[FEAT#67] neuron Phase 13 — HybridGraphLinear 의 Transformer FFN 통합 + RMSNorm by juhy0987 · Pull Request #68 · EinSofINTEREST/GraphLM

juhy0987 · 2026-05-26T10:33:27Z

연관 이슈

Closes [FEATURE] neuron Phase 13 — HybridGraphLinear 의 Transformer FFN 통합 + RMSNorm #67

구현 내용

Phase 12 에서 HybridGraphLinear (outer group + inner channel dual routing) 를 paradigm 의 ultimate 표현으로 확립. Phase 13 은 이를 실제 Transformer 의 FFN 위치 에 통합하고 RMSNorm 을 도입.

신규 모듈

src/graphlm/neuron/rms_norm.py — RMSNorm (LLaMA / Mistral 표준 norm)
- mean centering 생략 + bias 없음 → 파라미터 ↓
- weight=1 초기화 → function preservation
- mixed precision 안전성 위해 float32 cast 후 RMS 계산
src/graphlm/neuron/hybrid_transformer.py — Transformer block 통합
- HybridGraphFFN: HybridGraphLinear fc1 + fc2 + GELU
- HybridGraphTransformerBlock: pre-norm (RMSNorm + CausalSelfAttention + RMSNorm + FFN + residual)
- PlainFFN / PlainTransformerBlock: 공정 비교용 baseline (norm 은 RMSNorm 통일)
- make_block: 4 arch dispatch
src/graphlm/neuron/hybrid_transformer_demo.py — LM + 학습 helper
- HybridGraphTransformerLM: small char-LM (token+pos emb → N blocks → RMSNorm → lm_head)
- train_hybrid_transformer_lm: 1 run sweep unit
- _snapshot_adj: block 별 fc1/fc2 outer/inner snapshot (Phase 12 hierarchy 호환)
notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb — 4 arch × 2 seed sweep + 자동 verdict + 시각화

4 가지 arch

plain — RMSNorm + 표준 nn.Linear FFN (baseline)
hybrid_full_full — function preserving baseline
hybrid_full_around_one — Phase 11 channel-level scale-corrected
hybrid_around_one_around_one — outer/inner 둘 다 scale-corrected

identity outer 미지원 — FFN 은 rectangular (hidden ≠ ffn) 구조라 정방 identity 정의 불가. ValueError 명시.

테스트 (122 → 145, +23 신규)

test_rms_norm.py — 8 tests (shape / RMS unit / weight learnable / 입력 검증)
test_hybrid_transformer.py — 15 tests (FFN/block function preservation, gradient flow, 4 arch dispatch)
전체 145 tests all green, coverage 92.87%

0-init 금지 + magnitude rule 일관 적용

HybridGraphFFN 도 underlying HybridGraphLinear 의 ValueError 그대로 상속
RMSNorm 의 weight 는 1 로 시작 (0-init 회피, multiplier 위치 magnitude rule)

CI / 머지 게이트 점검

변경 영향 범위

영향 패키지/모듈: src/graphlm/neuron/ (3 신규 파일), notebooks/02-function-level/ (1 신규 노트북), tests/neuron/ (2 신규 테스트 파일)
위험도(택1): Low — 신규 모듈 추가만, 기존 모듈 변경 없음. __init__.py exports 만 갱신.

Required Status Checks

통과 확인 대상 (PR Checks 탭에서 확인):
- Commit Lint
- PR Title Lint
- Linked Issue Check
- Format Check (로컬 통과)
- Build (로컬 통과)
- Test (로컬 145 passed)
- Lint (로컬 통과)

롤백 계획

신규 파일만 추가했으므로 revert 시 영향 없음 (기존 모듈 변경 0)
Phase 12 노트북 / 모듈은 그대로 작동

다음 단계 (Phase 14 후보)

attention 의 qkv / out 을 HybridGraphLinear 로 교체 — function-level graph 가 attention 까지 확장
Net2Net / LiGO 식 growable Transformer — 학습 중 hidden_dim 또는 n_layers 증가

Summary by CodeRabbit

New Features
- Introduced four hybrid transformer variants for FFN-path experiments, plus a plain baseline.
- Added an RMS normalization layer for transformer models.
- Added training/demo utilities for small hybrid transformer language models with parameter counting, per-step loss tracking, and snapshotting.
- Included an experimental notebook demonstrating variant sweeps and visualizations on a character-level dataset.
Tests
- Expanded test coverage for hybrid transformer components and RMS normalization, including behavior, gradients, and numeric/dtype checks.

