Skip to content

replace fixed seed with true-random entropy from RTL-SDR via sdr-random#333

Closed
kordless wants to merge 6 commits intokarpathy:masterfrom
DeepBlueDynamics:master
Closed

replace fixed seed with true-random entropy from RTL-SDR via sdr-random#333
kordless wants to merge 6 commits intokarpathy:masterfrom
DeepBlueDynamics:master

Conversation

@kordless
Copy link

No description provided.

… on learning rate

Applies Weber's force law bracket W = 1 - v²/(2c²) + v·a/c² to both
AdamW and Muon parameter updates. Parameter 'velocity' (momentum) and
'acceleration' (momentum change) modify the effective learning rate per
element/matrix, analogous to how Weber's bracket modifies effective
inertial mass for spinning bodies.

New hyperparameter WEBER_C_SQ controls the correction scale.
- ferricula integration: remember/recall tools give agent persistent
  thermodynamic memory across context resets and runs
- Dockerfile: CUDA runtime image + uv + PyTorch
- docker-compose: train, agent, ferricula, prepare services
- agent gets --memory flag for ferricula URL
- Flash Attention 3 auto-detects and falls back to PyTorch SDPA on non-Hopper GPUs
- TORCHDYNAMO_DISABLE=1 on Windows (no Triton)
- torch.compile wrapped in try/except for eager mode fallback
- ROPE_BASE moved to runtime default (was failing at class definition)
- README: platform notes for H100, consumer GPUs, Windows
Copy link
Collaborator

@svlandeg svlandeg left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Considering things like

What's different in this fork

I assume you wanted to PR this to your fork, not to karpathy's master 😉

@svlandeg svlandeg closed this Mar 18, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants