unilab 1.3.4

What's Changed

• feat(isaacsim): expose material sharing and task-local G1 fixture by @TATP-233 in #1683
• chore: pin tensor runtime local workspace by @TATP-233 in #1685
• ci(tensor): skip uv cache maintenance on cuda runner by @TATP-233 in #1688
• feat(warpsac): declare tensor inference runtime knobs by @TATP-233 in #1689
• feat(training): monitor G1 FlashSAC soaks by @TATP-233 in #1687
• feat(training): consume tensor runtime schemas by @TATP-233 in #1696
• feat(training): consume abnormal shutdown diagnostics by @TATP-233 in #1697
• feat(training): consume tensor runtime bounds by @TATP-233 in #1698
• feat(training): fail closed on tensor platform matrix by @TATP-233 in #1699
• docs(training): add tensor runtime production guide by @TATP-233 in #1700
• docs(adr): define torch-only manager runtime by @TATP-233 in #1707
• feat(base): add tensor environment lifecycle by @TATP-233 in #1708
• feat(env): migrate manager lifecycle to torch by @TATP-233 in #1709
• feat(managers): tensorize metric accumulators by @TATP-233 in #1710
• feat(managers): tensorize reward and termination carriers by @TATP-233 in #1711
• feat(managers): tensorize action history by @TATP-233 in #1712
• feat(managers): publish observations as tensors by @TATP-233 in #1713
• feat(actions): tensorize base action processing by @TATP-233 in #1715
• feat(actions): tensorize Allegro incremental targets by @TATP-233 in #1716
• feat(entity): expose validated joint tensor reads by @TATP-233 in #1717
• feat(observations): tensorize non-temporal processing by @TATP-233 in #1719
• feat(entity): expose sensor and body tensor reads by @TATP-233 in #1720
• feat(env): pair packed scene reads and reset transactions by @TATP-233 in #1721
• feat(observations): tensorize temporal buffers by @TATP-233 in #1722
• feat(commands): publish command state as tensors by @TATP-233 in #1723
• feat(events): tensorize scheduling carriers by @TATP-233 in #1724
• bench(env): use tensor actions for env rollouts by @TATP-233 in #1725
• bench(env): consume public step timing without private probes by @TATP-233 in #1726
• bench(rl): keep active collector transitions in tensors by @TATP-233 in #1727
• bench(env): replace NpEnv shard throughput workload by @TATP-233 in #1728
• bench(rl): rename step timing metrics without NpEnv prefix by @TATP-233 in #1729
• fix(playback): keep off-policy policy IO in tensors by @TATP-233 in #1730
• fix(appo): keep Motrix playback observations in tensors by @TATP-233 in #1731
• refactor(rl): enforce TorchEnv tensor VecEnv boundary by @TATP-233 in #1732
• fix(play): consume tensor velocity commands by @TATP-233 in #1733
• refactor(registry): enforce TorchEnv factory output by @TATP-233 in #1734
• docs(runtime): describe the TorchEnv boundary by @TATP-233 in #1738
• docs(agents): point env contract at TorchEnv by @TATP-233 in #1739
• feat(observations): tensorize noise carriers by @TATP-233 in #1741
• feat(rl): round-trip TorchEnv training state by @TATP-233 in #1736
• feat(training): enforce tensor NaN diagnostics by @TATP-233 in #1735
• Remove the legacy NpEnv runtime by @TATP-233 in #1742
• feat(runtime): unify CUDA tensor inference placement by @TATP-233 in #1744
• fix(managers): slice per-env observation scales on reset rows by @TATP-233 in #1746
• fix(tasks): skip generic CPU reads in FlashSAC device-resident proxy by @TATP-233 in #1748
• fix(tasks): route FlashSAC Motrix through HOST_BRIDGE runtime by @TATP-233 in #1752
• fix(tasks): materialize IsaacGym before tensor negotiation by @TATP-233 in #1754
• fix(tasks): materialize lazy FlashSAC backends only by @TATP-233 in #1758