- src/graphlm/neuron/rms_norm.py 신규 — LLaMA / Mistral 표준 norm - mean centering 생략 + bias 없음 (LayerNorm 대비 파라미터 ↓, 안정성 동등) - weight=1 초기화로 function preservation 보장 - mixed precision 안전성 위해 float32 cast 후 RMS 계산 - 8 unit tests (shape / scaling / gradient / 입력 검증)

- src/graphlm/neuron/hybrid_transformer.py 신규 - HybridGraphFFN: HybridGraphLinear 로 fc1 + fc2 (GELU) - HybridGraphTransformerBlock: pre-norm (RMSNorm + attn + RMSNorm + FFN) - PlainFFN / PlainTransformerBlock: 공정 비교용 baseline (norm 은 RMSNorm 통일) - make_block: 4 arch dispatch (plain / hybrid_full_full / full_around_one / around_one_around_one) - identity outer 는 FFN 의 rectangular 구조상 미지원 (ValueError 명시) - 14 unit tests — FFN 동치성 / block 동치성 / gradient flow / arch dispatch - __init__.py exports 갱신

… helper - src/graphlm/neuron/hybrid_transformer_demo.py 신규 - HybridGraphTransformerLM: token+pos emb → N blocks → RMSNorm → lm_head - train_hybrid_transformer_lm: 1 run unit (4 arch dispatch) - count_parameters: arch 간 파라미터 수 reporting - _snapshot_adj: hybrid arch 의 block 별 FFN fc1/fc2 outer/inner snapshot (Phase 12 hierarchy 호환) - ruff format / unused import 정리 - 120 tests all green

…_around_one / around_one_around_one) - notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb 신규 - 4 arch × 2 seed = 8 run, max_steps=1500, TinyShakespeare char-LM - model: hidden=128, ffn=256, n_layers=4, n_heads=4, block_size=64 - 자동 verdict 4 가지: function preservation / scale-corrected ≤ plain / full around_one 안정성 / RMSNorm NaN 없음 - loss curve (mean ± σ rolling) + hybrid adj heatmap (outer / inner) - output figure → runs/notebook-neuron-phase13/{loss_curves,hybrid_adj}.png

coderabbitai · 2026-05-26T10:33:40Z

Warning

Review limit reached

@juhy0987, we couldn't start this review because you've reached your PR review rate limit.

More reviews will be available in 38 minutes and 9 seconds. Learn how PR review limits work.

Your organization has run out of usage credits. Purchase more in the billing tab.

⌛ How to resolve this issue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans include higher PR review limits than trial, open-source, and free plans. In all cases, reviews become available again over time. During sustained high-volume PR review activity, CodeRabbit may temporarily slow when the next review becomes available.

Please see our Fair Usage Limits Policy for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: b99751b0-6dca-4b44-beb5-c57c8d4771d3

📥 Commits

Reviewing files that changed from the base of the PR and between 153d381 and 428bacd.

📒 Files selected for processing (4)

notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb
src/graphlm/utils/__init__.py
src/graphlm/utils/metrics.py
tests/utils/test_metrics.py

📝 Walkthrough

Walkthrough

Phase 13 introduces RMSNorm and new transformer blocks that use HybridGraphLinear in the FFN path, plain baselines, a small HybridGraphTransformerLM and training harness, tests verifying function preservation and gradients, updated public exports, and a notebook that sweeps four architecture variants on TinyShakespeare reporting losses, verdicts, and adjacency visualizations.

Changes

Phase 13 Hybrid Transformer

Layer / File(s)	Summary
RMSNorm Layer Normalization `src/graphlm/neuron/rms_norm.py`, `tests/neuron/test_rms_norm.py`	RMSNorm implements float32-stable RMS normalization with per-feature learnable scaling, dtype round-trip for mixed precision, input/param validation, and tests for shapes, init, normalization, gradients, error handling, and dtype behavior.
Hybrid and Plain Transformer Blocks `src/graphlm/neuron/hybrid_transformer.py`	Adds HybridGraphFFN (two HybridGraphLinear layers + GELU + dropout), HybridGraphTransformerBlock (pre-norm attention + hybrid FFN residuals), PlainFFN/PlainTransformerBlock baselines, Arch literal, and make_block factory for four arch variants.
Transformer Block Tests & Function Preservation `tests/neuron/test_hybrid_transformer.py`	Tests assert shapes, reject invalid adj_outer_init, verify forward-equivalence to plain blocks for full/full init via weight-copying, ensure gradients reach weight/adj_outer/adj_inner, validate make_block dispatch, and check dropout behavior.
Hybrid Transformer LM and Training Loop `src/graphlm/neuron/hybrid_transformer_demo.py`	Introduces HybridTransformerTrainConfig dataclass, HybridGraphTransformerLM, train_hybrid_transformer_lm(config) training loop (seed control, AdamW, per-step losses, final_loss averaging), adjacency snapshotting for hybrid blocks, and count_parameters().
Module Integration `src/graphlm/neuron/__init__.py`	Imports and re-exports HybridGraphFFN, HybridGraphTransformerBlock, PlainTransformerBlock, RMSNorm, and make_block to expose Phase 13 components from the neuron package.
Phase 13 Experimental Sweep Notebook `notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb`	Notebook runs a four-architecture sweep (plain, hybrid_full_full, hybrid_full_around_one, hybrid_around_one_around_one) across seeds on TinyShakespeare, records per-step losses and final metrics, computes per-architecture statistics and automated verdicts, plots rolling-mean loss curves with ±std bands, and renders FC1 adjacency heatmaps for hybrid variants.

Sequence Diagram(s)

sequenceDiagram
  participant Trainer
  participant Model as HybridGraphTransformerLM
  participant Blocks as TransformerBlocks
  participant FFN as HybridGraphFFN
  participant Optimizer as AdamW
  Trainer->>Model: train_hybrid_transformer_lm(config)
  Model->>Model: set seed, build model, move to device
  loop for each batch (max_steps)
    Trainer->>Model: Forward pass (tokens)
    Model->>Blocks: token/pos embed -> block stack
    Blocks->>Blocks: RMSNorm -> CausalSelfAttention -> residual add
    Blocks->>FFN: RMSNorm -> HybridGraphFFN -> residual add
    FFN->>FFN: HybridGraphLinear(fc1 adj_outer/adj_inner) -> GELU -> HybridGraphLinear(fc2)
    Blocks-->>Model: block outputs
    Model->>Model: final RMSNorm -> lm_head -> logits
    Trainer->>Optimizer: loss -> backward -> step
    Optimizer->>FFN: update weight, adj_outer, adj_inner
    Trainer->>Model: record per-step loss
  end
  Trainer->>Model: snapshot_adj() (hybrid only)
  Model-->>Trainer: per-block adj_outer/adj_inner tensors

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

Possibly related PRs

EinSofINTEREST/GraphLM#66 — Phase 12 HybridGraphLinear implementation that this PR integrates into Transformer FFN paths.

"🐰 Phase thirteen hops in delight,
RMSNorm steady, hybrid FFNs take flight.
Four variants dance, seeds twirl around,
Loss curves hum and adj heatmaps are found.
Gradients wink — the small model leaps with might."

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 32.50% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The PR title clearly references Phase 13, the main integration of HybridGraphLinear into Transformer FFN, and RMSNorm adoption—all primary changes. It is specific and reflects the core objectives.
Linked Issues check	✅ Passed	The PR successfully implements all coding requirements from issue `#67`: RMSNorm module with unit tests, HybridGraphTransformerBlock with pre-norm structure and hybrid FFN, HybridGraphTransformerLM, function preservation validation (adj_outer=full/adj_inner=full), 4-architecture sweep, tests (152+), coverage ≥92%, notebook with visualization, and 0-init avoidance with magnitude rules applied consistently.
Out of Scope Changes check	✅ Passed	All changes are within scope: new RMSNorm, HybridGraphFFN/HybridGraphTransformerBlock, hybrid_transformer_demo, notebook sweep, module exports update, and comprehensive tests. No unrelated or unexpected modifications outside Phase 13 FFN integration objectives detected.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch feature/#67/neuron-phase13-hybrid-transformer

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

gemini-code-assist

Code Review

This pull request implements Phase 13 of the project, integrating the HybridGraphLinear layer into the standard Transformer FFN position and introducing RMSNorm as a modern normalization layer. It includes the core implementations of HybridGraphFFN, HybridGraphTransformerBlock, PlainTransformerBlock, and RMSNorm, along with a demo notebook, training helpers, and comprehensive unit tests. The review feedback highlights two important issues: a potential mixed-precision dtype mismatch in RMSNorm.forward when multiplying by self.weight, and a potential OverflowError in the notebook when calculating perplexity using math.exp on high loss values.

