Preserve source shards during streaming shuffle so four workers can share 19 shards

Score packed QSA within each document; 2.70× faster in an 8K synthetic selector benchmark

Preserve low-precision rounding in gated residual mixing

Bound PLE backward's extra workspace through chunked token reductions

Add Bridge-local Qwen4-Exp text-decoder support; GPU integration is pending

Make HF/Megatron comparison failures return a nonzero exit status

Score fan-out rollout samples as a flat group

Add order-preserving packing

Warm up NCCL before training

Pass through Muon Nesterov settings

Expose Muon coefficient selection

Restore K/V calibration in FP8 KV prefill, correcting silently mis-scaled attention outputs

Forward recompute flags; fix hybrid models

Keep Muon scale-invariant

Preserve HTTP errors across process boundaries
_Co-author

Prevent redundant backward-kernel recompilation with stable cache keys, without CPU–GPU sync
_{Solution adopted by the PR author}

Clarify Qwen3 parser boundary tokens in custom grammars

Remove full-weight copies during MoE loading; conversion peak 7.88 → 3.94 GiB in the exact-shape TP2 benchmark

Preserve native Adam step counters across checkpoint restoration

Reuse packed chunkwise CP metadata across GDN and KDA layers

Fuse GDN Q/K normalization to remove an extra backward activation buffer

Enable selective Mamba mixer recompute to save activation memory without full-layer recomputation

Route GDN input projections to Adam

Exclude GDN input projections from global clipping

Skip gradient clipping for Muon

Render evaluation prompts as complete conversations

Unify worker selection through configuration

Bypass driver tensor materialization in distillation; 4.4–5.3× faster transfers in a controlled Ray benchmark

Sanitize non-finite async log probabilities

Reject developer messages silently dropped by templates

Publish Mamba cache updates before dependent batches capture stale KV mappings

Explain cold MXFP4 JIT startup

Honor weight-check exclusions during reset

Track response truncation across context limits

Validate actor FSDP strategy

Control partial-stream continuation for batch evaluation

Clarify API retry budgets and streaming defaults