_contig geohot 2026-08-05 18:56:26 -07:00
• b6d531426a clean up geohot 2026-08-05 18:48:26 -07:00
• 14b386c7f1 no user contig on custom_kernel geohot 2026-08-05 18:35:41 -07:00
• b4372df9c6 revert wrong custom kernel fix (#17427) George HotzandGitHub 2026-08-05 18:31:35 -07:00
• d51e55aa17 remove some pm_fold_cast_const [pr] (#17426) chenyuandGitHub 2026-08-05 21:28:38 -04:00
• eec82b0db4 revert wrong custom kernel fix revert_wrong_ck_fix geohot 2026-08-05 18:27:40 -07:00
• be25207a7a scope variable names inside CALLs (#17424) sirhcmandGitHub 2026-08-05 17:59:09 -07:00
• d726e5f7f3 split pm_fold_cast_const [PR] (#17425) chenyuandGitHub 2026-08-05 19:56:04 -04:00
• b09e879e95 remove benchmark_qwen_amd, use extra/benchmark_llm.py geohot 2026-08-05 16:36:34 -07:00
• db7e683d91 Merge remote-tracking branch 'origin/master' into qwen36_27b_amd_900 geohot 2026-08-05 16:26:04 -07:00
• 470c032a5e fix slice + non contig kernels (#17423) George HotzandGitHub 2026-08-05 16:18:17 -07:00
• 92acbc3ac4 schedule: realize custom kernel inputs that don't resolve to a buffer state qwen36_amd_bugfixes geohot 2026-08-05 16:11:42 -07:00
• a8074f6e1b movement: resolve negative int slice bounds against symbolic sizes geohot 2026-08-05 16:11:07 -07:00
• a8a8030bc9 add benchmark_llm script geohot 2026-08-05 15:59:51 -07:00
• 20cc62a29a Merge branch 'master' into qwen36_27b_amd_900 George HotzandGitHub 2026-08-05 14:41:53 -07:00
• 581bfdd94f merge track_rewrites and profile_matches into rewrite_group [PR] (#17420) George HotzandGitHub 2026-08-05 14:41:38 -07:00
• 07ac911665 few weak and decomp tweaks [PR] (#17419) chenyuandGitHub 2026-08-05 15:55:04 -04:00
• c2f1e5ae2a fix weak cast to strong dtype [pr] (#17418) chenyuandGitHub 2026-08-05 15:32:23 -04:00
• 179b25b593 test: tolerate 1 ulp in the lazy custom kernel srcs test geohot 2026-08-05 19:23:12 +00:00
• ec1f94e55c llm: chunk size 32 everywhere geohot 2026-08-05 19:23:12 +00:00
• f2b873f884 llm: gate the custom AMD kernels to RDNA3 (gfx11) geohot 2026-08-05 19:23:00 +00:00
• 6435b7e627 tensor graph is a single sink tensor_graph_single_sink geohot 2026-08-05 11:58:21 -07:00
• 757a727808 move callify into tensor (#17416) George HotzandGitHub 2026-08-05 11:48:56 -07:00
• 31e6770783 Merge branch 'master' into qwen36_27b_amd_900 George HotzandGitHub 2026-08-05 10:50:06 -07:00
• 2cce85a606 chat: display reasoning_content from streamed responses (#17414) George HotzandGitHub 2026-08-05 10:49:24 -07:00
• 39b08d6fba fix slice bug geohot 2026-08-05 17:25:40 +00:00
• 9b27ea8523 hcq2: cleaner (#17413) nimlgenandGitHub 2026-08-05 19:45:04 +03:00
• 40ccfe6c11 llm: tighten the generate loop body geohot 2026-08-05 16:29:40 +00:00
• 6cb419b9b7 regression test for bert nan with weak (#17412) chenyuandGitHub 2026-08-05 12:14:26 -04:00
• 0077d75565 llm: simplify GatedDeltaNetBlock._attention and _init_state geohot 2026-08-05 15:51:56 +00:00
• b1b0d982bc test: regression test for custom kernel inputs that don't resolve to a buffer state geohot 2026-08-05 15:51:56 +00:00
• fe30c86908 llm: restore master form in TransformerBlock._attention geohot 2026-08-05 15:06:43 +00:00
• 10274e29b4 schedule: realize custom kernel inputs that don't resolve to a buffer state geohot 2026-08-05 15:06:43 +00:00
• 5b0b68ec55 remove debug from test (#17410) nimlgenandGitHub 2026-08-05 15:47:40 +03:00
• ad32bd272b viz/cli: faster and more complete rewrites print (#17411) qazalandGitHub 2026-08-05 19:02:27 +08:00
• 6455a7fc6f llm: match warmup chunk size to generate default for non-recurrent models geohot 2026-08-05 08:25:50 +00:00
• 874d33128b hcq2 benchmark (#17235) nimlgenandGitHub 2026-08-05 10:00:42 +03:00
• 3bf9e70b19 Revert "don't cast weak in _broadcasted [pr] (#17408)" (#17409) chenyuandGitHub 2026-08-05 02:40:33 -04:00
• 1c4a3134d5 Revert "don't cast weak in _broadcasted [pr] (#17408)" revert-17408-_broadcasted-weak chenyuandGitHub 2026-08-05 02:40:21 -04:00
• 77e124e455 fix AMD WMMA emulation and test in CI (#17184) 3 people 2026-08-05 02:25:32 -04:00
• ee7d441606 fixes prefill_qwen_fast geohot 2026-08-04 23:23:10 -07:00
• b45058b5ec don't cast weak in _broadcasted [pr] (#17408) chenyuandGitHub 2026-08-05 02:19:31 -04:00
• 46f0003776 more KernelCountException (#17407) George HotzandGitHub 2026-08-04 22:55:38 -07:00
• 1796f6bb5b more KernelCountException more_kernelcount geohot 2026-08-04 21:37:31 -07:00