Copilot

Pull request overview

이 PR은 Phase 12에서 확립된 HybridGraphLinear를 Transformer FFN 위치에 통합하고, norm을 RMSNorm으로 교체한 Phase 13 실험용 모듈/데모/테스트/노트북을 추가합니다. plain 대비 hybrid_* 아키텍처를 동일한 블록 구조에서 공정 비교할 수 있게 구성되어 neuron 실험 라인을 Transformer 백본으로 확장하는 변경입니다.

Changes:

RMSNorm 신규 구현 및 단위 테스트 추가
HybridGraphFFN/HybridGraphTransformerBlock 및 make_block(4-arch dispatch) 신규 추가 + 동치성/그라디언트 테스트 추가
Phase 13 sweep 실행을 위한 demo LM/train helper 및 실험 노트북 추가, neuron.__init__ export 갱신

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
`src/graphlm/neuron/rms_norm.py`	Phase 13용 RMSNorm 구현 추가
`src/graphlm/neuron/hybrid_transformer.py`	Hybrid FFN + pre-norm Transformer block + arch dispatch 추가
`src/graphlm/neuron/hybrid_transformer_demo.py`	char-LM 데모 모델/학습 헬퍼 및 adj 스냅샷 유틸 추가
`src/graphlm/neuron/__init__.py`	Phase 13 신규 심볼 export 추가
`tests/neuron/test_rms_norm.py`	RMSNorm 기능/검증 테스트 추가
`tests/neuron/test_hybrid_transformer.py`	FFN/블록 동치성, 그라디언트, dispatch 테스트 추가
`notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb`	4 arch × 2 seed sweep 및 시각화/자동 verdict 노트북 추가

- src/graphlm/neuron/rms_norm.py — RMSNorm.forward 에서 self.weight 를 x_dtype 으로 cast (gemini #3303153077): mixed precision (FP16/BF16) 입력 시 weight 가 float32 로 남아 있어 출력 dtype 이 강제로 float32 promotion 되던 문제 — residual connection dtype mismatch 회피 - notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb — safe_perplexity 헬퍼 도입 (gemini #3303153101): 학습 초반 큰 loss 에서 math.exp OverflowError 회피 (cap=20.0)

coderabbitai

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/graphlm/neuron/hybrid_transformer_demo.py`:
- Around line 113-130: Replace the long parameter list on
train_hybrid_transformer_lm with a single dataclass (e.g.,
HybridTransformerTrainConfig) that contains the current scalars: dataset:
TinyShakespeareDataset, vocab_size: int, seed: int, arch: Arch, hidden_dim: int,
n_heads: int, ffn_dim: int, n_layers: int, group_size: int, block_size: int,
batch_size: int, lr: float, max_steps: int, device: str = "cpu", dropout: float
= 0.0; update the function signature to accept that config object
(train_hybrid_transformer_lm(config: HybridTransformerTrainConfig)) and change
all internal references from individual parameters (hidden_dim, n_heads, lr,
etc.) to config.hidden_dim, config.n_heads, config.lr, etc.; add the dataclass
import and the new type to annotations and update any call sites/tests to
construct and pass the new config instance when calling
train_hybrid_transformer_lm.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 351e0665-b8a9-4bc2-8c1b-f623db17cc70

📥 Commits

Reviewing files that changed from the base of the PR and between bbbfe97 and c04a86f.

