fix: reduce H100 replay buffer VRAM fraction 55% → 45% — OOM from IQN 8.6GB + Phase 2

IQN with 64 quantiles × batch=16384 allocates 8.6GB for tiled intermediates.
Combined with attention (250MB), Phase 2 aux workspaces (64MB), backward
branch scratch (16MB), the 55% replay fraction left only 349MB free —
not enough for experience collector's 192MB next_states allocation.

45% frees ~8GB headroom: replay buffer 19.4M → ~15.9M (still 60× batch).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-18 18:42:06 +02:00
parent 32151ba8bc
commit b4cf279b73

View File

@@ -4,7 +4,7 @@ batch_size = 16384
num_atoms = 52
buffer_size = 500000
hidden_dim_base = 256
replay_buffer_vram_fraction = 0.55 # state_dim=96, 86 weight tensors — leaves ~33GB for compute
replay_buffer_vram_fraction = 0.45 # state_dim=96, 86 tensors + IQN 8.6GB + attention 250MB + Phase 2 128MB
[experience]
gpu_timesteps_per_episode = 5000