model-architecture
Instantiate and debug the attention-is-all-you-need-pytorch Transformer architecture, masks, attention layers, feed-forward blocks, weight sharing, tensor shapes, and scheduled optimizer.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- model-architecture
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Instantiate and debug the attention-is-all-you-need-pytorch Transformer architecture, masks, attention layers, feed-forward blocks, weight sharing, tensor shapes, and scheduled optimizer.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md8.8 KB
- 参考文档references/architecture-notes.md8.0 KB
- 参考文档references/troubleshooting.md8.7 KB
- 脚本scripts/architecture_smoke_check.py11.5 KB
03 SKILL.md 正文
Model Architecture
Use this sub-skill when you need to construct, inspect, or debug the repository's Transformer model classes rather than preprocess data, run training CLIs, or perform checkpoint translation.
Use This For
- Building
Transformer,Encoder,Decoder,EncoderLayer,DecoderLayer,MultiHeadAttention,PositionwiseFeedForward,ScaledDotProductAttention,PositionalEncoding, orScheduledOptiminstances. - Checking source/target mask shapes, boolean semantics, causal target masks, device placement, attention tensor shapes, and flattened model logits.
- Choosing safe tiny-model settings for unit tests, especially weight-sharing
flags and
d_model == d_word_vec. - Debugging architecture assertions, projection/embedding scaling, incompatible shared vocabularies, and Noam-style warmup scheduling.
Route Elsewhere
- Dataset fields, vocab building, BPE, spaCy, or pickle schemas belong to the data-preparation sub-skill.
- Training command-line options, logging, checkpoints, label smoothing, and loss handling belong to the training sub-skill.
- Checkpoint loading,
Translator, beam search, and decoding workflows belong to the translation sub-skill; this sub-skill only covers class-level model shapes that those workflows rely on.
Primary References
- Start with references/api-reference.md for constructor signatures, forward contracts, return shapes, and optimizer API.
- Use references/architecture-notes.md for end-to-end tensor flow, mask broadcasting, tiny-model recipes, and weight sharing rules.
- Use references/troubleshooting.md for common assertion, mask, CUDA/device, flattened-output, sharing, and scheduler failures.
Smoke Check
Run the bundled architecture check from any current directory by pointing it at an explicit repository checkout:
python scripts/architecture_smoke_check.py --repo-root /path/to/attention-is-all-you-need-pytorch
Useful variants:
python scripts/architecture_smoke_check.py --repo-root /path/to/repo --device cpu
python scripts/architecture_smoke_check.py --repo-root /path/to/repo --device cuda
python scripts/architecture_smoke_check.py --repo-root /path/to/repo --skip-negative-checks
The script constructs a tiny deterministic Transformer with sharing disabled,
checks pad and causal masks, verifies multi-head attention broadcasting, probes
ScheduledOptim warmup/decay behavior, and optionally confirms that known-bad
architecture settings fail for the expected reasons.
Fast Operating Pattern
- Keep
d_modelequal tod_word_vec; the model asserts this after parameter initialization because residual connections require a common hidden width. - If source and target vocabularies are not the same shared vocabulary with the
same token ids, set
emb_src_trg_weight_sharing=False. For tiny tests with different vocab sizes, also settrg_emb_prj_weight_sharing=Falseunless you specifically need target embedding/projection sharing. - Use
scale_emb_or_prjonly as one ofemb,prj, ornone; scaling is applied only when target embedding/projection sharing is enabled. - Treat masks as boolean keep-masks:
Truemeans a key position is visible andFalsemeans it is masked before softmax. - Remember that
Transformer.forward(src_seq, trg_seq)returns flattened logits with shape[batch * target_length, n_trg_vocab], not[batch, target_length, n_trg_vocab].
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).