[Fix] Reduce routed-expert trace memory with uint16 storage - #2046

Open

matrix72c wants to merge 46 commits into

InternLM:mainfrom

matrix72c:fix/compact-routed-experts-uint16

Commits on Aug 18, 2026

feat(rl): support sandbox dependency groups
matrix72c
committed
fbf6ade
style(rl): format sandbox pool test
matrix72c
committed
a4533e2
Merge branch 'main' into feat/sandbox-pool-dependencies
matrix72c
authored
8715365

Commits on Aug 24, 2026

feat(trace): support external OTLP collectors
matrix72c
committed
4a032dd
fix(trace): normalize external collector paths
matrix72c
committed
0d1920d
fix(trace): narrow external endpoint types
matrix72c
committed
7681fde

Commits on Aug 25, 2026

feat(trace): instrument sandbox agents and SessionServer
matrix72c
committed
482a44b
fix(trace): defer external JSONL loading
matrix72c
committed
d296013
test(trace): avoid integration import conflict
matrix72c
committed
9da4e53
fix(trace): normalize external JSONL path for mypy
matrix72c
committed
1578a77
feat(model): derive scalar value models for RL critics
matrix72c
committed
9678323
fix(model): initialize the critic value head under no_grad
matrix72c
committed
dff82ab
fix(model): initialize missing value head on every rank
matrix72c
committed
30a382f
feat(rl): add value-function loss for a PPO critic
matrix72c
committed
3a9f744
test(rl): make value loss tests device-agnostic
matrix72c
committed
2b0ddc5
refactor(rl): extract per-token KL divergence estimator
matrix72c
committed
ff326df
feat(rl): add a learning-rate scheduler to the RL training worker
matrix72c
committed
66aaacb
feat(rl): add GAE advantage estimation for PPO
matrix72c
committed
ee55de6
refactor(rl): extract _train_actor from the RL worker's fit
matrix72c
committed
f424848
feat(rl): train a PPO critic alongside the actor
matrix72c
committed
3becea4
fix(rl): allow the actor onload transition from ACTOR_READY
matrix72c
committed
b4ac1d5
fix(rl): surface critic health sums through the extra-info whitelist
matrix72c
committed
1cd9be8
feat(rl): add KL-in-reward and critic warmup for PPO
matrix72c
committed
db0e2ee
docs(rl): add a PPO example config for Qwen3.5 on GSM8K
matrix72c
committed
01aec5c
docs(rl): make the PPO example config smoke-tunable
matrix72c
committed
bd749cb
docs(rl): let the PPO example emit machine-readable metrics
matrix72c
committed
32683a1
docs(rl): make the PPO example's per-worker CPU request tunable
matrix72c
committed
82a388e
fix(rl): keep critic metric collectives rank-consistent
matrix72c
committed
fe16af6
fix(rl): store routed experts as uint16
matrix72c
committed
7909d9c

Commits on Aug 26, 2026

Commits on Sep 4, 2026

Merge branch 'feat/sandbox-pool-dependencies' into integration
matrix72c
committed
25307b5
Merge branch 'feat/external-otel-collector' into integration
matrix72c
committed
baac803
Merge branch 'feat/sandbox-otel-instrumentation' into integration
matrix72c
committed
ea7bbfe
Merge integration into feat/ppo
matrix72c
committed
b2d170e
Merge branch 'feat/sandbox-pool-dependencies' into integration
matrix72c
committed
9d1a459
Merge branch 'feat/external-otel-collector' into integration
matrix72c
committed
701140c
Merge branch 'feat/sandbox-otel-instrumentation' into integration
matrix72c
committed
3b909f1
Merge branch 'feat/ppo' into integration
matrix72c
committed
0b6922d
Merge integration into compact routed experts topic
matrix72c
committed
e817562