AGENT SKILL

acceleration-backends

Diagnose and choose TurboDiffusion CUDA custom-op, INT8/FastNorm, SLA, and SageSLA acceleration backends.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
acceleration-backends
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Diagnose and choose TurboDiffusion CUDA custom-op, INT8/FastNorm, SLA, and SageSLA acceleration backends.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Acceleration Backends

Use this sub-skill when the task concerns TurboDiffusion backend readiness or acceleration choices: CUDA extension builds, turbo_diffusion_ops, INT8 Linear, FastRMSNorm/FastLayerNorm, SLA/SageSLA, attention_type, sla_topk, quant_linear, default_norm, or source-layout import quirks.

Route first

  • For backend installation, extension build requirements, CUDA_HOME, nvcc, CUTLASS, and custom-op import checks, read backend-build.md.
  • For API and option semantics around SparseLinearAttention, SageSparseLinearAttention, Int8Linear, FastNorm modules, and model replacement helpers, read acceleration-api.md.
  • For symptom-to-fix matrices covering CUDA headers, turbo_diffusion_ops, flash-attn, missing SpargeAttn, SAGESLA_ENABLED = false, source-layout imports, and tiny SLA smoke warnings, read troubleshooting.md.
  • To inspect a user's current environment without downloads or model execution, run check_acceleration_backend.py.

Boundaries

  • User-facing T2V/I2V launch construction belongs to the sibling video-inference sub-skill.
  • Interactive TUI serving belongs to the sibling interactive-serving sub-skill.
  • Training launch, checkpoint conversion, and quantized checkpoint export command construction belong to the sibling training-and-checkpoints sub-skill.
  • TurboT2AV's TileLang W8A8/FastNorm stack and Pixi/LTX-2 environment belong to the sibling turbot2av-extension sub-skill; only compare the concepts here.

Operating procedure

  1. Establish whether the user is checking an installed package, a source build, or runtime flag choices. Do not ask them to download checkpoints for backend-only checks.

  2. Run the safe diagnostic script when backend state is unknown:

    python sub-skills/acceleration-backends/scripts/check_acceleration_backend.py
    

    Add --require-cuda only when the requested capability truly needs CUDA custom ops.

  3. If attention_type=sagesla is requested, require SpargeAttn-backed SageSparseLinearAttention; otherwise route to the sla or original fallback decision in acceleration-api.md.

  4. If a quantized checkpoint is used, require both --quant_linear and working turbo_diffusion_ops; if an unquantized checkpoint is used, do not add --quant_linear by habit.

  5. If a command is run from a source checkout and imports fail for top-level modules such as imaginaire, rcm, ops, SLA, or serve, apply the source-layout PYTHONPATH note in backend-build.md instead of hard-coding local paths.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351