Repository navigation

All pull requests

feat: add TileLang backend support with Ascend SwiGLU
#1572xvxuopop opened Oct 28, 2026Contributor4 tasks
fix(fused_linear_kl_div): avoid common-offset loss cancellation
#1569Levius-Fubuki opened Oct 28, 20261 of 3 tasks
1
test(convergence): make MoE node comparison flip-aware
#1568Yulong-Cauli opened Oct 27, 2026Contributor5 of 6 tasks
misc(npu): pin triton-ascend 3.6.0 and cap cv workspace to one slot
#1567zheliuyu opened Oct 27, 2026Contributor4 of 5 tasks
fix(group_norm): accept 2-D inputs and fix extra_repr
#1563anh-nguyenn opened Oct 26, 2026Contributor
docs(README): use Huawei Ascend in the NPU section
#1561lltiaor opened Oct 26, 2026
fix(cute): use local scratch for SM90 split-K dH
#1559ReinforcedKnowledge opened Oct 25, 20263 tasks done
feat(grpo): support top_entropy_quantile in fused linear GRPO loss
#1558nvdm59 opened Oct 24, 2026Contributor1 of 3 tasks
feat(grpo): add off-policy sequence masking (OPSM) to fused linear GRPO loss
#1557nvdm59 opened Oct 24, 2026Contributor1 of 3 tasks
Fix softmax collapse in CE, GRPO, KL/JSD and softmax
#1555kashif opened Oct 23, 2026Contributor1 of 3 tasks
1
perf(flce): skip the projection GEMMs for ignored rows
#1554gf239 opened Oct 23, 2026Contributor3 tasks done
feat(chunked_loss): native Triton execution for fused linear preference losses (DPO/ORPO/SimPO)
#1549piyushumate opened Oct 20, 2026Contributor1 task done
feat(ops, ppo): add selective logprob Triton kernel and accelerate fused linear PPO/GRPO
#1548piyushumate opened Oct 20, 2026Contributor1 task done
fix(compile): use tl.cast for scalar kernel params so torch.compile does not fall back
#1544AckerMan5659 opened Oct 19, 2026
1
[NPU] Optimize GroupNorm: fused 2D fast path and launch overhead reduction
#1542fbossing-123 opened Oct 16, 20263 tasks
2
fix(fused_linear_kl_div): compute chunk logits in fp32 for bf16/fp16 inputs
#1539jayden1711 opened Oct 15, 2026Contributor1 of 3 tasks
feat(ascend): add fused linear GRPO loss path
#1528LZC-BELIEVER opened Oct 11, 20261 of 3 tasks
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
Fix disabled NLL normalization in fused preference loss
#1521taking-lying-flat opened Oct 8, 2026Contributor3 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

Navigation Menu

Repository navigation

All pull requests

feat: add TileLang backend support with Ascend SwiGLU

fix(fused_linear_kl_div): avoid common-offset loss cancellation

test(convergence): make MoE node comparison flip-aware

misc(npu): pin triton-ascend 3.6.0 and cap cv workspace to one slot

fix(group_norm): accept 2-D inputs and fix extra_repr

docs(README): use Huawei Ascend in the NPU section

fix(cute): use local scratch for SM90 split-K dH

feat(grpo): support top_entropy_quantile in fused linear GRPO loss

feat(grpo): add off-policy sequence masking (OPSM) to fused linear GRPO loss

Fix softmax collapse in CE, GRPO, KL/JSD and softmax

perf(flce): skip the projection GEMMs for ignored rows

feat(chunked_loss): native Triton execution for fused linear preference losses (DPO/ORPO/SimPO)

feat(ops, ppo): add selective logprob Triton kernel and accelerate fused linear PPO/GRPO

fix(compile): use tl.cast for scalar kernel params so torch.compile does not fall back

[NPU] Optimize GroupNorm: fused 2D fast path and launch overhead reduction

fix(fused_linear_kl_div): compute chunk logits in fp32 for bf16/fp16 inputs

feat(ascend): add fused linear GRPO loss path

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

Fix disabled NLL normalization in fused preference loss

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

Pull requests