📒 Files selected for processing (7)

notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb
src/graphlm/neuron/__init__.py
src/graphlm/neuron/hybrid_transformer.py
src/graphlm/neuron/hybrid_transformer_demo.py
src/graphlm/neuron/rms_norm.py
tests/neuron/test_hybrid_transformer.py
tests/neuron/test_rms_norm.py

- src/graphlm/neuron/rms_norm.py (Copilot #3303168589): floating-point 입력 검증 - int 등 비-floating 입력 silent cast 회피 — TypeError 명시 (LayerNorm 정책과 동일) - src/graphlm/neuron/hybrid_transformer.py (Copilot #3303168649 / #3303168686): - HybridGraphFFN / PlainFFN 둘 다 dropout 인자 추가 — backbone.FFN 의 fc2-뒤-dropout 패턴 일관 - HybridGraphTransformerBlock / PlainTransformerBlock 의 dropout 인자가 attention 만이 아니라 FFN 까지 전달 - 7 신규 unit tests — non-floating 거부 / dtype roundtrip / FFN dropout / block dropout 전파 - 152 tests all green

- CodeRabbit #3303186824 — project rule "5+ args → dataclass" 적용 - src/graphlm/neuron/hybrid_transformer_demo.py: - HybridTransformerTrainConfig (frozen dataclass) 신규 — data/model/train/runtime 4 그룹 - train_hybrid_transformer_lm(config) 단일 인자로 단순화 (기존 14 kwargs) - notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb §2 — config 객체 구성 후 train 호출 - 152 tests all green

coderabbitai

Actionable comments posted: 2

🧹 Nitpick comments (1)

src/graphlm/neuron/hybrid_transformer_demo.py (1)
179-179: ⚡ Quick win

Extract magic number to a named constant.

The hardcoded 100 violates the guideline "Replace magic numbers with named constants" and makes the averaging window less discoverable.

As per coding guidelines, magic numbers should be replaced with named constants.
♻️ Proposed fix

Add a module-level constant near the top of the file (after imports):
# 학습 마지막 N step 의 loss 평균 계산 window
_FINAL_LOSS_AVERAGING_WINDOW = 100
Then update line 179:
-    n_last = min(100, len(losses))
+    n_last = min(_FINAL_LOSS_AVERAGING_WINDOW, len(losses))
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/graphlm/neuron/hybrid_transformer_demo.py` at line 179, Replace the magic
number 100 used when computing n_last (n_last = min(100, len(losses))) with a
module-level named constant (e.g., _FINAL_LOSS_AVERAGING_WINDOW) declared near
the top of src/graphlm/neuron/hybrid_transformer_demo.py (after imports); then
change the assignment to n_last = min(_FINAL_LOSS_AVERAGING_WINDOW, len(losses))
so the averaging window is discoverable and configurable while keeping the
existing behavior of using the smaller of the window and len(losses).

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb`:
- Line 43: Extract the safe_perplexity function into a reusable module under the
graphlm package (e.g., create graphlm.utils or graphlm.metrics) keeping the same
signature (safe_perplexity(loss: float, cap: float = 20.0) -> float) and
tests/docstring, export it, then remove the in-notebook definition and replace
it with an import like `from graphlm.utils import safe_perplexity` in the
notebook; ensure any references to safe_perplexity (e.g., calls in plotting or
evaluation code) continue to work unchanged and update package __init__ if
needed to expose the symbol.
- Line 131: The stability checks currently use math.isnan which treats ±inf as
valid; update the verdict logic to use math.isfinite for numeric-finiteness
checks: change the full stability check (variable verdict_3 that tests aa_loss
with not math.isnan(aa_loss) and aa_loss < plain_loss + 0.5) to require
math.isfinite(aa_loss) and the final RMSNorm stability aggregator (all_finite
computed with not math.isnan(out["final_loss"])) to use
math.isfinite(out["final_loss"]) so infinities fail the stability verdicts.

---

Nitpick comments:
In `@src/graphlm/neuron/hybrid_transformer_demo.py`:
- Line 179: Replace the magic number 100 used when computing n_last (n_last =
min(100, len(losses))) with a module-level named constant (e.g.,
_FINAL_LOSS_AVERAGING_WINDOW) declared near the top of
src/graphlm/neuron/hybrid_transformer_demo.py (after imports); then change the
assignment to n_last = min(_FINAL_LOSS_AVERAGING_WINDOW, len(losses)) so the
averaging window is discoverable and configurable while keeping the existing
behavior of using the smaller of the window and len(losses).

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 04e28d5a-75f8-49d4-a6b3-55104e41e5e1

📥 Commits

Reviewing files that changed from the base of the PR and between c04a86f and 153d381.

