cd /news/machine-learning/my-latest-ablation-run-integrating-e… · home topics machine-learning article
[ARTICLE · art-61370] src=discuss.huggingface.co ↗ pub= topic=machine-learning verified=true sentiment=· neutral

My latest ablation run: integrating Engram onto two backbones

A researcher at OLMo-core ran a 200-step, ~1.7B-parameter ablation comparing Engram on a standard Transformer versus a GDN hybrid backbone, finding that the Transformer achieved slightly lower loss and higher MFU. After fixing indexing bugs in the Engram implementation, the performance gap between the two architectures narrowed significantly, revealing that the initial differences were partly execution artifacts rather than pure architectural behavior.

read1 min views1 publishedJul 16, 2026

Engram + Transformer reached slightly lower loss than Engram + GDN hybrid — but the more important result was how much of the original gap disappeared after I fixed an indexing edge case.

I ran a 200-step, ~1.7B-parameter ablation in OLMo-core comparing Engram on two backbones:

• a standard attention Transformer

• a 3 GDN-layer + 1 attention-layer hybrid

Main observations:

• Transformer + Engram reached slightly lower CE loss and perplexity

• Transformer showed larger isolated gradient spikes; the hybrid had a choppier overall gradient profile

• Transformer achieved higher MFU; the hybrid processed more tokens per second

• GPU memory stayed flat for both runs, even when the hybrid produced elevated CUDA allocation retries

The central systems tradeoff came from parameter parity.

Matching the two models pushed the hybrid toward d_model=1280, which was less friendly to the FLA/Triton/compiler path than 1024. The result was lower MFU, lower TPS, and more allocator churn.

Then the Engram implementation surfaced two indexing failures:

• signed overflow in the hash path produced negative lookup indices

• padded token IDs exceeded the real Engram lookup range

Fixing those paths — along with moving static tensors into registered buffers — lowered loss for both backbones and substantially narrowed the apparent architecture gap.

That changed the scientific interpretation:

What initially looked like architecture behavior was partly execution artifact.

Working inside OLMo-core made the engineering bar very visible. It pushed me to treat throughput, memory behavior, compiler constraints, and clean integration as part of the research itself — not implementation plumbing after the math.

I wrote up the full training-dynamics and engineering story here:

[“When Memory Meets Memory: Engram on Transformers vs GDN Hybrids”](https://medium.com/@jenwei0312/when-memory-meets-memory-engram-on-transformers-vs-gdn-hybrids-6dbefa1b6d23?sharedUserId=jenwei0312)

Repo: [GitHub - JenWei0312/OLMo-core at feature/engram-poc · GitHub](https://github.com/JenWei0312/OLMo-core/tree/feature/engram-poc)

Runs/W&B link: [Weights & Biases](https://wandb.ai/jenwei0312/olmo3-engram-experiments)
── more in #machine-learning 4 stories · sorted by recency
── more on @engram 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/my-latest-ablation-r…] indexed:0 read:1min 2026-07-16 ·