Issues

fix(source-resolution): emit device .cu for @compile_ops/torch.ops launchers so high-E2E editable kernels reach the optimizer

#734

· iraj465 opened

on Jun 12, 2026

fix(splitter): recognize SGLang step[...] per-forward annotations so inference traces split through the standard steady-state path

#733

· iraj465 opened

on Jun 12, 2026

fused-MoE expert kernel emits empty input shapes (Input Dims absent on pybind built-in), blocking shape-anchored kernel-opt dispatch

#731

· Umangatamd opened

on Jun 11, 2026

Triton fused-MoE GEMM (invoke_fused_moe_kernel) falls into un-roofline'd `other` category → no optimization candidate

#726

· Umangatamd opened

on Jun 9, 2026

Trace splitter ignores sglang step[DECODE]/step[EXTEND] annotations → 0 steps, no steady-state split

#723

· Umangatamd opened

on Jun 9, 2026

Evals 06-05: Agent Flakiness Issue — Eval Pipeline

#721

· Ahmedhasssan-aig opened

on Jun 5, 2026

Move redundant logic to common util for trace splitting

Task

#702

· devalshahamd opened

on Jun 2, 2026

Add a TraceLens blog page

#700

· ajassani opened

on Jun 2, 2026

Move arch specs into shared TraceLens utilities

#697

· ajassani opened

on Jun 2, 2026

PyTorch trace parser over-counts kernel wall-time when dur overflows the next same-track slice (ROCm rocprof artifact)

#669

· joyjie818 opened

on May 28, 2026

Build nightly TraceLens processing dashboard for private large-trace perf tracking

enhancement

#658

· ajassani opened

on May 26, 2026

Add perf models for aiter::(wrapper_)fmha_v3_varlen_{fwd,bwd} + aten::_flash_attention_forward + fix training-context categorization (Wan 2.2 coverage)

effort: M

enhancement

perf_model

#650

· gphuang opened

on May 19, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(source-resolution): emit device .cu for @compile_ops/torch.ops launchers so high-E2E editable kernels reach the optimizer

fix(splitter): recognize SGLang step[...] per-forward annotations so inference traces split through the standard steady-state path

fused-MoE expert kernel emits empty input shapes (Input Dims absent on pybind built-in), blocking shape-anchored kernel-opt dispatch

Triton fused-MoE GEMM (invoke_fused_moe_kernel) falls into un-roofline'd `other` category → no optimization candidate

Trace splitter ignores sglang step[DECODE]/step[EXTEND] annotations → 0 steps, no steady-state split

Evals 06-05: Agent Flakiness Issue — Eval Pipeline

Move redundant logic to common util for trace splitting

Add a TraceLens blog page

Move arch specs into shared TraceLens utilities

PyTorch trace parser over-counts kernel wall-time when dur overflows the next same-track slice (ROCm rocprof artifact)

Build nightly TraceLens processing dashboard for private large-trace perf tracking

Add perf models for aiter::(wrapper_)fmha_v3_varlen_{fwd,bwd} + aten::_flash_attention_forward + fix training-context categorization (Wan 2.2 coverage)

Search results