• fix(env): restore legacy IsaacSim CPU training wire by @TATP-233 in #1760
• fix(tasks): release FlashSAC IsaacGym views before backend close by @TATP-233 in #1759
• perf(entity): remove repeated packed tensor read validation by @TATP-233 in #1761
• fix: satisfy typed tensor runtime checks by @TATP-233 in #1763
• fix(env): dispatch external CUDA IPC selected resets to tensor boundary by @TATP-233 in #1766
• fix(env): stagger Manager initial episode counters by @TATP-233 in #1767
• fix(config): restore legacy G1 walk collector penalty schedule by @TATP-233 in #1768
• test: align device-resident reset sequence with staggered counters by @TATP-233 in #1769
• fix(tasks): cross FlashSAC selected-reset readiness boundary by @TATP-233 in #1770
• fix(tasks): refresh device-resident named sensors each read by @TATP-233 in #1773
• perf(actions): keep joint position control on tensor plane by @TATP-233 in #1775
• fix(tasks): keep G1 contact proxy reward-only by @TATP-233 in #1774
• feat(config): productionize FlashSAC IsaacSim mapped CUDA IPC owner by @TATP-233 in #1777
• fix(tasks): keep HOST_BRIDGE reset on its paired read boundary by @TATP-233 in #1780
• perf(mjwarp): dispatch selected resets through tensor boundary by @TATP-233 in #1781
• perf(mjwarp): keep manager reads on packed device plane by @TATP-233 in #1783
• perf(mjwarp): reduce action and reset boundary overhead by @TATP-233 in #1785
• perf(managers): batch reset and state validation boundaries by @TATP-233 in #1787
• fix(config): make SAC mjbatch owner select the packed tensor runtime by @TATP-233 in #1791
• feat(training): make CUDA visibility rank-local by @TATP-233 in #1798
• fix(env): negotiate task sensors and restore reward diagnostics by @TATP-233 in #1797
• fix(offpolicy): inject the tensor nan guard across the collector boundary by @TATP-233 in #1799
• perf(rng): own torch CUDA randomization in G1 collector by @TATP-233 in #1801
• fix(env): keep device-resident command sensors canonical by @TATP-233 in #1803
• feat(training): make CUDA visibility the only GPU topology by @TATP-233 in #1806
• perf(mjwarp): avoid reset readiness physics step by @TATP-233 in #1805
• perf(manager): keep G1 update state device-resident by @TATP-233 in #1807
• chore(build): remove uv version pin by @TATP-233 in #1810
• perf(manager): cut stable-view host boundaries on slow-CPU servers by @TATP-233 in #1808
• refactor(env): keep episode counters tensor-resident by @TATP-233 in #1812
• refactor(recorder): publish reset rows as tensors by @TATP-233 in #1813
• refactor(command): publish reset rows as tensors by @TATP-233 in #1814
• refactor(events): publish reset rows as tensors by @TATP-233 in #1815
• fix(locomotion): consume CUDA reset rows safely by @TATP-233 in #1816
• perf(command): draw resample timers with Torch RNG by @TATP-233 in #1817
• perf(events): draw interval timers with Torch RNG by @TATP-233 in #1818
• perf(observations): sample IMU bias with Torch RNG by @TATP-233 in #1819
• perf(command): sample tensor velocity commands on device by @TATP-233 in #1820
• perf(command): sample pose values with Torch RNG by @TATP-233 in #1821
• perf(events): sample impulses with Torch RNG by @TATP-233 in #1822
• refactor(env): remove stale Manager host boundaries by @TATP-233 in #1823
• test(entities): exercise tensor action consumer by @TATP-233 in #1824
• perf(events): sample velocity pushes with Torch RNG by @TATP-233 in #1825
• perf(events): sample encoder bias with Torch RNG by @TATP-233 in #1826
• fix(rng): expose Manager facade to CUDA owners by @TATP-233 in #1827
