fix: H100 VRAM fraction 0.70→0.55 — OOM with state_dim=96 + 86 tensors

batch_size auto-scaled to 16384 consuming 78GB, leaving only 3GB
for experience collector → OOM. Reducing replay buffer fraction
gives more headroom. state_dim grew 88→96 (plan features), weight
tensors grew 68→86 (ISV + plan head).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-17 09:23:51 +02:00
parent 3ce8ed40f1
commit f1be4619e0

View File

@@ -4,7 +4,7 @@ batch_size = 8192
num_atoms = 51
buffer_size = 500000
hidden_dim_base = 256
replay_buffer_vram_fraction = 0.70
replay_buffer_vram_fraction = 0.55 # Reduced from 0.70: state_dim grew 88→96, 86 weight tensors
[experience]
gpu_timesteps_per_episode = 5000