Repository navigation

All pull requests

Labels (23)Milestones (0)New pull request

Open

(324)

Closed

(1,666)

perf(r3): avoid full-size copies in routed-experts validation
#2531dogeONE-bit opened Nov 10, 2026
Reuse the logits storage for the chunked log-prob gradient
#2530R0k1e opened Nov 10, 2026
Add --logits-dtype to keep Megatron actor logits in bf16
#2529R0k1e opened Nov 10, 2026
[Fix] Keep consumed_train_samples in sync with the LR scheduler on resume
#2528SongyuQi-Francisco opened Nov 10, 2026
docs: fix invalid Usage JSONL examples
#2527GoldArowana opened Nov 10, 2026Contributor
fix(ppo): stabilize distributed advantage whitening
#2526LingYi-Liang opened Nov 9, 2026
fix(ppo): exclude Megatron's padded vocab from log-prob and entropy normalization
#2525raspberryice opened Nov 8, 2026
1
perf(agent): defer sampler replay concatenation
#2524Hang-get opened Nov 8, 2026
fix(fully-async): snapshot unfinished groups instead of draining them
#2523wangyanli3630 opened Nov 8, 2026
fix(pipeline-rl): disable radix cache when --flush-cache-interval <= 0
#2522wangyanli3630 opened Nov 8, 2026
fix(coding_agent_rl): stop the patch fallback from reverting applied hunks
#2521wangxuzhou666-arch opened Nov 5, 2026
1
fix: Fix placeholder IPC handles when a rollout engine spans multiple nodes
#2520leondada opened Nov 5, 2026
1
fix(reward): guard expensive powers and factorials
#2517chaoworks opened Nov 3, 2026Contributor
docs: update coding-agent example to the revised MiMo dataset
#2513Syycc-code opened Nov 1, 2026
fix(launcher): quote Ray runtime environment JSON
#2512chaoworks opened Oct 31, 2026Contributor
fix(rollout): release timestamps for consumed buffer groups
#2511GoldArowana opened Oct 31, 2026Contributor
fix(reward): preserve large integer precision
#2510lll-hhh opened Oct 30, 2026
4
fix(ppo): reduce cancellation in k3 KL estimates
#2507LingYi-Liang opened Oct 29, 2026
[ROCm] Support ROCm 7.15 on MI300/MI325
#2506jun-amd opened Oct 29, 2026
fix: retain GPU locks while canceling command groups
#2505bvolpato opened Oct 28, 2026
1
test: reuse the single-rank whitening reference
#2504bvolpato opened Oct 28, 2026
1
ci: batch CPU tests to reuse dependency setup
#2503bvolpato opened Oct 28, 2026
1
test: reuse Gloo workers across metric reductions
#2502bvolpato opened Oct 28, 2026
1
test: reuse Gloo workers in score-centering checks
#2501bvolpato opened Oct 28, 2026
1
Send the generated response when continuing a multimodal partial rollout
#2500apex-mochen opened Oct 28, 2026

Navigation Menu

Repository navigation

All pull requests

perf(r3): avoid full-size copies in routed-experts validation

Reuse the logits storage for the chunked log-prob gradient

Add `--logits-dtype` to keep Megatron actor logits in bf16

[Fix] Keep consumed_train_samples in sync with the LR scheduler on resume

docs: fix invalid Usage JSONL examples

fix(ppo): stabilize distributed advantage whitening

fix(ppo): exclude Megatron's padded vocab from log-prob and entropy normalization

perf(agent): defer sampler replay concatenation

fix(fully-async): snapshot unfinished groups instead of draining them

fix(pipeline-rl): disable radix cache when --flush-cache-interval <= 0

fix(coding_agent_rl): stop the patch fallback from reverting applied hunks

fix: Fix placeholder IPC handles when a rollout engine spans multiple nodes

fix(reward): guard expensive powers and factorials

docs: update coding-agent example to the revised MiMo dataset

fix(launcher): quote Ray runtime environment JSON

fix(rollout): release timestamps for consumed buffer groups

fix(reward): preserve large integer precision

fix(ppo): reduce cancellation in k3 KL estimates

[ROCm] Support ROCm 7.15 on MI300/MI325

fix: retain GPU locks while canceling command groups

test: reuse the single-rank whitening reference

ci: batch CPU tests to reuse dependency setup

test: reuse Gloo workers across metric reductions

test: reuse Gloo workers in score-centering checks

Send the generated response when continuing a multimodal partial rollout

Pull requests