⛔ Files ignored due to path filters (2)

docs/figures/neuron/phase13/hybrid_adj.png is excluded by !**/*.png
docs/figures/neuron/phase13/loss_curves.png is excluded by !**/*.png

📒 Files selected for processing (6)

notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb
src/graphlm/neuron/hybrid_transformer.py
src/graphlm/neuron/hybrid_transformer_demo.py
src/graphlm/neuron/rms_norm.py
tests/neuron/test_hybrid_transformer.py
tests/neuron/test_rms_norm.py

- CodeRabbit #3304306120 — safe_perplexity 를 노트북 → src/graphlm/utils/metrics.py 로 이전 (project rule: 노트북은 analysis 만, 로직은 src/) - graphlm.utils 에서 export, 8 unit tests 추가 - 노트북은 `from graphlm.utils import safe_perplexity` 만 사용 - CodeRabbit #3304306127 — verdict 안정성 검사를 isnan → isfinite 로 교체 (verdict 3, 4) - isnan 만 쓰면 ±inf 가 silent PASS — 학습 발산 미감지 위험 - 152 → 160 tests all green

juhy0987 added 5 commits May 26, 2026 19:24

[CHORE]: Phase 13 노트북 ruff format 적용 + nb-clean

c04a86f

Copilot AI review requested due to automatic review settings May 26, 2026 10:33

juhy0987 added the enhancement New feature or request label May 26, 2026

Copilot started reviewing on behalf of juhy0987 May 26, 2026 10:33 View session

gemini-code-assist Bot reviewed May 26, 2026

View reviewed changes

Comment thread src/graphlm/neuron/rms_norm.py Outdated

Comment thread notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb Outdated

Copilot AI reviewed May 26, 2026

View reviewed changes

Comment thread src/graphlm/neuron/rms_norm.py Outdated

Comment thread src/graphlm/neuron/hybrid_transformer.py

Comment thread src/graphlm/neuron/hybrid_transformer.py

coderabbitai Bot reviewed May 26, 2026

View reviewed changes

Comment thread src/graphlm/neuron/hybrid_transformer_demo.py Outdated

juhy0987 added 4 commits May 26, 2026 19:42

[CHORE]: CI 재트리거 nudge (GitHub Actions webhook 지연 회피)

b8ee88f

[CHORE]: Phase 13 figure 자산 추가 (Notion 임베드용)

153d381

coderabbitai Bot reviewed May 26, 2026

View reviewed changes

Comment thread notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb Outdated

Comment thread notebooks/02-function-level/12-phase13-hybrid-transformer.ipynb Outdated

juhy0987 self-assigned this May 26, 2026

juhy0987 merged commit d7eb107 into main May 26, 2026
9 of 10 checks passed

This was referenced May 26, 2026

[FEATURE] neuron Phase 14 — attention 도 HybridGraphLinear 로 (full graph block) #69

Closed

[FEAT#69] neuron Phase 14 — attention 도 HybridGraphLinear 로 (full graph block) #70

Merged

Navigation Menu

Uh oh!

[FEAT#67] neuron Phase 13 — HybridGraphLinear 의 Transformer FFN 통합 + RMSNorm - #68

[FEAT#67] neuron Phase 13 — HybridGraphLinear 의 Transformer FFN 통합 + RMSNorm#68
juhy0987 merged 11 commits into
mainfrom
feature/#67/neuron-phase13-hybrid-transformer

juhy0987 commented May 26, 2026 •

edited by coderabbitai Bot

Loading

Uh oh!

coderabbitai Bot commented May 26, 2026 •

edited

Loading

Review limit reached

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Possibly related PRs

❌ Failed checks (1 warning)

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

juhy0987 commented May 26, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

연관 이슈

구현 내용

신규 모듈

4 가지 arch

테스트 (122 → 145, +23 신규)

0-init 금지 + magnitude rule 일관 적용

CI / 머지 게이트 점검

변경 영향 범위

Required Status Checks

롤백 계획

다음 단계 (Phase 14 후보)

Summary by CodeRabbit

Uh oh!

coderabbitai Bot commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review limit reached

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Possibly related PRs

❌ Failed checks (1 warning)

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

juhy0987 commented May 26, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented May 26, 2026 •

edited

Loading