TraceLens release notes

TraceLens release notes#

2026-08-21

2 min read time

This topic summarizes the features available in each TraceLens release. For the hardware and software versions validated for a release, see the Compatibility matrix.

TraceLens 1.0.0#

The release expands performance-model coverage; extends profiling support to newer vLLM and SGLang versions along with xDiT diffusion and Genesis physics-simulation workloads; strengthens the TraceLens Analysis Agent with a portable skill layout and comparative graph-capture support; and overhauls TraceDiff’s name disambiguation. It also adds production-grade unit-test coverage.

Performance models#

This release adds the following performance models and classifications:

  • DeepSeek-V4 support: Sparse paged-decode modes (sliding-window, compressed, and hybrid) and a sparse-prefill model, plus the manifold-constrained Hyper-Connection operator family.

  • Additional operators: New models for nearest-neighbor upsampling (1D/2D/3D) and fused RMSNorm with quantization, efficient-attention classification, and an FP8 MoE precision fix for FP8/BF16 inputs.

  • Block-scaled GEMM detection: Detection of block-scaled MXFP4/MXFP6 GEMM paths for roofline modeling.

Frameworks and profiling#

This release extends framework and profiling support:

  • vLLM and SGLang: Profiling support through recent vLLM and SGLang releases, using stock upstream images where a patched image is no longer required.

  • xDiT diffusion: Diffusion-model profiling using the same capture-merge infrastructure as vLLM and SGLang.

  • Genesis physics simulation: A TraceLens_generate_perf_report_genesis report generator for Genesis/Taichi workloads that isolates the steady-state simulation window from JIT and build overhead.

TraceLens Analysis Agent#

This release includes the following agent improvements:

  • Portable skill layout: Agent orchestrator, sub-agent, and template specs ship in a portable skills/ layout discoverable by any agentic runner that supports skill-file discovery.

  • Comparative graph-capture analysis: Comparative mode accepts a capture folder per trace, merging call-stack and shape information into both graph-replay trees before analysis.

  • Local evaluation: Repeatability evals can run against a local serving backend for fully offline eval loops.

TraceDiff#

This release overhauls trace-tree matching:

  • Name disambiguation: Same-named sibling nodes are matched by structure rather than position, with an additional aggressive matching pass for slightly-differing Python-function names.

  • Correctness fixes: The same-name candidate-matching path is rewritten to fix several matching correctness issues.

Test coverage#

This release strengthens automated testing:

  • Unit coverage: New unit-test suites span every SDK subpackage, including Trace2Tree, TreePerf, PerfModel, NcclAnalyser, TraceFusion, TraceDiff, EventReplay, Reporting, and the shared utilities.

  • Coverage reporting: Coverage tracking and a coverage-regression gate run in continuous integration to guard against regressions.

Contents