feat: rjob rl with cyber env - #80

Open

sys555 wants to merge 26 commits into

v2from

feat/cyber-rl

Commits on Sep 9, 2026

feat(rl): enhance patcheval RL pipeline, gateway admission control, and data manager

root
and
cursoragent
committed
858466e
chore: untrack config.yaml to avoid leaking credentials; add config.yaml.example template

root
and
cursoragent
committed
e7b8749
feat(rl): switch GDN packed-seq fix from bshd padding to runtime monkey-patch

root
and
cursoragent
committed
ab80038
fix(rl): tune patcheval recompute layers/max-tokens and fix gdn_packed_seq import path

root
and
cursoragent
committed
fc2b3bc
feat(rl): add trajectory truncation patch, colocate/CPU-offload support, and multi-node NCCL settings

root
and
cursoragent
committed
88d6859
feat(sz): switch PatchEval to non-colocate mode and add Harbor RL example

root
and
cursoragent
committed
28cf847
feature(sz): chat template layer
root
committed
9fb1d3d
chore(gateway): drop noisy per-record telemetry logs; fix chat template import fallback; reuse buffer_server run dir

root
and
cursoragent
committed
d5da1a3
feat(rl): add patcheval RJob example scripts for qwen3.5-9b and qwen3.8-27b

root
and
cursoragent
committed
3101b3f
revert: restore gateway/telemetry.py logs deleted in 35ac8a5

root
and
cursoragent
committed
4dc39aa
feat(rl): make timing_log switchable via module-level enable flag

root
and
cursoragent
committed
bccdf82
refactor(gateway): extract _emit_llm_step helper on TelemetryRecorder

root
and
cursoragent
committed
a11522a
feat(gateway,manager): move default_max_tokens into GatewayConfig (32K); add eval_elapsed_s to episode timing

root
and
cursoragent
committed
23c100e
feat(manager): add rjob_running_ts absolute timestamp to episode metrics

root
and
cursoragent
committed
f517d40
refactor(rl): replace step-id lookback cursor with env-id cursor for buffer fetch

root
and
cursoragent
committed
ecdb87b
refactor(rl): simplify env-id cursor — move two-phase fetch into buffer_server, drop strategy changes

root
and
cursoragent
committed
3e44087

Commits on Sep 10, 2026

fix(patcheval-rl): unblock training — relax reward pre-gate, fix env/gateway config, drop dead cursor code

root
and
cursoragent
committed
2ba46ef
chore(patcheval-rl): lower group_size to 2, merge docs, ignore local scripts

root
and
cursoragent
committed
351b152
fix(patcheval-rl): lower RL_GLOBAL_BATCH_SIZE 64->8 to fix lr_decay_steps assert

root
and
cursoragent
committed
ab20a0b
refactor(patcheval): rename strict_runner.py -> runner.py

root
and
cursoragent
committed
b44afde
fix(patcheval-rl): set RL_GLOBAL_BATCH_SIZE=4 (= rollout_batch*group_size)

root
and
cursoragent
committed
a85610e
chore(patcheval): drop duplicate qwen3_8_27b scripts

root
and
cursoragent
committed
d3ccbde
chore(patcheval): drop unused train_qwen3_5_9b.sh and run_eval_one.sh

root
and
cursoragent
committed
a378132
chore(patcheval): drop unused eval gateway scripts + leaked secret

root
and
cursoragent
committed
89e2d51
chore(rl): untrack local-only utility/experiment scripts

root
and
cursoragent
committed
c484ced
chore(patcheval): untrack push_patcheval_done.txt

root
and
cursoragent
committed
dc0a573

Navigation Menu

Uh oh!

feat: rjob rl with cyber env - #80

feat: rjob rl with cyber env #80
sys555 wants to merge 26 commits into
v2from
feat/cyber-rl

Commits on Sep 9, 2026

feat(rl): enhance patcheval RL pipeline, gateway admission control, and data manager

chore: untrack config.yaml to avoid leaking credentials; add config.yaml.example template

feat(rl): switch GDN packed-seq fix from bshd padding to runtime monkey-patch

fix(rl): tune patcheval recompute layers/max-tokens and fix gdn_packed_seq import path

feat(rl): add trajectory truncation patch, colocate/CPU-offload support, and multi-node NCCL settings

feat(sz): switch PatchEval to non-colocate mode and add Harbor RL example

feature(sz): chat template layer

chore(gateway): drop noisy per-record telemetry logs; fix chat template import fallback; reuse buffer_server run dir

feat(rl): add patcheval RJob example scripts for qwen3.5-9b and qwen3.8-27b

revert: restore gateway/telemetry.py logs deleted in 35ac8a5

feat(rl): make timing_log switchable via module-level enable flag

refactor(gateway): extract _emit_llm_step helper on TelemetryRecorder

feat(gateway,manager): move default_max_tokens into GatewayConfig (32K); add eval_elapsed_s to episode timing

feat(manager): add rjob_running_ts absolute timestamp to episode metrics

refactor(rl): replace step-id lookback cursor with env-id cursor for buffer fetch

refactor(rl): simplify env-id cursor — move two-phase fetch into buffer_server, drop strategy changes

Commits on Sep 10, 2026

fix(patcheval-rl): unblock training — relax reward pre-gate, fix env/gateway config, drop dead cursor code

chore(patcheval-rl): lower group_size to 2, merge docs, ignore local scripts

fix(patcheval-rl): lower RL_GLOBAL_BATCH_SIZE 64->8 to fix lr_decay_steps assert

refactor(patcheval): rename strict_runner.py -> runner.py

fix(patcheval-rl): set RL_GLOBAL_BATCH_SIZE=4 (= rollout_batch*group_size)

chore(patcheval): drop duplicate qwen3_8_27b scripts

chore(patcheval): drop unused train_qwen3_5_9b.sh and run_eval_one.sh

chore(patcheval): drop unused eval gateway scripts + leaked secret

chore(rl): untrack local-only utility/experiment scripts

chore(patcheval): untrack push_patcheval_done.txt