Repository navigation

All pull requests

Fix Triton RoPE token stride when seq_len is 1
#1527KimbingNg opened Oct 10, 2026
[Fix][Softmax] Fix RuntimeError on rows wider than 65536 and NaN in _softmax_multi_block_forward_kernel
#1525arxgy opened Oct 9, 20261 of 3 tasks
Fix CuTe DSL JSD underflow while preserving fast probability-space math
#1524cyanseek opened Oct 9, 20261 of 3 tasks
Guard fused linear KL launches with device_context and pin kernel scalars to fp32
#1523LaelaZorana opened Oct 8, 2026
Update README.md
#1522vaibhavjindal opened Oct 8, 2026Collaborator3 tasks
Fix disabled NLL normalization in fused preference loss
#1521taking-lying-flat opened Oct 8, 2026Contributor3 tasks
Fix tiled MLP parameter gradients with frozen inputs
#1520taking-lying-flat opened Oct 8, 2026Contributor5 of 6 tasks
fix(grpo): normalize noncontiguous advantages before kernel dispatch
#1519taking-lying-flat opened Oct 8, 2026Contributor1 of 3 tasks
fix: handle leading masked blocks in cross entropy
#1518taking-lying-flat opened Oct 8, 2026Contributor1 of 3 tasks
fix: compute softmax backward products in fp32
#1517taking-lying-flat opened Oct 8, 2026Contributor1 of 3 tasks
perf(distillation): skip dW/dB calculation when parameters are frozen
#1487djl11 opened Sep 24, 2026Contributor1 of 3 tasks
5
docs: add Halo training cookbook
#1480advpropsys opened Sep 21, 2026
1
fix(mlp): preserve packed activations across backward replay
#1460arde171 opened Sep 11, 2026Collaborator3 tasks
fix(cutile): respect native shape and backward contracts
#1459arde171 opened Sep 11, 2026Collaborator3 tasks
feat(cutile): register six backends and harden CE/SwiGLU contracts
#1456arde171 opened Sep 10, 2026Collaborator3 tasks
5
feat(cutile): chunk FLCE with Triton sizing and cuBLAS GEMMs
#1455vaibhavjindal opened Sep 10, 2026Collaborator2 of 3 tasks
feat(cutedsl): add CuTe-DSL Newton-Schulz operator and LigerMuon optimizer
#1449piyushumate opened Sep 9, 2026Contributor2 tasks done
[WIP] Add Qwen4Exp text model support
#1447alifurkanstahl opened Sep 9, 20263 of 7 tasks
[Draft] Apple Silicon (MPS) support via triton-ext
#1442zheliuyu opened Sep 4, 2026Contributor1 of 3 tasks
10
feat(flce): expose direct chunk-size control via LIGER_FLCE_CHUNK_SIZE
#1439jin-five opened Sep 4, 20263 tasks done
5
perf(flce): write the weight-gradient product directly into grad_weight
#1429ananthv26 opened Sep 1, 20262 of 3 tasks
fix(peft): patch the LayerNorm copies, not the ModulesToSaveWrapper itself
#1425yurekami opened Aug 30, 2026
Add CCE loss
#1405kashif opened Aug 25, 2026Contributor1 of 3 tasks
perf: tile wide GeGLU rows on SM103
#1403heiheiha798 opened Aug 22, 20261 of 3 tasks
perf: fuse mHC coefficient backward epilogue
#1402heiheiha798 opened Aug 22, 20261 of 3 tasks

Navigation Menu

Repository navigation

All pull requests

Fix Triton RoPE token stride when seq_len is 1

[Fix][Softmax] Fix RuntimeError on rows wider than 65536 and NaN in `_softmax_multi_block_forward_kernel`

Fix CuTe DSL JSD underflow while preserving fast probability-space math

Guard fused linear KL launches with device_context and pin kernel scalars to fp32

Update README.md

Fix disabled NLL normalization in fused preference loss

Fix tiled MLP parameter gradients with frozen inputs

fix(grpo): normalize noncontiguous advantages before kernel dispatch

fix: handle leading masked blocks in cross entropy

fix: compute softmax backward products in fp32

perf(distillation): skip dW/dB calculation when parameters are frozen

docs: add Halo training cookbook

fix(mlp): preserve packed activations across backward replay

fix(cutile): respect native shape and backward contracts

feat(cutile): register six backends and harden CE/SwiGLU contracts

feat(cutile): chunk FLCE with Triton sizing and cuBLAS GEMMs

feat(cutedsl): add CuTe-DSL Newton-Schulz operator and LigerMuon optimizer

[WIP] Add Qwen4Exp text model support

[Draft] Apple Silicon (MPS) support via triton-ext

feat(flce): expose direct chunk-size control via LIGER_FLCE_CHUNK_SIZE

perf(flce): write the weight-gradient product directly into grad_weight

fix(peft): patch the LayerNorm copies, not the ModulesToSaveWrapper itself

Add CCE loss

perf: tile wide GeGLU rows on SM103

perf: fuse mHC coefficient backward epilogue

Pull requests