~/fba-lab/lab/speedrun/journey/tyler-01

FBALab

Code · architecture · study mode

AboutRoadmapSpeedrun
FBALab

Study mode — no GPU required.

Interactive LLM training & inference lab.

Qwen CAboutContactTermsPrivacyCookiesCommunity

© 2026 FBA Lab

Contact · contact@bubblspace.com · +91 75061 55016

Speedrun›Teaching arc›#1 Baseline
Act 1Model Training FundamentalsAct 2AI Systems OptimizationAct 3World-Record Training Optimization
Step 2 of 28Skill: Model Training Fundamentals
← Welcome to FBA Lab#2.1 Architecture Tweaks →
TRAINING SIMULATION

Speedrun — #1 Baseline

running
◷train_gpt2.py▸GPT model◎learner⌁01-baseline$no GPU
active
baseline
GPT-2 d12
Starting architecture and training loop.
step 01
train
optimize
Speedrun step
Architecture diffs and optimizer changes.
steps 02–05
loss
target
val loss 3.28
Cumulative loss timeline and finale.
step 06
1/4
Blocks
Quick summary

GPT-2 depth-12 with custom RMSNorm (no affine), GELU MLP, learned position embeddings.

Full explanation below the code →

fba-lab — train_gpt2.py · GPT modelexecuting
// block: GPT model · lines 91–153$ study train_gpt2.py --block gpt_modelGPT-2 depth-12 with custom RMSNorm (no affine), GELU MLP, learned position embeddings. ✓
Explanation

GPT-2 depth-12 with custom RMSNorm (no affine), GELU MLP, learned position embeddings.

Think about

How many parameters in the d12 preset? What norm is used?

// architecture

Live diagram

100%
speedrun journey013.28→023.28→033.28→043.28→053.27→063.27val loss10.99 → 3.28 (-7.71)BaselineStarting point: GPT-2 d12, DDP on FineWe…b shards, target val loss 3.28FineWeb.bin shardsDDP ranksrank 0 / 1Batch(B, T)GPT-2 d12RMSNormAttentionFFN (GELU)L5LM headlogitsval loss11.82.51target 3.28train minuteslatest: 3.28 @ 488 mingrad_accum · cosine LR · val every N steps
← Welcome to FBA Lab#2.1 Architecture Tweaks →

What changed

Baseline step — no previous snapshot to diff against.

What it bought

Learn this step, your way

WatchReadDeep Dive

Navigate by roadmap

StepTopicThis lesson
Getting startedWelcome to FBA Lab—
#1 BaselineThe baseline training runYou are here
#2.1 Architecture TweaksRoPE, ReLU² MLP, scaled attention output, torch.compile.Coming soon
← Back to Roadmap

Baseline run — the first beat

After you finish reading, use the code panel above to walk the four blocks — GPT model, DistributedDataLoader, Training loop, and DDP setup — and match each one to the diagram.