• perf(scene): consume aggregate body views by @TATP-233 in #1828
• perf(motion): migrate MJWarp anchor observations to tensor reads by @TATP-233 in #1829
• fix(gait): consume tensor contact reads and Torch reset rows by @TATP-233 in #1830
• perf(gait): keep air-time state tensor-resident by @TATP-233 in #1831
• perf(gait): reduce regular contact groups in one kernel by @TATP-233 in #1832
• perf(env): remove redundant state finite reduction by @TATP-233 in #1833
• fix(allegro): keep observation bounds on NumPy carrier by @TATP-233 in #1834
• perf(env): attribute update-state manager phases by @TATP-233 in #1835
• test(g1): align Manager term fixtures with tensor row selectors by @TATP-233 in #1836
• perf(gait): evaluate selected Bezier branch by @TATP-233 in #1837
• perf(gait): fuse Tensor foot target kernels by @TATP-233 in #1838
• perf(gait): fuse feet-phase reward reads by @TATP-233 in #1839
• fix(gait): declare termination body entity owner by @TATP-233 in #1840
• perf(scene): serve single-body reads without projection by @TATP-233 in #1841
• perf(env): expose update-state queue drain by @TATP-233 in #1842
• perf(env): attribute update-state publication by @TATP-233 in #1843
• perf(gait): share phase across reward terms by @TATP-233 in #1844
• perf(env): attribute update-state residual by @TATP-233 in #1845
• fix(env): time update-state residual from phase start by @TATP-233 in #1846
• fix(env): exclude drain once from residual by @TATP-233 in #1847
• fix(env): normalize drain units for residual by @TATP-233 in #1848
• fix(env): use one clock scale for residual by @TATP-233 in #1849
• feat(reset): stage tensor joint state rows by @TATP-233 in #1850
• fix(types): reset defaults through validated rows by @TATP-233 in #1851
• perf(benchmark): interleave in-process variant blocks by @TATP-233 in #1852
• chore(managers): type tensor reset selectors by @TATP-233 in #1853
• chore(tasks): type tensor reset selectors by @TATP-233 in #1854
• chore(types): close manager carrier seams by @TATP-233 in #1855
• feat(managers): bind commands after read-plan compile by @TATP-233 in #1856
• refactor(motion): split command read-phase binding by @TATP-233 in #1857
• feat(motion): declare command carrier protocol by @TATP-233 in #1858
• feat(motion): dispatch body rewards by carrier by @TATP-233 in #1859
• feat(motion): dispatch scalar rewards by carrier by @TATP-233 in #1860
• feat(motion): dispatch terminations by carrier by @TATP-233 in #1861
• feat(motion): expose tensor observation carriers by @TATP-233 in #1862
• fix(env): publish tensor manager outputs directly by @TATP-233 in #1863
• feat(env): compile command body read declarations by @TATP-233 in #1864
• perf(motion): keep tensor actions on control plane by @TATP-233 in #1865
• feat(motion): add tensor command carrier by @TATP-233 in #1866
• fix(motion): support tensor command Manager construction by @TATP-233 in #1867
• feat(motion): register Manager MJWarp runtime by @TATP-233 in #1868
• feat(motion): register Manager Genesis runtime by @TATP-233 in #1869
• feat(motion): register Manager flip runtime by @TATP-233 in #1870
• feat(scope): gate shelved tensor backends by @TATP-233 in #1871
• feat(runtime): remove tensor runtime switch by @TATP-233 in #1872
• feat(motion): remove direct registration by @TATP-233 in #1873
• refactor(motion): internalize direct runtime by @TATP-233 in #1875
• test(env): enforce selected reset publication by @TATP-233 in #1876
• feat(motion): tensor host command carrier by @TATP-233 in #1877
• fix(motion): publish tensor sampler frames by @TATP-233 in #1878
• fix(motion): tensor sampler carrier state by @TATP-233 in #1879
