advanced-model-configuration
Inspect and troubleshoot Protenix configs, model internals, kernel backends, TFG guidance, confidence utilities, metrics, and safe runtime fallbacks.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- advanced-model-configuration
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Inspect and troubleshoot Protenix configs, model internals, kernel backends, TFG guidance, confidence utilities, metrics, and safe runtime fallbacks.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/config-api.md6.6 KB
- 参考文档references/kernels-and-backends.md7.4 KB
- 参考文档references/tfg-confidence-and-metrics.md8.7 KB
- 参考文档references/troubleshooting.md8.8 KB
- 脚本scripts/protenix_runtime_doctor.py14.0 KB
03 SKILL.md 正文
Protenix Advanced Model Configuration
Use this sub-skill for expert Protenix internals work: configuration parsing, model-type defaults, optional CUDA/Triton/cuEquivariance/DeepSpeed paths, conservative backend fallbacks, Training-Free Guidance (TFG), confidence post-processing, and direct metrics calls.
Start Here
- Run
python scripts/protenix_runtime_doctor.py --jsonfor a read-only report of package metadata, PyTorch/CUDA, optional backend imports, relevant environment switches, CLI availability, and external tool availability. - Read
references/config-api.mdbefore usingparse_configs, saving configs, composing dotted overrides, or comparing model-name-specific defaults. - Read
references/kernels-and-backends.mdbefore changingLAYERNORM_TYPE,triangle_attention,triangle_multiplicative,--triatt_kernel,--trimul_kernel, TF32, cache, fusion, dtype, cuEquivariance, Triton, or DeepSpeed behavior. - Read
references/tfg-confidence-and-metrics.mdbefore enabling TFG, editing potential terms, calling confidence helpers, interpreting pTM/gPDE/PAE outputs, or using RMSD/LDDT metrics directly. - Read
references/troubleshooting.mdwhen optional dependency imports, CUDA visibility, ABI mismatches, kernel-image failures, config overrides, TFG feature validation, or metric tensor shapes fail.
Bundled Tool
scripts/protenix_runtime_doctor.py: read-only runtime checker with text and--jsonoutput. It does not run kernels, compile extensions intentionally, download checkpoints, run inference, run training, or inspect local source paths.- Add
--include-model-importsonly when import probing is the goal; model-module imports may trigger fast layer norm behavior unlessLAYERNORM_TYPE=torchis already set.
Route Elsewhere
- Routine prediction command assembly, checkpoint/cache download behavior, output folders, and CLI inference workflows belong to
../cli-and-inference/SKILL.md. - Protenix input JSON authoring, ligands, bonds, constraints, seed fields, and feature-file paths belong to
../input-data-and-features/SKILL.md. - Training data preparation, database layout, fine-tuning launches, optimizer/scheduler usage, and distributed training belong to
../training-and-data-pipeline/SKILL.md.
Safety Rules
- Prefer reversible runtime/config overrides before code changes:
triangle_attention="torch",triangle_multiplicative="torch",--triatt_kernel torch,--trimul_kernel torch, andLAYERNORM_TYPE=torchare the first isolation path. - Set
LAYERNORM_TYPE=torchbefore importing model modules when the task is config inspection, CPU debugging, tensor-shape debugging, or fallback triage rather than proving the fused layer norm extension works. - Treat cuEquivariance, Triton, and DeepSpeed as optional acceleration paths. Do not promise they work on every GPU or ABI combination; verify imports, CUDA visibility, and selected config values first.
- Treat TFG, confidence, and metrics symbols as current implementation entry points for expert debugging, not as a stable public API contract.
04 同仓库更多 Skills
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.
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.
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.
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.
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.
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).