v0.2.65

Reactant v0.2.65

Diff since v0.2.64

Merged pull requests:

• feat: optimize then pad (#1127) (@avik-pal)
• [ReactantCUDAExt] Don't initialise during precompilation (#1128) (@giordano)
• [ReactantExtra] Update XLA (#1129) (@giordano)
• feat: shardy progagation options [JLL] (#1131) (@avik-pal)
• feat: expose more control over shardy pipeline (#1132) (@avik-pal)
• Regenerate MLIR Bindings (#1133) (@github-actions[bot])
• updates vitepress config (#1135) (@lazarusA)
• More optimizaation passes (pending jll) (#1138) (@wsmoses)
• Enable optimization string (#1140) (@wsmoses)
• fix: better errors (#1141) (@avik-pal)
• refactor: print proper error msg and prevent stack overflow (#1143) (@avik-pal)

Closed issues:

• ArgumentError: reducing over an empty collection is not allowed; consider supplyinginit to the reducer (#1130)

v0.2.64

Reactant v0.2.64

Diff since v0.2.63

Merged pull requests:

• [Compiler] Add new passes (#1125) (@giordano)

v0.2.63

Reactant v0.2.63

Diff since v0.2.62

Merged pull requests:

• [CI] Disambiguate names of MLIR artifacts (#1119) (@giordano)
• feat: padded sharding (#1120) (@avik-pal)
• New jll passes (#1121) (@wsmoses)
• feat: init directly on device (#1123) (@avik-pal)
• Add slice_broadcast (#1124) (@wsmoses)

Closed issues:

• Non-Divisible Sharding on Reactant End (#1109)

v0.2.62

Reactant v0.2.62

Diff since v0.2.61

Merged pull requests:

• Add sort memory (#1113) (@giordano)

v0.2.61

Reactant v0.2.61

Diff since v0.2.60

Merged pull requests:

• trying out some more nospecialize infers (#1078) (@wsmoses)
• Dump MLIR modules also before running passes and XLA compilation (#1086) (@giordano)
• Format code of branch "main" (#1087) (@github-actions[bot])
• fix: actually do donation (#1088) (@avik-pal)
• feat: codegen without embedding runtime pointers into body (#1089) (@avik-pal)
• fix: allow not passing a memory kind (#1090) (@avik-pal)
• feat: empty memory kind jll (#1091) (@avik-pal)
• Fix nospecialize of varargs (#1092) (@wsmoses)
• fix: kernel abstractions (#1094) (@avik-pal)
• Update Project.toml (#1097) (@wsmoses)
• Update XLA to include bugfix and remove build_bazel_rules_apple (#1098) (@giordano)
• feat: move transposes up (and then down) (#1099) (@avik-pal)
• Update BUILD (#1101) (@avik-pal)
• ci: upload failing mlir (#1102) (@avik-pal)
• Dump failed lockstep flag (#1104) (@wsmoses)
• Add missing optimizations (#1105) (@wsmoses)
• Update WORKSPACE (#1107) (@wsmoses)
• fix: let shardy run everything (#1108) (@avik-pal)
• fix: aggressive checks for buffer donation (#1110) (@avik-pal)
• Add while_deadresult (#1111) (@wsmoses)
• feat: give nice names to dumped mlir files (#1114) (@avik-pal)
• fix: codegen to fix aliasing issues (#1115) (@avik-pal)

v0.2.60

Reactant v0.2.60

Diff since v0.2.59

Merged pull requests:

• feat: allow configuring precision & algorithm for dot_general (#1042) (@avik-pal)
• fix: shardy passes (#1073) (@avik-pal)
• feat: allocator stats for IFRT [JLL Changes] (#1074) (@avik-pal)
• feat: allocator stats for IFRT (#1075) (@avik-pal)
• fix: run finalizer on array (#1076) (@avik-pal)
• fix: remove ifrt::HloSharding handling (#1077) (@avik-pal)
• fix: remove symbol (#1081) (@avik-pal)
• fix: assert non-allocating (#1082) (@avik-pal)
• Regenerate MLIR Bindings (#1084) (@github-actions[bot])

v0.2.59

Reactant v0.2.59

Diff since v0.2.58

Merged pull requests:

• feat: replicas + better APIs for IFRT (#1049) (@avik-pal)
• feat: expose more ifrt API + abi changes for compile (#1056) (@avik-pal)
• feat: add a flag to detect resharding (#1058) (@avik-pal)
• fix: logical vs physical device handling (#1060) (@avik-pal)
• Add an option to always dump MLIR module to be compiled (#1061) (@giordano)
• Add improved arg info (#1062) (@wsmoses)
• Feature: add check for non allocating (#1065) (@wsmoses)
• fix arg error (#1067) (@wsmoses)
• Fix #1066 (#1068) (@mofeing)
• Update Reactant_jll and bump version number (#1069) (@giordano)
• Fix logic for dumping MLIR modules (#1070) (@giordano)
• Bump vendored xla (#1071) (@wsmoses)

Closed issues:

• Construct Sharded Array with Data on Different Processes (#934)
• Error when sharding ConcreteRNumber (#1066)

v0.2.58

Reactant v0.2.58

Diff since v0.2.57

Merged pull requests:

• Run passes on a single PM (#1045) (@Pangoraw)
• Fix tpu (#1054) (@wsmoses)
• New Reactant_jll and bump version (#1055) (@giordano)

Closed issues:

• TPU error (#1053)

v0.2.57

Reactant v0.2.57

Diff since v0.2.56

Merged pull requests:

• JLL bump (#1052) (@wsmoses)

Closed issues:

• Changes to while tracing broke recurrent layers in Lux (#1050)

v0.2.56

Reactant v0.2.56

Diff since v0.2.55

Merged pull requests:

• Fix pathing issues during tracing (#1041) (@wsmoses)
• Update JLL and bump version number (#1046) (@giordano)
• Libdevice ordering (#1047) (@wsmoses)
• fix: hook to disable gather_getindex (#1048) (@avik-pal)

Closed issues:

• make_tracer of dictionaries (#1043)