David Gornshtein
Founder
David Gornshtein leads the technical direction behind MegaCpp and the public engineering narrative around SLM Ensemble.
- Primary byline
- David Gornshtein
- Canonical URL
- /people/david-gornshtein/
- Identity links
- 2
Articles attributed to David Gornshtein
These links are derived from the local blog index and use the same author-name mapping as MegaCpp bylines and article schema.
Torch/XLA 2.11 expectations vs TPU reality
What MegaCpp expected from the Torch/XLA 2.11 line on TPU, what the shipped stack actually looked like in practice, and how that changed our bringup strategy.
vLLM on GB10: the overlay, the registration fixes, and the paths we kept off
How MegaCpp stabilized a GB10-oriented vLLM lane with an on-disk overlay, text-only model registration, and a deliberate keep-disabled list for serving paths that were not yet honest.
Attention Validity and Structure-Aware Attention
A packed-row validity regression, the clustered-sparse follow-up it forced, and the structure-aware attention plan we are integrating into the MegaCpp training stack.
Dynamo and torch.compile Breakage on a Mamba-3 Hybrid
Graph breaks, recompile storms, guard explosions, and cache-hygiene rules we landed while keeping torch.compile useful on MegaCpp's hybrid Mamba-3 + Transformer stack.
External library glitches we fixed
A catalog of upstream bugs we hit while training our hybrid Mamba-3 plus DSA recipe, grouped by library: what broke, what we patched locally, and what we prepared upstream.
Flash Attention 4 in practice: what we shipped and what we cut
Our hybrid stack's applicability matrix for Flash Attention 4, the validation profiles, the dense-full rollout gates, and the regressions that killed the first FA4 variants before they reached deployment.