• perf(motion): pack tensor motion ingest by @TATP-233 in #1880
• perf(motion): cache tensor motion features by @TATP-233 in #1881
• perf(motion): fuse Manager reward pack by @TATP-233 in #1882
• perf(motion): step tensor sampler on device by @TATP-233 in #1883
• perf(motion): fuse tensor anchor observations by @TATP-233 in #1884
• perf(env): reduce state finite diagnostics by @TATP-233 in #1885
• perf(env): reduce reset finite diagnostics by @TATP-233 in #1886
• perf(env): attribute Manager reset phases by @TATP-233 in #1887
• perf(command): reduce reset metric diagnostics by @TATP-233 in #1888
• perf(env): attribute selected reset reads by @TATP-233 in #1889
• perf(command): reduce reset validation boundaries by @TATP-233 in #1890
• perf(command): attribute reset sampling phases by @TATP-233 in #1891
• fix(command): publish reset term timings by @TATP-233 in #1892
• perf(motion): reuse selected reset root scratch by @TATP-233 in #1893
• perf(motion): sample tensor reset frames on device by @TATP-233 in #1894
• fix(motion): scope tensor sampler mirror rows by @TATP-233 in #1895
• revert(motion): remove regressive reset sampling slices by @TATP-233 in #1896
• perf(obs): reuse tensor uniform noise scratch by @TATP-233 in #1897
• revert(obs): remove regressive noise scratch by @TATP-233 in #1898
• test(benchmark): cover paired variant blocks by @TATP-233 in #1899
• perf(benchmark): report paired variant ratios by @TATP-233 in #1900
• perf(motion): avoid reset row host mirror by @TATP-233 in #1901
• perf(motion): gather refresh frames on device by @TATP-233 in #1902
• perf(env): attribute reset command post-compute by @TATP-233 in #1903
• perf(motion): attribute post-compute refresh phases by @TATP-233 in #1904
• fix(motion): publish post-compute refresh timings by @TATP-233 in #1905
• perf(motion): compile relative refresh kernel by @TATP-233 in #1906
• perf(motion): compile packet ingest by @TATP-233 in #1907
• perf(motion): compile robot state refresh by @TATP-233 in #1908
• perf(motion): attribute reset resample phases by @TATP-233 in #1909
• fix(motion): publish reset resample timings by @TATP-233 in #1910
• fix(command): preserve term reset timings by @TATP-233 in #1911
• perf(motion): attribute reset tensor writes by @TATP-233 in #1912
• fix(command): time only reset validation by @TATP-233 in #1915
• refactor(managers): require tensor action terms by @TATP-233 in #1916
• docs(support): scope production matrix to tensor backends by @TATP-233 in #1917
• ci(tensor): drop shelved motrix extra by @TATP-233 in #1918
• audit(sim2sim): scope contract pairs by @TATP-233 in #1919
• benchmark(rl): scope collector defaults by @TATP-233 in #1920
• benchmark(env): scope step matrix backends by @TATP-233 in #1921
• test(go2): scope owner contract by @TATP-233 in #1922
• test(g1): scope owner contract by @TATP-233 in #1923
• test(motion): scope owner contract by @TATP-233 in #1924
• test(reward): scope config injection by @TATP-233 in #1925
• test(config): scope system contracts by @TATP-233 in #1926
• test(training): scope backend adapter contracts by @TATP-233 in #1927
• test(scripts): scope training owner cases by @TATP-233 in #1928
• test(cli): scope owner routing by @TATP-233 in #1929
• fix(stewart): make mujoco owner self-contained by @TATP-233 in #1930
• test(sim2sim): scope resolver owners by @TATP-233 in #1931
• test(flashsac): scope owner fingerprint by @TATP-233 in #1932
• test(benchmark): scope flashsac backend owners by @TATP-233 in #1933
• feat(config): remove shelved task owners by @TATP-233 in #1934
• refactor(obs): tensor-native delay state by @TATP-233 in #1935
• refactor(obs): tensor-native noise models by @TATP-233 in #1936
• perf(command): publish uniform velocity tensor by @TATP-233 in #1937
• perf(command): keep G1 resampling on device by @TATP-233 in #1938
• feat(managers): add reset-row observation protocol by @TATP-233 in #1939
• feat(reset): publish call-graph diagnostics by @TATP-233 in #1942
• feat(motion): add tensor-native sampler by @TATP-233 in #1943
• feat(managers): declare reset owner contract by @TATP-233 in #1944
• feat(managers): expose reset state boundaries by @TATP-233 in #1945
• feat(reset): execute reset owner lifecycle by @TATP-233 in #1946
• perf(motion): own command action metric reset by @TATP-233 in #1947
• perf(motion): own observation reset state by @TATP-233 in #1948
• perf(obs): row-scope motion anchor reset by @TATP-233 in #1949
• perf(motion): fuse actor observation carrier by @TATP-233 in #1950
• perf(motion): fuse reset command validation by @TATP-233 in #1951
• perf(motion): fuse selected reset state write by @TATP-233 in #1952
• refactor(motion): remove direct runtime remnant by @TATP-233 in #1953
• perf(curriculum): remove reset host carriers by @TATP-233 in #1954
• refactor(metrics): require tensor carriers by @TATP-233 in #1955
• feat(reset): carry tensor randomization by @TATP-233 in #1956
• feat(sac): own tensor motion runtime by @TATP-233 in #1957
• perf(read): reuse stable device views by @TATP-233 in #1958
• perf(motion): fuse post-compute refresh by @TATP-233 in #1959
• refactor(cli): gate shelved backends early by @TATP-233 in #1960
• docs(scope): limit installation backends by @TATP-233 in #1961
• docs(scope): gate backend guide surface by @TATP-233 in #1962
• test(scope): gate benchmark backends by @TATP-233 in #1963
• docs(scope): label historical benchmarks by @TATP-233 in #1964
• build(scope): limit docker backends by @TATP-233 in #1965
• ci(scope): gate backend extras by @TATP-233 in #1966
• chore(scope): label historical tools by @TATP-233 in #1967
• perf(motion): reduce reset host synchronization by @TATP-233 in #1968
• perf(reset): fuse motion payload diagnostics by @TATP-233 in #1969
• perf(command): scope reset refresh validation by @TATP-233 in #1970
• perf(motion): fuse critic observation carrier by @TATP-233 in #1971
• perf(motion): remove reset owner sync by @TATP-233 in #1972
• feat(sampler): support selected-row stepping by @TATP-233 in #1973
• perf(reward): defer per-step log host publication by @TATP-233 in #1974
• perf(motion): fuse penalty reward carrier by @TATP-233 in #1975
• perf(motion): compile metric refresh by @TATP-233 in #1976
• perf(motion): compile reward carriers by @TATP-233 in #1977
• perf(env): reuse public manager tensors by @TATP-233 in #1978
• perf(env): split update-state drain attribution by @TATP-233 in #1979
• perf(benchmark): collect update-state CPU attribution by @TATP-233 in #1980
• perf(motion): publish full sampler step by @TATP-233 in #1981
• perf(env): attribute update-state events by @TATP-233 in #1982
• perf(env): attribute update-state boundaries by @TATP-233 in #1983
• perf(env): attribute update-state residual by @TATP-233 in #1984
• perf(env): reconcile update-state attribution by @TATP-233 in #1985
• perf(env): split command epilogue attribution by @TATP-233 in #1986
• perf(motion): attribute post-compute phases by @TATP-233 in #1987
• perf(motion): fuse failure terminations by @TATP-233 in #1988
• perf(motion): attribute per-step command phases by @TATP-233 in #1989
• fix(motion): publish command step timing by @TATP-233 in #1990
• fix(timing): separate motion step keys by @TATP-233 in #1991
• fix(motion): initialize step timing carrier by @TATP-233 in #1992
• fix(timing): snapshot update-state phases by @TATP-233 in #1993
• perf(command): attribute per-step manager phases by @TATP-233 in #1994
• fix(command): clear step timing before compute by @TATP-233 in #1995
• perf(command): attribute dt validation by @TATP-233 in #1996
• perf(command): attribute manager loop phases by @TATP-233 in #1997
• perf(termination): attribute manager phases by @TATP-233 in #1998
• fix(timing): preserve termination phase snapshot by @TATP-233 in #1999
• perf(termination): accept transient tensor owners by @TATP-233 in #2000
• perf(reward): attribute manager phases by @TATP-233 in #2001
• fix(benchmark): sample reward attribution phases by @TATP-233 in #2002
• perf(observation): attribute manager phases by @TATP-233 in #2003
• perf(reward): accept transient tensor owners by @TATP-233 in #2004
• perf(noise): cache segmentwise tensor bounds by @TATP-233 in #2005
• fix(observation): exclude child phases from residual by @TATP-233 in #2006
• perf(observation): bypass one-term concatenation by @TATP-233 in #2007
• perf(reward): attribute motion pack calls by @TATP-233 in #2008
• perf(sampler): keep failure alpha on device by @TATP-233 in #2009
• fix(motion): recognize mjwarp owner v16 by @TATP-233 in #2010
• perf(env): bypass public autoreset row validation by @TATP-233 in #2012
• perf(motion): preserve reset row identity by @TATP-233 in #2013
• perf(reset): keep payload validation on device by @TATP-233 in #2014
• perf(motion): fuse reset value construction by @TATP-233 in #2015
• test(motion): cover fused reset construction by @TATP-233 in #2016
• perf(env): skip duplicate autoreset observation scatter by @TATP-233 in #2017
• perf(motion): fuse sampler reset dispatch by @TATP-233 in #2018
• revert(sampler): keep dispatch eager by @TATP-233 in #2019
• perf(sampler): index clip ends by frame by @TATP-233 in #2020
• perf(motion): attribute sampler mirror sync by @TATP-233 in #2021
• perf(velocity): make command owner tensor-native by @TATP-233 in #2024
• perf(walk): select tensor reset events on mjwarp by @TATP-233 in #2025
• perf(gait): make phase reset tensor-native by @TATP-233 in #2026
• perf(curriculum): track episode lengths on device by @TATP-233 in #2027
• fix(env): fail closed on missing device tensor sensors by @TATP-233 in #2028
• perf(benchmark): report diagnostic tail throughput by @TATP-233 in #2029
• perf(walk): fuse G1 tensor reset owner by @TATP-233 in #2030
• perf(action): bypass indexed full-width control write by @TATP-233 in #2031
• perf(action): keep device encoder bias resident by @TATP-233 in #2032
• perf(action): remove duplicate finite synchronization by @TATP-233 in #2033
• perf(action): fuse full-width joint target publication by @TATP-233 in #2034
• perf(reset): key cached tensor reset tables on host by @TATP-233 in #2035
• perf(reward): fuse G1 walk reward pack by @TATP-233 in #2036
• chore(scope): shelve Genesis tensor runtime by @TATP-233 in #2037
• docs(scope): regenerate Genesis-free support matrix by @TATP-233 in #2038
• test(scope): align suites with shelved Genesis runtime by @TATP-233 in #2039
• perf(reset): fuse tensor root sampling graph by @TATP-233 in #2040
• perf(gait): fuse G1 walk termination pack by @TATP-233 in #2041
• perf(reset): validate tensor root payload without host sync by @TATP-233 in #2042
• merge: tensor-only Manager runtime and MJWarp tensor-native closeout by @TATP-233 in #2044
• feat(genesis): restore canonical tensor owners by @TATP-233 in #2046
