Repository navigation

All pull requests

Labels (21)Milestones (0)New pull request

Open

(113)

Closed

(1,111)

fix(fused_linear): pass the column stride when scaling KL and JSD gradients
#1613bhargavchintam opened Nov 9, 20261 of 3 tasks
fix(utils): restore torch cross-entropy state in revert helpers (#1587)
#1611YuvrajSinghBhadoria2 opened Nov 9, 2026
fix(utils): skip device_context guard under torch.compile (#1590)
#1610YuvrajSinghBhadoria2 opened Nov 9, 2026
fix(sparsemax): preserve low-precision gradients by subtracting the mean in fp32
#1609wqh-creator opened Nov 9, 20261 of 3 tasks
fix(group_norm): keep the backward statistics and reductions in fp32
#1608bhargavchintam opened Nov 9, 20261 of 3 tasks
1
fix(dyt): put the row count on grid dimension 0 in the forward launch
#1607bhargavchintam opened Nov 9, 20261 of 3 tasks
1
feat: add SwiGLU support for Qwen3-VL models
#1603YuvrajSinghBhadoria2 opened Nov 6, 2026
fix: warn when kwargs are silently dropped in _apply_liger_kernel
#1602YuvrajSinghBhadoria2 opened Nov 6, 2026
[BugFix] int64-widen MHC kernel offsets to prevent int32 overflow on large N
#1600selimfedakar opened Nov 5, 20262 of 3 tasks
[BugFix] int64-widen attn_res kernel offsets to prevent int32 overflow on large shapes
#1599selimfedakar opened Nov 5, 20262 of 3 tasks
Fix NaNs in the DiscoPOP exponential term
#1598taking-lying-flat opened Nov 5, 2026Contributor
[CuTe DSL] Add native Blackwell fused scaled cross entropy
#1597yueyiming2009 opened Nov 4, 2026Collaborator1 of 3 tasks
fix(monkey_patch): warn on kwargs dropped by the dispatcher
#1596ASHMEET555 opened Nov 3, 20263 tasks
ops: validate targets in LigerVocabParallelCEFunction
#1593ASHMEET555 opened Nov 3, 20263 tasks
1
fix(cutile): guard the JSD mixture log against fp32 underflow
#15910z5a opened Nov 3, 20261 of 3 tasks
1
fix(utils): skip device_context guard under torch.compile
#15900z5a opened Nov 3, 20261 of 3 tasks
1
fix(flce): preserve head gradients with frozen inputs
#1588taking-lying-flat opened Nov 2, 2026Contributor
Restore torch cross-entropy state in revert_liger_kernel helpers
#1587reginaldalfret opened Nov 1, 2026
feat(chunked_loss): add fused linear IRPO (Iterative Reasoning Preference Optimization) loss
#1586Yulong-Cauli opened Nov 1, 2026Contributor
feat(chunked_loss): add fused linear BCO (Binary Classifier Optimization) loss
#1585Yulong-Cauli opened Nov 1, 2026Contributor
docs: per-kernel roofline/bandwidth benchmark (RMSNorm, SwiGLU)
#1584ragulk143 opened Nov 1, 20261 of 3 tasks
3
fix(grpo): keep the Triton gradient when delta < 1 + epsilon_high
#1583arun-elr opened Nov 1, 20261 of 3 tasks
fix(grpo): mask the luspo loss elementwise to match TRL
#1582arun-elr opened Nov 1, 20261 of 3 tasks
Add apply_liger_kernel_to_glm5_next (GLM-5.3-Flash): fused linear cross-entropy and RMSNorm
#1580eilamc14 opened Oct 31, 20263 tasks done
docs: add Simplified Chinese translation (README.zh.md)
#1578JasonYeYuhe opened Oct 31, 20261 of 3 tasks
1

Pull requests