Skip to content

v0.7.0

Latest

Choose a tag to compare

@FindHao FindHao released this 10 Nov 19:01
· 5 commits to main since this release

TritonParse Release Notes v0.7.0 (30 commits)

  • Date range: 2026-09-21 — 2026-11-10
  • Scope: Feature release - CuTeDSL tracing as a new backend, virtual-target capture, autotune session splitting, and constexpr/tuple argument capture and replay. 95 files changed, 11,111 insertions / 1,129 deletions.

Highlights

  • 🧩 CuTeDSL Tracing Support: TritonParse now traces CuTeDSL kernels alongside Triton kernels, writing compilation and launch events in the standard trace format so parsing and the viewer work unchanged. Opt-in via init(enable_cutedsl_tracing=True) or TRITONPARSE_CUTEDSL_TRACING=1. The website ships a bundled Hopper GEMM example trace.

  • 🎯 Bounded Virtual-Target Capture: New capture command compiles a script for another GPU target (H100/B200, MI300X/MI350X) without loading or executing the target kernels. Events record virtual-target provenance. This is cross-target code generation and inspection, not emulation or a correctness check.

  • 🔀 Autotune Sessions Split by Launch Key: Autotune sessions previously collapsed repeated invocations at the same call site with different key arguments into one session. Sessions now split into per-key sub-sessions at analysis time, each resolving its own winner. Unsplit sessions are unchanged.

  • 📦 Constexpr and Tuple Launch Arguments: Compile-time-constant and tuple launch arguments are now captured with their values and replayed faithfully in reproducers, instead of being truncated to a text representation. Older traces only holding a text representation use best-effort literal recovery for plain scalars.

  • 🔍 Pre-Optimization IR and Compiler Identity: Pre-optimization IR is now a first-class stage (ttir_pre_opt / ttgir_pre_opt, disable with TRITONPARSE_DUMP_SOURCE_IR), and every compilation event records a process-cached compiler identity (package version and source-tree hash) with no subprocess probes.

  • 🪝 Triton HookChain Compatibility (#447): TritonParse now registers into Triton's hook chain when present and falls back to legacy hooks otherwise. Cleanup removes only its own registrations, leaving other tools' hooks intact.

Changes by Area

🧩 CuTeDSL Support

  • Tracing and adapter: New tracing core plus cutedsl backend adapter (MLIR, clean MLIR, PTX, SASS) wired through init() / clear() and the context manager. Launch configuration and tensor metadata are extracted from compiled artifacts.
  • PTX fallback and viewer: Source mapping now falls back to the first entry block for marker-less NVVM PTX, and the viewer displays CuTeDSL stage names, in-memory compilation payloads, and object launch arguments correctly.
  • Coverage: CPU compilation/launch tests plus a Hopper end-to-end GEMM workload checked against a PyTorch reference.

🎯 Virtual-Target Capture

  • Runtime: Fixed target profiles, capture-only launcher, isolated cache, and explicit rejection of measurement-dependent autotuning and conflicting targets.
  • Verification: GPU tests compile for cuda:90, cuda:100, hip:gfx942, and hip:gfx950 from an H100 host and prove target kernels were captured, not executed.

🔀 Parsing and Reproducer

  • Autotune split: Launches are grouped by runtime-argument signature; each sub-session resolves its own winner.
  • Argument capture: Constant wrappers, recursive tuple/named-tuple values, and later-defined helper functions are now preserved in traces and reproducers. Unsupported nested values raise an explicit error instead of becoming None.
  • Timestamp fix: Trace timestamps no longer carry a literal %f; records now emit real YYYY-MM-DDTHH:MM:SS.ffffffZ timestamps.

🛠️ Bisect

  • LLVM descriptors: Reads both legacy and JSON descriptor formats and compares normalized source hashes, so artifact-only updates never start a source bisect.
  • Outcomes and logs: Skipped boundaries produce explicit found/ambiguous/aborted/error outcomes with per-attempt result and replay files. Logger instances are now isolated (#439), and quoted Git result terms from newer Git versions are parsed.

🖥️ Website

  • File Diff labels (#407): Editable label per side, persisted in the URL (label_a / label_b) and mirrored into editor accessibility labels.

🏗️ Infrastructure and CI

  • H100 lane (#448): New H100 nightly coverage alongside the existing A10G and MI350 lanes.
  • Examples and deps (#440, #445, #446): Triton and inductor example traces regenerated (matmul now splits into two per-key sub-sessions); website dependencies refreshed with newer TypeScript type checking.
  • Contract test: The stage-parser contract test now loads its fixture via package resources and fails explicitly when missing instead of silently skipping.

Compatibility Notes

  • No breaking changes. CuTeDSL tracing, virtual-target capture, and pre-optimization IR are additive or opt-in; new trace fields extend existing events without removals or renames.
  • Autotune session ids may multiply. Call sites exercised with multiple key sets now produce one <site>:<key> sub-session per set. Re-parse existing traces to pick this up.
  • Re-capture recommended for traces with constant or tuple arguments if you generate reproducers from them.

Upgrade Guidance

  1. Standard upgrade:

    pip install --upgrade tritonparse
  2. Trace a CuTeDSL kernel:

    import tritonparse.structured_logging as tsl
    
    tsl.init("./logs", enable_cutedsl_tracing=True, enable_sass_dump=True)
  3. Capture for another GPU target without executing it on the host:

    tritonparseoss capture --target cuda:100 --out ./capture_out script.py
  4. Re-parse existing traces for per-key autotune sub-sessions:

    tritonparseoss parse ./logs --out ./parsed_output --overwrite