• merge: Genesis canonical tensor runtime by @TATP-233 in #2047
• feat(newton): restore canonical tensor owners by @TATP-233 in #2049
• test(newton): harden canonical selected-reset lifecycle by @TATP-233 in #2050
• test(newton): route legacy conformance through device helper by @TATP-233 in #2051
• roadmap: enable Newton canonical tensor runtime by @TATP-233 in #2052
• fix(onnx): align export examples with playback module device by @TATP-233 in #2053
• feat(motrix): restore canonical packed host-bridge owners by @TATP-233 in #2056
• test(motrix): align flashsac owner scope with canonical runtime by @TATP-233 in #2057
• test(motrix): scope tensor event assertion to Motrix owner by @TATP-233 in #2058
• test(support): allow absent Motrix support cells by @TATP-233 in #2059
• roadmap: restore Motrix canonical packed host-bridge runtime by @TATP-233 in #2060
• feat(macos): enable mujoco/mjbatch tensor-runtime off-policy training on macOS by @TATP-233 in #2061
• chore(workspace): bump unilab-rl pin for terminal reward fix by @TATP-233 in #2062
• refactor(tasks): remove g1_flip_tracking task by @TATP-233 in #2064
• feat(training): add single-rank CUDA MPS mode by @TATP-233 in #2065
• feat(drake): scope Go2 owner into tensor runtime by @TATP-233 in #2067
• feat(training): enable single-host multi-GPU MPS by @TATP-233 in #2069
• feat(superdex): restore TorchEnv CPU host-bridge owners by @TATP-233 in #2070
• fix(sac): adjust G1 WBT reset and push velocities by @mothinf in #1714
• chore(deps): consume published tensor runtime packages by @TATP-233 in #2071
• feat: remove retired task owners and WarpSAC by @TATP-233 in #2072
• chore(deps): adopt unilab-rl 1.4.6 sac naming by @TATP-233 in #2074
• feat(tasks): unify G1 motion tracking identity by @TATP-233 in #2075
• chore(motion): remove unused FlashSAC owner contract by @TATP-233 in #2076
• docs(logging): describe grouped terminal metrics and width tiers by @TATP-233 in #2078
• feat(appo): restore Go2 Motrix owner on tensor runtime by @TATP-233 in #2077
• fix(config): restore current-task Motrix owners by @TATP-233 in #2081
• chore: remove checked-in drift baseline artifacts by @TATP-233 in #2082
• fix(runtime): honor host-bridge manager Torch devices by @TATP-233 in #2084
• refactor(motion): remove G1 fused owner terms by @TATP-233 in #2086
• chore(deps): update unilab-rl to 1.4.9 by @TATP-233 in #2088
• fix(tests): keep MPS sockets within macOS path limit by @TATP-233 in #2090
• feat(cli): manage user-owned CUDA MPS daemons by @TATP-233 in #2085
• fix(cli): keep default MPS runtime within socket limits by @TATP-233 in #2092
• fix(runtime): require explicit host-bridge manager devices by @TATP-233 in #2094
• chore(deps): update unisim-core to 1.7.12 by @TATP-233 in #2097
• feat(training): select recorded CUDA MPS daemon by @TATP-233 in #2095
• fix(training): stabilize APPO and MJWarp tests by @TATP-233 in #2098
• chore(deps): update unilab-rl to 1.4.10 by @TATP-233 in #2099
• fix(training): keep MuJoCo collector tensors on CPU by default by @TATP-233 in #2101
• fix(motion): gather reference frames from device carrier in full refresh by @TATP-233 in #2105
• chore(deps): bump unilab-rl to 1.4.11 by @TATP-233 in #2112
• chore(deps): bump unisim-core to 1.7.13 by @TATP-233 in #2115
• feat(flashsac): expose learner CUDA Graph controls by @EtherealTide in #2117
• refactor(motion): move MimicLite ownership downstream by @TATP-233 in #2120
• feat(motion): migrate MimicLite observations to tensor runtime by @TATP-233 in #2107

New Contributors

• @mothinf made their first contribution in #1714

Full Changelog: v1.3.3...v1.3.4